1.3. LANGUAGE GAMES 15

1.3 LANGUAGE GAMES

The traditional unit of analysis chosen by linguists is the isolated sentence, often inventedby the researcher for illustrating a particular linguistic phenomenon. This unfortunatelyleaves out some essential ingredients of verbal communication: the context, the speaker andthe hearer, the communicative intention, and the linguistic community of which speakerand hearer are a part. Given that Fluid Construction Grammar wants to tackle languagein actual situated usage by embodied autonomous individuals, another unit of analysis isdesirable. We will use the notion of a language game. This section introduces languagegames, provides an initial typology for classifying language games and surveys methods thathave been used to study language games.

1.3.1 What is a language game?

There are many things we do with language. But they all boil to the same thing: we saythings to get someone else to do something for us. For example, we are sitting at a table, Iask you to give me some coffee and you poor it into my cup. Or, we are watching a footballmatch, I shout “Ronaldinho”, and you switch your attention to this player just in time tosee him make a goal. Or, two people A and B are building a wall. A is the builder and B theassistant. A needs blocks, pillars, slabs, and beams and B has to pass the relevant stones.A uses the words “block”, “pillar”, “slab”, “beam” and B hands over the stone being asked.

Language utterances never directly state the goals or actions that the speaker wants to seeachieved by the hearer. Instead they express conceptualisations of the world and the heareris supposed to figure out for himself how to react to them. For example, if I want coffee, Imay say “coffee” which is the name of a class of liquids, hoping that you then know thatI want some of this liquid. Or I may say “the pot”, which identifies a class of cylindricalcontainers with a handle, of which coffeepots are typical examples, hoping that you willpoor some of the contents of the specific pot on the table into my cup. Finding the rightconceptualisation to achieve a particular goal or interpreting a conceptualisation to figureout the intended goal of a speaker forms an integral part of language communication andis maybe even more complex and mysterious than formulating or parsing the utterancesinvolved.

Following in the footsteps of Wittgenstein, we call a grounded, situated, verbal interaction alanguage game. A language game involves at least two individuals that are embodied in theworld and are drawn from a population which shares enough background and context to makecommunication relevant and feasable. A language game may involve just one interaction, ora sequence of turn-taking interactions between speaker and hearer.

Even the most basic language game involves the following ingredients:

1. At least two individuals: a speaker and a hearer.

15

1. LANGUAGE GAMES 16

2. A shared situation and cooperative action in the world: We, sitting around the tablehaving breakfast.

3. A goal that requires the hearer’s cooperation: getting coffee.

4. An utterance: “Could you give me the coffee please?”.

5. Some non-verbal communicative gestures: nodding the head, pointing.

6. An action that satisfies the intention of the speaker: pooring the coffee.

A communicative goal is about achieving something in the world whereby help from thehearer is desirable. The language game succeeds if the hearer performs the intended action.As Wittgenstein put it: “the term ’language game’ is meant to bring into prominence thefact that the ’speaking’ of language is part of an activity, or form of life.”19 When I showyou the king in a game of chess and tell you this is “the king”, you cannot know what Imean unless you know the rules of the game of chess. If you don’t, your first reaction mightbe to think of Elvis.

Language is only a small part of a cooperative exchange, and the utterance may not evenbe necessary to reach the speaker’s goal. I could just show my empty cup and if you aresmart enough, in a cooperative mood, and not devoid of social intelligence, you will give mecoffee. Very often there is not an immediate physical action. The speaker simply assumesthat the hearer stores the information suggested by the utterance, and this information maythen play a role in later actions, for example in giving the right answer to a question.

It is entirely up to the speaker and the hearer to decide whether their communication wassuccessful. Success may happen, even if the speaker made a mistake, or produced only partof a sentence or an ungrammatical sentence. And it may happen even if the utterance isonly half heard by the hearer or the meaning the hearer reconstructed from the utteranceis not exactly the same as what the speaker tried to convey. Speaker and hearer alwayssee the world from their own perspective anyway and they surely conceptualise the world indifferent ways because the concepts and categories they use are based on learning processesembedded in their own individual history of interactions with the world. So communicativesuccess needs to be defined in terms of achieving the goal intended by the speaker and notin terms of sharing exactly the same meanings.

1.3.2 Typology of language games

Language games can be characterised along a number of different dimensions. There are twodimensions that are particularly useful for our present purpose: communicative goal andsemantic indeterminacy.

Communicative Goal19 Wittgenstein, L. (1965) Philosophical Investigations. The Macmillan Company, New York. nr. 23

16

1.3. LANGUAGE GAMES 17

The first useful classification of language games is in terms of what they are trying to achieve.Here we can see three main classes: reference games, description games and action games.

1. Reference games (also called games of reference) are about getting the hearer to payattention to a particular object in the world by expressing distinctive features aboutthat object in the present context. We call this object the topic of the game. Thenotion of object is to be interpreted in the broadest possible way here. A referencegame can be about a physical object, but also about an event, or an action, or situation.Here is a typical example: The speaker says “Where is the vacuum cleaner?” and thehearer points or looks at the corner of the room where the vacuum cleaner can befound or opens a cupboard and fetches it.

2. Action games are about getting the hearer to do something in the world. For example,the speaker says “raise your left arm” and the hearer does this. The topic of an actiongame is the action that the speaker wants to see performed. Reference games can alsolead to actions, but then this happened because the speaker inferred what cooperativeactions might be relevant given the present context and attention to a particular objector situation in the present context. For example, the speaker says “coffee”, drawingattention to a particular object in the present context, and the hearer infers that heshould poor coffee in the cup.

3. Description games involve a turn-taking exchange. The initiator requests informationabout a particular situation and the respondent replies with the requested information.For example, the speaker says “When did John come home?” and the hearer replies“John came home around five.”. The topic in a description game is the informationbeing requested, in this case the time of the event. The information produced bythe initiator need not necessarily be a precise answer or of immediate relevance. Inthat case the dialog partner is assumed to store it nevertheless or use it as contextfor further interactions. Even hours later, speaker or hearer may refer back to thissituation and ask more questions about it or relate it to other new experiences.

In real discourse, these different types of games all intermix. For example, when you say“Could you give me the red ball” you are not only playing an action game, asking thehearer to give you an object, but also playing a reference game in order to get the hearerto pay attention to the red ball. There are of course still many other kinds of languagegