Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
1
Property Effects in Inductive Inference
A dissertation presented
by
Nadezda Vasilyeva
to
The Department of Psychology
In partial fulfillment of the requirements for the degree of
Doctor of Philosophy
in the field of
Psychology
Northeastern University
Boston, Massachusetts
2
Abstract
This project examined how people generate inductive inferences – probabilistic
hypotheses rendered plausible by the given evidence but not guaranteed by it. The hypotheses
people generate are often fine-tuned to specific features of the problem. For example, when
one is asked to project a property from a known case to unknown – e.g., given that ducks have
gene X, what else is likely to share the gene? – the nature of the projected property can have a
profound effect on the generated hypotheses. Intrinsic properties, such as having a gene, tend
to be projected to members of the same class – sparrows, other birds. Contextual, or
environmentally-transmitted properties, such as having a parasite – tend to be projected to
entities that interact or co-occur with the known case – otters, other aquatic animals;
predators of ducks. Such changes in inductive inferences based on the projected property –
“property effects” - are ubiquitous in induction, but the theoretical accounts of induction
capable of addressing the underlying psychological mechanism are lacking.
To address this gap, I proposed a simple two-component model of inductive inference,
consisting of retrieval of knowledge about the premise category (duck) and property (gene),
and generation of an inductive hypothesis. I hypothesized that property effects stem from
selective retrieval of knowledge about premise categories in the context of different
properties. One set of results suggested that the knowledge about premise category and
property is combined interactively in order to form an inductive hypothesis: the relationship
between participants’ knowledge about animals and the inferences they generated about these
animals varied depending on the property. For example, salient ecological knowledge about
animals promoted ecological inferences about them (projections of property to ecologically
related species), but only when participants were reasoning about a contextual property.
3
Likewise, salient categorical knowledge about animals suppressed ecological inferences, but
only when participants were reasoning about intrinsic properties. However, a study examining
the time course of early knowledge activation in the context of inductive inference found no
evidence for interactive retrieval: despite systematic differences in the use of knowledge in
inferences, the activation of knowledge about the premise category was not affected by the
property.
These results suggest that premise category and property knowledge are combined
interactively, but not during retrieval. Because these results were not compatible with the
proposed retrieval-based model of property effects, they required a revision of the model. The
revised proposal introduced explanation of the evidence to the model. On this account,
inference generation involves first, explaining the evidence provided by the premise (the
combination of premise category and the property) by identifying a larger regularity that it
belongs to, and second, formulating a guess about other entities that might share the property
by virtue of belonging to same regularity. Within this account, different properties can affect
inferences by triggering different types of explanations – formal, based on category
membership, or causal, based on a sequence of enabling events, or teleological, based on ends
and goals. Different explanations, or identifying an observation as a part of different types of
regularities can in turn lead to different generalization hypotheses. Preliminary examination of
explanations for premise information that participants spontaneously generated during the
inference-generation task showed that the dominant type of explanation indeed varied with
the property participants were reasoning about, providing the first support for the explanation-
based account of property effects.
4
In sum, this project was the first to examine the processing details of property-
sensitive induction. It contributed towards understanding the mechanism of property effects in
induction in two ways. First, it suggested how property effects do not work, by demonstrating
that property does not influence retrieval of knowledge about premise categories. Second, it
introduced property-driven explanations as a possible source of property effects and provided
preliminary evidence for this proposal, opening up a new promising line of research.
5
Table of Contents Abstract 2 Table of Contents 5 Chapter I. Property Effects: Overview 7 Introduction 7 Roadmap 8 What is induction, and (how) does it differ from deduction? 9 Property effects in evaluation of inductive arguments 13
Projectable vs. non-projectable properties: role of category homogeneity 14 Projectable properties: flexible reasoning and the role of structured knowledge 17 Other types of property effects 25
Argument evaluation vs. inference generation 28 Test knowledge domain: folkbiology 31 Chapter II. Theories and models of induction 33 Existing theories and models of Induction 33
Similarity-based models 33 Hypothesis-based models 35 Bayesian models 36
A (minimalistic) computational account of inference generation 39 Ingredients: sources and types of knowledge 40 Recipe, or mechanism 43
Knowledge retrieval vs. hypothesis generation: where do we start looking for property effects?
43
Model of induction, v.1.0: knowledge retrieval as a locus of property effects 45 Overview of the Studies 50 Chapter III. “Ingredients”: Knowledge types and sources 52 Study 1. Property knowledge: distributional beliefs 52
Method 56 Results 58 Discussion 60
Study 2. Premise category knowledge 62 Method 66 Results 70 Discussion 73
Summary 73 Chapter IV. Study 3. Property Effects and Knowledge Recruitment in Induction 75 Background 75 Relations among property, knowledge, and inference 77
Property only 77 Independent recruitment of property and premise knowledge 78 Interactive recruitment of property and premise knowledge 79
Method 81 Results 85
6
Discussion 97 Taxonomic vs. ecological inferences 97 Property effects on inferences 98 Recruitment of property knowledge 99 Recruitment of premise category knowledge 99 Recruitment of property and premise category knowledge: independence or interaction? 103
Chapter V. Study 4. Property Effects over the Time Course of Premise Category Knowledge Retrieval in Induction.
106
Study 4 106 Property effects on knowledge retrieval: how do you know them when you see them? 109 Non-monotonicity in knowledge retrieval 113 Method 118 Results 122 Discussion 141
Chapter VI. Revising the Model of Induction: Explaining Evidence as the Engine of Property Effects in Induction (Study 5)
146
What is an explanation, and what does it have to do with induction? 147 Hypothesis generation as a locus of property effects 150 Study 5: Property effects on spontaneous explanations 152
Coding 153 Results 156 Discussion 159
Chapter VII. General discussion 167 Overview 167 Model 1.0: Property effects via knowledge retrieval 167 The Ingredients 169 The Recipe 170 Implications for Model 1.0 171 Model 2.1: Property effects via explaining the evidence 172 Conclusions and Implications 173 Future Directions 176 References 178 Appendix A. General Instructions Used in Study 1. 185 Appendix B. Selection of Animal Stimuli for Studies 2 and 3: Animal Typicality Ratings
186
Appendix C. Stimuli Used in Study 4. 189 Appendix D. Study 4: Association Norming. 191 Appendix E. Full text of instructions for the Study 4. 194
7
Chapter I. Property Effects: Overview
Introduction
Imagine you are watching the news and the presenter announces in a dramatic voice
that a certain deadly virus was found in chicken, and the numbers of chicken casualties are
growing. What have you learned? On the face of it, exactly that – there is a virus affecting
chickens. But the cognitive consequences of learning this fact can go far beyond chickens.
You may start hypothesizing about the implications of this news. Should you give a call to
your auntie who has a little chicken coop and ask her how she’s been feeling lately? Should
you be worried yourself – after all, you had that big bowl of chicken soup just a couple of
days ago; is it time to get a health check up, just in case? Should you nix your plan of cooking
corn chowder for dinner – in the end, corn is in everything, including chicken food, maybe
that’s where chicken got the virus from? Should you stay away from farms and farm animals
on the coming weekend trip to the countryside?
Generating such hypotheses about uncertain outcomes from limited evidence –
inductive inference - is a pervasive cognitive activity. We rely on it heavily both in everyday
life and science, we tend to engage in it spontaneously, and we do it well enough to avoid
persistent failures1 in navigating a world of uncertainty.
Although seemingly effortless2, this is a sophisticated cognitive activity: the
hypotheses we generate are often fine-tuned to specific aspects of the problem. For example,
had you heard the news presenter announce, perhaps in a less dramatic voice, that instead of a
virus a certain genetic disorder is found in chickens, you would be much less likely to worry
that you’ve got the problem gene now after eating that bowl of soup, or call your auntie to 1 With some individual variability 2 Perhaps somewhat less so in science than in everyday life
8
check on her in case she caught the gene cleaning the coop. Instead, you might speculate that
other birds – pigeons out in the street, or your grandpa’s parrot (who, by the way, has never
seen a single chicken in his life) – might share the defective gene.
Such sensitivity of inferences to the specifics of the property (virus vs. genetic
disorder) has been documented in many natural and laboratory contexts, across many
domains, in many test paradigms. Despite that, little is known about the psychological
underpinnings of property’s influence on inferences. The goal of this project is to fill the gap
by proposing and evaluating the mechanism and processing details of property effects in
hypothesis generation.
Roadmap
Chapter 1 introduces the notions of induction as well as property effects in induction
and reviews related previous work. The bottom line of the review is that the phenomenon of
property effects in induction is well-established, but the theoretical work addressing the
underlying psychological mechanism is lacking.
In Chapter II, we review existing models and theories of induction, and evaluate their
(in)ability to account for property effects. To address the lack of adequate mechanistic
account of property effects, we propose a model of induction based on the claim that property
affects inferences by moderating retrieval of knowledge about premise categories.
Chapter III prepares the ground for testing the proposed model. It reports results of
two studies that describe the sources and types of knowledge available at the outset of
generating an inference (Studies 1 and 2).
Chapters IV and V present two studies testing the main claim of the model, that
property moderates knowledge retrieval. Study 3 provides initial constraints on knowledge
9
retrieval mechanism by examining relation between available salient knowledge and
outcome inferences. Study 4 examines knowledge retrieval in a more direct manner, by
mapping out the activation of different knowledge types across time. Chapter V concludes
with the discussion of possible practical and theoretical issues with the Study 4.
In Chapter VI, we discuss a possibility that property affects inferences by triggering
different types of explanation of the initial evidence provided by the inductive problem. We
propose a revised model of induction, report a preliminary analysis that supports it (Study 5),
and suggest further ways to test it.
Chapter VII reviews the findings as well as absence thereof from the conducted
studies, and summarizes what we have learned about the mechanism of property effects in
induction.
What is Induction, and (How) Does it Differ from Deduction?
Induction is classically defined by contrast to deduction. There are two major
approaches to distinguishing induction from deduction. The Problem approach defines
induction vs. deduction in terms of corresponding reasoning problems, assuming that it should
be possible to classify a question or an argument as deductive or inductive based on their
formal properties. The Process approach treats induction and deduction as two possible kinds
of psychological processes, assuming that people could approach any given problem using
one or the other, or a mixture of both (Heit, 2007).
There are problems with each of these approaches. Under the Problem approach, one
attempt to define inductive and deductive problems contrasts them in terms of the direction of
reasoning – general to specific (deduction) vs. specific to general (induction); however, there
are plenty of exceptions to this rule (see Heit, 2007, for review). Another attempt that falls
10
under the Problem approach is to define deductively valid arguments as following the rules
of a well-specified logic, and inductive arguments as everything else; although this defines
deduction and it can do the job of distinguishing the two, it does not say much about what
induction is (see also Mill, 1963, for a philosophical discussion of deduction-induction
distinction).
Designating problems or tasks as deductive or inductive cannot guarantee that people
will reason deductively about one and inductively about the other. According to the Process
approach, deduction and induction should be specified in terms of underlying psychological
processes. On some accounts, both induction and deduction are handled by the same type of
reasoning (Johnson-Laird, 1994; Oaksford & Hahn, 2007; Osherson, Smith, Wilkie, Lopez, &
Shafir, 1990; Rips, 2001; Sloman, 1993). In contrast, two-process accounts emphasize the
distinction between two kinds of reasoning – one deliberative and analytic or rule-based, and
another one fast, intuitive and heavily influenced by context and associations. As Heit (2007)
points out, these two systems may not directly correspond to deduction and induction, and in
fact, this traditional distinction may not be optimal for psychology. However the two systems
are defined, there is some evidence that asking people to make deductive vs. inductive
judgments may lead to qualitatively different results, supporting the idea that induction and
deduction may involve different processes (Goel, Gold, Kapur, & Houle, 1997; Rips, 2001;
Rotello & Heit, 2009). Yet, although this proposal makes a claim about the number of
involved processes (>1, it appears), at the moment it is not very well developed and needs
more work in terms of specifying the underlying processes and deriving specific predictions
(Heit, 2007).
11
Thus, as of now, the distinction between induction and deduction appears to be a
useful approximation. There seem to exist two clusters of problems and corresponding
approaches to solving them – deductive and inductive -- and it is not entirely clear whether
they are handled by two distinct reasoning processes, or they could both be reduced to one;
and if there are two processes, it is not clear what kind of processes they are, and where to
draw a precise line between them.
At this point, it may be reasonable to focus on describing regularities in deductive and
inductive reasoning, using some (imperfect) combination of problem- and process-
approaches to operationally distinguish between the two. Accumulated descriptions of such
regularities may either help to distinguish them further and to refine the definitions, or it may
demonstrate that the distinction between induction and deduction is superficial and suggest
that they should be combined under one name.
Nowadays, the term “induction” is used in reference to a variety of cognitive
activities, including concept learning, analogy, generation and testing of hypotheses, and
many others. An element of inductive reasoning, using the known to predict unknown, is
found in a wide range of activities, “from problem solving to social interaction to motor
control” (Heit, 2000, p. 569). Our proposal focuses on a narrower range of phenomena,
investigating how people generate inferences based on given premise information, and how
the nature of projected property affects the kinds of inferences they generate.
As a tentative basis for an operational definition of induction for our studies, we
propose that, in addition to the probabilistic nature of inductive inference, the set of relevant
knowledge for solving an inductive problem is not clearly defined. In a deductive argument,
in contrast, the relevant knowledge is restricted to the premise(s) and logical rules. All the
12
other knowledge a person may have is, strictly speaking, irrelevant. For example, in the
case of a valid but a non sound deductive argument “All elephants can fly. Socrates is an
elephant. | Socrates can fly”, the person’s prior knowledge about elephants, flying and
Socrates is irrelevant for determining the argument’s validity3. In contrast, in the argument
“Adam is interested in the distinction between deduction and induction | Nancy is interested
in the distinction between deduction and induction”, the set of relevant knowledge is not
clearly defined (is it the common profession that matters? Gender? Age? Nationality?
Political affiliation? Something else?), and argument strength heavily depends on what subset
of knowledge is recruited to inform a particular judgment.
Based on this definition, the property effects – different conclusions or different
strength of a given conclusion as a function of projected property – should only be observed
when people reason inductively (i.e. assume that relevance of any given piece of knowledge is
a matter of their subjective opinion), but not when they reason deductively (i.e. only rely on
given information and logical rules). In a deductive problem, varying a property should not
affect validity judgments (imagine replacing “flying” with “sleeping” or “writing poems” in
the Socrates argument above – the deductive validity should not be affected). Whereas in an
inductive argument, varying a property may affect what knowledge is deemed relevant
leading to changes in perceived argument strength (imagine replacing “is interested in the
distinction between induction and deduction” with “supports new Massachusetts law about
primary school education” or “likes peanut butter”).
3 Not all background knowledge is irrelevant to estimating validity of a deductive argument – a person at least should be able to understand the language and may need to know something about category membership of objects that figure in an argument. The claim is that such set of relevant knowledge is relatively well-‐defined in deduction.
13
Within the limits of this project, we are focusing on studying property effects in
induction only. The format of the questions in the presented studies (e.g. “Flu X5 is found in
ducks. What else is likely to have flu X5? Why?”) invites participants to use their background
knowledge to generate novel hypotheses, and the guesses they generate are inherently
probabilistic. This gives us reasons to believe that participants in these studies will be using
inductive reasoning, if such a thing exists.
By describing in some detail property effects in induction we hope to make a small but
novel contribution to the body of knowledge about regularities that govern inductive
reasoning. Further studies may show that our findings (if any) are not restricted to induction –
but that will require either a replication of property effects with deductive problems that
clearly define the set of relevant knowledge, or a demonstration that the questions in our
studies were non-inductive.
Property Effects in Evaluation of Inductive Arguments
A vast body of empirical evidence demonstrates that people make systematically
different inferences when they project different properties. A classical method of studying
property effects involves evaluation of complete inductive arguments. Typically, participants
are presented with one or more premises in which a property is attributed to a category, along
with a conclusion in which the property is attributed to a different or superordinate conclusion
category, and asked to evaluate the argument, or rate the likelihood that the conclusion is true
given that the premises are true. Within this paradigm, property effects are defined as
qualitative changes in selected conclusions, or quantitative changes in evaluated strength of a
complete argument, associated with manipulation of a projected property.
14
The properties used in studies of inductive reasoning can be roughly divided into
three classes: first, idiosyncratic or non-projectable, such as fell on the floor this morning, has
a drop of water on it. Then, there are two kinds of projectable properties: uninformative, or
blank properties, e.g. has property X or has sarca, and informative biasing properties, that are
systematically projected based on a particular relation (e.g., anatomically biasing property has
a two-chambered liver; behaviorally biasing property travels in a zig-zag path; or a property
biasing towards ecological interaction relation sick with disease X).
Most research on property effects falls into one of two broad categories. One subset of
work examines how property generalization is rooted in people’s beliefs about homogeneity
of members of the conclusion category (comparisons of projectable vs. non-projectable
properties). Another subset of studies demonstrates that different projectable properties may
tap into different subsets of knowledge, or different ways to define similarity between
premises and conclusions.
Projectable vs. non-projectable properties: role of category homogeneity. Some
properties appears to be more projectable, or generalizable from one category to another, than
others. To borrow an example from Heit (2000), if one home in the neighborhood is
burglarized, it increases the perceived probability that the home next to it might be as well
(i.e. the property of being broken into generalizes from one home to the next). But if one
home is painted blue, it does not increase the perceived probability that a home next to it
would be painted blue as well (i.e. the property of being painted blue does not generalize).
The discussion of what makes a property projectable or not has a long history. Nelson
Goodman (1955) proposed that in order to be able to generalize at all, people must possess
overhypotheses, or abstract beliefs describing the scope of properties. Goodman explains this
15
term on the example of bags containing marbles. If one examines a series of bags with
marbles, and notes that some contain all black marbles, others contain all white marbles, one
could form an overhypothesis about color distribution: each bag contains marbles uniform in
color. When a new bag is encountered, and a single marble drawn from it turns out to be blue,
the overhypothesis about color distribution in bags allows formulating a specific hypothesis,
an inductive guess about the bag content based on a single novel observation: the rest of the
marbles in the new bag are all blue. Having an overhypothesis about uniform color
distribution is crucial for making such a prediction; without it, after extracting a single blue
marble from the bag, a person would have no basis for a preference for any one of such
hypotheses as “all blue”, “one is blue and all the others are white”, “one black, one white, one
red, three green..”, etc. An overhypothesis provides abstract knowledge that sets up a
hypothesis space at a less abstract level (Kemp, Perfors & Tenenbaum, 2007), constraining it
to such possible specific hypotheses as “all blue”, “all white”, “all black”, etc.
Translating the marble examples to real world categories, having an overhypothesis
that homes in close proximity are likely to be under same kinds of threat justifies
generalization of a “burglary threat” property from one home to other homes belonging to the
“same bag”. And if one does not have an overhypotheses that houses in the same
neighborhood tend to be of the same color, they would not generalize blue paint on one house
to others.
One of the early pieces of empirical evidence for differences in property projectability
comes from work by Nisbett, Krantz, Jepson, & Kunda (1983). They showed that adults treat
some properties (e.g. skin color) as more generalizable than others (e.g. obesity). For example,
after being told about skin color of one member of a certain tribe, participants expected over
16
90% of that tribe to have the same skin color; but when they were told about one obese
tribe member, their estimate of obese people in the tribe was only 40%. Participants justified
their responses saying that tribe members are likely to be homogeneous in skin color but
heterogeneous with respect to body weight. Thus, the beliefs about homogeneity of properties
determined their projectability: for a given category, homogeneous properties were expected
to be more projectable than heterogeneous properties.
Developmental work shows that children, like adults, have systematic expectations
about how projectable different properties are. As demonstrated by Gelman (1988), by 4 years
of age children distinguish between generalizable (permanent and intrinsic, such as made of
cellulose, eats alfalfa) and non-generalizable properties (transient, accidental and
idiosyncratic, such as fell on the floor this morning, has gum stuck on the bottom, is dirty). In
this study, children projected generalizable properties dependent on the similarity between
premises and conclusions, and did not project non-generalizable properties regardless of the
similarity. Similarly, in a study by Heyman & Gelman (2000) young children were willing to
project novel behavioral properties such as likes to play tibbits, relying on shared personality
traits between characters, but did not project transient properties such as feeling thirsty.
There is evidence that such discrimination between properties is likely based on
children’s expectations about homogeneity among the members of a category. As shown by
Gelman & Coley (1990), for 2 ½ -year-old children a common category label (“is a bird”)
applied to premise and conclusion promoted projections of a behavioral property (says tweet-
tweet) between them. In contrast, when the objects were told to share a transient attribute (is
dirty / hungry / far away), children were not any more willing to project a behavioral property
between them than in the absence of any common label. Presumably, common category
17
information conveyed to children that the premise and conclusion were “from the same
bag”, like marbles, triggering their intuitions about category homogeneity. But a transient
common label does not bring about such an overhypothesis that would promote generalization
(See also Gutheil & Gelman, 1997; Macario, Shipley, & Billman, 1990; Waxman, Lynch,
Casey, & Baer, 1997).
In a study by Springer (1992) children were asked to make inferences between
biologically related animals (a mother and a child) and between perceptually and socially
related animals (playmates). Children were at chance for idiosyncratic properties (very dirty
from playing in mud), but tended to project biological properties (hairy ears) between
biologically related animals more than between perceptually and socially related pairs,
suggesting that not only can children distinguish between projectable and non-projectable
properties, but they also have expectations about relations appropriate for projection of a
particular property – the topic that we turn to in the next section.
Projectable properties: flexible reasoning and the role of structured knowledge.
The studies reviewed in the previous section suggest that “not all properties are created
equal”: some generalize easily even from one observation, others are seen as non-
generalizable. However, the picture is more complex and interesting than some properties
simply being inherently more projectable than others. Depending on a specific pairing of the
premise and conclusion properties, one and the same property can behave either as
projectable, or non-projectable. In other words, different properties can be projectable to
different kinds of categories; what matters is the match between the property and category
type.
18
Going back to Goodman’s overhypotheses, the example of bags with marbles has
its limits. The categorical structure of human knowledge is notably more complex than bags
of marbles. Every entity belongs to “multiple bags”, multiple cross-cutting categories, that in
turn belong to cross-cutting category systems (also known as knowledge structures). A duck
belongs to a category of birds, which is a part of a taxonomic classification of animals. A
duck is also an aquatic animal, one of the classes of contextual classification of animals.
There is evidence that induction is sensitive to the match between property type and category
system, or, more broadly, a type of relation between premise and conclusion.
A classic demonstration of this effect comes from the study by Heit & Rubinstein
(1994). Participants were shown pairs of animals related anatomically, such as a bat and a
mouse (two mammals) or behaviorally, such as a sparrow and a bat (both fly), and asked to
estimate the probability that they might share an anatomical (has a liver with two chambers
that act as one) or behavioral (travels in back and forth, or zig-zag, trajectory) property. They
found that the strength of the argument was determined by the agreement between property
and the relation shared by the premise and conclusion. Arguments in which the premise and
conclusion were anatomically similar were judged stronger for an anatomical property than
for a behavioral property, whereas arguments with a behaviorally similar premise and
conclusion were judged stronger for behavioral than anatomical properties.
Heit and Rubinstein proposed that strength of an argument depends on a flexible
measure of similarity between premises and conclusions, and the relevant dimensions of
assessing similarity are determined by the projected property. A single similarity metric
19
would be unable to account for the results of this study4. In support of their claim, a
separate analysis showed that reasoning about different properties was related to different sets
of similarity ratings: reasoning about anatomical properties was predicted by anatomic
similarity alone, whereas reasoning about behavioral properties was predicted by both
anatomic and behavioral similarity.
This study, although one of the earliest systematic investigations of property effects,
came closest to outlining a possible mechanism: Heit & Rubinstein proposed that “the
property being considered influences which features [of the categories] are important for
induction” (p. 411) and, on a different occasion, mention context-dependent retrieval of
information from semantic memory (Barsalou, 1989). Unfortunately, with regards to
specifying the mechanism of property effects, this is as specific as it gets.
Work of Shafto & Coley (2003) on expertise effects in induction highlighted the
interaction between property and structured knowledge, and, most revealingly, took the idea
of relying on different relations beyond similarity. In this study, the participants - biological
novices (undergraduate students) and experts (commercial fishermen) – were tested in an
induction task about marine creatures. When the projected property was uninformative
(property X), experts and novices in marine biology alike relied on taxonomic, similarity-
based relations. When, however, the property was informative and ecologically-biasing
(disease X), although novices continued to rely heavily on taxonomic relations, experts
changed their strategy and relied on causal and ecological relations between marine species.
4 Relatedly, Goodman (1972) argued that induction based on unconstrained similarity is not possible at all, since any category has a potentially infinite set of features and can be infinitely similar to any other category. Thus it is a necessary logical requirement for inductive inference to impose some initial constraints on a subset of relevant features.
20
Thus, in the absence of detailed knowledge of alternative knowledge structures, or in the
absence of ability to match a particular property with the appropriate distribution, people may
rely on default distributions of properties across highly available knowledge structures such as
taxonomic categories. But when both knowledge of alternative relations and ability to
selectively match certain properties to these relations are available, the reasoning becomes
flexible, and we start seeing property effects in induction.
Ross and Murphy (1999) extended the study of property effects to the domain of food.
They compared taxonomic categories of food based on shared features or composition (such
as fruit or dairy) to script categories (breakfast foods), which are determined by time, location,
or setting in which the foods are eaten. Participants were presented with triads consisting of a
target food (e.g. cereal) and two alternatives, one taxonomic match (noodles, another member
of the “breads & grains” category), and one script match (milk, another “breakfast food”).
They were taught that the target food (cereal) had a biochemical property (a novel enzyme) or
a situational property (eaten in a novel culture at a particular ceremony), and asked to project
that property to one of the alternatives. Participants, like in Heit & Rubinstein (1994) study,
showed inductive selectivity: they preferred taxonomic relations when projecting biochemical
properties, and script relations when projecting situational properties. The authors emphasize
how this finding arises from the complex organization of knowledge based on two cross-
cutting systems of categories – taxonomic and script.
However, Ross and Murphy do not discuss the mechanism of property effects beyond
its general reliance on structured knowledge. Such treatment of property effects is, in fact,
fairly characteristic of the reviewed work (with a few exceptions). A vast majority of studies
treat property effects as a useful tool to examine other aspects of cognitive functioning, such
21
as cross-classification, reliance on statistical information, or development of early
categorical representations and understanding of domain differences – but they do not ask the
question of how this tool actually works.
Vitkin, Coley & Feigin (2005) replicated and extended the findings of Ross and
Murphy (1999) using a task in which taxonomic and script inferences were not mutually
exclusive. In their study, participants were asked to confirm or disconfirm projections to foods
whose taxonomic or script relations to the premise were manipulated orthogonally.
Participants reasoned about a compositional property (has enzyme X) intended to render
taxonomic categories relevant, or about a situational property (appears on the menu at X’s
new restaurant), intended to render relational (script) categories relevant. Vitkin et al. (2005)
found that taxonomic relations supported inferences broadly: inferences to taxonomic targets
were strong for both compositional and situational properties. In contrast, script categories
supported inferences more narrowly: inferences to script targets were strong for situational
properties but weak for compositional properties. These results suggest that script categories
were selectively recruited to reason about menus, whereas taxonomic categories supported
reasoning about both properties. The fact that taxonomic and relational categories were
differentially susceptible to the property manipulation demonstrates the interplay between
baseline availability of knowledge (taxonomic knowledge being more available) and property
guiding selective recruitment of knowledge.
In addition, analysis of projections to foods related to premises via both taxonomic
and script categories showed an interesting asymmetry: common taxonomic categories had a
much larger influence on relational inferences than script categories had on taxonomic
inferences. Specifically, the increase in inductive likelihood conferred by taxonomic category
22
membership in the menu condition was consistently stronger than the parallel increase
conferred by script category membership in the enzyme condition. Likewise, according to
regression analyses, taxonomic category membership was more predictive of responses in the
menu condition than script category membership was of responses in the enzyme condition.
This asymmetry may be viewed as taxonomic categories having a stronger effect on relational
inferences than relational categories have on taxonomic inferences; but it can also be viewed
as another level of property effects on induction: property could cue not only the kind of
category (taxonomic or relational) exclusively relevant to a particular inference, but also
whether a single category system or multiple category systems might be relevant.
And again, this ability to differentiate between different types of biasing properties (in
addition to the ability to distinguish between projectable and non-projectable properties) can
be observed at a young age. Using an experimental paradigm similar to Ross and Murphy
(1999), Nguyen and Murphy (2003) showed that property information guides children’s
inferences about food: at the age of 7, they make taxonomic inferences when projecting a
biochemical property (e.g. they expect a novel enzyme found in cereal to be also found in
noodles, another member of “breads & grains” category), but they prefer contextual
inferences when they reason about a situational property (e.g., they expect two breakfast
foods – cereal and milk - to appear together at a novel ceremony).
Coley and colleagues (Coley, 2005; Coley, Vitkin, Seaton & Yopchick, 2005; Coley,
in press) examined the development of property-sensitive reasoning about the biological
world. School-aged children were presented with biological triads of a target species and two
alternatives. A taxonomic alternative was a species from the same superordinate class but
ecologically unrelated to the target; an ecological alternative was related to the target via
23
habitat or predation, but belonged to a different taxonomic class. An example of a triad
could be a banana tree (target), a calla lily (taxonomic match) and a monkey (ecological
match). Children were taught either a taxonomically or ecologically biasing property (stuff
inside, an internal and plausibly stable property, or a disease, a property transmittable through
an interaction in an ecosystem) about the target and asked to pick one of the alternatives that
shared the property. Children were clearly sensitive to the type of property they were asked
about: they picked taxonomic matches for stuff inside and ecological matches for diseases at
above-chance levels. Importantly, follow-up work by Coley (2009) and Coley, Muratore, &
Vasilyeva (2009) relating children’s and adult inferences to adult beliefs about relatedness of
animals in the triads showed that ecological inferences about a disease property were
predicted by the salience of ecological relations in the triad, but unrelated to the salience of
taxonomic relations. Inferences about the stuff inside were sensitive to salience of taxonomic
and ecological relations in a triad, although this pattern varied with age and amount of
experience with nature. This again suggests that different properties lead to recruitment of
different knowledge structures.
Developmental evidence on property effects is not restricted to reasoning about
animals and food. For example, Kalish & Gelman (1992) presented 4-year-old children with
novel artifacts that had double labels consisting of a material and an object kind, such as
“glass scissors”, and asked to generalize different novel properties to other items, such as
metal scissors and a glass bottle. When a property was dispositional (will get fractured if put
in really cold water), 4-year-old children’s judgments were based on the material part of the
label; when the property was functional (used for partitioning), children relied on the object
24
kind part of the label. These results showed that young children have expectations about
what specific information about objects is relevant for generalizing different kinds of
properties.
Mandler &McDonough (1996, 1998) reported sensitivity to property at an even
younger age. Fourteen-month old children distinguished between actions appropriate for
animals (being given a drink) and for vehicles (being open with a key), and were less willing
to imitate actions that were demonstrated on a “wrong” item than when the actions matched
the items – another piece of evidence showing that from an early age, children distinguish
between different kinds of properties and associate them, in this case, with different
knowledge domains (or, as shown above, with different knowledge structure within a single
domain).
Importantly, in addition to argument evaluation and triad selection, property effects
have been observed in an open-ended experimental paradigm. When people are provided with
the premises and the property to be projected, and need to generate conclusions on their own,
the type of conclusions and related justifications depend heavily on the property. For example,
Coley and colleagues (Vitkin, Coley & Hu, 2005; Vitkin, Vasilyeva & Coley, 2007;
Vasilyeva & Coley, 2009) taught school-age children that a pair of animals had a novel stuff
inside or a novel disease, and were asked to make guesses about what other things might
share the property and explain why. Children’s responses were coded as taxonomic (based on
similarity or common kind membership) or ecological (based on interaction or shared
ecosystem). When children were thinking about internal substances, their responses were
more category and similarity based; when they were considering diseases, their responses
tended to be based on ecological interaction. Thus, even when children were not restricted to
25
only a few alternatives provided by the experimenter, their open-ended responses were
guided by different kinds of knowledge based on what property they were reasoning about.
Coley & Vasilyeva (2010; see also Coley, Vasilyeva & Muratore, 2008) took the
open-ended induction paradigm one step further beyond documenting property effects, and
examined how salient relations between animals in paired premises predicted inferences. This
study demonstrated that salient relations can promote inferences of a corresponding type, and
in some cases they can inhibit inferences of mismatching types. With regard to property, not
only did different properties produce different inferences from the very same pair of animals,
but they also influenced to what extent relations among premise categories were driving
inferences. Although no consistent pattern of how exactly property moderates recruitment of
premise relations emerged, and authors acknowledged that “any detailed explanation of this
pattern of results would be a speculation” (p. 215), these results established that property can
systematically affect open-ended inference generation, not just evaluation of complete
arguments, and emphasized the need for a more detailed account of the interplay between
attributes of premise categories and projected properties.
Other types of property effects. A separate line of research focused on studying
inductive problems that involve revising beliefs about properties that are associated with
certain levels of strength or intensity of related attributes, such as can bite through wire
(expected to require certain amount of strength). Smith, Shafir, & Osherson (1993) showed
that many principles describing reasoning about blank properties – premise-conclusion
similarity and premise typicality associated with higher projection likelihood estimates – can
be violated when people reason about certain types of familiar properties. In their study,
participants were more willing to project such property as can bite through wire from a
26
weaker atypical dog species to a stronger one (from toy poodles to German shepherds) than
vice versa or between similar highly typical dogs of roughly equal strength (from Dobermans
to German shepherds). Presumably, since people have an initial belief about how hard it is to
bite through wire, when they are presented with the fact that even weak dogs can do it, this
requires some belief revision to cover the discovered “gap” between the premise and the
property: “biting through wire must not require that much strength, after all”. This finding
adds to the evidence demonstrating that different properties can yield different reasoning
patterns, and shows that inductive inference is a highly interactive process in which a person
may strive to bring all elements of the problem – premises, property, conclusions – to a
maximum congruence, reinterpreting and modifying them to have a coherent story justifying
an inference. However, this line of research departs in many respects from the other work on
property effects, since it focuses on revision of beliefs about the property itself, and the
findings are very specific to the types of properties and arguments used. Acknowledging that
interpretation of any given property may not always be fixed, we will assume for the purposes
of this project that there are cases that do not involve such revision of belief, and that
description of general processing mechanism based on such cases will generalize to other
cases as well.
Another somewhat stand-alone perspective on property effects is represented by work
of Sloman (1994), who focused on properties that allowed for rather elaborate reasoning
involving construction of multiple explanations for a property, and examined how such
explanations influence adults’ estimates of projection likelihood. When there was a single
likely explanation of a property being true of both premise and conclusion, the premise
increased belief in the conclusion (relative to evaluating the conclusion statement in isolation;
27
for example, “P: “many furniture movers have a hard time financing a house”, C: “many
secretaries have a hard time financing a house” [in both cases, likely due to low income]). But
when the conclusion statement suggested a different explanation for the property, its
subjective likelihood decreased (P: “many furniture movers have bad backs” [as a result of
heavy lifting], C: “many secretaries have bad backs” [from sitting over a desk for too long]).
Sloman discusses this effect in the context of explanation discounting – the idea that people
prefer a single explanation for a given phenomenon; when alternative explanations are
present, they compete with each other and mutually decrease each other’s credibility. For the
purpose of our project, however, we would like to point out that Sloman’s work demonstrates
that explanation of the evidence may play an important role in judgments of argument
strength, although the exact role of explanation is not clear. We will come back to this idea
in Chapter 6.
To summarize this section, property effects have been demonstrated in reasoning
about a wide range of domains: biological, social, food, artifacts. Both adults and children
seem to have a notion that properties come in different kinds, and that the relevant criteria for
projecting these properties differ accordingly. Whether a property is considered non-
projectable, or projectable but biasing toward a particular projection direction has a
considerable effect on the outcome of inference. When a property is idiosyncratic and is not
systematically distributed across any known system of categories, information about one
instance having that property does not create the expectation that the same property is true of
other entities as well. However, when a property is expected to be homogeneous within
classes of a certain category system, information about one instance of a class having that
property promotes the generalization to other members of the class. Moreover, different
28
properties can generalize to different classes: based on property, the projections from one
and the same entity can vary dramatically.
Although property effects are pervasive, there has not been much research on the
underlying mechanism. A large majority of studies involving property effects use them as a
tool to study other cognitive phenomena, without questioning how this tool works. The few
studies that discuss the workings of property effects make general proposals about property
affecting recruitment of different subsets of knowledge (Coley, Muratore, & Vasilyeva, 2009;
Coley & Vasilyeva, 2010; Heit & Rubinstein, 1994), with a general reference that it might
have something to do with context-dependent knowledge retrieval (Heit & Rubinstein, 1994).
However, these ideas have not been tied together in a testable proposal about the mechanism
of property effects.
Argument Evaluation vs. Inference Generation
As the above review illustrates, most previous research in inductive reasoning has
involved the evaluation of complete inductive arguments in one form or another. Relatively
recent work started exploring another way of studying induction – by using an open-ended
task of hypothesis generation.
Argument evaluation differs from hypothesis generation in a number of important
ways. In an argument evaluation task participants evaluate hypotheses given to them. This is
akin to a recognition memory task or a multiple-choice exam where people choose the best
answer from a pool of prefabricated answers. In the open-ended paradigm, participants are
presented with the premise and asked to generate their own conclusions (e.g., As have
property X. What else do you think would have property X, and why?). This approach is more
like a recall memory task or an essay exam, and it has a number of advantages.
29
First, we would like to argue that adding the open-ended format to the toolbox of
inferential tasks increases ecological validity of research on induction. The context mimicked
by argument evaluation tasks is not so frequent in everyday life: how often do we have
someone providing us with possible conclusions that can be drawn from the given evidence?
When one day you got sick from eating a cheesecake, did anyone approach you with a survey
packet and ask you to make judgments about potential foods to avoid: “Carrots? Ice-cream?
How about garlic salt?” Such situations are, at best, infrequent. Typically, it is our own task to
generate these hypotheses from the evidence we have. Studying induction exclusively via
argument evaluation potentially misses a large part of the spectrum of induction.
Second, interpretation of argument evaluation results depends on the participant
recognizing and evaluating specific hypotheses intended by the experimenter. In contrast,
open-ended inferences provide greater opportunity to participants to draw on any knowledge
they spontaneously deem relevant. In addition, it allows participants to base their inferences
on multiple salient relations, instead of restraining participants to choosing one or two of
them.
Third, as demonstrated by Coley & Vasilyeva (2010) this approach allows participants
to utilize relatively abstract or inchoate knowledge to guide inductive inference. As Keil
(2003) has shown, our explanatory knowledge and understanding of causal mechanisms are
often much more superficial and vague than we know or would like to admit. Nevertheless,
inductive inferences can bridge gaps in specific factual knowledge—indeed, making uncertain
guesses about the unknown is what induction is all about. Many researchers have suggested
that fairly abstract principles (domain theories, schemata) provide inductive constraints in
concept learning (e.g., Heibeck & Markman, 1987; Keil, 1981) language acquisition
30
(Chomsky, 1980), and inductive inference (e.g., Coley, Hayes, Lawson, & Moloney, 2004;
Coley, Medin, & Atran, 1997; Goodman, 1955). As demonstrated by Kemp, Perfors, and
Tenenbaum (2007), it is possible to learn abstract knowledge from observations before
acquiring specific knowledge at lower levels of abstraction. If so, inductive reasoning may be
especially likely to rely on abstract ideas in the absence of specific knowledge.
For example, if a participant is told that macaws are sick with a certain disease, she
might guess that anything that eats them might be sick with the same disease, even if she has
no idea what might do so. In an open-ended task, she might still be able to articulate an
inference – “Anything that preys on macaws can get this disease from eating the contaminated
meat”. But if she is asked to judge the strength of a complete argument “Macaws have disease
X, therefore monitor lizards have disease X”, she may not know that monitor lizards prey on
macaws, and lizards and macaws don’t seem otherwise similar, so she may rate the argument
as relatively unlikely. Because the participant was unable to apply her relatively abstract
knowledge about disease transmission, she rated the argument as weak when in fact she
believed it to be strong, but just didn’t realize it.
The experimenter in this case would note that arguments based on ecological relations
get low ratings, and may to conclude that it is similarity that drives this participant’s
inferences. Indeed, when Coley & Vasilyeva (2010) used an open-ended format that did not
constrain participants by lack of specific knowledge nor, indeed, “by facts or reality” (p. 220),
the inferences generated by participants showed notable increase in reliance on ecological
relations relative to argument-evaluation studies, which typically demonstrate that taxonomic,
or similarity-based relations dominate in induction. If anything, salient ecological relations
among premises had a stronger effect on taxonomic inferences than vice versa. A sizeable
31
proportion of ecological inferences came from vague, generic inferences. E.g., projecting a
novel substance from the premise pair “leafcutter ant & anteater”, one participant said “an
animal that is a predator to an anteater. I can’t think of any ‘cause I’m not an animal expert.
Anteater eats ants, and they both have this substance. So I assume whatever eats an anteater
will have it too or receive it by eating it.” (p. 218). This example illustrates how people can
generate sophisticated inferences based on general, framework theories, even in the absence
of specific knowledge. The authors argued that “the relative paucity of ecological and causal
reasoning among folk-biologically naïve participants in previous research may be due in part
to the fact that they were being asked to recognize such relations, rather than generate them”
(p. 220).
In sum, this inference generation task is a relatively sensitive measurement instrument
that is able to detect participants’ spontaneous use of different kinds of knowledge that they
deem relevant, in the context of an ecologically valid inductive problem. If our goal is to
examine how property moderates knowledge recruitment in the course of induction, and if the
recruited knowledge is likely to be abstract and/or vague, we may want to prefer a task that is
sensitive to these types of knowledge rather than ignores or masks them. Thus, in this project
we focus on open-ended inference generation as a more promising context for investigating
the underlying mechanism of property effects.
Test Knowledge Domain: Folkbiology
As prior work illustrates, property effects have been documented in multiple
knowledge domains. We have selected one of them as a field for studying property effects -
the domain of folkbiological reasoning, i.e. non-expert, everyday thinking about plants and
animals. This domain is organized by two relatively salient cross-cutting knowledge
32
structures – a taxonomic classification of living things (mammals, birds, reptiles, insects,
fish, etc.) and an orthogonal system of ecological classes (based on habitats – forest animals,
aquatic creatures – or on ecological interaction – predation chains, symbiotic groups). These
two knowledge structures have been shown to be both distinct and psychologically real: as
early as five years of age, children can sort animals into taxonomic groups as well as based on
habitat (Vitkin, Coley & Kane, 2005), and time pressure selectively impairs recruitment of
ecological but not taxonomic knowledge (Shafto et al., 2007) suggesting that these are indeed
two distinct kinds of knowledge that differ both in content and accessibility. Presence of a
limited number of clearly identifiable knowledge structures that cover a large proportion of
knowledge about this domain5, and existence of prior work on property effects in reasoning
about the natural world make folkbiology an attractive domain for investigating the
mechanism of property effects.
Next, we turn to the discussion of existing models and theories of induction and of
how they (fail to) account for the vast evidence on property effects that we have just
reviewed.
5 See Study 2 for additional empirical support for this claim
33
Chapter II. Theories and models of induction
The biggest gap in our understanding of inductive inference concerns its mechanism.
What happens between the time an inductive premise is presented and the moment an
inductive inference is eagerly spit out by a cooperative participant? A number of models of
induction have been proposed, mainly designed to account for basic induction phenomena
such as premise-conclusion similarity (higher similarity between premises and conclusions is
associated with higher perceived argument strength), typicality (premises typical in a lowest
superordinate category that includes premises and conclusion provide better evidence for
conclusion), diversity (diverse premises provide stronger support to a conclusion than non-
diverse premises), and other phenomena (for a review, see Heit, 1997, 2007, 2008).
Unfortunately, until recently, property effects have not been on the list of the phenomena a
model of induction should be able to account for. And even taking into account the most
recent developments, the proposed models are predominantly designed to model inference
outcomes rather than process, and none of them address the question of how property-
sensitive inductive hypotheses are generated. In what follows, we briefly review some
prominent induction models, focusing on their take on property effects, and where available,
review related theory.
Existing Theories and Models of Induction
Similarity-based models. The first formal model of induction was proposed by Rips
in 1975. The model was based on a multidimensional scaling solution for a set of animal
categories and derived similarity and typicality measures that were used to predict strength of
inductive arguments with these animal categories figuring as premises and conclusions. The
model does not specifically address property effects; in fact, since the model relies on a fixed
34
multidimensional scaling solution, the derived similarity measure is also fixed, and cannot
vary with the projected property (as discussed in Heit, 2000).
Osherson, Smith, Wilkie, Lopez, & Shafir (1990) proposed the Similarity-Coverage
Model of induction (SCM), which has been very influential in guiding research on inductive
inference. The SCM describes the strength of an argument as determined by two factors:
similarity between premise and conclusion categories, and the degree of “coverage” provided
by the premises with respect to the lowest-level category that includes both premise and
conclusion categories. The SCM does not consider knowledge structures other than similarity
(such as causal, interaction-based relations), and does not model generalization with different
property types, offering no account of property’s potential to promote or limit generalization.
Sloman’s (1993) Feature-Based Induction (FBI) model also relies on similarity
between premise and conclusion category as the basis for induction. However, it sidesteps the
retrieval of a superordinate category for premises and conclusions and describes projection of
a novel property in terms of calculating featural overlap between known features of premises
and conclusions. Generalization of a novel property is proportional to the number of
conclusion features also observed in the premises. It can account for varying argument
strength if shared causal features are assumed to have higher weights, however, as proposed,
it has no mechanism for flexible computation of similarity based on the nature of projected
property.
Sloutsky & Fisher’s (2004) Similarity, Induction, Naming, and Categorization (SINC)
model is another similarity-based proposal, in which similarity is based both on perceptual
features and common labels shared by premise and conclusion terms. Like other similarity-
35
based models, it does not provide any account of how property knowledge could influence
induction.
Hypothesis-based models. The Hypothesis-Assessment Model of Categorical
Argument Strength by McDonald, Samuels, & Rispoli (1996) proposes that people “actively
construct hypotheses in response to uncertain conditions and evaluate the plausibility of these
hypotheses in light of available evidence” (p. 202). Premises of an inductive argument are
viewed as evidence, and the conclusion as a hypothesis. The authors propose that the strength
of an inductive argument is determined by the number of competing hypotheses brought to
mind by the evidence: the more alternatives are available, the weaker is any given hypothesis.
Although this model is open to the possibility of property effects in induction (presumably by
virtue of affecting the number of hypotheses generated from the evidence), it does not provide
a specific explanation of how property effects would arise. It does, however, emphasize the
role of hypothesis generation in induction (see also Feeney, Coley, & Crisp, 2010, for
supporting evidence that people spontaneously construct and evaluate hypotheses as they are
processing premises of an inductive argument).
Medin, Coley, Storms, & Hayes (2003) proposed the Relevance Theory of Induction
based on the idea that people assume premises to be relevant to the conclusion(s), and treat
distinctive properties of premise categories as likely bases for induction. In contrast to
similarity-based models relying on fixed notion of similarity that assume reasoning about
blank predicates (property X), the idea of flexible reliance on variable relevant subsets of
knowledge about premise categories, be it similarity-based or causal or of some other kind,
lies at the core of this proposal. In fact, it suggests that even with uninformative properties,
people might seek to interpret them and extract some meaning for the property from the
36
salient features of premise categories. Thus, this theory has everything necessary to
accommodate reasoning about informative predicates, although it does not provide a process-
level description for it.
Bayesian models. A class of Bayesian models makes a useful contribution by
explicitly distinguishing prior knowledge that needs to be in place for property-sensitive
induction to take place from the operations over that knowledge.
The key assumption of a Bayesian model proposed in Heit (1998) is that in projecting
a novel property people rely on prior knowledge about distribution of familiar properties. If a
large number of shared features is already known for the premise and conclusion in question,
this supplies high prior probability to a hypothesis that a novel property is shared as well. The
information about premises in an argument is treated as a new piece of evidence, which is
used to revise prior beliefs using Bayes’ theorem. After the beliefs are revised, the plausibility
of the conclusion given the premise is evaluated.
In contrast to the similarity-based models reviewed above, this Bayesian model allows
for property effects, although it does not specifically model them. Heit proposed the following
mechanism for property effects: the projected property serves as a cue for retrieving familiar
properties. Retrieving different subsets of properties from long-term memory provides
different priors, potentially resulting in different inferences. For example, if a projected
property sounds vaguely biological and internal (“has an omentum inside” or “has a two-
chambered heart”, as in Heit & Rubinstein, 1994), the retrieved familiar properties will also
be of the same nature, and inferences to biologically related conclusions (from hawks to
chickens) will become likely – because people already know a relatively large number of
biological internal properties true of both hawks and chickens. In contrast, if the projected
37
property is behavioral, such as “prefers to feed at night”, the prior hypotheses will be
drawn from distribution of familiar behavioral properties, promoting inferences to
behaviorally similar animals, such as from hawks to tigers, again, because people already
know of other behavioral properties shared by tigers and hawks.
Although driven by the goal to explain patterns observed in human data, this model, as
noted by Heit, was designed to provide a computational-level analysis of normative reasoning
with a hypothesis space, rather than a processing account of inductive inference.
The model proposed by Tenenbaum, Kemp & Shafto (2007) takes the Bayesian
approach a step further by focusing on the contribution of structured knowledge. This
modeling work explicitly specifies how a single general-purpose Bayesian reasoning
mechanism combined with different knowledge structures can lead to different patterns of
generalization behavior in different inductive contexts. This proposal emphasizes the role of
different structural representations, or relational systems of categories such as taxonomic
hierarchies or food webs in folkbiology. Knowledge in this form provides prior probabilities
for a domain-general Bayesian inference engine which drives inductive inference. Since
people can draw on different prior knowledge structures within a single domain, depending on
which of these knowledge structures is triggered, very different patterns of inference arise.
Tenenbaum et al. demonstrated that their model can strongly predict people’s
judgments in reasoning about two different types of properties, but only when the model is
given an appropriate knowledge structure matched with the property. To predict people’s
reasoning about generic biological (anatomical and physiological) properties, the Bayesian
inference engine had to draw upon a taxonomic knowledge base; while when the same
general purpose inferential engine was given an ecological knowledge base, it predicted
38
people’s reasoning about diseases. Kemp, Perfors & Tenenbaum (2007) extended this work
by demonstrating that the multiple knowledge structures do not have to be specified in
advance but can be learned by a hierarchical Bayesian model from data. However, as Kemp et
al. (2007) mention, the explanation provided by such modeling work is at the level of
computational theory (Marr, 1982), and does not specify the psychological mechanisms by
which the model could be implemented.
In sum, similarity-based models like SCM, FBI or SINC focus on accounting for a
limited set of phenomena and ignore property knowledge: it is not clear how such knowledge
could be incorporated into these models without major revisions. Hypothesis-based
approaches allow for property effects, but do not specify how they can be implemented.
Bayesian models offer a promising perspective and allow property to moderate induction by
positing separable knowledge structures that make contact with informative properties,
yielding different reasoning in different induction contexts (Tenenbaum et al., 2007, Kemp et
al., 2007). However, Heit (1998), Tenenbaum et al. (2007) and Kemp et al. (2007) are explicit
in that their models are computational-level models that make no commitments about
processing details such as the kind and time-course of involved psychological processes.
Tenenbaum et al. (2007) write that it remains an open question how theories “are acquired
and selected for use in the particular context” (p.200), and “the most immediate gap in our
model is that we have not specified how to decide which theory is appropriate for a given
argument” (p.200). These gaps and open questions represent the current state of
understanding of context-sensitive inductive inference.
This absence of process models of inference generation, and of proposals about
psychological mechanisms giving rise to property effects in inference generation is the
39
starting point for this project. The following is an attempt to fill these gaps. We propose a
model of inductive inference generation and use it to guide investigation of processing
underlying property effects.
The main question of this project is how does property exert its effect on inferences?
What is the relevant mechanism? We tackle this question at two levels. First, starting from the
higher level of description, we propose a (fairly minimalistic) computational account of
hypothesis-generation, addressing the question of what tasks need to be accomplished in order
to generate an inductive inference. Second, we examine the mechanism, i.e. describe how
these tasks are accomplished by the cognitive system. We break this step into two sub-tasks:
first, to catalog the “ingredients” that go into preparing an inference, i.e. available knowledge,
its sources and types; second, describe the “recipe”, i.e. propose and evaluate operations over
this knowledge.
A (Minimalistic) Computational Account of Inference Generation
What goals need to be accomplished in order to generate an inductive inference?
When a person is presented with some starting evidence, e.g. that a certain parasite is found in
ducks (a certain property is true of a certain premise category), and asked what else might
have the disease and why, what tasks does the cognitive system need to solve in order to
arrive to a formulated inductive hypothesis at the output: “foxes, because they eat ducks”, or
“eagles, because they are also birds”? This proposal is based on the starting assumption that
the hypotheses are related in a systematic way to the inductive problems (see Coley &
Vasilyeva, 2010 for supporting evidence). This assumption shapes the first task for the
inference-generating system: it needs to understand the incoming evidence (what “duck”
means; what “parasite” means; etc.). Second, in order for the outcome to count as a novel
40
hypothesis rather than a restatement of the question, it needs to go beyond the given
information, i.e. to attribute the property in question to an entity or set of entities not already
claimed to have this property in the premise (“foxes”, or “eagles”, or “other birds”). These
two goals define two major components of the proposal: knowledge retrieval and hypothesis
generation (see Figure 1).
Figure 1. Tasks that a cognitive system needs to accomplish in order to generate an inductive inference6.
Before examining the mechanism, we will describe the “ingredients” that go into
preparing an inference, i.e. available knowledge, its sources and types. Next, we will describe
the “recipe”, i.e. propose and evaluate operations over this knowledge. At this stage, we will
make more specific claims about how “knowledge retrieval” and “hypothesis generation” may
be accomplished.
Ingredients: Sources and Types of knowledge
Hypothesis generation is knowledge driven; when one learns that A has a novel
property X, one uses what they know about A and its relations to other things to form guesses
6 This part of the proposal makes no claims about the mechanism (stages, order, sequential or parallel processing, feedback, etc).
41
about what else is likely to have the property. As soon as the “A” in this example is
replaced with a familiar term: duck, lawyer, or bacon – large amounts of information become
available as potential guides for inductive inference. In other words, one source of knowledge
that serves as input to inductive inference is knowledge about premise categories. When such
categorical knowledge is accessed, a probabilistically determined subset of features and
relations that comprise the representation of that concept becomes available as a raw material
for the inference. For example, if A turned out to be a duck, such features as “is a bird”,
“flies”, “lives in ponds”, “quacks”, “poops around Fenway”, “eaten by foxes”, “tasty when
grilled with apples” may come to mind.
Although there are many different kinds of knowledge, researchers hav divided
knowledge about living things into two broad classes: taxonomic categories based on intrinsic
similarity of members, e.g. “birds”, and contextual categories, based on extrinsic relations
between members and other entities (including extrinsic similarity and causal interactions),
e.g. “aquatic animal”, “prey”. For example, ducks at the same time belong to a taxonomic
category of birds and contextual categories of aquatic animals and prey to their predators.
Each of these types of knowledge can serve as a basis for projection from ducks – to other
birds, or other aquatic animals, or things that eat ducks. In fact, the problem is typically not
that there are not enough bases for inferences, but that there are too many, and describing
selection among many possible bases for projection is one of the major issues in psychology
of induction. (Goodman, 1955, 1972; Keil, 1981).
In addition to the premise category, knowledge about the property can also serve as a
source of information. If X in the example above is replaced with an informative property –
having a certain disease, personality trait, or fat content – this adds more information to the
42
problem. Independently of what we know about ducks, we also know something about
diseases: what they are like, how they are transmitted, etc. Somewhat counterintuitively, this
does not make the problem worse but helps to constrain the set of relevant information. The
critical point here is that this information comes from a different source, and different sources
of information can constrain each other, if they match in type. In keeping with the existing
literature, we will classify properties as taxonomic or contextual. As such, the property type
(e.g. disease, an ecological property) could indicate which subset of knowledge about ducks is
most relevant for inference generation.
As mentioned above, existing similarity-based models do not take into account
property as an additional source of knowledge, and/or ignore that knowledge comes in
different types (e.g. the Osherson et al. SCM addresses premise category knowledge only, and
only one type of premise knowledge, taxonomic; Sloman’s FBI and Sloutsky & Fisher’s
SINC do not account for different types of knowledge).
In sum, we propose to expand the focus of attention and to consider two orthogonal
attributes of knowledge: first, their source (premise category vs. property), and second, their
type (taxonomic vs. contextual).
Recipe, or Mechanism
Knowledge retrieval vs. Hypothesis generation: where do we start looking for
property effects? The two tasks identified earlier - knowledge retrieval and hypothesis
generation – offer two possible loci of property effects (not mutually exclusive). On the first
look, hypothesis generation is an appealing candidate - after all, the property effects are
observed in the produced hypotheses. However on the other hand, the state of knowledge
about the process of hypothesis generation can be most optimistically described as “close to
43
nothing”. This is not to claim that there has been no work done on hypothesis generation –
there has been. However, the specific ways the questions are framed and addressed in this
research make it less applicable to our case of interest. In most cases, what is studied under
the label of “hypothesis generation” actually applies to rule discovery in contexts with limited
set of applicable rules and a correct answer, such as the Wason selection task (e.g., Farris &
Revlin, 1989). Other questions studied under the rubric of “hypothesis generation” include
choice of hypothesis testing strategies and a related topic of confirmation bias, hypothesis
evaluation, subjects’ meta judgments about exhaustiveness of their sampling of hypothesis
space, completeness of generated hypotheses, subjects’ confidence judgments (Gettys &
Fisher, 1979; Klayman & Ha, 1989; McDonald, 1990; Mehle, 1982; Mehle, Gettys, Manning,
Baca, & Fisher, 1981) - pretty much everything except the process of hypothesis generation
itself. This leaves the hypothesis generation component of our model a rather mysterious
entity, and makes it less attractive in terms of its explanatory potential.
In contrast, knowledge retrieval is known to be affected by context. As McElree,
Murphy & Ochoa (2006) discuss, it is impossible to retrieve all known facts about every word
every time it is encountered; and even if it were possible, most of this information would be
irrelevant in any given situation and could impede reasoning rather than aid it. Indeed, is has
been shown that people tend to retrieve conceptual information selectively, based on the
context.
For example, Tabossi & Johnson-Laird (1980) presented participants with sentences
that emphasized different aspects of meaning of an embedded word: The goldsmith cut the
glass with the diamond or The mirror dispersed the light from the diamond. After reading one
of the sentences, participants were presented with a feature-verification task. Context-relevant
44
features (such as hard as a feature of diamonds, after reading the first sentence) were
verified faster than other true, but context-irrelevant features (e.g. brilliant).
In a study of text memory by Barclay, Bransford, Franks, McCarrell & Nitsch (1974),
after participants had been presented with such sentences as The man lifted the piano or The
man tuned the piano, in a cued memory test, something heavy was a better cue for the first
sentence than was something with a nice sound, and the reverse was true when the second
sentence was cued. This suggests that during encoding of sentences, one context caused
participants to selectively retrieve the weight of the piano, and the other context caused them
to retrieve the sound.
Work by Barsalou (1982) demonstrated context effects on knowledge retrieval in tasks
involving making judgments about categories, such as attribute verification and similarity
judgments. Results showed that people retrieve a subset of that category’s features, and this
subset is determined by the context. For example, roofs do have an attribute “can be walked
upon”, but whether it is retrieved or not when a person hears “roof” largely depends on the
context in which roof is discussed, as suggested by the attribute verification times varying
with the context.
Such context-sensitive retrieval of information has been well-documented in a variety
of experimental paradigms: sentence verification (McKoon & Ratcliff, 1988; Tabossi, 1982;
Tabossi & Johnson-Laird, 1980, Glucksberg & Estes, 2000), lexical decision (Tabossi, 1988;
Swinney, 1979) and naming (Hess, Foss, & Carroll, 1995), all showing that context focuses
memory access on relevant features of a concept.
In an inductive problem that consists of a premise category(-ies) and a to-be-projected
property, property could be viewed as context for the premise(s). A premise category label, as
45
any word, is connected to a vast amount of conceptual knowledge. When the information
about a premise category is retrieved from long-term memory, this process may be no
exception to context effects described above. This opens a possibility that property effects, or
differential projections dependent on the property, are an outcome of property-guided
selective retrieval of information about premises. For example, when an inductive problem
presents “duck” in the context of a taxonomic property (“ducks have gene X”), one may be
more likely to retrieve taxonomic knowledge about ducks (bird, have feathers); while in the
context of an ecological property (“ducks have parasite X”), one may be more likely to
retrieve ecological knowledge about ducks (aquatic, prey to foxes).
Thus, in the quest for the mechanism of property effects, our first-choice strategy is to
avoid attributing any major explanatory role to some mysterious processes taking place inside
the black box of hypothesis generation, and try to formulate and test a minimal account in
which property exerts its effect on inference by moderating knowledge retrieval. This
proposal is consistent with the intuitions outlined by Heit & Rubinstein (1994) and the general
idea of the Bayesian model proposed by Heit (1998). We hope to advance our understanding
of property effects by making our proposal more specific than that of Heit & Rubinstein
(1994) and stating and testing it in the context of a descriptive process model of inductive
reasoning.
Model of induction, v.1.0: Knowledge retrieval as a locus of property effects.
Knowledge retrieval. As discussed above, when the premise is presented, one needs to
understand it enough to generate a hypothesis. Understanding requires retrieval of prior
knowledge about premise categories and/or properties from long-term memory. By retrieval
of knowledge about a premise category or property we mean activation of a corresponding
46
lexical node and, via spreading activation, subsequent activation of nodes representing
related knowledge, with the likelihood of passing the activation threshold determined by the
strength of connection to the starting lexical node, and, as outlined above, the context of
retrieval. The specifics of context-moderated retrieval – via changes in baseline activation of
context-congruent nodes or inhibition of context-incongruent nodes, or both – will be
examined in this project, as discussed below.
In the following, we discuss both what is retrieved and how knowledge retrieval may
proceed.
Presumably, one would attend to both components of the premise statement and
retrieve both premise category and property knowledge. In this case, the retrieval of premise
and property knowledge may proceed either independently or interactively.
In the independent retrieval, the way the premise category knowledge is accessed does
not depend on the accompanying property, and vice versa. In this case, the same knowledge
about ducks would be accessed regardless of whether they were said to have a flu, a gene,
some unspecified property X, or were presented in isolation, and the same knowledge about
flu would be accessed regardless of whether it was stated true of ducks, elephants, people or
crickets, or were presented in isolation. The outcome of retrieval can be described by a simple
additive function combining salient knowledge about premise, and the salient knowledge
about property.
In contrast, in interactive retrieval, accessing knowledge about premise category is
moderated by the property, and/or vice versa. E.g., the extrinsic and ecologically transmitted
property of having a flu may direct one more towards selective retrieval of knowledge about
duck’s habitat and interaction with other animals than the intrinsic property of having a
47
certain gene would. Moreover, the interaction could involve inhibition of incongruent
knowledge (e.g. taxonomic vs. contextual). In this case, the outcome of retrieval will be
described by a multiplicative function reflecting potential mutual enhancement between
congruent property and premise knowledge or attenuation between incongruent property and
premise knowledge (knowledge about premise x knowledge about property).
Alternatively, one may retrieve property knowledge only (what they know about flu
infections), ignoring the premise. Such retrieval strategy would yield property effects in
inferences, but inferences would be unrelated to knowledge about specific premise categories.
Finally, in principle, even incomplete understanding of the premise may be sufficient
for inference generation. That is, if one told that a flu is found in ducks and asked to guess
what else may have it, one may focus on premise category knowledge only and retrieve what
they know about ducks, ignoring the property. If participants were consistently doing this, we
would observe inference showing no signs of property effects. Since extensive evidence
suggest that inferences are sensitive to property, and since our goal here is to propose a model
that would be able to account for property effects, we are not going to consider this scenario
here.
At this point, we need to make a distinction between knowledge recruitment and
knowledge retrieval. Knowledge recruitment describes what knowledge actually ended up
being used in the inference. It describes the relation between the input and the outcome.
Knowledge retrieval, in contrast, refers to the process of activating knowledge relative to
baseline, regardless of the consequences this knowledge had (or did not have) on the outcome
hypotheses. This distinction is supported by work showing that effects of constraining context
on knowledge retrieval can be delayed, resulting in initial activation and later discarding of
48
knowledge subsets (Swinney, 1979). The scenarios above are about knowledge retrieval.
Knowledge retrieval, however, necessarily constrains knowledge recruitment. Thus, in the
following studies, we examine both knowledge recruitment (Study 3) and retrieval (Study 4)
in order to determine which retrieval strategy is used.
The knowledge retrieval step leaves one with some active (although possibly not
detailed and/or vague) knowledge about premise features (a duck is a bird with a yellow bill,
and is eaten by foxes), the property (a flu is an infection that spreads via contact; a gene is
some tiny thing inside that determines how one looks), or both. This knowledge is then used
to generate an inference.
Hypothesis generation. In section 4.1 we stated that our first-choice approach is to
look for property effects in knowledge retrieval, rather than in hypothesis generation. In other
words, our current proposal is that there is not much action during hypothesis generation that
can enlighten us about the mechanism of property effects. At this point it may be a good time
to clarify a. what exactly we mean by “not much action”, and b. that there are alternatives that
we are aware of but not considering (yet).
Generating a hypothesis involves going beyond what is given by making a prediction
about the unknown. In our minimal account, this may be achieved without much extra work –
the active knowledge about the premise and/or property can be directly translated into an
inference using a formula “whatever else has the same known features as the premise” (other
birds; other things with webbed feet), or “whatever can be related to the premise in the
property-relevant way” (“things that come in contact with ducks”, “animals that eat ducks”).
When hypotheses are generated by such “direct translation”, we expect the resulting
inferences to reflect the knowledge about the premise retrieved earlier. If both the property
49
and premise category knowledge has been retrieved, we expect the inferences to be a
function of both components of the premise – additive or interactive. If property knowledge
only has been retrieved, we expect the inferences to be predominantly generic, i.e. not specific
to the given premise category (e.g. “whatever eats them or come in contact with them”).
These three possible strategies of knowledge retrieval in the course of induction are shown in
Figure 2.
Figure 2. Model of induction, v.1.0: Knowledge retrieval as a locus of property effects. (The “premise category only” retrieval strategy is not shown, since it would not yield property effects).
When it comes to alternatives to simple “direct translation” during hypothesis
generation, the way in which retrieved knowledge is used to generate a hypothesis could be
rather complicated. As suggested by Sloman’s (1994) work, explanation of evidence may play
an important role in evaluating inductive arguments. Generation of inductive inferences may
not be an exception. But we will save this discussion until Chapter VI.
The core two studies of this project investigating property effects in induction –
Studies 3 & 4 - were driven by the proposed model stating that inductive hypotheses are
50
produced in two steps: first knowledge is retrieved, and then it is used to formulated the
hypotheses. With regard to property effects, the main claim of the model is that they take
place in knowledge retrieval: property moderates what subset of knowledge about premise
categories gets retrieved, and then this biased set of retrieved knowledge is used to generate
an inference in a relatively straightforward way via direct translation. Within this model, we
distinguish three possible scenarios of knowledge retrieval: only property knowledge is
retrieved, both property and premise knowledge is retrieved independently, both property and
premise knowledge is retrieved interactively. Staying within the limits of the model, we are
going to test all three scenarios.
The overall approach is to see how far we can push this relatively simple model
(plausibly localizing the source of property effects in knowledge retrieval, and postulating the
simplest possible mechanism of hypothesis generation) in accounting for property effects in
inference generation – until it either provides a satisfactory description of the phenomenon, or
fails to do so, which would require revision of the theory.
Overview of the Studies
Studies 1 and 2 together provide a detailed description of the knowledge recruited by
the inferences in terms of its sources (property, premise category) and type (taxonomic,
contextual (ecological)). These two studies address the question “What knowledge serves as
input to inference-generation processes?”.
Study 3 documents property effects in inference generation and examines the
relationship between the knowledge about premises (input) and the generated inferences
(outcome). In other words, the goal of this study is to specify the function relating stimuli to
51
responses (knowledge recruitment), which would provide initial constraints on the retrieval
strategy that could have produced such a function.
Study 4 examines the time-course of knowledge retrieval in the course of inference
generation, focusing on premise category knowledge.
Study 5 subjects to test a revised proposal about the locus of property effects.
52
Chapter III. “Ingredients”: Knowledge types and sources
This chapter addresses the question “What knowledge serves as an input to inference-
generation processes?” Studies 1 and 2 describe the ingredients available at the outset of an
inference: the types and sources of knowledge brought about by initial evidence, i.e. the
inductive problem. Because we focus here on inference generation, this will consist of a
premise made up of a category and a property, for example “Ducks have gene X”. The goal is
to provide a detailed description of knowledge recruited by the inferences in terms of its
sources (property, premise category) and type (taxonomic, ecological).
Study 1. Property knowledge: distributional beliefs
Prior research reviewed in Chapter I shows that people tend to reason differently about
such properties as possessing a gene vs. disease: genes tend to be projected to other members
of taxonomic categories, while diseases tend to be projected to ecologically related targets.
Based on such inference patterns, properties like gene have been labeled as taxonomic, while
properties like disease have been labeled ecological, or contextual. However, such definitions
are circular: it is not clear what makes a property taxonomic or ecological, besides the
inferences people make about them. What is the source of the taxonomic or ecological bias
brought about by the property?
To clarify from the beginning, we are not talking about the cases when knowing the
property would give away the answer; if a person happens to know whether pigeons actually
have a liver with two chambers that act as one or not, we are not dealing with inductive
inference, or reasoning under uncertainty, anymore. The cases we are interested in are when a
person knows something about the property, but does not know the answer to the inductive
question.
53
Another possibility we would like to exclude from the outset is that property is
likely to suggest potential projection targets directly. In other words, if the property of having
a gene gets projected from a duck to a pigeon, and the property of having a disease gets
projected from a duck to an otter, we claim that in the majority of such cases, the reason is not
that gene was directly associated with pigeons, and diseases were directly associated with
otters.
Instead, we would like to test the possibility that what property brings to the table is
information about its distribution. And crucially, people’s distributional beliefs about different
properties can tie them to different category systems, such as a classification of animals that
groups them into species, taxa, etc., and an ecological classification of animals that groups
animals from the same habitat and/or animals that participate in the same food chain into the
same category. For example, people can believe that genes are homogeneous within
taxonomic classes (fish tend to have similar genes), while diseases are homogeneous within
ecological classes (jungle animals tend to share diseases, or predators and prey tend to share
diseases). This association between properties and classification systems can, in turn, be the
source of bias that properties exert on inferences.
One part of this idea – about properties being associated with distributional beliefs –
comes from Goodman’s (1955) proposal about overhypotheses, or abstract beliefs describing
the scope of properties (see Chapter I for a more detailed discussion). According to
Goodman, having a belief that category members tend to be homogeneous with respect to
some property is a prerequisite for generalization. Nisbett, Crantz, Jepson, & Kunda (1983)
provided empirical support for this claim, by demonstrating that when people believe that a
category is homogeneous with respect to one property type, they generalize willingly from
54
one observation (e.g. skin color is generalized to the entire tribe based on one observation
of a brown-skinned tribesman), but they treat a property as non-projectable if the absence of a
belief that such property would be homogeneous within a given category (e.g. they do not
generalize the property of being obese from one observation to the whole tribe). This,
however, speaks only to what makes a property projectable or not, but not to how different
projectable properties can be biased towards one type of categories (e.g. taxonomic classes)
rather than others (e.g. ecological categories).
Yet, in principle, could overhypotheses explain a taxonomic or ecological bias among
different projectable properties? Yes, if we allow for multiple-category membership - which is
characteristic of most categorical knowledge. For example, a duck belongs to a category
birds, which is a part of a taxonomic classification of animals. A duck is also an aquatic
animal, which is an ecological class. If we allow for a belief that some properties are more
homogeneous within taxonomic classes, and other properties are more homogeneous within
ecological classes, such distributional beliefs would provide a bridge between properties and
inductive selectivity in inferences.
Indeed, the relatively recent modeling work of Tenenbaum, Kemp & Shafto (2007)
reviewed in Chapter I suggested that people might draw upon different subsets of their
categorical knowledge in reasoning about different properties. When an inductive problem
contains an informative property, it may serve as a cue towards the appropriate distributional
overhypothesis, indicating which knowledge structure is relevant for projection. As
Tenenbaum et al. report, a general-purpose Bayesian model provided with a taxonomic
category system succeeded in predicting human reasoning about a gene (an intrinsic property)
but failed to predict human reasoning about a disease (a causally transmitted property),
55
whereas a model provided with an ecological system of categories based on food-web
relations between the very same set of animal species succeeded in predicting human
reasoning about a disease but failed to predict reasoning about gene. Kemp, Perfors, &
Tenenbaum (2007) demonstrated that orthogonal knowledge structures can be learned by a
hierarchical Bayesian model from observations, and upon learning such model was able to
accurately predict property-sensitive reasoning of participants making inferences about such
properties of social categories as skin color vs. body weight (data reported in Nisbett, Krantz,
Jepson & Kunda (1983)). Critically, the learning success of the model was conditional on its
hierarchical component, having an extra layer of hypotheses – the overhypothesis level. An
alternative model that was unable to incorporate overhypotheses was making identical
inferences about skin color and body weight.
In sum, we have theoretical description of what sort of information provides a basis
for property projectability (Goodman’s “overhypotheses”) with initial supporting evidence
(Nisbett et al., 1983), and we have a set of modeling results suggesting that human reasoning
can be modeled if different properties are allowed to be associated with different
distributional beliefs. However, the latter claim is an inference from a modeling result that has
not been previously demonstrated directly.
This study addresses this gap by empirically testing whether people have specific
beliefs about the distribution of hypothetical properties and whether properties differ in terms
of their expected distribution across taxonomic and ecological category systems. To address
these questions, we presented participants with a range of novel properties. Participants were
told that a certain (unspecified) animal species has one of these properties, and asked to
estimate a) the percentage of other animals belonging to the same biological family, and b)
56
the percentage of other animals living in the same area that are likely to share that property.
We chose to phrase the ecological estimate question broadly in terms of shared habitat, rather
than in terms of a variety of more specific ecological relations – predation-based, symbiotic,
parasitic, etc. For most ecological relations between animals, common location is a
prerequisite; thus whichever specific ecological relations affect estimates of ecological
distribution of properties, the shared habitat is likely to capture them.
If people believe that distribution of a particular property corresponds to taxonomic
categories (i.e. members of a given biological class tend to be relatively uniform in their
properties), then upon learning about one species possessing a property, participants should
estimate that a high percentage of other animals in the same biological family have this
property as well. If people believe that a particular property clusters within eco-systems (i.e.
animals within the same eco-system tend to share properties), then participants should
estimate a high percentage of animals in the same area to share the property. If people believe
that properties can differ systematically in their distribution, with some properties being
taxonomically-biased (distributed predominantly along taxonomic category system), while
other properties are ecologically-biased (distributed predominantly along ecological category
system), the former should receive higher estimates of being shared by animals in the same
biological family than animals living in the same area, and the pattern should be reversed for
the latter.
Method
Participants. Twenty-six Northeastern University undergraduates were recruited from
introductory psychology classes and participated in exchange for course credit.
57
Materials. We selected a range of properties that could be plausibly attributed to
animals (since we limited the domain of investigation to reasoning about biology). Some of
these properties were predicted to be taxonomic (either based on prior research or on
subjective judgment of the experimenters), some were predicted to be ecological, and some
neutral. There was a total of 13 properties: 4 expected to be taxonomic (has gene X, has X-
cells, has enzyme X, has hormone X), 6 expected to be ecological (has disease X, has bacteria
X, has flu X, as infection X, has parasite X, has virus X) and 3 neutral, or non-biasing (has
property X, has substance X in the bloodstream, a fact X is true of [one animal species]).
Design. Every participant was presented with all 13 properties, each combined with an
arbitrary letter-digit code, such as Gene X5, Disease Z7, Property A3. Participants provided
one estimate of taxonomic and one estimate of ecological distribution for each property. The
order of properties and the estimates was randomized for every participant.
Procedure. The experiment was presented as a computerized survey. General
instructions introduced participants to the format of the task, explained main terms used in the
questions (species, biological family, area where an animal lives), and instructed to treat each
question as independent. In the main part of the experiment, participants were presented with
26 questions, one at a time. Every property was presented twice: once in the context of a
taxonomic estimate question, once in an ecological estimate question. The questions were
phrased as follows: “Imagine that one animal species has gene Z8. What percentage of
animals IN THE SAME BIOLOGICAL FAMILY is likely to have gene Z8? Enter a number
between 0 & 100.”; “Imagine that one animal species has Q1-cells. What percentage of
animals LIVING IN THE SAME AREA is likely to have Q1-cells? Enter a number between 0
& 100.”, etc (for the general instructions introducing participants to the format of the task and
58
clarifying the main terms – animal species, biological family, area where an animal lives,
see Appendix A). Total time to complete the experiment was 10-15 minutes.
Results
This study examined distributional beliefs about properties. The main question was
whether people believe that some properties are relatively homogeneous within taxonomic
categories, while other properties are relatively homogeneous within ecological categories. To
address this question, we calculated two average percentage estimates for every property, one
indicating the percentage of animals in the same biological family and the other one the
percentage of animals living in the same area likely to share that property. The average
percentage estimates for all properties are shown in Table 1.
Table 1. Mean percentage estimates of animals in the same biological category and same habitat that share provided properties (*p<.05, **p<.01, ***p<.001, ****p<.0001). Comparisons to 50% (⇑ or ⇓ - above or below 50%, p<.05; ↑ - above 50%, p<.08) % shared by
animals in the same biological
family
% shared by animals living in
the same area
Taxonomically distributed properties Gene 67.6⇑ 31.4⇓ **** Cells 69.5⇑ 33.0⇓ **** Hormone 67.5⇑ 34.5⇓ **** Enzyme 66.6⇑ 44.4 ** Fact 56.9 37.7⇓ * Ecologically distributed properties Flu 30.0⇓ 60.2↑ **** Infection 30.0⇓ 58.9↑ **** Parasite 33.3⇓ 56.4 *** Virus 32.1⇓ 53.9 *** Bacteria 38.1 57.8↑ ** Disease 33.2⇓ 49.3 * Non-biased properties Property 59.2↑ 49.0 n.s. Substance 52.5 44.4 n.s.
59
As one way to determine whether for each property there was a systematic bias of
higher homogeneity estimates for taxonomic categories than ecological or vice versa, we
compared the two estimates for each property in a series of paired tests. For the following five
properties, taxonomic estimates were higher than ecological: having a gene (t(25)=5.58,
p<.0001, Cohen’s d=1.1), having cells (t(25)=7.54, p<.0001, Cohen’s d=1.48), having a
hormone (t(25)=5.36, p<.0001, Cohen’s d=1.05), having an enzyme (t(25)=2.96, p=.007,
Cohen’s d=.58), and, surprisingly, a fact true of a given species (t(25)=2.69, p=.013, Cohen’s
d=.53). For six properties, ecological estimates were higher than taxonomic: having a flu
(t(25)=6.78, p<.0001, Cohen’s d=1.33), having an infection (t(25)=5.22, p<.0001, Cohen’s
d=1.02), having a parasite (t(25)=4.34, p=.0002, Cohen’s d=.85), having a virus (t(25)=3.93,
p=.0006, Cohen’s d=.77), having bacteria (t(25)=3.42, p=.002, Cohen’s d=.67), and having a
disease (t(25)=2.63, p=.015, Cohen’s d=.52). And finally, for two properties the biological
family estimates did not differ from habitat estimates: having a property (t(25)=1.41, p=.17,
Cohen’s d=.28), having a substance in the bloodstream (t(25)=1.07, p=.29, Cohen’s d=.21).
As another way to determine taxonomic or ecological bias, we compared the estimates
to 50% mark. Of course, there is no a priori chance level of property occurrence in any given
biological category. Depending on within-category variability, even the most prevalent
properties may not be present in 100% of all members. However, since 50% can serve as a
psychological marker for being present in a majority of category members, we took 50% as a
reasonable anchor to compare assigned percentages to. The overall pattern of the comparisons
to 50% was consistent with the paired tests reported above (although some comparisons did
not reach significance). As shown in Table 1, among the five properties that had higher
60
taxonomic than ecological estimates – having certain cells, gene, hormone, enzyme, or a
fact true of a species – 4 out of 5 properties received taxonomic estimates above 50% and the
ecological estimates below 50%. Among the six properties that had higher ecological than
taxonomic estimates – having a flu, infection, parasite, virus, bacteria or disease – 5 out of 6
were had taxonomic estimates below 50% and 3 out of 6 had ecological estimates marginally
above 50%. The two properties that received non-differing taxonomic and ecological
estimates – having an unspecified property and having a substance in the bloodstream – had
estimates that did not differ from 50%, although an unspecified property was believed to be
shared by 59% of animals in the same biological family, which was marginally above 50%
(one sample t-test, t(25)=1.87, p=.07, Cohen’s d=.367).
Based on these two analyses, we categorized one set of properties as taxonomically
biased, or more likely to be shared by the animals in the same biological family (>50%) than
by animals in the same habitat (<50%): having certain cells, gene, hormone, enzyme, or a fact
true of a species. Another set of properties was categorized as ecologically biased, or more
likely to be shared by the animals in the same habitat (>50%) than in the same biological
family (<50%): having a flu, infection, parasite, virus, bacteria or disease. Finally, two
properties were categorized as neutral, or non-biasing, i.e. equally likely to be distributed
along taxonomic as well as ecological category systems (for each category, expected to be
shared by about half the category members): having an unspecified property and having a
substance in the bloodstream.
Discussion
The main goal of this experiment was to describe one source of knowledge available
at a starting point of generating an inference: property. The central question was what
61
information can properties bring to the table when they are presented as a part of an
inductive problem. A theoretically plausible candidate for the knowledge that a property
might contribute is a set of beliefs about property distribution. To address this possibility, we
examined people’ beliefs about distribution of a range of properties by asking participants to
estimate percentage of animals in the same biological family or percentage of animals living
in the same area sharing a given property.
Three classes of properties emerged from the analyses of the estimates. For one set,
scores for “same biological family” were above 50%, higher than the scores for “same
habitat”, which were below 50%. This suggests that participants have strong a priori beliefs
that these properties tend to be taxonomically distributed. For another set, scores for “same
habitat” were marginally above 50%, higher than the scores for the “same biological family”,
which were below 50%. This suggests that participants believe these properties to be
ecologically distributed. Finally, for the last set of properties, the “same family” and “same
habitat” did not differ from each other or from 50%, suggesting that these properties are
neutral, or non-biased with respect to either taxonomic or ecological distribution.
Presence of such overhypothesis-like beliefs had been suggested by the modeling
work of Tenenbaum and colleagues (2007), but has not been previously demonstrated
directly. This study demonstrated that people do indeed possess beliefs that map distribution
of different biological properties on different category systems (taxonomic and ecological, in
this case). Given that nearly every object belongs to multiple categories, limiting the number
of relevant categories can provide necessary constraints on hypothesis space required in order
to generalize meaningfully. These findings are compatible with Goodman’s overhypotheses
62
and suggest to us that his proposal can be extended to distributional beliefs over multiple
category systems, which, as we argue, underlie property effects in induction.
Based on the findings of this study, we argue that the nature of the property may bias
people towards attending to a limited subset of relations or category systems most relevant for
reasoning about that specific property by virtue of activating a specific set of distributional
beliefs. Specifically, in the context of an inductive problem, the properties we classified as
taxonomic might activate the taxonomic distributional beliefs, which in turn may bias people
towards attending selectively to taxonomic relations of the premise category. In contrast, the
properties we classified as ecological may activate ecological distributional beliefs and guide
attention to the ecological relations the premise category participates in.
Study 2. Premise category knowledge
Study 1 examined the contribution of property to the inductive problem. It
demonstrated that knowledge about properties includes distributional beliefs, which could
serve as a potential source of a taxonomic or ecological bias in induction. The other basic
input to an inductive inference discussed in Chapter II is knowledge about premise category.
The main goal of this study is to describe two types of knowledge about premise categories:
taxonomic and ecological.
First, to clarify the terms, in these studies we use “knowledge” in the sense of “belief
that something is true”, regardless of veridicality of the belief. Since true and false statements
are phenomenologically indistinguishable from the perspective of a person who believes them
both to be true, we do not expect such statements to have different cognitive effects.
Next, by taxonomic knowledge about animals we mean beliefs about membership in
taxonomic categories (“..is a bird”) and inherent attributes of an animal, such as intrinsic
63
qualities, behaviors, abilities and parts, whether observable or not (“is yellow”, “has a
beak”, “has a cold blood”, “walks on four legs”). By ecological knowledge about animals we
mean beliefs about animals’ typical environment, dietary habits, interactions with other
entities in the environment via predation and non-predatory behaviors (“lives in water”, “eats
gazelles”, “pollinates flowers”).
In this study, we aim to describe the knowledge about animals that is likely to come to
a person’s mind when she is presented with the animal name. Since our model of property
effects makes claims about moderating effect of property on retrieval of knowledge about
premise categories, it is important to describe the initial “playfield” for the property, i.e. the
state of affairs with category knowledge when it is not affected by the property. When
property moderates retrieval of knowledge, we cannot assume that it acts upon a “level
ground”, with all pieces of knowledge having equal baseline probability of coming to mind. If
we were to assume such “level ground” as a starting point, we would always expect biasing
properties to introduce inequalities by promoting activation of some subsets of knowledge
and suppressing activation of others. However, if we know from the outset which subsets of
knowledge are more available than others, we will be able to adjust our predictions about
specific effects of biasing properties on retrieval of that knowledge: for example, if we knew
that ecological knowledge is initially less likely to come to mind than taxonomic, we could
expect the facilitating effect of an ecological property to reduce the difference between
reliance on ecological and taxonomic knowledge in induction.
There are many factors that can affect availability of knowledge about a category. One
established view of the representation of categorical knowledge describes categories as sets of
features organized by theories specifying how the features are related to each other (Murphy,
64
2002; Murphy & Medin, 1985). We assume that when knowledge about a category is
accessed, the likelihood of retrieving any given feature may be affected by a large number of
chronic and acute factors, ranging from feature importance within that category’s theory to
recency of last access, emotional significance, etc. (Ahn, Kim, Lassaline, & Dennis, 2000;
Buchanan, 2007; Ebbinghaus, 1962; Shafto, Coley, & Vitkin, 2007). Here we are focusing on
the type of knowledge – taxonomic, describing intrinsic properties, and ecological, or
contextual, describing animal’s relations to other entities – as a potential source of differences
in availability.
Prior work demonstrates that in many reasoning tasks taxonomic knowledge
dominates ecological, serving as a default kind of knowledge (see Shafto, Coley, & Vitkin,
2007, for review; however, see Coley & Vasilyeva, 2010, for counter-evidence). For example,
priming ecological relations can increases similarity ratings for ecologically related items, but
priming taxonomic relations does not affect similarity ratings for ecologically related items,
suggesting that taxonomic relations are chronically more available (Ross & Murphy, 1999;
Shafto, Coley, & Vitkin, 2007). Likewise, time pressure in induction selectively impairs
reliance on ecological knowledge but not on taxonomic knowledge, consistent with the idea
that ecological knowledge is less spontaneously available, and accessing it is a more fragile
process than accessing taxonomic knowledge (Shafto, Coley, & Baldwin, 2007). Importantly,
however, such demonstrations of differences in knowledge availability are inferences based
on the outcomes of inductive or other reasoning tasks. If our aim is to describe knowledge that
serves as an input to induction, we would like to use a more direct measure of available
knowledge, that does not involve a reasoning task. This determined our choice of the
65
measurement instrument of amount of salient taxonomic and ecological knowledge about
animals – a feature listing task.
In order to measure salient available knowledge about animal categories, we presented
participants with a set of animal names and asked them to write all the attributes of each
animal that came to their mind. The mean number of taxonomic and ecological features they
listed per animal was taken as a measure of amount of salient taxonomic and ecological
knowledge about the animal.
This measure (mean number of listed features) is likely to tap both into knowledge
amount (number of known facts or beliefs) and its salience (likelihood with which a given
piece of knowledge comes to mind). Knowing many taxonomic facts does not necessarily
preclude one from knowing just as many or more ecological facts, but salient pieces of
knowledge have to compete for limited access to working memory. In a task like this, where
participants are free to list as many features as they want, the total number of listed taxonomic
and ecological features inevitably reflects some mix of knowledge amount and salience (thus
justifying our label for this measure, the “amount of salient knowledge”). In our studies, we
view this combination of knowledge amount and salience as an advantage of the measure
rather than a confound. For our purposes, “amount of salient knowledge” serves as a
comprehensive descriptor of the knowledge available to a participant when they are presented
with the animal name (in this study, in isolation; in the subsequent studies, in the context of
an inductive problem).
Besides having the theoretical reasons to describe possible differences in baseline
availability of taxonomic and ecological knowledge, this study also serves a practical purpose
in the context of this project: to characterize the knowledge that a given population
66
(undergraduate students at Northeastern University) has about a specific set of animals that
will be used as stimuli in subsequent studies.
Method
Participants. Twenty nine Northeastern university undergraduates (15 females),
participated in exchange for course credit. Additional data from one subject were discarded,
after she mentioned to the experimenter that since she was not familiar with some of the
animals, she googled their attributes during the experiment using her smart phone.
Materials. The experiment was run using Superlab 4.0.4 software running on a
MacBook laptop. Forty-two animals, all relatively good members of their respective
taxonomic and ecological categories (based on typicality ratings described in Appendix B),
were selected as stimuli. However, feature listing data for one animal were lost due to a
software error, leaving 41 animals. A full list of stimuli can be found in the Appendix B.
Procedure. Participants were tested individually. The instructions were as follows:
This is a very simple experiment to find out what characteristics and attributes come
to people's mind when they think of different animals. You will be presented with 42
animal names, one at a time. Below the animal name, there will be a space where you
can type. Please type in anything you can think of that is generally true of that animal.
There are no right or wrong kinds of attributes; if there's something you know or
believe about the animal and it came to your mind, type it in. However, try not to just
free associate - for example, if an animal just happens to remind you of your spring
break last year, DON'T write down "spring break". For every animal, please type as
many characteristics and attributes as you can.
67
The experiment was self-paced; on each trial a participant saw one animal name at
a time, in singular form, presented at the top of the screen, with a space to type below. There
was no limitation on how much text participants could type in. The order of animals was
randomized for each participant. Participants could move on to the next animal by clicking on
the “Next” button at the bottom of the screen; there was no option to go back and revise
previous responses.
Data processing & Coding. Each response consisted of a list of participant-generated
features. Most participants typed features as a list, with one feature per line. For the small
number of participants who wrote descriptions of animals as complete sentences, the
responses were converted into feature lists by the experimenter (e.g. “it’s a big green reptile”
was converted into “big”, “green”, and “reptile”).
In order to systematically quantify participants’ responses, we developed a coding
system characterizing each feature. Based on theoretical principles discussed above, we
divided features into two broad classes. In general, taxonomic features referred to category
membership, perceptual features, or non-interactive aspects of behaviors and physiology. In
contrast, ecological features involved references to animals’ diet, habitat, behavioral,
physiological, or other unspecified interactions with other entities in their environment7.
Coding categories were not mutually exclusive; a given response could receive multiple
7 Although this may be an atypical way to code feature content, it was important to account for animal features referring to interactions or potential bases for interactions, since interactions are a likely basis for inferences about these animals. Our objective was to apply the same codes to features (this study) and inferences (Studies 3 & 4), in order to be able to examine the relationship between features and inferences in a coherent manner. The coding system was originally developed for quantifying open-ended inferences from one animal to another animal or class of animals, and with minor modifications adapted to feature coding. The detailed description of the inference coding system can be found in the Study 3.
68
codes8. Four or five trained coders coded each response independently, and later met to
discuss and resolve cases on which their original code assignment did not match. Agreement
was achieved on 99% of the cases. The coding scheme is summarized in Table 2.
Table 2. Coding scheme for animal features
Feature Type Definition Example Taxonomic Features Category membership Reference to a biological class
or category [beaver] mammal
Perceptual feature Reference to superficial surface appearance
[butterfly] colorful
Behavior (non-interactive) Reference to behaviors that do not imply interaction with other entities
[duck] poops a lot
Physiology (non-interactive) Reference to specific internal organs or systems
[alligator] cold-blooded
Ecological Features Diet Reference to animal’s diet or
role in a food chain [whale] eats plankton [seal] eaten by whales
Habitat Reference to animal’s habitat [deer] forest animal Behavior (interactive) Reference to behaviors
involving interaction with other entities
[bee] pollinates flowers, stings people
Physiology (interactive) Reference to physiological interactions (often involving exchange of bodily liquids)
[rattlesnake] can inject poison through its fangs
General Interaction General reference to interaction with other entities
[koala] likes trees [bee] contact with honey
General similarity Reference to similarity without
specifying details [deer] horse-like
Other (uncodable) No applicable code, or the
response is too vague to be assigned a specific code
[clownfish] Nemo [bear] camping
8 Multiple-code assignments did not represent ambiguous responses, but rather responses in which participants invoked different kinds of reasons to support a particular inference “because it is also a bird, and it lives in the same area”, or listed a complex characteristic (e.g. “walks on four feet” refers both to the behavior of “walking” and to a structural attribute of having “four feet”).
69
To examine taxonomic and ecological knowledge about animals, the initial coding
categories were collapsed into two broad classes: taxonomic (combining references to
category, perceptual characteristics, non-interactive aspects of behavior and physiology) and
ecological (combining references to animals’ diet, habitat, interactive aspects of behaviors and
physiology, and general interaction). Features coded as general similarity were not included
into either of the categories since such features cannot be unambiguously classified as either
referring to taxonomic or ecological attributes. The precise make up of these categories can be
found in Table 2. If a given feature was coded as any of the component feature types, it was
scored as taxonomic or ecological. Again, these categories were not mutually exclusive.
Table 3. Mean number of features listed per animal (averaged over subjects), by feature type.
Feature Type Mean(St.Dev.) Range Taxonomic 3.10(0.67) 1.58-4.82
Category membership 0.32(0.20) 0.03-0.68
Perceptual 1.49(0.59) 0.42-3.24 Behavior (non-interactive) 1.32(0.59) 0.17-2.72 Physiology (non-interactive) 0.13(0.19) 0-0.79 Ecological 1.04(0.41) 0.06-2
Diet 0.20(0.16) 0-0.58
Habitat 0.52(0.23) 0.03-1.03 Behavior (interactive) 0.20(0.19) 0-1.10 Physiology (interactive) 0.03(0.11) 0-0.65 General interaction 0.08(0.09) 0-0.48 General Similarity 0.14(0.11) 0-0.51
Other 0.17(0.11) 0-0.58
70
Results
Data were scored and analyzed by item. To examine the amount of salient taxonomic
and ecological knowledge, we calculated average number of each type of feature generated
for each animal (first, separately for each participant, then averaged across participants). To
examine knowledge salience more directly, via relative order in which taxonomic and
ecological features tend to come to mind, we computed two indices. First, for each animal we
calculated the mean serial position in the participant-generated feature lists at which a
taxonomic feature first appeared, and the mean serial position at which an ecological feature
first appeared on the list. Second, for each animal we calculated the proportion of subjects
whose first feature on the list was taxonomic, and proportion of subjects whose first feature
was ecological9.
On average, 4.2 features were listed for each animal. If taxonomic knowledge is more
abundant and salient than ecological, we expect to see participants listing more taxonomic
than ecological features. This was supported by the data: the mean number of taxonomic
features listed per animal, 3.1 (range across animals, averaged over subjects 1.6-4.8) was
higher than the mean number of ecological features, 1.0 (range 0.1-2.0), paired samples
t(40)=17.53, p<.001, Cohen’s d=2.74). This pattern, of mean number of taxonomic features
exceeding ecological, held true for every animals in the dataset. This result suggests that
within the set of knowledge that becomes available upon the presentation of animal name,
taxonomic information is likely to dominate ecological.
9 These indices are not redundant because, in principle, the proportions of taxonomic and ecological first features do not have to sum to 1, since one feature could receive more than one code (although this did not happen frequently), and some codes (other, general similarity) were not included into either category.
71
As mentioned in the introduction to this study, the mean number of listed
taxonomic and ecological features reflects both amount and salience of participants’
knowledge about animals (which serves our larger purposes well). The next analyses,
however, tap more directly into knowledge salience, via examining the relative order in which
taxonomic and ecological features come to mind.
The first index of the relative order was calculated as the mean serial position of the
first taxonomic and ecological features in the participant-generated lists. The analysis of this
measure had to be qualified by the fact that oftentimes, participants listed no ecological
features for animals at all, making calculation of a serial position impossible (for a given
animal, on average, 11 participants out of 29 (38%) did not list any ecological features (with
animals ranging between having 5 to 27 participants listing no ecological features), compared
to on average, less than 1 (.78, or 2.7%) participant listing no taxonomic features per animal
(ranging from 0 to 4 participants, across animals)). Excluding the cases when participants
listed no ecological or no taxonomic features for a given animal, on average taxonomic
features appeared reliably earlier in the list than ecological (taxonomic 1.18 (range 1-1.69),
ecological 2.48 (range 1-4.43), paired samples t(41)=12.04, p<.001, Cohen’s d=1.88),
suggesting that taxonomic knowledge comes to mind more easily than ecological.
This claim was supported by the analysis of first-listed features (the second index of
the relative order in which taxonomic and ecological features come to mind). The first listed
feature was more likely to be taxonomic than ecological: .83 of the first features were
taxonomic, while only .18 of the first features were ecological (paired samples t(40)=18.59,
p<.001, Cohen’s d=2.9). Moreover, this result could not be accounted by a random sampling
from a pool of equally salient taxonomic and ecological features in memory, with the former
72
being simply more abundant than the latter. The proportion of taxonomic first features was
larger than expected based on the overall proportion of generated taxonomic features (one
sample test against .74510 t(40)=5.47, p<.001, Cohen’s d=.85); likewise, the proportion of
ecological first features was lower than expected based on the overall proportion of generated
ecological features (one sample test against .2511 t(40)=-3.12, p=.003, Cohen’s d=.49),
suggesting that taxonomic knowledge is not only more abundant, but also more salient than
ecological knowledge (more likely to come to mind early).
The last analysis examined the relationship between the mean number of taxonomic
and ecological features listed per animal in order to addresses two potential concerns with the
experimental procedure and stimuli set. First, participants might have felt constrained to list a
more or less constant number of features per animal, which would force them to choose
among taxonomic and ecological features to fill the spots. This would interfere with our
intention to use the mean number of listed taxonomic and ecological features as the measure
reflecting both natural amount and salience of participants’ knowledge about animals. If
participants were choosing between taxonomic and ecological features to fill the limited
number of slots on the list, we would observe a negative correlation between the number of
taxonomic and ecological features listed per animal. The second concern has to do with the
stimuli selection. If some animals in the stimuli set were more familiar to participants than
others and elicited more features in general, we would expect to see a positive correlation
between the number of taxonomic and ecological features. Luckily, the mean number of
10 Calculated as proportion of mean number of taxonomic features (3.11) out of mean number of features listed per animal (4.18)
11 Calculated as proportion of mean number of ecological features (1.05) out of mean number of features listed per animal (4.18)
73
taxonomic and ecological features listed per animal showed no relationship (r=.122,
p=.448), suggesting that there were no task demands imposing restrictions on feature list
length, and the animals in the stimuli set were roughly equally familiar to the participants.
Discussion
The main goal of this study was to describe the salient knowledge about animals.
Examination of the participant-generated lists of animal features yielded two main findings.
First, participants possess both taxonomic and ecological knowledge about animals: a sizeable
proportion, about a quarter of all listed features, were ecological, the rest were taxonomic.
Second, taxonomic and ecological knowledge are not equal. For the set of 41 animals from a
variety of biological categories and habitats, taxonomic knowledge appears to be more
abundant and salient than ecological knowledge: in a free feature listing task people listed
more taxonomic than ecological attributes for animals, and they tend to list taxonomic
attributes before ecological attributes. These results support the proposal about taxonomic
knowledge dominating ecological knowledge.
These findings have implications that go beyond description of relative amount and
salience of taxonomic and ecological knowledge. In the context of testing our model of
property effects, once we translate the general proposal about the moderating effects of
property on retrieval of knowledge about premises to retrieval of taxonomic and ecological
knowledge, the expectations about what knowledge will be more active at any given time will
have to be qualified by the baseline differences in availability of these two knowledge types.
Summary
The goal of Studies 1 & 2 was to catalog knowledge recruited by the inferences in
terms of its sources (property, premise category) and type (taxonomic, ecological). Study 1
74
demonstrated that property can have a relatively taxonomic or ecological content, likely
based on the statistical beliefs about property distribution across relevant systems of
categories (taxonomic or ecological). Study 2 showed that participants possess both
taxonomic and ecological knowledge about animals, with the former being more abundant
and salient. Now that we have outlined what knowledge serves as an input to inference-
generation processes, we can move on to characterizing these processes.
75
Chapter IV. Study 3. Property Effects and Knowledge Recruitment in
Induction
In this chapter, we take a step from documenting the ingredients towards describing
the recipe, and examine how information about the premise categories and properties is
combined to produce an inductive inference. Study 3 focuses on the relationship between the
knowledge about premises and the generated inferences, and addresses the question of
whether (and how) this relationship is moderated by the property.
Background
On the proposed account of property effects in inference generation, there are three
retrieval strategies that can yield property effects: retrieving property knowledge only,
retrieving both property and premise category knowledge independently, and retrieving
property and premise category knowledge interactively. The nature of the relationship
between the input (available knowledge) and the output (generated inferences) can be
informative in determining which strategy is used.
As mentioned in Chapter II, knowledge recruitment (what ended up being used to
construct an inference) can be distinguished from knowledge retrieval (what knowledge
stored in long-term memory was activated to detectable levels). Although retrieval (process)
constrains recruitment (outcome), recruitment does not tell a full story about retrieval. Some
knowledge might have been retrieved but not recruited. This issue will be addressed by Study
4. For now we examine the relationship between stored prior knowledge and its recruitment
by inferences as a first rough constraint on the possible retrieval process that might have
generated such a relationship.
76
The goals of this study are to examine how knowledge from different sources is
combined in the course of inference generation, and how taxonomic and contextual
(ecological) knowledge about premise categories and properties interact in the course of
inference generation. In addition, this study documents property effects in a new version of
the open-ended hypothesis generation paradigm, in which the premise of inductive questions
contains a single category (e.g. “gene X is found in ducks”). There is an important difference
between this version of the task and the paired-premise paradigm (“gene X is found in ducks
and otters”) that Coley, Vasilyeva & Muratore, 2008 and Coley & Vasilyeva, 2010 used to
document property effects in inference generation. In a paired-premise paradigm, each
premise category can be viewed as a context for the other, which makes it suboptimal as a
first attempt to examine moderating effects of property on knowledge retrieval: we would
need to take into account not only influence of property, but also (possibly interactive)
knowledge retrieval about 2 premise categories. Another new feature of this study is a within-
subject manipulation of the property. Coley, Vasilyeva & Muratore, 2008 and Coley &
Vasilyeva, 2010 used a between-subject design. It remains an open question whether
participants will still vary the generated hypotheses based on the property when they do not
have the opportunity to enter a task set of generating one type of inferences repeatedly.
In this study, participants are asked to generate inferences about 42 animals in an
open-ended inference generation task with single premise categories and a within-subject
manipulation of the property. The generated inferences were examined in two ways that
address two questions. First, we examine whether people generate different inferences about
different properties. Second, to examine whether property affects recruitment of knowledge
about premise categories, we use the feature-listing data collected in Study 2 (the measure of
77
available salient knowledge about a set of animals) to predict inferences about the same
animals presented in the context of different properties. This analysis will tell us whether
knowledge about animals predicts inferences about them, whether this relationship between
knowledge and inferences varies with the property, and finally, whether the relationship
between knowledge and inference is based on facilitation on congruent inferences
(taxonomic-taxonomic, or ecological-ecological), or on inhibition of incongruent inferences
(taxonomic-ecological and vice versa), or both.
Relations Among Property, Knowledge, and Inference
When a person is presented with an inductive problem “Parasite Z9 is found in ducks.
What else is likely to have parasite Z9? Why?”, potentially relevant knowledge may include
taxonomic and ecological knowledge about the premise category ducks, and taxonomic and
ecological knowledge about the property has parasite Z9 (beliefs about how parasites are
distributed across taxonomic and ecological systems of classes). Starting from that, the
participant has options of making a taxonomic inference, an ecological inference, both or
neither.
Property only. If participants recruit property knowledge, we expect inferences to
follow the distributional beliefs about the properties, resulting in property effects, i.e.
property-congruent inferences. In the scenario where the property knowledge is the only
knowledge recruited, we expect to see no relationship between prior knowledge about animals
and inferences. However, the absence of such a relationship would not necessarily imply the
“property only” strategy (we come back to this point in the discussion for this chapter). A
more reliable indication of a “property only” strategy would be frequent “generic” inferences,
i.e. inferences without reference to specific features of animals, e.g. “whatever lives in the
78
same area with ducks” or “anything else from the same family” (as opposed to “other pond
animals”, “other birds”).
Independent recruitment of property and premise knowledge. In contrast, if
participants recruit both property and premise category knowledge, we expect to see not only
a significant effect of property on likelihood of producing a given inference type, but also a
significant predictive relationship between knowledge and inferences. As shown by Coley &
Vasilyeva (2010) in the paired premise paradigm (inductive questions of the format “Ducks
and otters have disease X, what else?”), beliefs about taxonomic and ecological relatedness of
premise species can predict inferences about them. Specifically, beliefs about taxonomic
relatedness can facilitate taxonomic and inhibit ecological inferences, and beliefs about
ecological relatedness can facilitate ecological and inhibit taxonomic inferences. If in our
single-premise paradigm inferences draw upon knowledge about premise animals, we can
predict a similar pattern. First, we expect category knowledge and inferences of a matching
type to be related via congruent facilitation, where amounts of salient knowledge of a given
type are associated with higher likelihood of congruent inferences (e.g. having a larger
amount of ecological knowledge about ducks relative to toads should be associated with
larger likelihood of making an ecological inference about ducks than toads). When a category
representation is accessed as a part of processing the premise information, the salient
knowledge about that category becomes available, and that available knowledge is likely to be
translated into the inference. Thus, to the extent that ecological knowledge is available, we
can expect it to translate into ecological inferences. Second, we expect category knowledge
and inferences of mismatching types to be related via incongruent inhibition: taxonomic
knowledge could inhibit ecological inferences and/or ecological knowledge could inhibit
79
taxonomic inferences. Such inhibitory effects could arise from direct suppression between
different knowledge structures (taxonomic and ecological), or from competition among
taxonomic and ecological pieces of knowledge for the working memory space. (Of course,
knowledge could have both facilitatory and inhibitory effects). If the retrieval of property and
premise category knowledge proceeds independently, there should be no interaction between
knowledge-inference relationship and property.
Interactive recruitment of property and premise knowledge. If recruitment of
category knowledge varies depending on the context provided by the property knowledge, the
property and premise category knowledge are recruited interactively. Of course, we may also
consider the reverse case of interactive recruitment: recruitment of property knowledge
affected by category knowledge, or both knowledge sources affecting each other mutually.
For this project, we selected taxonomic and ecological properties as biasing and not open to
interpretation as possible, and focused on examining unidirectional effects of property on
retrieval of category knowledge, saving detailed investigation other cases for later studies.
If recruitment of premise category knowledge is affected by the property, we expect
the relation of premise category knowledge to inferences to vary depending on the property.
Knowledge recruitment by inferences in the context of a neutral property (property X,
substance in the bloodstream) can be viewed as an unbiased baseline of knowledge
recruitment. This relationship can be positive (congruent facilitation, when knowledge and
inference type match) or negative (incongruent inhibition, when knowledge and inference
types mismatch). When the property is informative, it could strengthen the influence of
congruent knowledge (relative to neutral property), by guiding selective recruitment of that
knowledge to inform inferences. This can be manifested by a stronger positive relationship
80
(for matching inference type) or a stronger negative relationship (for mismatching
inference type). For example, if the property is ecological, it may increase the positive
predictive power of eco-knowledge for eco-inferences, whereas a taxonomic property may
increase negative predictive power of taxonomic knowledge for eco-inferences (relative to a
neutral property).
Likewise, an informative property may weaken the effects of incongruent knowledge
on inferences, and thus reduce its predictive power – because in the light of the property,
incongruent knowledge is likely to appear less relevant. For example, if a property is clearly
taxonomic, it may indicate to the participant that ecological features of the animal are
irrelevant to projecting this property, thus minimizing recruitment of ecological knowledge
and its effect on inferences. This can be manifested by weakening a positive relationship (in
case of congruent facilitation between knowledge and matching inference type), or by
weakening a negative relationship (in case of incongruent inhibition between knowledge and
mismatching inference type).
Figure 3 summarizes the questions and predictions about the possible relationship
between knowledge about animals and inferences about them, alongside with possible ways a
property can moderate this relationship.
Finally, we can expect the answers to all these questions to depend on the specific type
of knowledge in question. As shown by Study 2, taxonomic knowledge is more abundant and
salient than ecological (see also Shafto, Coley, & Vitkin (2007), for a review on differences
between taxonomic and ecological knowledge). Thus, we may expect disproportionally strong
effects of taxonomic knowledge on both taxonomic and ecological inferences, compared to
corresponding effects of ecological knowledge.
81
Figure 3. Summary of questions and predictions about the relationship between knowledge about animals and inferences about them, as well as moderating effect of property on this relationship. (Only the ecological inferences are shown). See text for explanation. Note: within this diagram, “knowledge” refers to knowledge about premise categories.
Method
Participants
One hundred Northeastern University undergraduates (65 females) participated in
exchange for course credit.
Materials
The experiment was run using Superlab 4.0.4 software running on a MacBook laptop.
The stimuli were open-ended inductive questions of the format “GENE X5 is found in
DUCKS. What else is likely to have gene X5? Why?” We used the same set of 42 animals as
in Study 2. All animals were relatively good members of their respective taxonomic and
82
ecological categories (see Appendix B). The stimuli were constructed by combining each
of the forty-two animals, with each of the six properties that were selected based on
participants’ distribution estimates, as described below. Since properties repeated across
animals, each time a property appeared it was accompanied by a unique alphanumeric code
(X5, A3), to encourage participants to treat questions as independent.
Selection of properties. Based on participants’ distribution estimates collected in the
Study 1, we selected the following six properties: has gene X and has X-cells (taxonomically
biasing properties), has flu X and has parasite X (ecologically biasing properties), and has
property X and has substance X in the bloodstream (neutral, or non-biasing properties). To
confirm that the properties were indeed biasing or neutral as expected, we conducted two
additional one-way ANOVAs, one on same biological family estimates, the other one on same
habitat estimates, comparing mean estimates for selected taxonomic (average of gene and
cells), ecological (average of flu and parasite) and neutral (average of property X and
substance). Animals in the same biological family were estimated to share taxonomic
properties (69%) more than neutral properties (56%), which were shared more than ecological
properties (32%) (F=36.98, p<.0001, η2partial =.60, all pairwise Tukey/Kramer comparisons
p<.05). In contrast, animals living in the same area were expected to share ecological
properties (58%) more than neutral properties (47%) which were shared more than taxonomic
properties (32%) (F=17.96, p<.0001, η2partial =.42, all pairwise Tukey/Kramer comparisons
p<.05).
In addition, for every one of the six properties, we counted the number of participants
whose percentage estimates for the animals in the same biological family were higher than for
the animals in the same habitat, indicating that the property was viewed as distributed more
83
along taxonomic lines relative to ecological lines, as well as the number of participants for
whom the direction of difference was reversed, indicating ecological distribution of the
property. For cell, 23 out of 26 participants gave higher estimates in the “biological family”
question than in the “habitat” question, which significantly deviated from chance indicating
preference for taxonomic distribution (χ2=15.39, p<.0001). Gene was also viewed as
taxonomically distributed by the majority of participants, 21 out of 26 (1 participant assigned
equal percentages12) (χ2=11.56, p=.0007). In contrast, for flu and parasite, the majority of
participants assigned higher ratings in “habitat” question than “biological family” question
(flu: 23 out of 26, χ2=15.39, p<.0001; parasite: 19 out of 26, and 1 participant assigned equal
percentages (χ2=6.76, p=.0093); in other words, most participants believed that animals in the
same habitat were more likely to be homogeneous in terms of a novel flu or parasite than
animals in the same biological family. Finally, for substance and property X, participants did
not consistently assign higher estimates in one question or the other: for both properties, two
participants assigned equal percentages; for Substance, 15 out 26 (χ2=1.5, p=.22), and for
property X, 14 out of 26 (χ2=.67, p=.41) gave higher estimates in “biological family” than
“habitat” question, suggesting that across our participants there was no systematic preference
of taxonomic over ecological distribution for these properties.
Thus, gene and cells were biased taxonomically, both in relative (taxonomic >
ecological estimates) and, as shown in Study 1, absolute (compared to 50%) sense. Flu and
parasite were biased ecologically, both in relative (ecological > taxonomic estimates) and
12 Participants assigning equal percentages in response to both questions were not included in the analyses.
84
absolute (compared to 50%) sense. Property and substance proved to be neutral, or non-
biasing in terms of distributional beliefs about them.
Design
The main independent variable, property, had six levels, nested in three property types
(two ecological properties: flu, parasite; two taxonomic properties: gene, cell; two neutral
properties: substance, property) and was manipulated within subjects. Henceforth, we discuss
the property manipulation in terms of the three property types: ecological, taxonomic, and
neutral. Each participant was presented with 42 questions, one about each animal, seven
questions per property. Six blocks of seven animals rotated through the property in a Latin
square design, creating six versions of the experiment. Thus, within a version each subject
was presented with all animals and all properties, and across versions, each animal was
presented with each property. The order of questions was randomized for each participant.
The dependent variables were the frequencies of taxonomic and ecological inferences, based
on coded open-ended inferences and justifications generated by participants.
Procedure
Participants were tested individually. The general instructions read as follows:
In this experiment, you will read a number of statements about different kinds of animals. Your task will be to come up with other animals or general kinds of organisms that these statements could apply to.
For example, in one question, you may learn that one kind of animals has a certain gene. In a different question, you may learn that another kind of animal has a certain parasite, or a substance in its bloodstream, or cells, or a flu, or just some unspecified property. ALL you know about the gene, or cells, or the parasite, or the substance in blood, or the flu, or the property, is that the listed animal has it. You will be asked to
85
list other animals or kinds of organisms you think might also have it, as well as reasons for your answers.
There are no right or wrong answers; we are only interested in what you think. Please consider each question independently. That is, the information provided for an animal in one question should not affect your judgments about a different question.
You will be asked 42 questions total. You will see one question at a time, and below the question, there will be a space where you can type your response. After listing the answers, remember to give a one- or two-sentence reason for each of them, so we know what you were thinking while responding to the question.
During the experiment, participants saw one question at a time, with a space to type in
their responses. The questions were phrased as follows: “GENE T5 is found in DUCKS. What
else is likely to have gene T5? Why?” Neither the amount of text they could type nor time to
spend on a given question were limited. Participants could move through questions by
clicking on the “Next” button on the screen; they could not go back and change their previous
responses.
Results
Data processing & Coding
Each response consisted of a list of participant-generated conclusion categories and an
explanation for why a property true of the premise category was likely to be shared by those
categories. Inferences were coded in two ways. First, based on the type of relation between
the premise and conclusion, inferences were coded as taxonomic and/or ecological. Second,
based on the absence or presence of references to specific attributes of the premise animal,
inferences were coded as generic or non-generic.
Taxonomic vs. ecological inferences. In order to systematically quantify participants’
responses, we used a coding system to characterize the basis of each inference. Responses
were coded based on the justification generated by the participant specifying relevant relation
86
between experimenter-generated premise category and participant-generated conclusion
(not based on the animals that participant mentioned). Responses were initially coded into
fine-grained categories described in Table 4. Coding categories were not mutually exclusive;
a given response could receive multiple codes13. Four or five trained coders coded each
response independently, and later met to discuss and resolve cases on which their original
code assignment did not match. Consensus (defined as agreement between N-1 coders) was
reached on over 99% of the codes. During the coding, for any given inference the coders had
access to the premise animal, but not to the property condition.
Table 4. Coding scheme for characterizing basis of inference.
Basis of Inference Definition Example
Taxonomic Inferences
Category membership P and C belong to the same class or category
[Bee] Other insects such as mosquitoes and ladybugs
Perceptual similarity P and C are similar with respect to some aspect of superficial surface appearance
[Panda] Zebras – black and white.
Behavioral similarity P and C are similar with respect to some aspect of behavior
[Butterfly] Moths because both fly and stick to walls.
Physiological similarity P and C are similar with respect to specific organs or systems
[Monkey] Apes. They (…) have similar immune systems.
Ecological Inferences
Similar diet P and C are similar with respect to diet or eating the same kind of thing
[Panda] Sheep might also have Flu D6 because they both consume mostly plants.
Similar habitat P and C share similar or the same habitat without specification that the property is transmitted via
[Camel] Rattlesnakes, lizards, because they all live in the desert
13 Multiple-‐code assignment did not represent ambiguous responses, but rather responses in which participants invoked different kinds of reasons to support a particular inference.
87
habitat
Predatory interaction P and C interact with respect to predation, that is, P eats or is eaten by C
[Ant] An anteater because it would eat the ant with the parasite
[Bear] berries, small animals – food sources for bear
Habitat interaction P and C share or pass a property by coming into contact through the same habitat
[Beaver] Fish. If the beaver were to leave the substance in the water it was working in the fish could contract the substance.
Behavioral interaction P and C interact via some aspect of behavior
[Dragonfly] Big animals that dragonflies land on
Physiological Interaction P and C interact va some aspect of physiology (usually involving exchange of bodily liquids)
[Bee] Humans. Bee’s venom transfers cells to the human.
[Kangaroo] Kangaroo milk that is fed to the young because it passes through the mothers body
General Interaction P and C interact without further specification of the nature of the interaction
[Rattlesnake] I think flu G5 would also be found in humans, because humans sometimes have contact with rattlesnakes, and the flu may be able to pass between species.
[Ants] Every organism that comes into contact with the ants... because ants are everywhere.
General similarity P and C are alike or have similarities without further specifying the nature of the similarity
[Hyena] Panther because they are similar.
Other (uncodable) No explicit justification for the inference / no applicable code for the justification / the response is too vague to be assigned a specific code / participant does not know the animal
[Shark] Stingrays.
[Monkey] People? Scientists say so..
[Scorpion] Beetles and spiders. They both come to mind when I think of scorpions
Note: P = (given) premise category; C = (participant generated) conclusion categories
To examine the effects of property on the two broad classes of inferences discussed
above, we collapsed the initial coding categories into those that reflect reasoning based on
88
common category membership, appearance, or other shared intrinsic features (henceforth,
we will refer to these as taxonomic inferences), those that reflect reasoning based on extrinsic
similarities, shared contextual features like similar diet or habitat, and on some causal
mechanism including co-occurrence in space and time or direct contact through behavior or
predation (ecological inferences). The precise makeup of these broad categories can be found
in Table 4. If a given response was coded as any of the component inference types, it was
scored as a taxonomic or ecological inference. Again, these categories were not mutually
exclusive. For example, a response based on category membership and shared habitat
interaction would be counted both as a taxonomic inference and an ecological inference. For
example, a projection from duck to "geese, because they are also birds" would be classified as
taxonomic; to "geese, because they also live in water" would be classified as ecological; to
"geese, because they are also aquatic birds" would be classified as both ecological and
taxonomic; and finally, an projection to “geese” with no further justification would get coded
as “other” (because it would not be possible to tell if the participant was relying on taxonomic
relations, ecological relations, or free-associating). The two general categories, taxonomic and
ecological inferences, accounted for 88% of all responses.
For each animal, relative frequency of taxonomic inferences was calculated as
percentage of subjects making a taxonomic inference about that animal, separately for each
property type condition. The three property types were taxonomic (collapsing over gene and
cell), ecological (collapsing over flu and parasite), and neutral (collapsing over property X
and substance in the bloodstream). Relative frequency of ecological inferences was computed
in the same way. All the analyses on proportions reported below were conducted on arcsine-
transformed data, while the reported means are non-transformed and presented as percentages.
89
Since items (animals) were held constant across Studies 3 (inference-generation) and 2
(feature listing) and participants varied, all analyses were done by-item, unless noted
otherwise.
Table 5. Mean percentage of participants making each type of inference, per animal, by property type. Standard deviations are shown in brackets.
Overall Ecological
Property Neutral Property
Taxonomic Property
Taxonomic Inference 59 (10)
47 (13)
59 (13)
70 (12)
Category 41 (15)
35 (18)
39 (20)
49 (17)
Perceptual similarity 16 (8)
11 (9)
17 (13)
21 (13)
Behavioral similarity 13 (9)
10 (9)
16 (13)
14 (12)
Physiological similarity 6 (5)
4 (5)
7 (9)
7 (8)
Ecological Inference 48 (12)
64 (14)
47 (17)
33 (14)
Similar diet 8 (5)
9 (9)
10 (10)
4 (6)
Similar habitat 29 (13)
36 (17)
27 (17)
24 (16)
Predatory interaction 13 (4)
20 (12)
14 (11)
5 (5)
Habitat interaction 1 (1)
3 (4)
0 (2)
0 (1)
Behavioral interaction 2 (2)
3 (5)
1 (3)
1 (2)
Physiological interaction 0 (0)
0 (1)
0 (1)
0 (1)
General interaction 1 (1)
3 (4)
0 (1)
0 (1)
General similarity 9 (5)
8 (7)
9 (9)
11 (9)
Other 4 (2)
3 (4)
3 (5)
5 (5)
Note: relative frequencies of composite taxonomic and ecological inferences do not necessarily equal the sum of frequencies of their components, because one inference could receive multiple scores.
90
Generic inferences. Inferences were coded as generic if they did not contain any
references to the specific attributes of the premise category, i.e. inferences that do not show
signs of recruiting knowledge about specific animal, such as “anything that eats them”,
“whatever lives in the same place”, “anything from the same biological family”. If by looking
at an inference alone it was impossible to tell what the premise animal was, it got coded as
generic. All inferences (taxonomic, ecological, based on general similarity or “other”) were
coded for generic format.
For each animal, relative frequency of generic inferences was calculated as percentage
of subjects making a generic inference about that animal – overall and separately for each
property type condition.
Relative Frequency of Inferences
Undergraduates generated inferences based on both taxonomic and ecological
relations, although the former were more frequent. Overall, 59% of inferences were
taxonomic (e.g., an inference from a bee to “other insects such as mosquitoes and ladybugs”),
and 48% of inferences were ecological (e.g., an inference from an ant to “an anteater because
it would eat the ant with the parasite”) (paired samples t(41)=3.47, p=.001, Cohen’s d= .54).
Generic inferences were infrequent, and accounted for only 3.8% of all inferences.
Generic inferences were more frequent for ecological inferences (6.3% of which were
generic) than for taxonomic inferences (only 1.8% generic; paired samples t(41)=6.508,
p<.001, Cohen’s d=1.00).
Effects of Property on Inferences
Property affected the relative frequency with which participants generated taxonomic
and ecological inferences (see Figure 4). A one-way repeated measures ANOVA on relative
91
frequency of taxonomic inferences showed a significant effect of property type (F(2,
82)=64.64, p<.001, η2partial=.61): taxonomic inferences were most frequent when the property
was taxonomic, followed by neutral and ecological properties (all planned pairwise
comparisons p’s<.001). Property also affected relative frequency of ecological inferences
(one-way repeated measures ANOVA F(2, 82)=95.05, p<.001, η2partial =.70): ecological
inferences were most frequent when the property was ecological, followed by neutral and
taxonomic properties (all planned pairwise comparisons p’s<.001). Moreover, planned
comparisons across inference type indicated that for ecological properties, ecological
inferences were more frequent than taxonomic (paired samples t(41)=4.41, p<.001, Cohen’s
d=.68), whereas for taxonomic properties the difference was reversed (paired samples
t(41)=10.47, p<.001, Cohen’s d=1.62). When the property was neutral, taxonomic inferences
were more likely than ecological (paired samples t(41)=3.00, p=.005, Cohen’s d=.46),
although the gap between inference type frequencies was smaller than when the property was
taxonomic.
92
Figure 4. Property effects on relative frequency (percentage of participants, per animal) of ecological and taxonomic inferences. Error bars represent one standard error of the mean14.
We did not have specific expectations about the effect of property on the frequency of
generic inferences, but we found that it did vary with the property (repeated measures
ANOVA F(2,82)=3.221, p=.045, η2partial =.073). Generic inferences about ecological
properties were more frequent than taxonomic properties (5.3% vs. 2.7%, paired samples
t(41)=2.315, p=.026, Cohen’s d=.36). Generic inferences about neutral properties fell
inbetween and did not differ from the other properties (4.5%, paired samples vs. ecological
t(41)=.944, p=.351; vs. taxonomic t(41)=1.62, p=.113).
Relationship Between Premise Category Knowledge and Inferences
In this analysis we used data from the feature-listing task (Study 2) to predict
inferences collected in Study 3, in order to determine whether there is a relationship between
category knowledge and inferences. The mean number of participant-generated taxonomic
and ecological features of forty-one animals collected in Study 2 served as a measure of
14 The bars that appear not to differ across inference type, i.e., ecological inferences about ecological properties and taxonomic inferences about neutral properties indeed did not differ (paired samples t(41)=1.24, p=.223, Cohen’s d=.19). Likewise ecological inferences about neutral properties were as frequent as taxonomic inferences about an ecological property (paired samples t(41)=.01, p=.992, Cohen’s d=.001).
93
available salient knowledge, the predictor variables. The relative frequencies of participant-
generated taxonomic and ecological inferences about same forty-one animals (calculated
either collapsing across property or separately for ecological, neutral and taxonomic
properties) served as predicted values. The first set of analyses addressed the question of
whether knowledge predicts inferences.
Taxonomic inferences. In two simple regressions using mean number of taxonomic
and ecological features listed per animal as predictors, and relative frequency of taxonomic
inferences (collapsed across property), taxonomic inferences were not significantly predicted
by either taxonomic (R2=.010, β=.102, p=.526) or ecological (R2=.054, β= -.233, p=.142)
knowledge.
Ecological inferences. Ecological inferences showed a marginally significant
tendency to decrease (incongruent inhibition) with increasing amounts of taxonomic
knowledge about an animal (R2=.079, β= -.281, p=.075). At the same time, ecological
inferences were reliably predicted by ecological knowledge: the mean number of ecological
features listed for an animal positively predicted the relative frequency of ecological
inferences about it (congruent facilitation, R2=.116, β=.340, p=.030).
So far, the low frequency of generic inferences and the presence of a predictive
relationship between premise category knowledge and ecological inferences (albeit for
ecological inferences only) suggest that inferences recruited knowledge not only about
property, but also about premise animals. Next, we examine possible interactions between
knowledge from these two sources.
94
Interactions Between Property and Premise Category Knowledge
To examine the relationship between knowledge and inferences in more detail, we
conducted 12 simple linear regressions. In one triplet of regressions, ecological premise
category knowledge served as a predictor of ecological inferences, separately for ecological,
neutral and taxonomic property. The other triplets covered the three remaining combinations
between knowledge type and inference type, broken down by the property.
Ecological inferences. The standardized regression coefficients are shown in Figure
5. For ecological inferences (panel a.) the predictive power of knowledge appears to vary with
the property. Ecological knowledge was overall a positive, albeit non-significant, predictor of
ecological inferences when participants were reasoning about a neutral (R2=.057, β=.239,
p=.132) or taxonomic (R2=.034, β=.242, p=.128) property, but its contribution was
strengthened and became significant in the presence of an ecological property (R2=.124,
β=.352, p=.024).
When we examined the contribution of taxonomic knowledge to ecological inferences,
overall larger amounts of taxonomic knowledge were associated with lower frequency of
ecological inferences (all β’s are negative), and this relationship again varied with property.
Taxonomic knowledge inhibited eco-inferences marginally when the property was neutral
(R2=.076, β= -.276, p=.08), and reliably so when it was intensified by a taxonomic property
(R2=.12, β= -.346, p=.027). Relative to taxonomic and neutral properties, ecological property
largely neutralized the inhibitory effect of taxonomic knowledge on eco-inference (R2=.007,
β= -.083, p=.605).
Taxonomic inferences. Consistent with the overall regressions presented above,
taxonomic inferences were not significantly predicted by knowledge (all p’s >.121). However,
95
as shown in panel b. of Figure 5, the sign and ordering of predictors follow the predicted
pattern of strengthening effects of congruent knowledge and weakening effects of incongruent
knowledge (illustrated in Figure 3) with two exceptions – the regression coefficient for
taxonomic knowledge predicting taxonomic inferences in the context of an ecological
property is a very small negative (rather than positive) value, and the predictive value of
taxonomic knowledge for taxonomic inferences in the presence of taxonomic property drops
relative to neutral property.
To further test for moderating effects of property on knowledge-inference relationship,
we constructed 95% confidence intervals around the slopes for all 12 regressions. Although in
the analysis reported above some of the β’s within each property triplet came out significant,
and some did not, suggesting that property moderates the knowledge-inference relationship, in
this analysis the confidence intervals of all three slopes within each property triplet
overlapped, failing to support the claim about moderating effects of property on recruitment
of category knowledge.
96
Figure 5. Ecological and taxonomic knowledge about animals (measured as a mean number of ecological and taxonomic features) as predictors of relative frequency of taxonomic and ecological inferences about these animals, separately for ecological, neutral and taxonomic properties. *p<.05, +p<.1, °p<.15
a. Ecological inferences predicted by the mean number of ecological and taxonomic features.
b. Taxonomic inferences predicted by the mean number of ecological and taxonomic features.
97
Discussion
This study examined the relationship among property, premise category knowledge
and inferences as a first approach to explore how property exerts its effect on inferences.
Within the larger context of this project, the goal of this study is to examine recruitment of
knowledge by inferences in order to provide constraints on the underlying retrieval processes.
The question of main interest was how knowledge from different sources is recruited and
combined in inferences. We examined three possible scenarios: only property knowledge is
recruited; both property and premise knowledge are recruited independently; both property
and premise knowledge are recruited interactively. We tested these scenarios by examining
how taxonomic and ecological types of knowledge about premise categories and properties
interact in inferences about biological world.
To address these questions, we examined how the knowledge that participants have
about a set of animals (feature lists collected in the Study 2) is recruited by the inferences
another group of participants generated about this very same set of animals, and how property
might moderate this relationship. The inferences were collected in a novel open-ended
inference generation paradigm with single-category premises and property manipulated within
subjects. In this task, participants were forming hypotheses about other species of animals that
might have a property based on the evidence provided by the premise animal category and the
property.
Taxonomic vs. Ecological Inferences
Taxonomic inferences were more frequent than ecological inferences. This held for
overall percentages of inferences as well as for inferences about neutral properties. This fits
98
with the findings of Study 2 demonstrating that taxonomic knowledge about animals is
more abundant and salient.
Despite the dominance of taxonomic inferences, ecological reasoning was also quite
common. In part, this might be due to the open-ended format of the task. Coley & Vasilyeva
(2010) reported an increase in ecological reasoning in open-ended inference generation
relative to typical pattern reported in argument-evaluation studies of induction, and speculated
that this might be due to generic inferences. In this study we had the opportunity to quantify
this claim. Indeed, it appears that ecological inferences benefit from the possibility to
formulate generic inferences – more so than taxonomic inferences (6.3% of ecological
inferences were generic, compared to 1.8% of taxonomic inferences). Moreover, generic
inferences were more frequent when participants were reasoning about ecological than
taxonomic properties. It appears that, when participants intended to make an ecological
inference (perhaps, based on the property they were reasoning about) but lacked specific
ecological knowledge about the animal, they were able to make up for it with generic
inferences.
Property Effects on Inferences
Overall, property had a profound effect on the inferences participants generated:
reasoning about genes and cells biased participants towards taxonomic inferences, while flu
and parasite promoted ecological (interaction and spatial contiguity-based) inferences. This
replication of property effects is particularly impressive in this novel experimental setup with
property varied within subjects. Even when participants did not have a chance to enter a task
set of producing the same type of inference about the same property repeatedly, they varied
their inferences based on property from one question to another. This suggests that the
99
variability in inferences was driven by the specific attributes of each inductive question that
participants were presented with. This brings us to the main question of the study: what
knowledge about the premise did participants recruit to make inferences?
Recruitment of Property Knowledge
The bare minimum that participants need to do in order to generate property-sensitive
inferences is to access their knowledge about properties. As discussed above, one possibility
that this is the only source of knowledge participants recruit. The “property knowledge only”
strategy predicts a high percentage of generic inferences and absence of predictive
relationship between animal knowledge and inferences. This strategy was ruled out by the low
frequency of generic inferences, and the presence of the relationship between knowledge
about animals and ecological inferences.
Recruitment of Premise Category Knowledge
When participants generated ecological inferences, they used both taxonomic and
ecological knowledge about premise animals. Ecological inferences were facilitated by
congruent knowledge about premise categories, and inhibited by incongruent knowledge. In
other words, higher amounts of salient ecological knowledge about animals facilitated
ecological inferences about them, and higher amounts of taxonomic knowledge about animals
inhibited ecological inferences.
The overall pattern of congruent facilitation and incongruent inhibition held for
taxonomic inferences as well, although it was weaker and did not reach significance. Even
though evidence of relationship between knowledge and inferences was present for ecological
inferences but absent for taxonomic inferences, we would like to argue that (1) premise
category knowledge did get recruited by taxonomic inferences, but our main measure of
100
recruitment failed to detect it, and (2) the positive finding about knowledge recruitment
by ecological (albeit not by taxonomic) inferences still provides a useful constraint on the
general underlying retrieval process.
First, even in the absence of statistical support of category knowledge recruitment by
taxonomic inferences, we have reasons to believe such knowledge was nevertheless recruited.
If participants were not using their knowledge of animals at all, their taxonomic inferences
would be predominantly generic, i.e. driven exclusively by property information. However,
only 1.8% of taxonomic inferences were generic, providing evidence against this possibility.
Moreover, closer examination of taxonomic inferences showed that a majority of them were
categorical, i.e. based on common membership in a taxonomic category. Prior literature on
conceptual representation shows that objects tend to be good members of only one taxonomic
category (within a given level of a taxonomic hierarchy) (Ross & Murphy, 1999). Moreover,
we took special effort to select animal stimuli that would be typical representatives of only
one taxonomic category. If taxonomic inferences were predominantly drawing upon category
membership (and less so on other types of taxonomic knowledge), and if number of
taxonomic categories does not vary much, it is not surprising that the variability in total
number of taxonomic features was not a good predictor of taxonomic inferences – even
though it appears to have been recruited.
Finally, we have reasons to believe that the analyses we conducted might have failed
to detect knowledge recruitment in some cases because the measure of knowledge we used
was not an optimal predictor of inferences. Although amount of salient knowledge about
animals does capture the knowledge that becomes available to a person as they process the
premise of inductive argument, variability in this measure may not necessarily result in
101
variability in inferences. Strictly speaking, insensitivity to amount of salient knowledge
(measured as number of features) does not mean that the knowledge is not used at all – it
could be always recruited to the same stable extent, regardless of how much of it is available
per animal. This could explain one puzzling finding: the weak or nearly absent predictive
value of prior knowledge about premise animals in the context of reasoning about a neutral
property. Neutral properties did not provide participants with much information that could be
used to guide inferences (“property X” is not helpful, and “substance X in the bloodstream” is
highly ambiguous – is it internally-produced or picked up from the environment?). If
participants did not reliably use their knowledge about animals to reason about such
properties, what could have they possibly used? One possibility is that they relied on some
other knowledge type, not accounted by taxonomic and ecological categories, which seems
unlikely given that most animal features spontaneously generated by participants were
successfully accounted for by taxonomic and ecological coding categories (on average, less
than one (.17) feature per animal was coded as neither). It seems likely that using the “mean
number of taxonomic and ecological features” as predictors of inferences might have missed
category knowledge recruitment in this case. Alongside with other suggestive evidence, the
lack of relationship between knowledge and inferences for uninformative properties points to
the same idea – that the category knowledge had been recruited, even though the regression
analyses failed to detect them.
In addition, we find the logical argument against “property only” retrieval strategy
more compelling than the argument in support of this strategy. Essentially, this strategy makes
two predictions: (a) absence of a relation between knowledge and inferences, and (b) a large
amount of generic inferences. Our results provide evidence for (a) but not (b). Evidence for
102
(a) supports the “property only” strategy via affirming the consequent fallacy15. But
evidence against (b) denies the “property only” strategy by a valid logic of modus tollens.
Thus, based on the two pieces of evidence about taxonomic inferences - absence of
relationship between knowledge and inferences (a) and paucity of generic inferences (not-b) -
we prefer to draw the overall conclusion that the “property only” strategy lacks support, and
the category knowledge is in fact recruited.
Finally, of course it is also possible that participants could selectively switch retrieval
strategy from question to question, sometimes retrieving property knowledge only (and
making taxonomic inferences showing no signs of category knowledge recruitment),
sometimes retrieving both property and premise category knowledge (and making ecological
inferences with clear signs of category knowledge recruitment). Although we cannot
completely discard this possibility, we do not find it very plausible.
In sum, we argue that, despite the absence of a predictive relationship between
premise category knowledge and taxonomic inferences, people recruited both knowledge of
the property and the premise category to generate inferences.
The second point that we would like to argue for, despite the absence of predictive
relationship between knowledge an taxonomic inferences, is that the positive evidence for
recruitment of premise category knowledge by ecological inferences still supports general
conclusions about retrieval. If we observe that some knowledge is recruited, this implies that
it must have been retrieved. The route from retrieval to recruitment, however, is less direct: a
piece of knowledge might have been retrieved, but it does not necessarily imply that we are
15 Not to question the entirety of experimental method here, which by and large accepts the logic of affirming the consequent.
103
going to see signs of its recruitment. For example, the knowledge could have been judged
irrelevant later in the processing and got discarded, not making it to the final product of an
inference (i.e. was not recruited). Or it might have made its way to the inference, but our
measure of recruitment could have failed to detect it. Thus it is not surprising that the
presence or absence of evidence for knowledge recruitment can vary with the outcome - in
our case, between ecological and taxonomic inferences. What matters in this case is the
“existence proof” of a relationship between category knowledge and inferences. Since we see
such a relationship of taxonomic and ecological category knowledge with ecological
inferences, we can claim that such knowledge was retrieved, whether or not we observe this
relationship for taxonomic inferences.
Recruitment of Property and Premise Category Knowledge: Independence or
Interaction?
The question of whether property moderates recruitment of knowledge about premise
categories remains somewhat open. There was a tendency of property to strengthen effects of
matching knowledge type on ecological inferences, and weaken effects of mismatching
knowledge type on ecological inferences. Ecological knowledge reached significance as a
positive predictor or ecological inferences only when the property was ecological as well, and
taxonomic knowledge only served as a significant negative predictor for ecological
knowledge when the property was taxonomic as well, reinforcing the inhibitory potential of
the taxonomic knowledge on ecological inferences. However, the statistical support for these
claims was not solid (the analysis of slope confidence intervals failed to show significant
differences in category knowledge--inference relationship based on property). In light of this
104
evidence any claims about interactive recruitment of property and premise category
knowledge will need to be interpreted with caution.
In sum, we see that inferences recruit prior knowledge both about properties and
premise animal categories, and we have tentative evidence that recruitment of category
knowledge might be moderated by the property: when a property matches categorical
knowledge in type, such knowledge has slightly higher chances to affect ecological
inferences. Such evidence of interactive recruitment of premise category and property
information by ecological inferences is consistent with the proposal about interactive retrieval
as a locus of property effects in inference generation (one of the three scenarios we outlined at
the outset of this discussion). Thus, with this study we take the first step towards specifying
the mechanism of knowledge retrieval in induction.
Study 3 raises some specific questions that examination of retrieval can address. For
example, in this study we see no relations between the amount of taxonomic knowledge and
the likelihood of inferences in the presence of an ecological property, and (to a lesser extent)
no relation between ecological knowledge and inferences in the presence of a taxonomic
property. Did property completely suppress retrieval of this knowledge, or is this knowledge
initially retrieved but is later judged to be irrelevant and is discarded? For example, when
participants are asked about ducks having a certain parasite, do they not retrieve the facts that
ducks are birds and they have wings and beaks and feathers (taxonomic information) from
their long-term memory at all, or do they retrieve this knowledge but discard it after they
judge it irrelevant for reasoning about a parasite? These and other questions about retrieval
will be examined more directly in the Study 4.
105
In sum, the amount of salient knowledge measured as the mean number of listed
features may have its limitations, and future studies may want to use a different measure of
knowledge to examine knowledge recruitment – perhaps based on a more labor-intensive
matching of specific features between generated feature lists and inferences. However, in this
study this measure served its purpose reasonably well, providing some suggestive evidence
that property and category knowledge interact in the process of inference generation. This
study examined knowledge recruitment by inferences, but left some important questions
about knowledge retrieval open. The following Study 4 addresses these questions by mapping
out activation of premise category knowledge in the course of inference generation “in real
time”.
106
Chapter V. Study 4. Property Effects over the Time Course of Premise
Category Knowledge Retrieval in Induction.
Study 3 examined the relationship between premise category knowledge and inductive
inferences, and found some, albeit not very convincing, signs of interactive recruitment:
property may moderate the contribution of animal knowledge to inferences, by promoting
recruitment of property-congruent knowledge and suppressing recruitment of property-
incongruent knowledge (although this effect was limited to knowledge recruitment by
ecological inferences). Study 4 is an attempt to localize this property-premise interaction in
retrieval. Specifically, the goal is to describe the contextual effect of property on retrieval of
property-congruent and –incongruent category knowledge by examining the time course of
knowledge activation during inference generation.
Study 4
In order to examine the time-course of knowledge retrieval, this experiment borrows
the cross-modal priming paradigm from Swinney’s (1979) study of the mechanism and time
course of context effects in ambiguity resolution during sentence comprehension. Since we
are using a similar experimental technique, we will describe this study in some detail here.
The main question was about retrieval of different meanings of ambiguous words: when such
words are encountered in a context strongly biasing towards one of the meanings, are the
inappropriate meanings retrieved at all? In Swinney (1979) participants listened to sentences
containing an ambiguous word. The first part of the sentence (context) either helped to
disambiguate the word or was uninformative (e.g., if the ambiguous word were “port”, the
preceding context would either suggest the “wine” meaning (“the waiter poured the port into
the glass”), or the “nautical” meaning (“the ship arrived at the port in the afternoon”), or be
107
neutral (“nobody knew where the port was”)). At the end of the ambiguous word (at
varying ISIs, interstimulus intervals) participants were visually presented with a lexical
decision task (i.e. shown a letter sequence and asked to decide whether it was a word or a
non-word). On critical (i.e. word) trials, the visually presented word could be contextually
related (“wine”, if the context suggested the “wine” meaning), contextually inappropriate
(“ship”), or unrelated (“book”). The decision time was taken as a measure of the activation of
each meaning of the ambiguous word during ambiguity resolution. Swinney found that
immediately after the ambiguous word was presented, recognition of visual targets related to
both appropriate and inappropriate readings of the ambiguity was facilitated relative to
baseline neutral context, even when the ambiguous word was preceded by a strong biasing
context. However, as soon as 700msec later, only the relevant meaning reaction time showed
advantage relative to inappropriate and unrelated word, that did not differ from each other.
This suggested that all meanings of an ambiguous word are initially activated, and context
aids selection of the relevant meaning, as opposed to guiding selective activation of the
relevant meaning from the beginning.
There are multiple parallels between the questions addressed by Swinney (1979) study
and knowledge retrieval in course of induction. Multiple meanings of an ambiguous word are
akin to multiple sets of knowledge about a premise, which provide potential bases for
projection in induction. In sentence comprehension, the biasing context is the sentence or
phrase containing the ambiguity, that can resolve ambiguity by indicating the correct
interpretation compatible with the rest of the sentence. In induction, the biasing context is the
property that can reduce number of relevant attributes of the premise and, consequently, the
number of plausible projection hypotheses.
108
However, the analogy between lexical access and selective induction has its
limit: ambiguity resolution in lexical access typically requires commitment to a single
meaning, whereas in induction, one can legitimately make a mixed inference, referring
to multiple kinds of knowledge. An projection to an ecologically related species -‐ a
predator or animal sharing a habitat -‐ is, strictly speaking, not incompatible with
projecting a gene, it is simply less likely. Although such probabilistic nature of property
information adds noise to selective knowledge retrieval, it does not invalidate the use of
Swinney’s approach to studying effects of biasing properties on induction.
To examine whether the property provides a context that biases retrieval of knowledge
about premises, and if yes, how this bias acts over time, this experiment employed an adapted
version of Swinney’s task. Participants were auditorily presented with a property and an
animal premise in an open ended inductive question, and were asked to generate possible
conclusions. The property varied between taxonomically biasing (gene, cells), ecologically
biasing (flu, parasite), or non-biasing (substance, property X). In addition, upon hearing the
property and animal, participants were presented with a lexical decision task involving targets
related to salient taxonomic or ecological knowledge about the premise animal. For example,
a participant might hear a property, gene, followed by the animal, duck, and, after a varying
time interval, see on the screen a taxonomic target bird, or an ecological target pond, or an
unrelated target sofa, or a non-word soach. Their task was to decide whether the letter
sequence is a word or a non-word. The time to respond to the related targets is taken as a
measure of activation of taxonomic or ecological knowledge about the premise animal. If
corresponding knowledge about duck is activated, we expect decisions about related targets
(bird and pond) to be faster than about unrelated targets (sofa). If property moderates
109
knowledge retrieval, we would expect decisions about ecologically related targets (pond)
to be faster in the presence of an ecological property (congruent facilitation) and/or slower in
the presence of a taxonomic property (incongruent inhibition) relative to a neutral, or
unbiasing property context. Similarly, we would expect decisions about taxonomically related
targets (bird) to be faster in the presence of a taxonomic and/or slower in the presence of an
ecological property compared to neutral. By varying the property and time interval between
the premise and the lexical decision task, we examine if, how and when properties influence
knowledge retrieval.
We use this method addresses three questions: Does property moderate knowledge
retrieval? If property moderates knowledge retrieval, is the mechanism based on facilitating
property-congruent knowledge, inhibiting property-incongruent knowledge, or both? And
finally, does inhibition of property-incongruent knowledge (if any) proceed monotonously, or
do we see knowledge initially retrieved (activated) but subsequently discarded (inhibited)?
Property Effects on Knowledge Retrieval: How do You Know Them When You See
Them?
In Study 3, ecological properties increased the predictive power of ecological
knowledge and decreased the predictive power of taxonomic knowledge about premise
categories, relative to neutral properties (for ecological inferences). If these property effects in
knowledge recruitment reflect property effects in knowledge retrieval, we can expect a
biasing property to facilitate retrieval of congruent knowledge, and inhibit retrieval of
incongruent knowledge. Because of likely differences in the strength of connection of
taxonomic and ecological knowledge to a target animal, we cannot make blanket predictions
110
about which kind of knowledge will be more active at any given moment16. However,
predictions can be made about increases or decreases in activation of property-congruent and
property-incongruent knowledge compared to the baseline, the neutral property context.
In the absence of a biasing context (i.e. in the context of a non-biasing property), we
expect related knowledge to rise in activation after the animal name is presented17, while
unrelated knowledge should maintain a relatively low level of activation. This is illustrated in
the Figure 6, panel a. By saying that knowledge has been retrieved, we mean that its
activation has risen to a detectable level above its baseline18. For example, after a person is
presented with the animal “duck”, related knowledge – “bird”, “pond” is more activated than
before presentation of “duck”, and more activated than unrelated and unprimed knowledge
(“sofa”). Moreover, there can be differences between the level of activation of different
related pieces of knowledge. The panel a. of Figure 6 shows the animal label “duck” to trigger
higher activation of the taxonomically related target (“bird”) than of an ecologically related
target (“pond”) - based on our finding of higher baseline salience of taxonomic relative to
ecological knowledge (feature-listing data, Study 2).
Next, we can add a biasing property to the picture. As shown in Figure 6, panel b. with
a biasing taxonomic property we can expect congruent facilitation of taxonomic knowledge
retrieval, and incongruent inhibition of ecological knowledge retrieval. A taxonomically
16 I.e., although we do not expect systematic differences in baseline activation between “bird”, “pond”, and “sofa”, we do expect that “duck” might prime “bird” more strongly than it primes “pond”.
17 We only consider a relatively short time window after the animal name is presented; later decay of knowledge activation is not discussed here
18 For the ease of display, all the knowledge types are shown having equal baseline level of activation, although we do not make such an assumption about the state of natural knowledge. For this experiment, however, we took effort to select stimuli that do not vary dramatically in baseline activation level (see the Method section, stimuli norming).
111
biasing property should facilitate the retrieval of congruent taxonomic knowledge and
inhibit the retrieval of incongruent ecological knowledge. Conversely, an ecologically biasing
property should facilitate the retrieval of congruent ecological knowledge and inhibit the
retrieval of incongruent taxonomic knowledge (not shown in the Figure 6). Such property-
biased retrieval would be consistent with the pattern observed in the Study 3 (most
pronounced for ecological inferences): property-congruent knowledge had stronger cognitive
effects on inferences, while property-incongruent knowledge had weaker cognitive effects on
inferences, relative to the same knowledge in the context of a neutral property.
If knowledge retrieval is moderated by property, we expect to see faster and more
accurate verification of targets when they are congruent with a biasing property than when the
property is neutral. We also expect to see slower and less accurate verification of targets when
they are property-incongruent than when the property is neutral.
112
Figure 6. Hypothesized curves of knowledge retrieval in unbiased and biased contexts. TAX/ECO/UNRtarg stand for taxonomic, ecological, and unrelated targets, respectively. NEU/TAXprop stand for neutral and taxonomic property, respectively.
a. knowledge activation in the context of a neutral property. Retrieval: activation of related knowledge relative to unrelated (r).
b. Knowledge activation in the context of a biasing taxonomic property. Congruent facilitation: activation advantage in the presence of a congruent property relative to a neutral property (f). Incongruent inhibition: activation decrease in the presence of an incongruent property relative to a neutral property (i).
113
Non-Monotonicity in Knowledge Retrieval.
Another question is how inhibition of property-incongruent knowledge unfolds over
time. In Swinney’s (1979) study of resolution of lexical ambiguity during sentence
comprehension, contextually inappropriate meaning of a word was initially activated, and
shortly afterwards suppressed. Thus, inhibition proceeded non-monotonically: first no signs
of inhibition, followed by a quick discarding of inappropriate meaning. If property serves as a
biased context for the animal name, it raises a question of whether property-incongruent
knowledge is inhibited non-monotonically.
Is it there? In the course of retrieval, does knowledge activation change
monotonically in the presence of an incongruent biasing property? If retrieval of knowledge
about animal premises in the context of a biasing property is not radically different from
retrieval of word meanings in the context of a disambiguating sentence, we may expect to see
initial activation followed by inhibition19. Based on the results of Study 3, we have two
candidate cases for such activation trajectory: taxonomic knowledge in the context of an
ecological property, and ecological knowledge in the context of a taxonomic property. In both
cases the knowledge appeared to be non-recruited by inferences (mean number of listed
features, taxonomic in the first case and ecological in the second case, did not predict
inferences). This raises the question of whether unrecruited knowledge is simply not retrieved
in the context of an incongruent property, or is it retrieved and later discarded?
19 Strictly speaking, non-‐monotonic retrieval could also refer to delayed facilitation (detectable changes in speed of activation driven by congruent context). This question lies beyond our interest here, and, as a practical point, change in speed of rising (delayed facilitation) would be harder to detect than a reversal from rise in activation to decline (delayed inhibition). Thus this discussion is limited to delayed inhibition of context-‐incongruent knowledge.
114
Figure 7 shows an activation trajectory of ecological knowledge initially retrieved
and then inhibited. For a given knowledge type retrieved in the context of an incongruent
property, sampling across time after the presentation of an animal, we can detect retrieval
non-monotonicity if we observe a detectable activation advantage at earlier times that
significantly declines or completely goes away at later times. The activation advantage has to
be detectable, otherwise we cannot claim that knowledge has been retrieved at all. The
advantage should be measured, ideally, relative to the baseline activation of that same
knowledge (activation of “pond” without being presented with “duck” before). But for the
sake of this study we take activation of unrelated targets (“sofa”) to represent common
baseline activation of all targets (“sofa”, “bird”, “pond” in the absence of “duck” prime)20. In
order to attribute this effect to the presence of a biasing incongruent property, we should also
see knowledge activation suppressed in the presence of a biasing incongruent property
relative to neutral property context.
Thus, the pattern of interest can be described as “non-monotonic retrieval of
knowledge inhibited by an incongruent property.” The signature of such a pattern would be
lower activation in an incongruent property context relative to a neutral property context (the
“inhibition” component of the pattern), accompanied by a rise-then-drop activation trajectory
in an incongruent property context relative to unrelated targets (the “non-monotonic”
component of the pattern). We need to demonstrate both parts of the pattern, because just
observing no inhibition at early times in retrieval followed by pronounced inhibition by itself
does not prove non-monotonicity: activation of knowledge takes time, and of course very
20 See the description of the norming of target words in the method section below for the justification of this assumption
115
early in retrieval we expect to see no difference between activation of related and
unrelated targets. We need to show inhibition accompanied by clear signs of initial retrieval –
hence, the two-component signature pattern of non-monotonic inhibition.
Figure 7. Knowledge activation in the context of a biasing taxonomic property. Incongruent inhibition: significant activation decrease in the presence of an incongruent property relative to a neutral property (i). Non-monotonic retrieval of property-incongruent knowledge: initial detectable activation advantage relative to unrelated knowledge (r) which significantly declines or completely goes away shortly afterwards (r1>r2)
Where (when) to look for it? In order to know when to look for such possible brief
activation followed by inhibition, we attempted to derive time course estimates from existing
sentence-comprehension and priming studies. The first estimate of the duration of such
“initial” process can be taken from Swinney (1979). When participants were asked to make
lexical decisions about the words related to one of the two meanings of an ambiguous word
presented in a biasing context, when the target for lexical decision was presented immediately
at the offset of the ambiguous word (0msec ISI measured from the offset of the previous
stimulus), the decisions about targets related to both meanings of ambiguous words were
116
equally facilitated relative to a baseline unrelated target. In the delayed condition (3
syllables after the offset, or 750-1000msec, according to Swinney (1979)), only the target
related to the context-appropriate meaning of the word was facilitated, suggesting that the
decision process selecting the meaning was completed by that time. The time period of 750-
1000msec is, of course, a generous estimate providing only the maximum duration of the
period when both context-appropriate and context-inappropriate meanings are “immediately
and momentarily accessed” (Swinney, 1979, p. 657).
A rather different study of Neely (1977) provides a similar estimate of the duration of
initial “automatic” retrieval period. Neely tested the attention theory of Posner & Snyder
(1975; as cited in Neely, 1977) postulating two distinct components of attention, a fast
automatic inhibitionless spreading activation process and a slow limited-capacity conscious-
attention mechanism. In the critical condition of Neely’s (1977) study, participants were
instructed to expect primes to be followed by targets which came from a specified category
other than the category represented by the prime itself. E.g., participants were told that the
prime “body” would be followed by furniture items such as “door”. ISI was varied between
250, 450 and 750 msec. The facilitation of the “body” prime on decisions about “door” target
was first evident only at the longest ISI of 750msec. Crucially, when participants were
presented with “unexpected” targets naturally related to the prime (“heart” following the
“body” prime), there was a short period of facilitation (250msec), which changed to inhibition
by 750 msec (with no facilitation or inhibition at the intermediate ISI of 400msec21). This
suggests a more stringent estimate: before 250-400msec, information access proceeds
relatively automatically and is not subject to additional contextual or goal-based constraints, 21 Reaction times in all ISI conditions were around 700-‐800msec.
117
and by 400msec, other selection or access mechanisms may begin to dominate
processing. However, such an estimate should be taken with caution: the upper boundary on
the time window of the initial retrieval process may be determined either by termination of
initial retrieval processes, or by initiation of next-step processes. The timing when such next-
step processes are initiated in Neely’s task (shifting attention to a prime category to a different
prespecified category) may differ from timing when selection processes in property-sensitive
induction may be initiated. But in the absence of studies of time-course of information
activation in induction22, we will take this estimate as the best available.
Thus, in order to detect temporary activation of property-incongruent knowledge
followed by inhibition, the best place to look is a window of 250-1000msec following the
animal name. This estimate determined our choice of SOA levels for this experiment, as
described in the method section.
The goal of the study was to examine whether property moderates knowledge
retrieval, and if it does – does it act by facilitating property-congruent knowledge, inhibiting
property-incongruent knowledge, or both? And finally, how does retrieval of property-
incongruent knowledge unfold overtime: is it steadily suppressed by the property, or is it
initially retrieved and then inhibited?
22 McElree, Murphy & Ochoa (2006) examined the dynamics of context-‐dependent feature retrieval using an SAT methodology which could provide a better and more appropriate estimate. Unfortunately for the purposes of the current study, McElree, Murhpy & Ochoa (2006) focused on emergent properties of combined concepts (soft being an emergent property of boiled celery). Since deriving such properties that are not true of the original concepts (celery is not soft) is likely to require additional steps in reasoning, it is not surprising they obtained much longer estimate of the non-‐context-‐sensitive time window -‐ as long as 2000msec – which is unlikely to apply to the proposed study
118
Method
Participants. One hundred eleven Northeastern undergraduates, all native speakers of
English, participated in the main experiment in exchange for a course credit (45 in SOA400,
32 in SOA900, 34 in SOA1650 condition). In addition, 29 participants, all native speakers of
English, were recruited for two norming studies, also in exchange for a course credit.
Materials.
Open-ended induction task. The stimuli set was composed of 36 animal premises,
each belonging to one salient taxonomic category (mammal, bird, reptile, fish, insect) and one
salient habitat-based ecological category (forest, desert, pond, ocean, savannah). Of the 36
animals, 30 were a subset of already normed and tested animal premises used in Studies 2 and
3, and six animals were new. In addition, 36 new animal names were used as fillers. Each of
the animals was presented in the context of an inductive problem about one of the six
properties from Study 3: flu, parasite, property X, substance, gene, and cell, presented with
unique alphanumeric codes (X5, Z9).
All the animal names and properties along with unique alphanumeric codes were
recorded in the voice of a female native speaker of English. The quality of recordings and
familiarity with the spoken animal labels were checked by asking a separate sample of 11
native English speakers (6 females) to listen to the recordings of animal names, write down
the names of the animals they heard23 and rate each animal on a 5-point scale (1= I don’t
know this animal, 2 = I’ve heard this animal name before but I don’t really know anything
about it, 3 = I know a little bit about this animal, 4 = I am moderately familiar with this
23 The accuracy on transcribing animal names from the audio recording was 100%, discounting (extremely abundant and diverse) spelling errors.
119
animal, 5 = I know a lot about this animal). The mean familiarity rating was 4.2
(SD=.54), suggesting that the tested population of Northeastern undergraduates was familiar
with the animals we selected.
Lexical decision task. Seventy-two letter strings were used as targets for the lexical
decision task. Half of these were words, and half were pronounceable non words. The word
targets were either taxonomically related, ecologically related, or unrelated to the animals
used as premise categories in the induction task. The source of target words was previously
collected open-ended responses to inductive problems referring to the animal premise
categories (Study 3) and feature listings (Study 2). The unrelated targets were nouns from a
non-biological domain, similar to the related targets in their consonant-vowel structure. The
full list of lexical decision targets is shown in Appendix C.
The strength of association of the taxonomic, ecological, and unrelated targets to the
corresponding animals, as well as lack of direct associations between properties and target
words, were verified with a separate group of 18 native speakers of English. The details of the
association norming are presented in the Appendix D. The ecological, taxonomic, and
unrelated targets did not differ in mean length (F(2,33)=.253, p=.780), frequency
(F(2,32)24=.093, p=.910), or number of orthographic neighbors (F(2,33)=.311, p=.735) (based
on the CELEX database, Medler & Binder, 2005). They also did not differ in non-primed
verification times (M=631, 640, 657msec, respectively, F(2,33)=.416, p=.663) or accuracy
(M=.96, .95, .97, respectively, F(2,33)=.5, p=.611) (based on the English Lexicon Project
database, http://elexicon.wustl.edu). These comparisons suggest that the baseline activation
levels of ecological, taxonomic and neutral targets in the absence of the animal name prime 24 Frequency data for one of the targets were not available
120
do not differ. This justifies using the unrelated items (“sofa”) as a general activation
baseline, as opposed to getting a separate baseline measure of activation of related items
(“pond”, “bird”) in the context of our task but not primed by an animal name (“duck”). Thus,
the forthcoming analyses of accuracy and response times will compare related items to
unrelated items (baseline), and the resulting differences (if any) will be taken to reflect
retrieval of related knowledge. The pronounceable non-word fillers were equated with the
word targets on mean length (t(70)=.771, p=.440) and number of orthographic neighbors
(t(70)=-.126, p=.900) (CELEX database, Medler & Binder, 2005).
Design. The main independent variable was property type (taxonomically biasing,
represented by gene and cell; ecologically biasing, represented by flu and parasite; and non-
biasing represented by substance and property). The second independent variable was the
target word type (taxonomic vs. ecological vs. unrelated). Each non-filler animal was yoked
to one target word type (taxonomic, ecological, or unrelated)25.
Finally, we varied stimulus onset asynchrony (SOA), the time between the onset of the
auditorily-presented animal name and the appearance of the visually-presented target word
(400, 900 or 1650msec)26. Mean duration of animal name was 640msec (SD=128.5; range
412-958msec), meaning that on average, at SOA=400 the target appeared 240msec before the
end of the animal name, at SOA=900, the target word appeared 260msec after the animal
25 Although originally a fully crossed set of stimuli was developed, in which each of the non-filler animals was matched with all 3 target words – taxonomic, ecological, and unrelated – we only report data from a subset of experimental conditions, in which each animal was yoked to one target word type.
26 We are using SOA rather than ISI in order to minimize noise from different durations of animal labels. Since lexical access begins before the complete word is presented, and in this experiment lexical access is facilitated because participants know to expect to hear animal names rather than random words, by 400msec into the recording of the animal name participants can unambiguously determine which animal is being presented to them, and the duration of “tail” of the word carries little extra information.
121
name, and at SOA=1650 the target appeared 1010msec after the end of the animal name,
thus covering the time-window of interest estimated based on Swinney (1979) and Neely
(1977). The property and target type were manipulated within subjects, and SOA between
subjects.
There were 72 trials total, of which 36 were experimental trials (six properties x three
target types with two animals per cell) and 36 were fillers (six properties with six non-word
targets per cell). The order of trials was randomized for every subject. Lexical decision
accuracy was measured as proportion of correct responses; reaction time was measured from
the moment the target word was presented until the response key was pressed.
Procedure. Participants were tested individually, on a MacBook laptop running
Superlab 4.0.4 software and set up with headphones and a microphone. The experiment
consisted of an open-ended induction task with intervening lexical decision task. Both tasks
were computerized.
Participants were instructed that they would be listening to utterances that would
introduce a property, followed by an animal that possesses that property, and their task is was
say out loud (at a cued moment) the first other thing that comes to their mind that is likely to
share that property, alongside with a short explanation. Participants were also informed that at
“random” moments a sequence of characters would appear on the screen, and their task was to
identify it as a word or a non-word as quickly as possible without sacrificing accuracy, using
the response buttons. Before starting the actual task, participants went through a practice
session that gradually introduced them to the format of the open-ended induction question, the
auditory format of the question, and allowed them to practice doing two tasks at the same time
122
– generating inferences and doing the lexical decision task. The full text of the
instructions is shown in the Appendix E.
Each trial began with a 2000msec pause; then a participant heard the property to be
projected (e.g., flu M3), followed by a pause of 1000msec, followed by the name of the
premise animal (e.g., bear), followed by a pause and a signal to start speaking. At varying
SOA’s (400, 900, or 1650msec from the onset of the animal name), a target for lexical
decision appeared on the screen and stayed there until the participant responded or for
3500msec, whichever came first. No accuracy feedback was provided. After a 2000msec
pause following the participant’s response or the end of lexical decision target presentation, a
short beep signaled that the participant could start saying their inference. Participants had 15
seconds to say their response, after which the experiment automatically moved on to the next
trial. If participants finished saying their inference sooner than in 15 seconds, they could
move on to the next trial by pressing a space bar.
Participants were instructed to keep their eyes at the center of the screen and their
index and middle fingers of the right hand on two response buttons throughout the experiment
(Yes (“o”) and No (“p”)). During the experiment, participants were given 4 breaks. Duration
of the breaks was not limited, participants could restart the experiment at any time by pressing
a space bar.
Results
Open-ended induction task. The inferences produced by participants in this task
were likely contaminated by the cues from the lexical decision task, and thus do not reflect
natural reasoning. Therefore, the data from this task only served as a check that participants
did indeed engage in inference generation, as instructed, and that the task format – auditory
123
presentation of the premise and vocalization of the inference – did not disrupt their ability
to reason.
Open-ended responses from a randomly selected subset of 70 subjects were
transcribed and coded by 4-5 trained coders using the same coding scheme as used in the
Experiment 1. All the coders were naïve to the hypothesis; while coding, the property that the
participant reasoned about on a given trial was concealed from the coders (unless the
participant mentioned it in their response).
Inferences from five subjects were discarded since they failed to provide justifications
for their inferences on more than half of the trials, making their responses uncodable. Using
the data from the remaining subjects, for each of the 36 non-filler animals we calculated the
proportion of subjects making taxonomic and ecological inferences for that animal, collapsing
across SOA conditions, split by property type (taxonomic, ecological, neutral). All the
analyses were conducted on arcsine transformed data; the reported statistics are based on
arcsine-transformed data, but the reported means are raw. The analyses were conducted both
by subjects (reported in the main text) and by items (indicated only if different from by
subjects).
Overall, participants generated more taxonomic (.64) than ecological (.43) inferences
(paired samples t(64)=7.16, p<.001, Cohen’s d=.89). The results confirmed that property did
affect inferences in the expected way. Taxonomic inferences increased in relative frequency
from ecological to neutral to taxonomic property (eco .53, neu .66, tax .74; repeated measures
ANOVA F(2,128)=18.741, p<.001, η2partial=.227; all pairwise comparisons p’s ≤.001).
Ecological inferences increased in frequency from taxonomic to neutral to ecological property
(tax .31, neu .40, eco .56; repeated measures ANOVA F(2,128)=21.881, p<.001, η2partial=.255;
124
all pairwise comparisons p’s ≤.004). This indicates that, despite a somewhat challenging
nature of the task – listening to open-ended inductive questions, making word/non-word
decisions about a letter string on the screen, and saying out loud the inferences – participants
were able to perform the task, and their performance overall matched what participants did in
a more standard paradigm (Study 3).
Next, we examined the effect of the target type on inferences. (Recall that a
taxonomically related, ecologically related, or unrelated word was visually presented in the
lexical decision task after the animal name was played to the participant.) The main function
of the target type manipulation in this study was to “probe” for activation of taxonomic and
ecological knowledge about the premise animals at different times in the course of inductive
inference, not to affect the outcome inferences. However, it is hard to have a “probe” that
would not have any consequences on the reasoning processes it is supposed to measure. We
expected that target type could have an effect on inferences, mostly for pragmatic rather than
theoretical reasons: participants might interpret the target words as (probabilistic) cues from
the experimenter to base their inferences on, and/or the target words could remind participants
of taxonomic or ecological features of the premise animals that they might not have thought
about otherwise, guiding their inferences. Thus, a main effect of target type on inferences
would not be of much theoretical interest. However, any interactions of target type with other
variables (property, SOA), meaning that our “probe” had differential effects on inferences
depending on levels of other variables, would require adjustments in interpreting the effects of
these variables (property, SOA) on accuracy and decision times about these targets that serve
as DVs in subsequent analyses.
125
As expected, target type had a main effect27 on participants’ taxonomic inferences
(F(2,128)=13.834, p<.001, η2partial=.178) and ecological inferences (F(2,128)=20.158, p<.001,
η2partial=.24). Taxonomic inferences were more frequent when the intervening lexical decision
target was taxonomic (.70) than when it was ecological (.58, paired samples t(64)=5.15,
p<.001, Cohen’s d=.64); taxonomic inferences in the presence of an unrelated target fell in
between (.62) and were significantly less frequent than taxonomic (t(64)=3.018, p=.004,
Cohen’s d=.37) and significantly more frequent than ecological inferences (t(64)=2.18,
p=.033, Cohen’s d=.27).28 Ecological inferences were more frequent when the intervening
lexical decision target was ecological (.51) than when it was taxonomic (.35, independent
samples t(64)=6.19, p<.001, Cohen’s d=.77); in the presence of an unrelated target the
inferences fell in between (.43) and was significantly different from both (related samples test
vs. tax t(64)=3.252, p=.002, Cohen’s d=.40); vs. eco t(64)=3.209, p=.002, Cohen’s d=.40)29.
The target type did not interact with property for taxonomic (F(4,256)=.497, p=.738)
or ecological (F(4,256)=.505, p=.732) inferences, as indicated by repeated measures 3 target
type (taxonomic, ecological, unrelated) x 3 property (taxonomic, ecological, neutral)
ANOVAs.
Nor did target type interact with SOA for taxonomic inferences, as indicated by a
mixed 3 target type (taxonomic, ecological, unrelated; within subjects) x 3 SOA (400, 900,
1650; between subjects) ANOVA on relative frequency of taxonomic inferences
27 Marginally significant in a by-item analysis, (F(2,33)=3.171, p=.055, η2
partial=.161) 28 In a by item analysis, frequency of taxonomic inferences in the presence of unrelated targets did not differ from either taxonomic (independent samples test t(22)=1.553, p=.135) or ecological targets (independent samples test t(22)=1.073, p=.295). 29 In a by item analysis, frequency of ecological inferences in the presence of unrelated targets did not differ from either taxonomic (independent samples test t(22)=1.776, p=.09) or ecological targets (independent samples test t(22)=1.236, p=.229).
126
(F(2,124)=1.576, p=.185). However, there was a significant interaction between the target
type and SOA for ecological inferences, as indicated by an analogous mixed ANOVA on
relative frequency of ecological inferences (F(4,124)=3.746, p=.007, η2partial=.108).
Figure 8. Relative frequency of ecological inferences as a function of target type and SOA. Error bars represent one standard error of the mean. Solid lines indicate significant difference at p<.05, dotted lines indicate significant difference at p<.1.
As shown in Figure 8, the overall pattern of target effect on ecological inferences held
across SOAs: relative to trials where an animal premise was accompanied by an unrelated
target (e.g., premise category “duck” followed by a lexical decision target “sofa”), ecological
inferences were more frequent on trials where an animal premise was followed by an
ecological target (e.g. “duck” followed by “pond”) less frequent on trials where an animal
premise was followed by a taxonomic target (e.g. “duck” followed by “bird”). However, this
pattern was more expressed at shorter SOA of 400msec (all pairwise comparisons between
target types p<.02) than longer SOAs of 900msec and 1650msec, as reflected by the
significant interaction between target type and SOA.
127
At an SOA of 400msec the target appears roughly in the middle of the auditory
presentation of the animal name. In other words, targets at this SOA are more likely to
overlap with knowledge retrieval about the animal, than targets with longer SOAs. The fact
that we see the most pronounced effect of target on inferences at this point suggests that we
might be observing context effects in lexical access. In light of this interaction, we can
speculate that the target provides additional retrieval context for the animal name (albeit
delayed relative to the beginning of the animal name), and this target-context affects category
knowledge retrieval by priming target-related information about the animal. On the one hand,
this (speculative) result is promising: if context in the form of a target word can affect
retrieval of category knowledge, we expect that so will the context in the form of a property.
On the other hand, this result might mean bad news: if the effects of property-context are
weaker than the effects of target-context (in other words, if our supposedly neutral “probe”
turned out to be more invasive with respect to knowledge retrieval than our main
manipulation), we have fewer chances to detect the effects of property on retrieval, if they
exist – especially at the shorter SOA of 400msec when the effect of the target is at its
strongest.
There were no significant main effects of SOA on either taxonomic (F(2,62)=.013,
p=.987) or ecological inferences (F(2,62)=1.75, p=.182), and there were no significant three-
way interactions between property, SOA, and target type (repeated measures ANOVA on
taxonomic inferences F(8,248)=1.063, p=.39; on ecological inferences F(8,248)=.716,
p=.677).
Lexical decision: accuracy. Overall accuracy was high, 97% correct on word and
96% correct on non-word targets. All the analyses were performed on arcsine-transformed
128
proportions; the reported statistics are based on arcsine-transformed data; the reported
means are raw. All the analyses on accuracy and RT in the lexical decision task were
conducted both by subjects (reported in the main text) and by items (indicated only if
different from by subjects).
Property and target effects on lexical decision accuracy. Does property moderate
retrieval of knowledge about premise animals? The question of main interest was whether
property type interacts with target type, by promoting accuracy of property-congruent targets
and interfering with it for property-incongruent targets. Specifically, we expected accuracy of
decisions about ecologically-related targets to be higher when the property is ecological,
lower when the property is taxonomic, and at the intermediate level when the property is
neutral. Similarly, we expected accuracy of decisions about taxonomically-related targets to
be higher when the property is taxonomic, lower when it is ecological, and in-between for the
neutral property.
The effects of property and target on accuracy were analyzed in a series of 3 property
(taxonomic, ecological, neutral) x 3 target (taxonomic, ecological, unrelated) repeated
measures ANOVAs on lexical decision accuracy, performed separately for each SOA
condition. Results are depicted in Figure 9. Contrary to our prediction, property did not
interact with target type at any SOA (SOA400: F(4,176)=1.456, p=.217; SOA900:
F(4,124)=.660, p=.621; SOA1650: F(4,132)=1.145, p=.338). Thus, if we take accuracy of
verifying targets related taxonomically and ecologically to the premise animal as the measure
of activation of taxonomic and ecological knowledge about that animal, then based on
accuracy results, property does not appear to moderate activation of premise category
knowledge.
129
Figure 9. Mean accuracy on the lexical decision task as a function of SOA, property and target. Error bars represent one standard error of the mean. Solid lines indicate significant difference at p<.05, dotted lines indicate significant difference at p<.1.
130
In order to examine whether premise animal name primed the two related targets –
taxonomic and ecological – more than the unrelated target, we examined the effect of target
type on accuracy. If hearing “duck” primes “pond” (ecological target) and “bird” (taxonomic
target) more than “sofa” (unrelated target), we would expect higher accuracy on ecological
and taxonomic targets than on unrelated targets. Based on the ANOVA analyses described
above, target type had a main effect on accuracy at SOA400 (F(2,88)=4.092, p=.02,
η2partial=.085) and at SOA900 (F(2,62)=3.325, p=.042, η2
partial=.097), but not at SOA1650
(F(2,66)=.376, p=.688)30. As shown in Figure 10, at SOAs 400 and 900, the pattern matched
our predictions: accuracy on related targets was overall higher than on unrelated, although not
all pairwise comparisons reached significance: at SOA400, accuracy on taxonomic but not
ecological items was higher than accuracy on unrelated targets (paired samples test, tax vs.
unrelated t(44)=3.026, p=.004, Cohen’s d=.45; eco vs. unrelated t(44)=1.337, p=.188); at
SOA 900, accuracy on ecological but not taxonomic items was higher than accuracy on
unrelated targets (paired samples test, eco vs. unrelated t(31)=2.996, p=.005, Cohen’s d=.53;
tax vs. unrelated t(31)=1.486, p=.147). At SOA1650, all pairwise comparisons were non-
significant (all p’s>.56). This suggests that at earlier SOAs, the animal name (“duck”)
sufficiently primed taxonomic (“bird”) and ecological (“pond”) targets to secure accuracy
advantage for them relative to unrelated targets (“sofa”), however, by 1650msec after the
beginning of the animal name, this effect is gone.
30 Despite the differences in significance of target type effect on inferences between SOA levels, the interaction between target type and SOA was not significant in a mixed 3 target type (within subjects) x 3 SOA (between subjects) ANOVA, F(4, 216)=1.223, p=.302).
131
Figure 10. Effect of target type on lexical decision accuracy, split by SOA. Error bars represent one standard error of the mean. Solid lines indicate significant difference at p<.05, dotted lines indicate significant difference at p<.1.
We did not predict any main effects of property on the lexical decision accuracy at any
SOA (unless reasoning about some properties turned out to be overall more taxing than about
others). Overall the results were consistent with this prediction (SOA900: F(2,62)=1.130,
p=.33; SOA1650: F(2,66)=.122, p=.886), except at SOA400, where property had a marginally
significant effect on accuracy (F(2,88)=3.082, p=.051, η2partial=.065). At SOA400, trials
involving a taxonomic property yielded a higher accuracy (.987) than trials with an ecological
(.970, paired samples t(44)=2.408, p=.02, Cohen’s d=.36) or neutral property (.972, paired
samples t(44)=2.462, p=.018, Cohen’s d=.37) which did not differ from each other (paired
samples t(44)=.404, p=.688).
We did not have predictions about main effect of SOA on accuracy, but there was a
marginal effect of SOA(one-way ANOVA F(2,108)=2.713, p=.071, η2partial=.048), caused by
a trend of decreasing accuracy at SOA of 1650msec (.96) relative to SOA of 400 (.98) and
900msec (.98) (independent samples 400 vs. 1650 t(77)=2.013, p=.048, Cohen’s d=.46; 900
vs. 1650 t(64)=2.021, p=.047, Cohen’s d=.51; 400 vs. 900 t(75)=.155, p=.877).
132
Changes in activation of premise category knowledge over time. The next
question we planned to address was whether the retrieval of property-incongruent knowledge
is subject to non-monotonic inhibition. In other words, for a given knowledge type retrieved
in the context of an incongruent property, do we see a detectable activation advantage at
earlier times that significantly declines or completely goes away at later times? (Advantage
measured relative to baseline (unrelated targets), as shown in Figure 7.)
Admittedly, it makes little sense to ask “how does incongruent property affect
knowledge retrieval?” in the absence of evidence that it does affects it at all, as indicated by
non-significant target x property interactions reported above. Nevertheless we examined the
results for the signs of non-monotonic retrieval.
In order to claim that non-monotonic inhibition of property-incongruent knowledge
takes place, property-incongruent premise category knowledge should show early signs of
retrieval that disappear later in processing. To examine whether such pattern is present, we
compare the accuracy of related but property-incongruent targets to the accuracy of unrelated
targets across at each SOA (schematically represented by r1 and r2, retrieval index at time1 and
time2 in Figures 11 and 12, panel a. In order to attribute the later decline in knowledge
activation to inhibition from an incongruent biasing property, such decline must bring this
knowledge below the activation level of the same knowledge type in the presence of a neutral
property (as represented by i, inhibition in Figures 11 and 12, panel a.)
The first case of interest was retrieval of taxonomic knowledge in the context of an
incongruent (ecological) property. The first part of the predicted pattern was examined by a
series of planned comparisons of taxonomic vs. unrelated target accuracy which showed that
the former was more accurate at SOA400 (.994 vs .944, t(44)=3.317, p=.002, Cohen’s d=.49)
133
, but not at SOA900 (.984 vs .977, t(31)=.442, p=.662) or 1650 (.963 vs .949, t(33)=.643,
p=.524) (Figure 11, panel b.). This suggests that property-incongruent taxonomic knowledge
showed early signs of retrieval that disappeared later in processing. However, this alone does
not constitute evidence for non-monotonic inhibition: we need to be able to attribute the later
decline in activation to inhibition from the incongruent property. The pairwise comparisons
of taxonomic target accuracy in neutral vs. incongruent biasing (ecological) property context
across SOA conditions were non-significant (SOA400 t(44)=1, p=.323; SOA900 t(31)=.570,
p=.572; SOA1650 t(33)=1.421, p=.165) yielding no evidence for incongruent inhibition. In
sum, we do not have sufficient evidence to claim that an ecological property non-
monotonically inhibits taxonomic knowledge about premise animals.
Figure 11. Mean accuracy of verifying taxonomic targets across SOA in the context of a biasing incongruent (ecological) property – contrasted with first, unrelated targets in an ecological property context, and second, taxonomic targets in a neutral property context. a. Predicted pattern for incongruent inhibition of taxonomic category knowledge (non-‐monotonic retrieval)
b. Results
The second case of interest was retrieval of ecological knowledge in the context of an
incongruent (taxonomic) property. Planned comparisons of ecological vs. unrelated target
accuracy showed that the former was more accurate at SOA900 (.992 vs .945, t(31)=2.396,
134
p=.023, Cohen’s d=.42) , but not at SOA400 (.989 vs .983, t(44)=.216, p=.83) or 1650
(.956 vs .956, t(33)=0, p=1) (Figure 12, panel b). This again demonstrates signs of retrieval
(detectable activation advantage) for ecological knowledge that is most evident at 900msec
SOA and then reduced. However, again there were no differences between ecological target
accuracy in neutral vs. incongruent biasing (taxonomic) property context across SOA
conditions (SOA400 t(44)=1.431, p=.16; SOA900 t(31)=.571, p=.572; SOA1650
t(33)=1.44, p=.16), suggesting that the late decline in activation in the first part of the pattern
cannot be attributed to property-incongruent inhibition. Again, the evidence does not support
the prediction that a taxonomic property might non-monotonically exhibit ecological
knowledge about premise animals.
Figure 12. Mean accuracy of verifying ecological targets across SOA in the context of a biasing incongruent (taxonomic) property – contrasted with first, unrelated targets in a taxonomic property context, and second, ecological targets in a neutral property context. a. Predicted pattern for incongruent inhibition of ecological category knowledge (non-‐monotonic retrieval)
b. Results
Accuracy on filler (non-word) items. Accuracy of negative decisions about fillers
(nonwords) was not affected by the property (F(2,216)=1.313, p=.271) or by SOA
(F(2,108)=.544, p=.582), and the two variables did not interact (F(4,216)=1.304, p=.269).
135
Lexical decision: reaction time. All analyses were conducted on raw RT data, as
well as with outliers eliminated by using Tukey’s 1.5-interquartile range method, Tukey’s 3-
interquartile range method, as well as by trimming 5% of the mean and trimming 10% of the
mean. Removing outliers did not affect the results. The analyses reported here were
performed on raw data.
Property and target effects on lexical decision RT. Just like for the accuracy analyses,
the first question is whether property moderates retrieval of knowledge about premise animals
by promoting speed of property-congruent targets and interfering with it for property-
incongruent targets. Such moderation would manifest itself as a significant interaction
between property type and target type. Specifically, we expected decisions about ecologically-
related targets to be faster when the property is ecological, slower when the property is
taxonomic, and at the intermediate level when the property is neutral. Similarly, we expected
decisions about taxonomically-related targets to be faster when the property is taxonomic,
slower when it is ecological, and at the intermediate level when the property is neutral.
The effects of property and target on RT were analyzed in a series of three property
(taxonomic, ecological, neutral) x three target (taxonomic, ecological, unrelated) repeated
measures ANOVAs, performed separately for each SOA condition. Mean RTs are shown as a
function of property, target and SOA in Figure 13. Contrary to our prediction, but consistent
with the accuracy analyses, property did not interact with target type at any SOA (SOA400:
F(4,176)=1.467, p=.214; SOA900 F(4,120)=1.056, p=.382); SOA1650: F(4,128)=.712,
p=.585).
136
Figure 13. Mean RT on the lexical decision task as a function of SOA, property and target. Error bars represent one standard error of the mean.
137
Although the interaction terms were not significant, we conducted planned
comparisons of simple effects of property for each target type (comparing taxonomic,
ecological, and neutral properties to each other), separately for each SOA condition. None of
them were significant (all p’s≥.162). Thus, if we take speed of verifying targets related
taxonomically and ecologically to the premise animal as the measure of activation of
taxonomic and ecological knowledge about that animal, then based on accuracy results,
property does not appear to moderate activation of premise category knowledge.
Next, we examined whether animal names primed the two related targets – taxonomic
and ecological – more than the unrelated target. If hearing “duck” primes “pond” (ecological
target) and “bird” (taxonomic target) more than “sofa” (unrelated target), we would expect
faster decisions about ecological and taxonomic targets than on unrelated targets. In
agreement with the accuracy analyses, RT was affected by the target type at SOA400
(F(2,88)=4.257, p=.017, η2partial=.088) and 900 (F(2,60)=3.41, p=.04, η2
partial=.102)31, but not
SOA1650 condition (F(2,64)=1.478, p=.236).
As shown in Figure 14, at SOAs 400 and 900, the pattern matched our predictions:
accuracy on related targets was overall higher than on unrelated, although not all pairwise
comparisons reached significance: at SOA400 decisions about ecological targets (1153msec)
were reliably faster than about unrelated targets (1233msec, related samples t(44)=-3.544,
p=.001, Cohen’s d=.53), but decisions about taxonomic targets (1210msec) did not differ
from unrelated targets (related samples t(44)=-.740, p=.463). Ecological targets were
marginally faster than taxonomic (1210msec, related samples t(22)=-1.855, p=.07, Cohen’s
d=.28). At SOA900, decisions about ecological (1183msec) but not taxonomic targets 31 Not significant by item, p=.507
138
(1207msec) were faster relative to unrelated targets (1244msec, paired samples eco vs.
unrelated t(31)=-2.604, p=.014, Cohen’s d=.46; tax vs. unrelated t(31)=-1.566, p=.128).
Decisions about ecological and taxonomic targets did not differ (paired samples test vs. eco
t(21)=-1.247, p=.222;). At SOA1650, all pairwise comparisons were non-significant (all
p’s>.229). Consistent with accuracy results, at shorter SOAs the animal names primed related
targets, but by SOA of 1650msec the effect was gone. Although the RT results only show
significant priming for ecological but not taxonomic targets, the fact that taxonomic targets
did show a trend towards priming in RT, and a significant priming in accuracy, suggests to us
that the taxonomic targets were most likely primed, although more weakly than ecological
targets.
Figure 14. Effect of target type on lexical decision RT, split by SOA. Error bars represent one standard error of the mean. Solid lines indicate significant difference at p<.05, dotted lines indicate significant difference at p<.1.
We did not expect property to have a main effect on the lexical decision RT at any
SOA, and indeed it did not (SOA400: F(2,88)=.375, p=.689; SOA900: F(2,60)=1.197,
p=.309; SOA1650: F(2,64)=.59, p=.557). SOA was not predicted to have a main effect on
139
RT, and although there was a trend of increasing RT for longer SOA, it was not
significant (F(2,108)=.652, p=.523).
Changes in activation of premise category knowledge over time. The next question
on the list was whether property-incongruent knowledge is inhibited non-monotonically. As
shown in Figure 15 panel b., ecological properties did not lead to changes in reaction time for
taxonomic targets relative to neutral targets (F(1,107)=1.092, p=.298) at any SOA
(interaction F(2,107)=.433, p=.650). Participants were not any faster to make decisions about
taxonomically related targets than about unrelated targets. Thus, based on the reaction time,
we have no evidence that taxonomic knowledge was retrieved in the context of an ecological
property. There was also no evidence for incongruent inhibition of taxonomic knowledge by
an ecological property (paired samples comparisons of taxonomic target RT for neutral vs.
ecological property, all p’s≥.155). In sum, there is no evidence that an ecological property
non-monotonically inhibits retrieval of taxonomic knowledge about premise animals.
Figure 15. Time to verify taxonomic targets across SOA conditions in the context of a biasing incongruent (ecological) property – contrasted with first, unrelated targets in an ecological property context, and second, taxonomic targets in a neutral property context. a. Predicted pattern for incongruent inhibition of taxonomic category knowledge (non-‐monotonic retrieval)
b. Results
140
Another potential place for non-monotonic inhibition is in retrieval of ecological
knowledge in the context of a taxonomic property. Planned comparisons showed that
decisions about ecological targets were faster than about unrelated targets at SOA900 (paired
t(30)=2.72, p=.011) but not at SOA400 (paired t(44)=1.675, p=.101) or SOA1650 (paired
t(33)=.256, p=.800) (see Figure 16, panel b.). This demonstrates an early activation advantage
for ecological knowledge that disappears with time. However, ecological target RT did not
differ for neutral vs. taxonomic property (all p’s≥.302) suggesting no non-monotonic
inhibition of ecological knowledge retrieval given a taxonomic property.
Figure 16. Time to verify ecological targets across SOA conditions in the context of a biasing incongruent (taxonomic) property – contrasted with first, unrelated targets in a taxonomic property context, and second, ecological targets in a neutral property context.
a. Predicted pattern for incongruent inhibition of ecological category knowledge (non-‐monotonic retrieval)
b. Results
Decision times on filler (non-word) items. Overall, it took participants longer to
correctly respond “No” to non-words (mean RT 1305msec) than to correctly respond “Yes” to
words (mean RT 1227msec; paired samples t(110)=6.213, p<.001, Cohen’s d=.59). To our
surprise and dismay, the speed of lexical decisions about filler items (non-words) was
141
consistently affected by the property (SOA400: F(2,88)=6.033, p=.004, η2partial=.121;
SOA900: F(2,62)=4.404, p=.016, η2partial=.124; SOA1650: F(2,66)=4.964, p=.01,
η2partial=.131). The order of the RT means at all SOAs was neutral (shortest RT) to taxonomic
to ecological property (longest RT), with ecological being significantly longer than neutral in
all SOA conditions (SOA400: paired samples t(44)=2.914, p=.006, Cohen’s d=.43; SOA900:
paired samples t(31)=2.865, p=.007, Cohen’s d=.51; SOA1650: paired t(33)=2.732, p=.01,
Cohen’s d=.47). In addition, the difference between taxonomic and neutral property
conditions was significant at SOA400 (paired samples t(44)=2.877, p=.006, Cohen’s d=.43)
and 1650 (paired samples t(33)=2.228, p=.033, Cohen’s d=.38). However, with Bonferroni
adjustment of significance for 9 non-planned comparisons, only two of the pairwise
comparisons were borderline significant (critical p=.006)
Discussion
The goal of this study was to examine the time course of knowledge retrieval while
participants were engaged in an open-ended induction task. Participants’ active knowledge
was probed across three time points by using lexical targets related to the premise animal via
taxonomy, ecology, or unrelated.
Based on the results of Study 3 we expected to find effects of property on knowledge
retrieval: specifically, facilitation of property-congruent knowledge and inhibition of
property-incongruent knowledge. However, property did not significantly affect accuracy or
RT of word target verification in the lexical decision task and did not interact with target type.
In other words, we did not obtain any evidence of property facilitating or inhibiting
knowledge retrieval.
142
Even in the absence of evidence of property moderating knowledge retrieval, we
attempted to examine targeted cases for non-monotonic retrieval of property-incongruent
knowledge (e.g., looking for cases when taxonomic knowledge might have been retrieved
originally, even though the property was ecological, and was subsequently inhibited). Not
surprisingly, although we did see some tenuous evidence of non-monotonic retrieval, it was
not accompanied by evidence of property inhibiting incongruent knowledge, which made
these findings as a whole uninformative with respect to the mechanism & time-course of
property effects on knowledge retrieval.
The effects of property that we did observe in this study were not predicted: property
affected accuracy and speed of decisions about filler (non-word) items. Ecological properties
(and to some extent taxonomic properties) slowed down correct decisions about non-words,
relative to neutral properties; accuracy of decisions dropped from neutral to taxonomic
properties as well (although at SOA1650 only). This may reflect overall cognitive load;
perhaps when participants are reasoning about informative properties, they have more
information to process and integrate with the animal information, and are thus doing “more
work” than when they are reasoning about neutral or uninformative properties. However, it’s
not clear why this would be restricted to filler items only.
The lack of interpretable results in Study 4 was not likely caused by faulty stimuli. As
indicated by target effects on accuracy and RT – consistent, if not always significant, pattern
of faster and more accurate verification of related vs. unrelated target words at SOA400 and
900 - animal names did prime related targets at 400 and 900msec SOA relative to unrelated
targets, which suggests that at earlier SOAs, the animal name (“duck”) likely triggered
retrieval of taxonomic and ecological knowledge about the animal. This shows that there were
143
no major problems with the stimuli, and the task itself was sensitive enough to detect
priming effects.
One interesting finding was that taxonomically related targets had a significant
accuracy advantage at an earlier SOA400, whereas ecologically related targets had a
significant advantage at a later SOA900. This might reflect the higher availability of
taxonomic knowledge that, as suggested by the feature listing results of Study 2, is likely to
come to mind earlier than ecological. This adds to the list of differences between taxonomic
and ecological knowledge: not only the first one is more abundant and likely to come to mind
overall, but this advantage is visible from the very first moments of knowledge retrieval.
Surprisingly, by SOA1650 the related items did not show any accuracy advantage
relative to the unrelated items. This was accompanied by two patterns: first, a significant drop
in overall accuracy by SOA1650 relative to shorter SOAs, and second, a non-significant trend
for increasing RT from shorter to longer SOAs. Perhaps by 1650msec after the beginning of
the animal name, the inference generation process enters the stage of hypothesis formulation,
and other factors override priming effects from animal name to related targets. This, however,
still leaves earlier SOA conditions tapping into the retrieval stage.
One unanticipated finding suggested that knowledge retrieval in this experiment was
open to context effects. The effect of intervening targets on participant-generated ecological
inferences was most pronounced at the shortest SOA of 400, the time most likely to overlap
with knowledge retrieval. This suggests that when participants were presented with an
inductive question “Gene X5 is found in ducks. What else is likely to have gene X5? Why”
(in the form of an auditory input “Gene X5.. duck”), their retrieval of knowledge about ducks
was significantly affected by a seemingly random word popping up on the screen (“bird” or
144
“pond”), but it was not systematically influenced by the property (“Gene”). This might be
attributed to modality (auditory for property, visual for target) or order differences in
presentation (property preceding, but target following the animal name). These possibilities
will need to be addressed in further experiments. Nevertheless, taken together, these results
show that it was possible to influence knowledge retrieval by context in this study – and yet,
property did not influence knowledge retrieval.
Another explanation of the lack of property effects in this experiment might have to do
with the inexplicably long RTs observed in this study. In Swinney’s (1979) study, reaction
times in a similar task ranged from 702 to 974msec whereas in this study they ranged from
1143-1317msec. Overall increase in reaction times is likely to come with increased
variability, masking the effects. This may explain our failure to find other meaningful patterns
in the data as well. Several features of this study may have invited longer reaction times: first,
overall cognitive load – after all, generating hypotheses is likely to be a more demanding task
than passive comprehension of sentences. This possibility can be addressed by measuring
duration of other well-calibrated tasks, and if we see a consistent increase in all RTs,
increasing a number of subjects in order to detect the effect. Second, longer reaction times
could have been due to a lack of motivation to respond fast. This can be addressed in future
experiments by introducing a monetary incentive for short RTs, and/or changing the
instructions so that lexical decision is introduced as a primary rather than secondary task for
the participant. Third, participants might have been slow due to overall fatigue: the
experiment started with an extended practice session followed by 72 open-ended induction
questions; indeed some participants remarked that the experiment lasted too long. This can be
addressed by cutting down the number of trials (indeed, we addressed this possibility in a
145
follow-up test not reported here; it did not change the results). Finally, another possibility
has to do with the procedural detail of this study: participants were cued to begin speaking
with a beep, which only came following a delay after the lexical target was presented. Perhaps
knowing that they would not be able to do anything else until the beep, participants delayed
their responses in the lexical decision task to fill in the gap (however, a follow-up test not
reported here ruled out this explanation). In sum, there may be several factors that contributed
to longer reaction times and decreased our chances to detect property effects. However, longer
reaction times cannot fully account for the absence of such effects on word targets; property
did affect reaction times on non-word (filler) items, even though the mean reaction time on
non-word trials was even longer than on words.
In sum, we cannot completely rule out the possibility that the lack of property effects
was caused by some procedural flaws of the study. However, because we did see property
effects on some aspects of the lexical decision task, we know that the method is in principle
capable of detecting property effects. And because the premise category did differentially
prime related vs. unrelated targets, we know that the method is capable of detecting
differential priming. Therefore, it is likely that we failed to see property effects on retrieval
because property does not moderate retrieval of knowledge. This would mean that we need to
revise our assumption about the locus of property effects in induction. It is this latter
possibility that we are going to explore a bit further here.
146
Chapter VI. Revising the Model of Induction: Explaining Evidence as the
Engine of Property Effects in Induction (Study 5)
Study 3 provided some, albeit tenuous, evidence, for interactive recruitment of
property and animal knowledge by inferences: ecological properties strengthened the potential
of ecological knowledge to promote ecological inferences, and taxonomic properties
strengthened the potential of taxonomic knowledge to inhibit ecological inferences.
Study 4 examined the moderating effects of property on knowledge retrieval in real
time, and found none. If this finding reflects the actual absence of property effects on
retrieval, rather than an experimental failure, this would require revision of the proposed
model of induction. We are going to discuss one such possible revision (which, in the end, is
not conditional on the success or failure of Study 4, as we argue below).
On the proposed account (see Figure 2), if property effects on inference do not take
place in retrieval, they must take place during hypothesis generation. As discussed earlier,
little is known about how hypotheses are generated. This fact lays ground for the new, revised
working model of induction shown in Figure 17.
Figure 17. Model of induction 2.0.
147
Upon closer examination, this version of the model turned out to be somewhat
unsatisfactory in terms of explaining the mechanism of property effects in induction. With
minor further elaboration, however, we arrived to a revised model of induction (version 2.1),
which is described next.
Admittedly, hypothesis generation is inherently a creative process, which makes it less
appealing as an object of systematic study. Previously, we made the simplifying assumption
that the hypothesis generation is a simple, direct transformation of currently active knowledge
about premise category (ducks are birds) into a hypothesis by the “other things like that”
function (other birds will have it). The revised model of induction is driven by a second
attempt to de-black-box the hypothesis generation component of inference generation. Our
patch to model 2.0 is to upgrade the hypothesis generation component by including
explanation of presented evidence in the hypothesis generation component.
What is an explanation, and what does it have to do with induction?
The topic of explanation has been in the focus of philosophers’ attention for centuries,
and recently it has started attracting increasing attention from cognitive psychologists. It is
beyond our goals here to review the philosophical debates about what constitutes an
explanation; instead, we focus on one proposal that appears promising in terms of advancing
our model of property effects in inductive inference generation.
On the subsumption account proposed by Williams & Lombrozo (2010), explaining an
observation involves identifying a larger pattern of which the observation is a part of.
“Subsumption (...) theories suggest the defining property of an explanation is that it shows
how the observation being explained is an instance of (subsumed by) a general pattern or
regularity” (Williams, Lombrozo, & Rehder, 2010, p.1); “Explanations tend to relate a
148
particular property or event being explained to broader principles or regularities”
(Lombrozo, in press, p. 20). In this sense, explanation identifies a relevant subset of
knowledge about the observation that can serve as a basis for generalizing to other cases.
There are multiple ways to explain an observation. Aristotle identified four modes of
explanation, or four ways to answer a “why” question: by reference to a formal, efficient,
final, or material cause. A formal explanation refers to the “form or properties that make
something what it is”. For example, a formal explanation of why a tiger has stripes would be
“because it is a tiger, and tigers have stripes”. Perhaps the most common mode of explanation,
an efficient, or causal explanation refers to the proximal mechanisms of change. For example,
one can explain why tigers have stripes by describing the biological mechanisms of gene
expression and protein synthesis resulting in the striped pattern. Final, or teleological
explanations refer to ends, goals or functions. A teleological explanation of tigers’ stripes
would refer to the functions they help to achieve, e.g. how they help this animal to
camouflage. Finally, a material explanation would refer to pigments in the fur (Lombrozo,
2006; Falcon, 2011).
Despite a long history of discussion of explanation and its types in philosophy, it is a
relatively recent topic in cognitive psychology. One relevant line of research focuses on
people’s use of different types of explanation. This work has documented that people
spontaneously engage in formal, causal and teleological explanations, and can selectively
deploy them depending on the specific attributes of the problem. For example, adults accept
teleological explanations selectively, for artifacts, biological parts (chairs are for sitting, feet
are for walking) but not for biological entities or phenomena (lions are not for anything)
149
(Kelemen, 1999), and they have systematic intuitions about appropriateness of formal
explanations (Prasada & Dillingham, 2006).
Another line of research focuses on the role that presence or absence of explanation
(rather than its type) plays in inductive inference. Evidence suggests that a readily available
explanation can undermine reliance on similarity between premises and conclusion. In
Rehder’s (2006) study of induction with artificial categories, featural overlap between
premises and conclusions drove induction ratings, except in conditions where participants
received an explanation of causal connections between features, which eliminated the
contribution of similarity to induction judgments. In a more realistic setting, Shafto & Coley
(2003) showed that while novices’ inferences about diseases shared by marine creatures are
strongly predicted by similarity between species, expert fishermen’s reasoning, in contrast,
shows less reliance on similarity and more on specific explanations about disease
transmission in marine ecosystems (see also Lopez, Atran, Coley, Medin, & Smith, 1997;
Proffitt, Coley, & Medin, 2000). Sloman (1994) used a somewhat different approach to show
that explanations have an effect on inductive inferences over and above similarity. His study
demonstrated that when both premise and conclusion could be explained by reference to the
same principle, the perceived strength of an inductive argument was higher than when
premise and conclusion statements required different explanations, even when the similarity
between premise and conclusion was held constant.
Explanation can also override the effects of premise diversity on argument strength.
Typically, more taxonomically diverse premises provide stronger basis for generalizing a
property to the entire superordinate class (Osherson et al, 1990). However, as shown by
Medin et al. (2003), an argument with diverse premises (skunks & stink bugs have property
150
X; therefore, all animals have property X) can be judged weaker than an argument with
less diverse premises (skunks & deer have property X; therefore, all animals have property X)
– if a more diverse argument invites a ready explanation that overrides diversity (in case of
skunks and stink bugs, property X may have something to do with smell).
In sum, the two lines of research briefly reviewed above suggest that people can
engage in different types of explanations, and the presence vs. absence of explanation affects
induction. Connecting these two lines of work, we pose new questions: does the type of
explanation inform inductive inference? Can different properties trigger different types of
explanation? And can different types of inferences stem from different types of explanations?
We start addressing these questions by revising our model of inference generation.
Hypothesis generation as a locus of property effects
We propose to make the contribution of explanation to induction official by including
it into the model of inference generation. There are several reasons for this proposal. First,
property effects are ubiquitous, and do not seem to be localized in knowledge retrieval.
Second, based on the evidence briefly reviewed above, explanation is involved in inference.
Finally, based on the subsumption account, explanation is a promising candidate for
accounting for property effects; by identifying different relevant patterns or regularities that
an observation is a part of, explanations can support selective generalization that is
characteristic of property effects.
We propose to unpack the black box of hypothesis generation and claim that it
involves explaining the available evidence (premise category + property)32, i.e. identifying the
32 Explanation is presented as a part of hypothesis generation, because it involves going beyond the given information (in contrast to knowledge retrieval).
151
larger pattern that this evidence exemplifies, and then formulating a guess about other
entities that might share the property by virtue of belonging to the same pattern or regularity.
This opens new territory for property effects: besides (or instead of) moderating retrieval of
knowledge, property can also influence hypothesis generation in a systematic way. On this
account of property effects, different properties can trigger different types of explanation. If
explaining consists of identifying an observation as a part of a larger pattern or regularity, and
different properties can trigger different explanations, then different properties can determine
whether a premise of an inductive argument is viewed as a part of one regularity or another.
As the outcome, such different explanations would in lead to different generalization
hypotheses.
This version of the model can explain the interactive nature of inferences in the
absence of evidence for interactive retrieval of knowledge. If a property systematically
triggers one or the other type of explanation (e.g. formal, in terms of taxonomic classes, or
causal, in terms of ecological interactions), it can strengthen recruitment of different subset of
knowledge by inferences, even if the knowledge about an animal premise and property had
been retrieved independently (see Figure 18).
Figure 18. Model of induction, v 2.1 “Hypothesis generation as a locus of property effects”
152
Although the revision of the model was invited by the failure of Study 4, the
proposal about the contribution of explanation to inference generation is not conditional on
the outcome of prior experiments. Conceptually, the idea that the property can promote
different types of explanations is not incompatible with the proposal that property can
moderate knowledge recruitment. However for now, as we are taking the first step towards
exploring the role of explanation in induction, we acknowledge but side-step possible
interactions between property-moderated knowledge retrieval and explanation.
Study 5: Property Effects on Spontaneous Explanations
The proposal about explanations makes a number of predictions. For example, if
participants are generating explanations, and if these explanations are responsible for property
effects on inferences, we should observe different types of explanations for inductive
questions involving different properties. To test the proposal that property affects inferences
by triggering different types of explanation for the provided evidence, we re-examined the
inferences collected in the Study 3. During original data processing, the coders noted that
participants often spontaneously came up with explanations for the provided evidence. For
example, participants might elaborate on the property: “this is a bird flu”; “perhaps B6-cells
defend deer from particular viruses that they are exposed to”; “maybe substance B7 allows
owls to store more glucose in their body which gives them more energy to move their large
body through the air”. In other cases, participants came up with explanations of why the
premise animal might have ended up with the property: “ vultures may get flu E5 from the
dead and decaying animals they feed off of”, “…this gene could have been developed as a
result of the environment that the birds live in”, or even “[the gene will be found in] fish since
pelicans eat them, the pelicans might develop that gene from the fish”.
153
To examine the plausibility of the explanation hypothesis, we conducted a follow-
up analysis, in which we examined the spontaneously generated explanations from Study 3 to
determine whether different properties trigger different types of explanations. We expected
taxonomic properties to trigger predominantly formal explanations, referring to classes of
objects (that would eventually translate into category-based, or taxonomic inferences) and
ecological properties to trigger predominantly causal explanations, referring to interactions
between animals and/or their environment (that would eventually translate into causal
inferences based on ecological interactions). We had no specific predictions about teleological
explanations; given Kelemen’s (1999) findings, if adult participants treat cells, genes, flu
viruses, substances occurring in the bloodstream as biological “parts”, they might attribute
purpose to them; if they treat them as self-standing biological phenomena, we would not
expect to see teleological explanations about them. Preliminary examination of inferences
suggested that material explanations were absent or nearly absent, so this explanation type
was not considered in this analysis.
Coding
Three trained coders independently re-coded all the inferences collected in Study 3 for
the presence of formal, causal and teleological explanations. The coding was done in two
steps. In the first round, the coders identified all responses that contained explanations of the
evidence (regardless of whether the participant actually made a codable projection to another
entity in their response). Explanations of how the animal got the property (“from eating its
prey”), elaborations on the type of property (“it’s a bird flu”), general statements about
property type and distribution (“viruses are transmitted via contact”) were counted as
154
explanations. Twelve percent of inferences (467 out of 3920 codable responses) contained
such spontaneous explanations.
In the second round of coding, the selected explanations were coded into four
categories: formal, causal, teleological, or other type of explanation. Explanations that
referred to kind membership (“mammal gene”, “forest parasite”) were coded as formal;
explanations describing a “story” of interactions between animals and other entities, or in
general a sequence of events that resulted in the premise category having the property
(“vultures got it from eating dead animals”) were coded as casual; explanations referring to
goals, functions or purposes of properties (“these cells are to protect them from the cold”)
were coded as teleological; idiosyncratic or vague explanations that could not be assigned to
any of the three categories were coded as “other”. Multiple codes could apply to a single
response if it contained more than one type of explanation (but not when the response was too
vague to be unambiguously assigned to one category – in which case it was coded as “other”).
Table 6 summarizes the coding scheme and illustrates the coding categories with examples.
Throughout coding, the coders did not have access to the property information (unless
the participant mentioned the property in the response).
155
Table 6. The explanation coding categories - formal, causal, teleological (with examples of premise explanations that participants spontaneously generated during the open-ended induction task).
Explanation type
Property Formal Causal Teleological
Ecological: Flu
Seagulls and birds. they have the bird flu.
Vultures may get flu E5 from the dead and decaying animals they feed off of.
Herbivores allows them to digest ?....
Ecological: Parasite
Anything else that lives in the three places in china that pandas still live... environmental.
Bugs since the bugs could have been on the trees that the beavers were taking wood from to make their shelter. The bugs could have given the parasite to the beavers.
Emus might also have parasite Q1, as both are hunted by the same type of animal. Parasite Q1 could help keep them both safe.
Neutral: Property X
Eagles or any bird because property b7 can be a property only prevalent in birds .
PROPERTY H7 could be in the water in which the turtles swim thus giving them the property.
Property R5 is probably a substance that allows camels to be better adapted to their environment. Therefore, I would guess that scorpions would also have this property R5, because they are also adapted to live in hot, dry climates.
Neutral: Substance in the bloodstream
I have no way of knowing what this substance is. It may be found in all animals or only in owls.
Certain trees might have [substance] K8 because beavers consume tree barks and so it would enter their bloodstream.
Those animals which ingest rubbish as Z1 can help to digest bacteria in the rubbish.\\\ Maybe substance B7 allows owls to store more glucose in their body which gives them more energy to move their large body through the air, so I would guess that other large birds, such as a eagle, would also have substance B7.
Taxonomic: Cell
Mammals, Cougars, Cheetah, Panthers - All the listed mammals are in the cat family, V*-Cell would be a common cell found in the list of mammals.
Anything that the lion has eaten most likely had R8-CELLS and passed it on to the lions.\\\ Roadrunner --> people who run because they are likely to have the same cells as a result of sharing the same activity.
I would guess that moose would also have B6- cells. Perhaps B6- cells defend deer from particular viruses that they are exposed to.
Taxonomic: Gene
Other species of fish because maybe this gene is prevalent in all types of fish.
Other tropical birds, because this gene could have been developed as a result of the environment that the birds live in.\\\ fish since pelicans eat them, the pelicans might develop that gene from the fish.
Birds. (…) this could be a gene related to their bone structure that allows them to fly.
156
Results
Scoring. The initial dataset contained inferences made by 100 subjects about 42
animals. Within this dataset, each subject made an inference about each animal paired with a
particular property, and animals rotated through the properties between subjects. After the
data from 5 participants were discarded as uncodable, it left 95 responses per animal, with a
varying number of subjects per property condition (30 to 34 responses per animal per property
type (taxonomic, ecological, neutral)). For each animal, we calculated the percentage of
subjects who generated each type of explanation out of total number of participants who
reasoned about that animal, separately for each property type (e.g. the proportion of subjects
generating formal explanations out of 30-34 subjects who reasoned about ducks having an
ecological property (flu or parasite)). This yielded 12 percentages, or relative frequency
scores, per animal (3 property types x 4 explanation types), that were arcsine-transformed for
the analyses33.
Analyses. First, we analyzed the percentage of subjects giving uncodable (“other”)
explanations as a function of the property. Uncodable explanations were, on average,
generated by 1.9% of participants per animal, and their frequency did not differ as a function
of the property (one-way repeated measures ANOVA F(2, 82)=1.21, p=.303)). Thus, they
were excluded from the following analysis.
The main question, whether different properties trigger different types of explanations,
was addressed by a 3 property type (ecological, neutral, taxonomic) x 3 explanation type
(formal, causal, teleological) repeated measures ANOVA on relative frequency of
explanations. The overall likelihood of providing an explanation did not differ with the 33 All the reported statistics are based on the arcsine-‐transformed data, all the reported means are raw
157
property (F(2, 82)=.044, p=.957). But the explanations of different types differed in
frequency (F(2,76)=18.836, p<.001, η2partial=.315): causal explanations were more frequent
(5.3%) than formal (2.8%, repeated samples t(41)=4.599, p<.001, Cohen’s d=.71) and
teleological (2.4%, repeated samples t(41)=5.017, p<.001, Cohen’s d=.77), that did not differ
from each other (repeated samples t(41)=1.267, p=.212, Cohen’s d=.2).
Of most theoretical interest was the significant interaction between the property and
explanation type, shown in Figure 19 (F(4,164)=34.442, p<.001, η2partial=.457). When the
property was ecological, out of the total of 13.4% of inferences containing explanations, the
majority of explanations (8.9%) were causal, followed by formal explanations (2.5%)34, with
only a few teleological explanations (0.4%), all significantly different from each other (causal
vs. formal t(41)=6.719, p<.001, Cohen’s d=1.04; causal vs. teleological t(41)=10.984, p<.001,
Cohen’s d=1.69; formal vs. teleological t(41)=4.524, p<.001, Cohen’s d=.7). When the
property was neutral, within a total of 11.6% of inferences containing explanations, causal
explanations (5.3%) still dominated (paired samples causal vs. formal t(41)=5.257, p<.001,
Cohen’s d=.81; causal vs. teleological t(41)=3.005, p=.005, Cohen’s d=.46), while formal
(1.6%) and teleological (3.0%) were equally frequent (t(41)=1.432, p=.16, Cohen’s d=.22).
When the property was taxonomic, within a total of 11.7% of inferences containing
explanations, two most frequent types of explanation were formal (3.7%) and teleological
(3.8%) (not different from each other, paired samples t(41)=.039, p=.969, Cohen’s d=.01) that
34 Most formal explanations about ecological properties were provided in the context of flu (“it’s a bird flu”, 3.6%), rather than parasite (“it’s a bird parasite”, 1.4%). Perhaps, the recent upsurge of avian and swine flu references the media made the species-‐specificity of viruses particularly salient to our participants. For the future studies of property effects, we would recommend replacing flu with another ecological property, such as infection.
158
were both more frequent than causal explanations (1.8%, paired samples vs. formal
t(41)=3.423, p=.001, Cohen’s d=.53; vs. teleological t(41)=3.812, p<.001, Cohen’s d=.59).
Figure 19. Percentage of participants spontaneously producing formal, causal and teleological explanations as a part of generating open-ended inferences about ecological, neutral and taxonomic properties.
Pairwise comparisons across property type confirmed that frequency of a given
explanation type varied with the property. Causal explanations were more frequent for
ecological than neutral than taxonomic properties (paired samples tests, ecological property
vs. neutral t(41)=3.532, p=.001, Cohen’s d=.54, eco vs. tax t(41)=9.659, p<.001, Cohen’s
d=1.49, tax vs. neutral t(41)=6, p<.001, Cohen’s d=.93 (all significant after Bonferroni
correction for 9 unplanned comparisons, critical significance level .006)). Frequency of
formal explanations was higher for taxonomic than neutral properties (t(41)=3.848, p<.001,
Cohen’s d=.59), but it did not differ between taxonomic and ecological (t(41)=2.315, p=.026,
Cohen’s d=.36, n.s. after Bonferroni correction) or ecological and neutral (t(41)=.983,
p=.332, Cohen’s d=.15). Finally, frequency of teleological explanations was significantly
159
higher for taxonomic properties that ecological (t(41)=7.327, p<.001, Cohen’s d=1.13), as
well as for neutral relative to ecological (t(41)=3.575, p=.001, Cohen’s d=.55), but did not
differ for taxonomic and neutral properties (t(41)=2.027, p=.049, Cohen’s d=.13, n.s. after
Bonferroni correction).
The revised proposal claims that the property affects inferences via explanation. The
analyses reported above demonstrate the link between property and explanation type. To
examine the second link, between explanation and inferences, we calculated percentage of
taxonomic and ecological inferences out of the subset of inferences containing explanations,
separately for each explanation type. Most inferences containing formal, teleological, or
uncodable explanations were taxonomic (formal explanations: taxonomic (86.7%) vs.
ecological (17.8%) inferences, paired samples t(41)=9.969, p<.001, Cohen’s d=1.54;
teleological explanations: taxonomic (63.8%) vs. ecological (31.5%) inferences, paired
samples t(41)=4.381, p<.001, Cohen’s d=.68 ; uncodable explanations: taxonomic (62.3%)
vs. ecological (24.8%) inferences, paired samples t(41)=4.295, p<.001, Cohen’s d=.66). In
contrast, the majority of the inferences containing causal explanations were ecological
(taxonomic (30.3%) vs. ecological (88.5%) inferences, paired samples t(41)=-8.608, p<.001,
Cohen’s d=1.33). Such systematic relationship between explanation type and inferences,
taken together with the evidence for the relationship between property and explanation type,
is consistent with the proposal that explanations mediate the effect of property on inferences.
Discussion
This chapter examined the explanation hypothesis by testing the proposal that different
properties trigger different types of explanations. We predicted taxonomic properties to
160
trigger predominantly formal explanations (referring to classes of objects), and ecological
properties to trigger predominantly causal explanations (referring to interactions between
animals and/or their environment). We had no specific predictions about teleological
explanations (referring to goals and functions of properties).
To test this proposal, we examined the premise explanations that participants
spontaneously generated in 12% of their inferences. Since we were interested in how
explanations of the given evidence can affect generated inferences, we focused on
explanations of the information provided to the participants, not on the justifications for
inferences. For example, when given the premise “parasite X is found in ducks”, a participant
might come up with an explanation “ducks must have gotten the parasite from the food” or
“that’s a bird parasite” or even “this parasite helps keep them safe from predators”, and then
generate an inference projecting the parasite to other animals. We were interested in whether
the property affects the type of explanation of the given evidence.
To address this question, we examined the explanations participants spontaneously
generated in Study 3. As predicted, different properties triggered different types of
explanations. When participants were reasoning about ecological properties, the majority of
explanations they provided were causal, referring to a mechanism of property transmission
that could have endowed the animal with the property (e.g., “Owls eat mice and could
contract the flu from the mice that it eats”). In contrast, when participants were reasoning
about taxonomic properties, they were less likely to use causal explanations, preferring formal
(“this cell could be specific to jaguars”) or teleological (“T5 is something to keep them
warm”) explanations. For neutral properties, causal explanations dominated, although they
161
were not as frequent than for ecological properties. In contrast, teleological explanations
were more frequent for neutral than ecological properties.
Our investigation of teleological explanations in this study was exploratory – we did
not have a priori predictions about their distribution across properties. Kelemen (1999)
demonstrated that adults accept teleological explanations for biological parts and organs
(eyes), but do not extend teleological explanations to biological entities or phenomena (lions,
rivers). The results of this study – abundance of teleological explanations about taxonomic
and neutral but not ecological properties - extend Kelemen’s finding of selective deployment
of teleological explanations based on entity type to selective deployment of teleological (as
well as causal and formal) explanations based on the projected property.
Finally, besides demonstrating the relationship between property and explanation type,
we also obtained evidence for a systematic relationship between explanation type and
inferences. Ecological inferences were associated with causal explanations, whereas
taxonomic inferences were associated with formal, teleological or uncodable explanations.
This finding is consistent with the idea that explanations serve as a mediator between
properties and inferences. Although of course, the analyses demonstrating the property-
explanation and explanation-inferences relationships were correlational, and more work is
needed to make stronger claims to determine the role of explanations in property-inference
relationship. One important question for future study is whether explanations play a causal
role in property effects on inferences. The current data do not readily lend themselves to
addressing this question directly. Nevertheless, the informal comparison of effect sizes
indicates that the mean effect size for meaningful explanations on inferences (d=1.18) is
larger than the mean effect size for properties on inferences (d=0.92). This suggests that
162
property-driven explanations are likely to affect inferences directly, rather than being a
mere correlate of properties. We come back to better ways of testing the questions about the
role of explanation in inference at the end of this chapter.
Admittedly, examining open-ended responses for the presence of explanations is not
the most straightforward way to test the proposal about property triggering different types of
explanations. We can identify at least a few objections to this analysis – that are, nevertheless,
not fatal for interpreting its results. First of all, the overall frequency of spontaneously
explanations was lower (12% of inferences) that one might expect if explanation were central
to inference generation. However, the absence of an explanation in the final hypothesis does
not mean that participants did not explain the evidence prior to verbalizing the hypothesis -
after all, participants were not explicitly asked to explain the premise in this task.
Another objection might be raised about the reliability of the pattern based on just
12% of the data. However, 12% of all inferences amounted to 467 explanations, which is not
such a small dataset. Thus we claim that the observed interaction between property and
explanation type is not likely to be a fluke of an extremely small sample.
We also argue that the observed interaction between property and explanation type is
not likely to be a side effect of liberal coding. Of course, open-ended responses are inherently
vague, and since participants were not particularly invested in conveying their message to the
experimenter accurately, their answers were often telegraphic and/or ambiguous. Thus, just
like in coding inferences, we took a rather conservative stance in coding ambiguous
explanations. For example, when a participant commented that a property “is an adaptation”,
it was hard to tell whether they relied on correct understanding of evolutionary theory and
meant that the current form of the trait is a product of natural selection (a causal explanation),
163
or whether they were driven by a common misconception about evolution as progress and
claimed that animal developed that trait in order to improve its performance in the
environment (a teleological explanation). The best way to deal with such ambiguous
responses was to assign them to the category of uncodable explanations. Since the frequency
of uncodable explanations did not vary with the property, the worst effect of response
ambiguity was loss of data, but it was not a likely confound for the observed results.
Finally, the observed interaction was not likely to pragmatic and linguistic factors
affecting likelihood of verbalizing explanations. People may view formal explanations as
self–evident and feel like they do not need to state them explicitly, or have no ready ways of
verbalizing them. For example, “other species of fish because maybe this gene is prevalent in
all types of fish” (involving a formal explanation referring to a “fish gene”) may seem more
redundant than “bugs since the bugs could have been on the trees that the beavers were taking
wood from to make their shelter. The bugs could have given the parasite to the beavers.”
(involving a causal explanation “bugs gave the parasite to beavers”). These factors could
account for overall difference in frequency of explanations, causal being more frequent than
formal and teleological, but not for the interaction between the property and explanation type.
The results of this study provide initial support for the proposal that different
properties invite participants to come up with different types of explanations for the evidence
they are provided with. In Study 3, we saw that inferences vary with property; Study 4
suggested that knowledge retrieval does not differ with property; and this follow-up analysis
shows that explanation type does vary with the property. So far, explanations appear to be the
most promising candidate for a mechanism connecting properties and their effect on
inferences. In sum, these results show that different properties can trigger different
164
explanations, which can potentially lead to different inferences. Although the idea of
explanations being involved in induction is not novel, the claim about property affecting the
type of explanation is an original proposal, which stands in contrast to prior (speculative)
accounts of property effects based on selective attention to different aspects of similarity
between premises and conclusions (Heit & Rubinstein, 1994; Heit, 1998).
Future directions: better ways to test the property-explanation-inductive
hypothesis link. To examine the link between property and different explanation types in
more detail, and to avoid the limitations of coding spontaneous explanations from an open-
ended inference-generation task, we plan on using a closed-end task, in which participants are
given initial evidence (“ducks have B5-cells”) and are asked to select an explanation from
several provided options: “because they are birds” (formal), “because their parents passed
these cells on to them” (causal), “because these cells serve some functions useful for ducks”
(teleological). If different properties are associated with different explanatory modes, we
would expect property to moderate choice of best explanation for the evidence: causal
explanations for ecological properties, formal and teleological for taxonomic properties.
Next, we will examine the link between explanation and inference generation more
directly. The main question is whether explaining the evidence is the cause of property-
sensitivity of final hypotheses. To address this question, we propose three experiments. In the
first two experiments, the main idea is to manipulate presence or absence of explanation, i.e.
prevent participants from explaining evidence in some conditions and see whether this
reduces property effects in inferences. One way to prevent participants from explaining is to
introduce time pressure after knowledge retrieval is completed. For example, participants can
be given unlimited time to study a display of multiple animals and properties (within the
165
limits of working memory), presumably completing retrieval of knowledge about them.
Then, after a button press, a connector between an animal and property would appear, and
participants will be forced to make a decision about a projection target (does it share the
property?) within a limited time period. It is of course possible that participants might attempt
to pre-generate explanations, but with as few as three properties and four animals, they would
need to generate 12 explanations – instead of waiting till they know the specific premise
combination, and generating just one explanation. Knowing that often participants are trying
to get through class-credit experiments with as little investment of effort as possible, we have
some faith that this manipulation might successfully impair explanation generation.
Another way to prevent participants from explaining could be by manipulating overall
framing of questions as an intentional “set up”: they can be told that someone was trying to
confuse the biologists on purpose, and introduced a certain gene/parasite/substance in ducks
that they studied”, and invited to make an inference. To the extent that these manipulations
impair generation of explanations, and if explanations play a causal role in property effects in
induction, the revised model predicts that they should lead to reduction of property effects in
the inferences.
The third experiment would examine the link between the type of explanation and
final hypotheses. Instead of relying on participants generating explanation, we can control the
type of explanation by providing participants with the explanation of the evidence – “ducks
got this property from the food they consumed” – and see whether this manipulation is a.
sufficient to drive differences in generated hypotheses, and b. strong enough to override the
initial intuitions about properties (in cases where provided explanation does not match the
typical explanation induced by that property type, e.g. teleological explanation for an
166
ecological property (“this parasite helps to keep away the predators”), or a causal
explanation for a taxonomic property (“this gene developed as a result of a mutation caused
by extreme heat in tropics”).
These studies that will examine the role of explanation in inductive inference,
alongside with modified versions of the Study 4 examining property effects on knowledge
retrieval (see the discussion of Study 4 for improvement suggestions) outline our plan for
further investigation of property effects in inference generation.
167
Chapter VII. General discussion
Overview
The goal of this project was to examine the mechanism of property effects in inductive
inference generation. Property effects - changes in argument strength or generated
conclusions based on the projected property – are ubiquitous in induction, but the theoretical
accounts of induction capable of addressing the underlying psychological mechanism are
lacking.
Model 1.0: Property Effects via Knowledge Retrieval
To address this gap, we proposed a simple model of inductive inference, which
identifies two components within inference generation: retrieval of knowledge about the
premise, and generation of an inductive hypothesis based on that information (and general
knowledge about the world). We decided to tackle the problem of specifying the mechanism
of property effects by targeting the knowledge retrieval component rather than the hypothesis
generation component because more work had been done on context-dependent retrieval of
information than on the complex and inherently creative process of hypothesis generation.
As a first approach to localizing property effects within theinference generation
process, we proposed that they arise during knowledge retrieval. We outlined three retrieval
strategies that could yield property effects. In the first, “minimalist” strategy, only the
knowledge about property is retrieved. In the second strategy, knowledge about property and
knowledge about premise category are retrieved autonomously and provide independent
contributions to the inferences formulated at the. In the third strategy, both property and
premise category knowledge are retrieved, but this time interactively; property serves as a
context which affects retrieval of knowledge about premise category, and such context-
168
dependent retrieval of information from semantic memory gives rise to the observed
property effects in inferences. To illustrate the latter strategy with an example, when a person
is presented with the premise “flu X is found in ducks”, and is asked to generate an inductive
hypothesis about what else might have the same flu, the epidemiological nature of the
property (flu) could influence which knowledge about the premise category (ducks) is
retrieved from the memory (such as that ducks are eaten by foxes, and could thereby transmit
flu to them). This strategy is consistent with the non-verbalized or vaguely verbalized
assumption in the field that “the property being considered influences which features [of the
categories] are important for induction” (Heit & Rubinstein, 1994, p. 411) and the property
effects have something to do with selective sampling of premise category features from
memory (Heit, 1998; Heit & Rubinstein, 1994).
We tested this proposal about retrieval-localized property effects on the example of
folkbiological reasoning, or reasoning about natural world. This knowledge domain provides
a good stage for investigating property effects in inductive reasoning, since it is structured by
two relatively well defined knowledge structures that can serve as bases for inferences –
taxonomic, based on hierarchy of biological classes, and ecological, based on shared habitat
and interactions between species. In addition, property effects have already been documented
in this knowledge domain, in a variety of experimental paradigms – from evaluation of
inductive arguments, to open-ended inference generation tasks that we used in this project.
We split the objective of examining property effects in folkbiological induction into
two tasks and addressed them in a series of four studies. First, we characterized the types and
sources of knowledge brought about by the premise of an inductive argument (description of
169
the “ingredients”; Studies 1 & 2). Second, we tested our proposal the retrieval-based
mechanism of property effects in Studies 3 & 4 (evaluating our proposal about the “recipe”).
The Ingredients
Study 1 demonstrated that knowledge about properties can contain systematically
biased beliefs about their distribution. Some properties are believed to be homogeneous
within taxonomic categories (e.g. fish tend to share similar genes), other properties are
believed to be homogeneous within ecological properties (e.g. jungle animals tend to share
infectious diseases). Such distributional beliefs can be the origin of the taxonomic or
ecological bias that properties have been shown to exert on inferences. Although the idea of
distributional beliefs about properties is not new (see Goodman’s discussion of
overhypotheses (1955)), the demonstration that people actually do possess such beliefs, and
that they can differ from property to property, is novel.
Study 2 examined taxonomic and ecological knowledge about premise categories.
Based on the features that participants listed for a set of animals, participants possess both
taxonomic and ecological knowledge about species, but taxonomic knowledge proved to be
more abundant and salient. Taken together, Studies 1 & 2 described what an inductive
premise like “flu X is found in ducks” brings to the table: abundant and salient taxonomic
knowledge about ducks (they are birds, they have beaks and feathers and webbed feet, they
quack, walk funny, and can fly), less abundant and less salient ecological knowledge about
ducks (they live in ponds, eat plants and insects, they are eaten by foxes and coyotes), and the
beliefs about the property distribution (the property of having a certain flu is likely to be
homogeneous within animals that share a habitat and/or interact ecologically).
170
The Recipe
The next two studies examined how this knowledge about premise category and
property is combined and translated into inferences. Study 3 replicated property effects in a
novel experimental setup with property varied within subjects. The fact that the same
participant could vary their inferences from question to question based on the property they
were reasoning about suggested that the property effects are indeed driven by processing of
specific attributes of each inductive question. To characterize how premise information is
translated into inferences, this study examined the recruitment of premise category
knowledge, or the relation between the available taxonomic and ecological knowledge about
animals and the inferences generated about them. It showed that salient ecological knowledge
about animals facilitates ecological inferences about them, and the salient taxonomic
knowledge inhibits ecological inferences. Most importantly, both facilitation and inhibition
were moderated by the property; ecological properties strengthened the facilitatory effect of
ecological knowledge on ecological inferences, while taxonomic properties strengthened the
inhibitory effect of taxonomic knowledge on ecological inferences. Although no such
property-moderated relationship to knowledge was observed for taxonomic inferences, we
took the results of this study to suggest overall that premise category knowledge is recruited
to generate an inference, and that such recruitment is moderated by the nature of the property.
This interactive recruitment of premise category knowledge & property knowledge provided
the first constraint on the process that might have generated property effects, suggesting the
interactive retrieval of property and category knowledge strategy (rather than retrieval of
property knowledge only, or independent retrieval of property and category knowledge).
171
Study 4 subjected the interactive retrieval strategy proposal to a further test, by
probing directly into the retrieval of category knowledge unfolding over time. This study
employed a combination of two tasks: the open-ended induction task insured participants
engaged in inference generation, while a lexical decision task served as a priming probe for
activation of taxonomic and ecological knowledge about premise animals. Even though the
method proved capable of detecting activation of information about the premise animals (as
reflected by accuracy of decisions about related vs. unrelated targets), the results provided no
evidence of property-moderated retrieval of category knowledge. This was surprising, given
the results of Study 3 which suggested that some interactive processing of property and
premise must have taken place, in order to yield the observed interactive recruitment of
knowledge by ecological inferences.
Implications for Model 1.0
In sum, by the end of Study 4, we had data from Study 3 saying that property and
category knowledge were combined interactively, and the data from Study 4 saying the
interaction was not likely taking place in retrieval. Based on the proposed retrieval-based
model of property effects, if both property and premise category knowledge are retrieved
independently, and the retrieved knowledge is directly translated into inferences, the resulting
inferences should reflect independent contributions of property and premise category. Yet, the
outcome inferences showed clear signs of interaction between property and category
knowledge: the effects of knowledge on inferences were strengthened in the presence of a
congruent property and weakened in the presence of an incongruent property. Such result was
not compatible with the Model 1.0 and required a model revision.
172
Model 2.1: Property Effects via Explaining the Evidence
The revised model had to be able to account for the pattern in which premise category
and property knowledge are combined interactively, but not during retrieval. Looking for such
interaction elsewhere within our model of induction, we introduced explanation of the
evidence to the model (within the hypothesis generation component) and nominated it as a
source of property effects on inferences. By explanation we mean identifying a larger pattern
or regularity that the observation is a part of (Williams & Lombrozo, 2010). The revised
proposal claims that first, knowledge about both property and premise category is retrieved,
and, since we have no evidence for interactive retrieval, we assume they are retrieved
independently. Next, knowledge retrieval is followed by hypothesis generation which
involves first, explaining the evidence provided by the premise (the combination of premise
category and the property) by identifying a larger regularity that it belongs to, and second,
formulating a guess about other entities that might share the property by virtue of belonging to
the same regularity. Within this account, different properties can affect inferences by
triggering different types of explanations – formal, based on category membership, or causal,
based on a sequence of enabling events. Different explanations, i.e., identifying an
observation as a part of different types of regularities, can in turn lead to the generation of
different hypotheses, and ultimately different inferences.
The revised proposal can accommodate the interactive relationship between premise
and property observed in outcome inferences (Study 3), and absence of evidence for such
interaction during retrieval (Study 4). The knowledge about premise category and property
may be retrieved via two independent processes, but a subsequent explanation of the evidence
can introduce the interactive relationship between knowledge and outcome inferences. The
173
variability in inference types associated with the property – formal explanations, in terms
of taxonomic classes, prevalent for taxonomic properties, and causal explanations, in terms of
ecological interactions, prevalent for ecological properties – can be driving the interactive
effects in the inferences.
To subject this proposal to a first test, we conducted Study 5 in which we explored the
explanations many participants spontaneously generated for the premises while making
inferences. The revised proposal predicted that the type of explanation should vary with the
property. The data supported this prediction: for ecological properties, causal explanations
dominated; for taxonomic properties, the majority of explanations were formal or teleological
(attributing ends or purposes to properties). Neutral properties showed a mixed pattern of
explanation types. Besides the link between property and explanation, this study also
demonstrated a systematic relationship between explanation and inferences: ecological
inferences were associated with causal explanations, while taxonomic inferences were
associated with and formal and teleological explanations. This evidence provides initial
support to our proposal localizing property effects within the hypothesis generation
component. Specifically, it shows that participants systematically vary in the kinds of
spontaneous explanations they provide for the evidence provided by the premise, depending
on the property, and these varying explanations are likely to translate in property-sensitive
inferences.
Conclusions and Implications
In sum, across the studies within this project, we have some evidence that – contrary
to our original hypothesis - property effects do not take place in retrieval. This questions the
existing, but not tested, assumption in the field about the mechanism of property effects based
174
on context-dependent retrieval of information from semantic memory (Heit & Rubinstein,
1994).
We also have some promising evidence that property effects may take place during
hypothesis generation. This potentially clarifies the connection between two lines of research
- on explanations and on induction. Most researchers agree that these are related, but very
little supporting empirical work exists, although it is increasingly acknowledged that presence
of available explanation can reduce reliance on overall similarity and override effects of
similarity and diversity of evidence on induction (see Lombrozo, in press, and Lombrozo,
2006, for a review). In this project we demonstrated that explanations do not just “mess up”
existing regularities in induction, but may in fact be a part of the mechanism of one such
established regularity – property effects in induction.
The revised model can also account for other phenomena in induction. For example, it
can explain why idiosyncratic properties discussed in the introduction, such as “has a gum
stuck on the bottom”, “fell on the floor this morning”, are deemed unprojectable. If an
explanation suggests that an observation is not a part of a larger regularity, such property
would not be generalized from one observation to other entities. The explanation component
of the model can also account for a finding from an unpublished study from our laboratory:
participants often refuse to generalize a property from a premise postulating two animals not
related either taxonomically or ecologically to share a property (e.g. “flu X is found in
peacocks and geckos”) – even though they are perfectly willing to make inferences if that
paired premise is split in two separate inductive questions (“flu X is found in peacocks”, “flu
Y is found in geckos”). On the proposed account, the absence of an obvious relationship
between peacocks and geckos suggests that a shared flu is more likely a coincidence rather
175
than a part of a larger regularity – preventing participants from generalizing it to other
entities.
The explanation-based proposal of property effects also fits well with the Relevance
Theory of Induction (Medin, Coley, Storms, & Hayes, 2003). This theory, as reviewed in the
introduction, claims that people assume that the premise is relevant to the conclusion and seek
to identify a relation between the premise and conclusion that would justify such an
assumption. We can extend this proposal to open-ended inference generation. When
participants are presented with a premise only, they can still ask, “why am I given this
particular information, this property plus this category?” (instead of “why is this premise
given with this particular conclusion?”). With this question, participants seek to identify a
relevant connection between the property and the category within the premise – in other
words, they attempt to explain the premise evidence. This goes in line with Medin et al.’s
(2003) proposal about the way people reason about uninformative properties, that people seek
to interpret them and extract some meaning for the property from the salient features of
premise categories. For example, when presented with the premise “penguins have Property
Y”, “people might expect that property Y is an adaptation to an Antarctic environment or
linked to swimming and waddling rather than flying” (p. 522). Indeed, as reported in Chapter
VI, for neutral, or uninformative properties, explicit spontaneous explanations were just as
frequent as for informative properties – consistent with Medin et al. (2003). In addition, our
results extend their proposal and demonstrate that in reasoning about neutral properties,
people use a variety of explanation types - formal, causal, or teleological – without a marked
preference for any single type, in contrast to reasoning about informative properties, that
apparently bias people towards one type of explanation or another.
176
Future Directions
This project by no means has exhausted the list of questions about property effects in
induction, and in fact, it raised some interesting new questions. As the first next step, we plan
on addressing some questions that the Study 4 left open. Although we did see priming effects
from animal names to related words presented as targets in the lexical decision task reflected
in accuracy of decisions (e.g. decisions about related words “bird” and “pond” following
“duck” were more accurate than decisions about an unrelated word “sofa”), the timing of
decisions was not affected. Thus, although we were able to extract useful (if tentative)
conclusions from that study – that property may not moderate retrieval of premise category
knowledge – we have not characterized retrieval process very well. Definitely, more work is
needed to describe how knowledge is retrieved in the context of inductive inference.
Introducing explanation into the account of property effects appears to be a step in the
right direction, but it also raises many new questions. In a sense, the question is shifted from
how inferences are generated to how explanations are generated. Lombrozo (in press)
concludes her review of explanation by saying that the question of how explanations are
generated remains largely unaddressed, and “explanation generation confronts many of the
most difficult questions in cognitive science concerning contents and structure of general
beliefs, and how these are retrieved in particular contexts” (p. 32). Thus, we need to explore
how explanations are generated, what factors affect the type of generated explanation (besides
the property), and what factors affect whether a person seeks explanation for the evidence
before generating a hypothesis or not.
One interesting avenue of research could examine the relationship between property
effects on inferences and individual’s need for cognition (Cohen, Stotland, & Wolfe,1955).
177
Individuals with higher need to find meaningful structure in information may tend to
explain the evidence provided by the premises more, and if explanations drive property
effects, this should result in a more pronounced pattern of property-driven differences in
inferences. Unless higher need for cognition is correlated with higher context-sensitivity of
retrieval (which can be tested separately), correlation between need for cognition and
inferences’ sensitivity to property differences would provide further support to our revised
proposal about the explanation-driven mechanism of property effects in induction.
To sum up, this project makes a step towards specifying the mechanism of property
effects in induction in two ways. First, it suggests that property effects do not work via
property-based retrieval of knowledge about premise categories from memory. Second, it
introduces property-driven explanations as a possible source of property effects. Of course,
these proposals are not mutually exclusive, and our main suggestion for the further research
would be not to abandon studying knowledge retrieval in induction, but to expand research on
the mechanism of property effects to include explanations.
178
References
Ahn, W., Kim, N. S., Lassaline, M. E., & Dennis, M. J. (2000). Causal status as a determinant of feature centrality. Cognitive Psychology, 41, 361-416.
Barclay, J.R., Bransford, J.D., Franks, J.J., McCarrell N.S., & Nitsch, K. (1974). Comprehension and semantic flexibility. Journal of Verbal Learning and Verbal Behavior, 13, 471-481.
Barsalou, L.W. (1982). Context-independent and context-dependent information in concepts. Memory & Cognition, 65, 15-32.
Barsalou, L.W. (1989). Intraconcept similarity and its implications for interconcept similatiry. In S. Vosniadou & A. Ortony (Eds.), Similarity and analogical reasoning (pp. 3-121). New York, NY: Cambridge University Press.
Buchanan, T.W. (2007). Retrieval of emotional memories. Psychological Bulletin, 133, 761-779
Chomsky, N. (1980). Rules and Representations. Oxford: Basil Blackwell.
Cohen, A.R., Stotland, E., & Wolfe, D.M. (1955). An experimental investigation of need for cognition. Journal of Abnormal and Social Psychology, 51(2), 291-294.
Coley, J.D. (2005). Relational properties in folk biological induction. Paper presented at the Biennial Meeting of the Society for Research in Child Development, Atlanta, GA.
Coley, J.D. (2009, June). Folk biological induction: Environmental and developmental differences in the use of taxonomic and ecological knowledge. Paper presented at the International Conference on Biological Understanding and Theory of Mind, Reims, France.
Coley, J.D. (in press). Experience and ecological reasoning: Environmental differences in the development of flexible category-based induction. Child Development.
Coley, J.D., Hayes, B., Lawson, C., & Moloney, M. (2004). Knowledge, expectations, and inductive inferences within conceptual hierarchies. Cognition, 90, 217-523.
Coley, J. D., Medin, D. L., & Atran, S. (1997). Does rank have its privilege? Inductive inferences within folkbiological taxonomies. Cognition, 64, 73-112.
Coley, J.D., Muratore, T.M., & Vasilyeva, N.V. (2009, April). Folk Biological Induction: Environmental and Developmental Differences in the Use of Taxonomic and Ecological Knowledge. Poster presented at the biennial meeting of the Society for Research in Child Development, Denver, CO.
Coley, J.D. & Vasilyeva, N.Y. (2010) Generating inductive inferences: premise relations and property effects. In Ross, B. (Ed.) The Psychology of Learning and
179
Motivation (53), pp.183-226. Burlington: Academic Press.
Coley, J.C., Vasilyeva, N.Y., & Muratore, T.M. (2008, August). Selective Use of Taxonomic and Ecological Knowledge in Open-ended Category Induction. Paper presented at the 6th International Conference on Thinking, Venice, Italy.
Coley, J.D., Vitkin, A.Z., Seaton, C. E, & Yopchick, J.E. (2005). Effects of experience on relational inferences in children: The case of folk biology. Paper presented at the 27th Annual Conference of the Cognitive Science Society, Stresa, Italy.
Ebbinghaus, H. (1885/1962). Memory: A contribution to experimental psychology. New York: Dover.
Falcon, A. (2011). Aristotle on Causality. In The Stanford Encyclopedia of Philosophy. Edward N. Zalta (ed.). Retrieved from <http://plato.stanford.edu/archives/spr2011/entries/aristotle-causality/>.
Farris, H., & Revlin, R. (1989). Sensible reasoning in two tasks: Rule discovery and hypothesis evaluation. Memory & Cognition, 17, 221-232.
Feeney, A., Coley, J.D. & Crisp, A.K. (2010). The relevance framework for category-based induction: Evidence from garden path arguments. Journal of Experimental Psychology: Learning, Memory and Cognition, 36, 906-916.
Gelman, S.A. (1988). The development of induction within natural kind and artifact categories. Cognitive Psychology, 20, 65-95.
Gelman, S.A. & Coley, J.D. (1990). The importance of knowing dodo is a bird: Categories and inferences in 2-year-old children. Developmental Psychology, 26, 796-804.
Gettys, C.F., & Fisher, S.D. (1979). Hypothesis plausibility and hypothesis generation. Organizational Behavior and Human Performance, 24, 93-110.
Glucksberg, S., & Estes, Z. (2000). Feature accessibility in conceptual combination: Effects of context-induced relevance. Psychonomic Bulletin & Review, 7, 510-515.
Goel, V., Gold, B., Kapur, S., & Houle, S. (1997). The seats of reason: A localization study of deductive and inductive reasoning using PET (O15) blood flow technique. NeuroReport, 8, 1305-1310.
Goodman, N. (1955). Fact, Fiction, and Forecast. Indianapolis, IN: Bobbs-Merrill,
Goodman, N. (1972). Seven structures on similarity. In Problems and Projects. New York, NY: Bobbs-Merrill.
Gutheil, G., & Gelman, S. A. (1997). Children’s use of sample size and diversity information within basic-level categories. Journal of Experimental Child
180
Psychology, 64, 159-174.
Heibeck, T.H. & Markman, E.M. (1987). Word learning in children: An examination of fast mapping. Chile Development, 58, 1021-1034.
Heit, E. (1997). Features of similarity and category-based induction. In Proceedings of
the Interdisciplinary Workshop on Similarity and Categorisation, 115-121. Edinburgh.
Heit, E. (1998). A Bayesian analysis of some forms of inductive reasoning. In M. Oaksford & N. Chater (Eds.), Rational models of cognition, (pp.248-274). Oxford: Oxford University Press.
Heit, E. (2000). Properties of inductive reasoning. Psychonomic Bulletin & Review, 7, 569-592.
Heit, E. (2007). What is induction and why study it? In Feeney, A. & Heit, E. (Eds.), Inductive reasoning. Cambridge University Press.
Heit, E. (2008). Models of inductive reasoning. In Sun, R. (Ed.), Cambridge handbook of computational psychology, 322-338. Cambridge University Press.
Heit, E. & Rubinstein, J. (1994). Similarity and property effects in inductive reasoning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 411-422.
Hess, D.J., Foss, D.J., & Carroll, P. (1995). Effects of global and local context on lexical processing during language comprehension. Journal of Experimental Psychology: General, 124, 62-92.
Heyman, G.D., & Gelman, S.A. (2000). Preschool children’s use of trait labels to make inductive inferences. Journal of Experimental Child Psychology, 77, 1-19.
Johnson-Laird, P.N. (1994). Mental models and probabilistic thinking. Cognition, 50, 189-209.
Kalish, C.W., & Gelman, S. A. (1992). On wooden pillows: Multiple classifications and children’s category-based inductions. Child Development, 63, 1536-1557.
Keil, F.C. (1981). Constraints on knowledge and cognitive development. Psychological Review, 88(3), 197-227.
Keil, F.C. (2003). Folkscience: Coarse interpretation of a complex reality. Trends in Cognitive Sciences, 7, 368-373.
Kelemen, D.J. (1999). Why are rocks pointy?: Children’s preference for teleological explanations of the natural world. Developmental Psychology, 35, 1440-1453.
Kemp, C., Perfors, A., and Tenenbaum, J. B. (2007). Learning overhypotheses with
181
hierarchical Bayesian models. Developmental Science 10(3), 307-321.
Klayman, J. & Ha, Y.W. (1989). Hypothesis testing in rule discovery: Strategy, structure and content. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 596-604.
Lombrozo, T. (2006). The structure and function of explanations. Trends in Cognitive Sciences, 10(10), 464-470.
Lombrozo, T. (in press). Explanation and abductive inference. K.J. Holyoak and R.G. Morrison (Eds.), Oxford Handbook of Thinking and Reasoning.
Lopez, A., Atran, S., Coley, J.D., Medin, D.L., & Smith, E.E. (1997). The tree of life: Universal and cultural features of folkbiological taxonomies and inductions. Cognitive Psychology, 32, 251-295.
Macario, J. F., Shipley, E. F., & Billman, D. O. (1990). Induction from a single instance: Formation of a novel category. Journal of Experimental Child Psychology, 50, 179-199.
Mandler, J.M., & McDonough, L. (1996). Drinking and driving don’t mix: Inductive generalization in infancy. Cognition, 59, 307-335.
Mandler, J.M., & McDonough, L. (1998). Studies in inductive inference in infancy. Cognitive Psychology, 37, 60-96.
Marr, D. (1982). Vision. New York: W.H. Freeman.
McDonald, J. (1990). Some situational determinants of hypothesis-testing strategies. Journal of Experimental Social Psychology, 26, 255-274.
McDonald, J., Samuels, M., & Rispoli, J. (1996). A hypothesis assessment model of categorical argument strength. Cognition, 59, 199-217.
McElree, B., Murphy G.L., & Ochoa T. (2006). Time course of retrieving conceptual information: A speed-accuracy trade-off study. Psychonomic Bulletin & Review, 13, 848-853.
McKoon, G., & Ratcliff, R. (1988). Contextually relevant aspects of meaning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 331-343.
Medin, D. L., Coley, J. D., Storms, G. & Hayes, B. K. (2003) A relevance theory of induction. Psychonomic Bulletin & Review, 10, 517-532.
Medler, D.A., & Binder, J.R. (2005). MCWord: An On-Line Orthographic Database of the English Language. http://www.neuro.mcw.edu/mcword/
Mehle, T. (1982). Hypothesis generation in an automobile malfunction inference task. Acta Psychologica, 52, 87-106.
182
Mehle, T., Gettys, S., Manning, C., Baca, S., & Fisher, S. (1981). The availability explanation of excessive plausibility assessments. Acta Psychologica, 49, 127–140.
Mill, J. S., Collected Works of John Stuart Mill, J. M. Robson (ed.), Toronto: University of Toronto Press, 1963
Murphy, G. L., & Medin, D. L. (1985). The role of theories in conceptual coherence. Psychological Review, 92, 289-316.
Murphy, G.L. (2002). The big book of concepts. Cambridge, MA: MIT Press.
Neely, J.H. (1977). Semantic priming and retrieval from lexical memory: roles of inhibitionless spreading activation and limited-capacity attention. Journal of Experimental Psychology: General, 106, 226-254.
Nguyen, S.P., & Murphy, G. L. (2003). An apple is more than just a fruit: Cross-classification in children’s concepts. Child Development, 6, 1783-1806.
Nisbett, R.E., Krantz, D.H., Jepson, C., & Kunda, Z. (1983). The use of statistical heuristics in everyday inductive reasoning. Psychological Review, 90, 339-363.
Oaksford, M. & Hahn, U. (2007). Induction, deduction and argument strength in human reasoning and argumentation. In Feeney, A. & Heit, E. (Eds.), Inductive reasoning. Cambridge University Press.
Osherson, D.N., Smith, E.E., Wilkie, O., Lopez, A., & Shafir, E. (1990). Category-based induction. Psychological Review, 97, 185-200.
Posner M. I. & Snyder, C. R. R. (1975). Attention and Cognitive Control. In Robert L. Solso (Ed.), Information Processing and Cognition: The Loyola Symposium. Lawrence Erlbaum.
Prasada, S., & Dillingham, E.M. (2006). Principled and statistical connections in common sense conception. Cognition, 99, 73-112.
Proffitt, J.B., Coley, D.D., & Medin, D.L. (2000). Expertise and category-based induction. Journal of Experimental Psychology: Learning, Memory, & Cognition, 26, 811-828.
Rehder, B. (2006) When similarity and causality compete in category-based property generalization. Memory and Cognition, 34, 3–16
Rips, L.J. (2001). Two kinds of reasoning. Psychological Science, 12, 129-134.
Ross, B.H., & Murphy, G.L. (1999). Food for thought: Cross-classification and category organization in a complex real-world domain. Cognitive Psychology, 38, 495-553.
Rotello, C.M., & Heit, E. (2009). Modeling the effects of argument length and validity
183
on inductive and deductive reasoning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35, 1317-1330.
Shafto, P., & Coley, J.D. (2003). Development of categorization and reasoning in the natural world: Novices to experts, naïve similarity to ecological knowledge. Journal of Experimental Psychology: Learning, Memory, and Cognition, 29, 641-649.
Shafto, P., Coley, J.D. & Vitkin, A. (2007). Availability in category-based induction. In A. Feeney & E. Heit (Eds.), Inductive reasoning: Experimental, developmental, and computational approaches. Cambridge University Press, 114-136.
Shafto, P., Coley, J.D., & Baldwin, D. (2007). Effects of time pressure on context-sensitive property induction. Psychonomic Bulletin & Review, 14, 890-894.
Sloman, S.A. (1993). Feature-based induction. Cognitive Psychology, 25, 213-280.
Sloman, S.A. (1994). When explanations compete: the role of explanatory coherence on judgments of likelihood. Cognition, 52, 1-21.
Sloutsky, V.M., & Fisher, A.V. (2004). Induction and categorization in young children: A similarity-based model. Journal of Experimental Psychology: General, 133, 166-188.
Smith, E.E., Shafir, E., & Osherson, D. (1993). Similarity, plausibility, and judgments of probability. Cognition, 49, 67-96.
Springer, K. (1992). Children’s awareness of the biological implications of kinship. Child Development, 63, 950-959.
Swinney, D. (1979). Lexical access during sentence comprenension: (Re)consideration of context effects. Journal of Verbal Learning & Verbal Behaior, 18, 645-659.
Tabossi, P. (1982). Sentential context and the interpretation of unambiguous words. Quarterly Journal of Experimental Psychology, 34, 79-90.
Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27, 324–340
Tabossi, P., & Johnson-Laird, P.N. (1980). Linguistic context and the priming of semantic information. Quarterly Journal of Experimental Psychology, 32, 595-603
Tenenbaum, J. B., Kemp, C., and Shafto, P. (2007). Theory-based Bayesian models of inductive reasoning. In Feeney, A. & Heit, E. (Eds.), Inductive reasoning. Cambridge University Press.
Vasilyeva, N.Y. & Coley, J.D. (2009, June). Effects of Development and Experience on Inductive Reasoning about Nature. Paper presented at the International
184
Conference on Biological Understanding and Theory of Mind, Reims, France.
Vitkin, A.Z, Coley, J.D., & Feigin, K. (2005). Accessibility of Taxonomic and Script Knowledge in the Domain of Food. Paper presented at the 46th Annual Meeting of the Psychonomic Society, Toronto.
Vitkin, A.Z., Coley, J.D., & Hu, R. (2005). Children’s use of relevance in open-ended induction in the domain of biology. Paper presented at the 27th Annual Conference of the Cognitive Science Society, Stresa, Italy.
Vitkin, A.Z., Coley, J.D., & Kane, R. (2005). Salience of taxonomic and ecological relations in children’s biological categorization. Paper presented at the Biennial Meeting of the Society for Research in Child Development, Atlanta, GA.
Vitkin, A.Z., Vasilyeva, N.Y., Coley, J.D. (2007, August). Experience and the Development of Flexible Inductive Reasoning in Biology. Paper presented at the annual meeting of British Psychological Society 2007 Developmental Section Conference, Plymouth, UK.
Waxman, S. R., Lynch, E. B., Casey, K. L., & Baer, L. (1997). Setters and samoyeds: The emergence of subordinate level categories as a basis for inductive inference in preschool-age children. Developmental Psychology, 33, 1074-1090.
Williams, J.J. & Lombrozo, T. (2010). The role of explanation in discovery and generalization: evidence from category learning. Cognitive Science, 34, 776-806.
Williams, J.J., Lombrozo, T., & Rehder, B. (2010). Why does explaining help learning? Insight from an explanation impairment effect. In S. Ohlsson & R. Catrambone (Eds.), Proceedings of the 32nd Annual Conference of the Cognitive Science Society (pp. 2906-2911). Austin, TX: Cognitive Science Society.
185
Appendix A. General Instructions Used in Study 1.
In this experiment, you will be asked to imagine that one animal species has a certain
property, and asked to estimate one of two things – 1. What percentage of other animals in the
same biological family is likely to have that property, or 2. What percentage of other animals
living in the same area is likely to have that property.
Here are a couple of examples:
Imagine that one animal species has slow metabolism.
- What percentage of animals in the same biological family is likely to have slow
metabolism?
Enter a number between 0 & 100 _____
or
Imagine that one animal species has thick fur.
- What percentage of animals living in the same area is likely to have thick fur?
Enter a number between 0 & 100 _____
Animal species are kinds of animals. Examples of animal species are monkey, deer,
trout, parrot, bluejay, scorpion, etc. When you are answering the questions, don’t try to guess
what exact animal species it refers to. Instead, read each question as referring to some
unspecified animal species. By biological families we mean more general classes of animals
that include many different species. Examples of biological families are mammal, bird, fish,
insect, reptile, etc. An area where an animal lives can refer to different habitats and/or
geographical locations, e.g. desert, forest, ocean, pond, savannah, Australia, Africa, etc.
186
Appendix B. Selection of Animal Stimuli for Studies 2 and 3: Animal Typicality
Ratings
Fifty-seven Northeastern University undergraduate students completed typicality
rating task in exchange for course credit. Typicality rating packets contained total of 97
animal names from a variety of biological classes and habitats. Each animal name was
accompanied by twelve 9-point typicality rating scales, asking to rate each animal on “how
good of an example the animal is of the category, 1 being not a good example at all and 9
being the best possible example”. Participants rated each animal on their goodness of
membership in the following thirteen categories: bird, fish, insect, mammal, reptile
(taxonomic categories) and Arctic animal, desert animal, forest animal, pond animal, ocean
animal, rainforest animal, savannah animal (ecological categories). The rating scales appeared
in fixed random order, same for all animals and across participants. Participants worked
through the rating packets at their own pace.
Based on these ratings, we selected 42 animals aiming for a diverse set that would
represent a variety of taxonomic and ecological categories, with all animals relatively good
members of their dominant categories (above the scale midpoint of 5). The selected animals
and their respective typicality ratings are shown in Table A1. Some cells remained naturally
uncovered, since there exist few desert fish or ocean insects that would be familiar to our
participants.
Mean taxonomic (7.69) and ecological (6.75) typicality ratings for the selected animal
set were significantly above the scale midpoint of 5 (ecological typicality t(41)=10.734,
p<.001; taxonomic typicality t(41)=20.480, p<.001), suggesting that animals were relatively
good members of both taxonomic and ecological categories.
187
Despite the effort to equate mean typicality in taxonomic and ecological
categories across the animals, this was not possible: the mean ecological typicality was lower
than the mean taxonomic typicality (paired samples t(41)=4.600, p<.001, Cohen’s d=.71).
This tendency of mean typicality in animal’s dominant taxonomic category to exceed mean
typicality in animal’s dominant ecological category held across the sample of 97 animals:
mean ecological typicality 6.3, mean taxonomic typicality 7.1, paired samples t(96)=3.980,
p<.001). This is consistent with well-established finding of contextual categories having more
graded internal structure (including a range of good to moderately-good members) than
taxonomic categories with their all-or-none membership (items are either good members or
non-members) (Ross & Murphy, 1999). Thus, our selection of stimuli is constrained by the
properties of natural categories.
188
Table A1. The final set of animals and their typicality rating in corresponding ecological and taxonomic categories (ecological typicality / taxonomic typicality).
Eco / Tax category Bird Fish Insect Mammal Reptile
Desert Animal roadrunner (5.67/5.33)
scorpion (7.36/5.52)
camel (7.59/7.24)
rattlesnake (6.53/8.31)
vulture (6.72/8.43)
Forest Animal owl (6.82/7.29)
ant (6.43/8.81)
monkey (8.07/8.44)
toad (5.57/5.09)
toucan (7.12/7.65)
bee (5.28/8.08)
bear (7.20/8.12)
iguana (6.00/7.89)
butterfly (6.63/7.91)
jaguar (5.29/7.53)
deer (8.22/8.10)
fox (7.90/8.27)
rabbit (7.37/8.10)
moose (7.57/9.00)
koala (5.14/7.86)
panda (4.71/8.86)
Ocean Animal gull (5.35/7.82)
tuna (8.11/8.41)
whale (7.75/7.63)
pelican (4.73/8.09)
shark (8.29/6.46)
seal (7.59/6.76)
salmon (5.86/9.00)
Pond Animal duck (8.11/7.93)
dragonfly (5.81/7.78)
otter (5.01/7.16)
turtle (7.47/7.19)
goose (6.18/7.71)
beaver (5.93/7.61)
Savannah Animal
ostrich (6.81/7.5)
lion (7.82/8.22)
zebra (7.81/7.79)
hyena (7.29/7.41)
giraffe (7.12/7.35)
gazelle (7.12/7.76)
elephant (7.80/8.30)
kangaroo (6.57/8.00)
Note: Forest category includes both forest and rainforest animals, the ecological typicality rating for forest animals is based on the highest of the two categories.
189
Appendix C. Stimuli Used in Study 4.
Table C1. The animals and target words used in the lexical decision task
Animal Target Taxonomic Ecological Unrelated Non-word filler Monkey primate Whale mammal Bee insect Tuna fish Scorpion bug Vulture bird Otter warm Coyote hair Fox fur Duck wing Flamingo feather Iguana reptile Bear forest Camel desert Shark ocean Dragonfly pond Anaconda jungle Beaver river Owl tree Giraffe safari Ant leaf Lion savannah Turtle lake Pelican coast Zebra plug Rabbit hotel Seal chapel Parrot dime Deer apron Hyena sofa Frog canvas Ostrich paper Butterfly wallet Salmon seducer Rattlesnake gum Goose desk Eagle knurdge Grasshopper deg Piranha sount Spider hork Robin lumf Trout beeg Lizard bipe
190
Koala gimits Moose scrame Ferret dascara Dolphin tir Hedgehog flunt Woodpecker lerd Seahorse malk Mosquito vank Salamander gly Penguin blord Eel bonch Gecko strolf Jaguar wrebe Squirrel chift Walrus thake Skunk jeech Elephant wook Toucan shround Tarantula moft Crab trosh Toad nonk Bluejay gruck Snail jume Weasel roop Panda lought Llama fice Gorilla scronth Kangaroo fruick Crocodile tich
Note: The table shows a subset of the full set of stimuli, in which each of the non-filler animals was matched with 3 target words – taxonomic, ecological, and unrelated. A small number of participants were tested with other versions of the experiment covering the full stimuli set, however, since these results are not reported here, we only present the relevant subset.
191
Appendix D. Study 4: Association Norming.
The goal of this norming was to determine the strength of association between the
animal names and target words selected for lexical decision task, as well as to verify lack of
direct association between properties used in the open-ended induction task and lexical
decision target words.
Method
Participants
Eighteen native speakers of English (12 females) participated in exchange for course
credit.
Materials
Six properties (gene, cell, property, substance, flu, parasite), thirty-six animal names
and corresponding taxonomic, ecological and unrelated targets served as stimuli (see
Appendix C, Table C1). The full set of targets, including those not shown in the table but
described in the note to the Table C1 (Appendix C), were used in this norming, yielding a
total of 108 lexical decision targets).
Design
There were three versions of the experiment, varied between participants. Each
participant was presented with 108 trials rating association of 36 animals to each of the three
targets, and 108 trials rating association of six properties to a subset (60) of the same target
words. The order of questions was randomized for each participant.
Procedure
Participants were tested in the lab individually. The experiment was run using
Qualtrics survey website. On each trial, participants were presented with a word (animal name
192
or property) on top of the screen, an arrow pointing down, and one lexical decision target
below, and a 9-point rating scale with five labeled points, shown in Table D1. The task was to
rate how much the top word brought to mind the bottom word. The general instructions read:
In this experiment, you will read pairs of words. Please rate how much the first word
brings to mind the second word, using the provided scale.
The relationship between the two words is not necessarily symmetrical. For example,
"dollar" may evoke "green" more than the reverse. In this task, please rate how much
the first word brings to mind the second, not vice versa.
Throughout the experiment, some words may repeat. Please treat each rating as
independent: that is, your rating for one pair of words should not depend on the rating
you gave to another pair.
Table D1. Association norming rating scale.
1 1.5 2 2.5 3 3.5 4 4.5 5 does not bring to mind at
all
weakly brings to
mind
moderately brings to mind
strongly brings to
mind
immediately brings to mind
Results
For the analyses, the scale was converted from a 1-5 to a 1-9 scale. Animals were
rated as bringing to mind the selected ecological targets and to taxonomic targets (mean
ratings of 6.35 and 6.91, both significantly above the scale midpoint of 5: ecological
t(11)=3.31, p=.006, taxonomic t(11)=4.72, p=.002)), but not the unrelated targets (mean rating
1.28, significantly below the scale midpoint t(11)=-67.40, p<.001).
193
Animals’ association to unrelated targets was rated as significantly weaker than to
ecological (t(22)=12.46, p<.001) or taxonomic (t(22)=6.46, p<.001) targets, while the latter
did not differ from each other (t(22)=.94, p=.34),
All six properties (gene, cell, flu, parasite, property, substance) were rated as not
bringing to mind the target words, all means significantly below the scale midpoint of 5 (all
p’s<.001, range 1.75-2.33).
Some properties varied in the extent to which they brought to mind ecological,
taxonomic, and unrelated targets. Gene, cell, and parasite brought to mind ecological and
taxonomic targets more than unrelated targets (independent samples pairwise comparisons, all
p’s<.016), probably because these are biological properties, and they are likely to be
associated to other biological terms (bird, jungle) more so than to non-biological unrelated
items (sofa, wallet). For cell, association to taxonomic targets (2.86) was stronger than to
ecological (2.08, t(20)=2.25, p=.036). However, the main point is that overall, properties were
rated as relatively unrelated to target words.
Discussion
The association norming validated the selected target words for the lexical decision
task: the animals’ names did bring to mind ecological and taxonomic targets equally, and did
not bring to mind unrelated targets. Hearing the property alone did not bring the lexical
decision targets to mind.
194
Appendix E. Full text of instructions for the Study 4.
General Instructions
In this experiment, you will be presented with a number of statements about different
kinds of animals. Your task will be to come up with other animals or general kinds of
organisms that these statements could apply to.
For example, in one question, you may learn that one kind of animal has a certain
gene. In a different question, you may learn that another kind of animal has a certain parasite,
or a substance in its bloodstream, or cells, or a flu, or just some unspecified property. ALL
you know about the gene, or cells, or the parasite, or the substance in blood, or the flu, or the
property, is that the listed animal has it. You will be asked to list other animals or kinds of
organisms you think might also have it, as well as reasons for your answers.
There are no right or wrong answers; we are only interested in what you think.
Please consider each question independently. That is, the information provided for an
animal in one question should not affect your judgments about a different question.
First, you will do a practice session that will familiarize you with the types of
questions and how they will be presented, and then you will begin the experiment.
If you have any questions at any point, feel free to ask the experimenter.
Press space bar to begin the practice session.
Practice session, Step I
You will now be presented with several questions. These questions are very similar to
the ones you will be answering during the experiment.
195
In this part of the practice session, you will see one question at a time, and below
the question, there will be a space where you can type your response. After listing the
answers, remember to give a short reason for each of them, so we know what you were
thinking while responding to the question. You can move on to the next question by clicking
on the "Next" button.
Press space bar to continue.
FLU H7 is found in RACCOONS.
What else is likely to have flu H7? Why?
[6 practice trials]
Practice session, Step II
Thank you for answering the practice questions. Now you have an idea of the
questions we'll be asking during the experiment.
However, instead of reading the questions and typing up your responses, YOU WILL
HEAR THE QUESTIONS IN THE HEADPHONES, AND YOU WILL SAY YOUR
RESPONSE OUT LOUD.
Also, to save time, every question will be presented in a shortened form. For example,
you may hear:
"FLU H7 - RACCOON"
This will stand for "FLU H7 is found in RACCOONS. What else is likely to have flu
H7? Why?"
After the question, following a short pause, you will hear a beep - this is a signal for
you to begin saying your response out loud. Please start speaking right after the beep, not
196
before, and not waiting till much after. Don't worry about how you say it, we are only
interested in the content of what you say.
You will have 15 seconds to say your response. At the end of 15 seconds, you will
hear another signal, and the program will automatically proceed to another question.
If you finish your response sooner than in 15 seconds, you may press SPACE BAR
ONCE to proceed to the next question.
If you have any questions at this point, please ask the experimenter.
Press space bar to practice listening to questions and saying your responses out loud.
Practice session, Step III
Now you should have a fairly good idea of how the experiment will proceed.
However, there is one more thing. In parallel with listening to questions and saying
your responses, you will also be doing a secondary task.
Alongside with the questions, at certain times a string of letters will appear on the
screen. Your task is to decide if the string is an English word (for example, "sock") or if it's a
non-word (for example, "glym"). If it's a word, press YES. If it's a non-word, press NO.
Please try to respond by pressing the key as fast as possible after the string appears on
the screen, but without sacrificing accuracy. You will not receive feedback on the correctness
of your responses.
Now you will practice listening to questions, saying your responses out loud, and
making decisions whether a string of letters is a word or not. Doing all this at once may be
confusing at first, so try to get a hang of it during the practice.
197
Now please place your index finger on YES key, and your middle finger on NO
key. Keep looking at the middle of the screen all the time. Press a space bar to proceed to the
practice.
Instructions reminder
You will now proceed to the main part of the experiment. You will be doing two tasks
simultaneously. First, you will be listening to the questions and after the cue signal, saying
your response. While listening to the questions, please pay attention both to the property and
the animal. Second, you will be looking for strings of letters appearing on the screen, and
deciding if they are words or non-words as quickly as possible.
If at this point you are not sure about what you will be doing in this experiment, please
find the experimenter and let them know. Otherwise, proceed to the main part of the
experiment.
There will be 72 questions total. Over the course of the experiment, you will be given
4 breaks. While the screen says "You may now take a break", you can rest for as long as you
need. But between the breaks please keep going through the questions without stopping. The
estimated time for completion of this experiment is under 40 min.
Remember that if you finish your response sooner than in 15 seconds, you may press
SPACE BAR ONCE to proceed to the next question.
Please keep looking at the center of the screen at all times. Also, keep your index and
middle fingers on the YES and NO keys at all times.
Press space bar to begin the experiment.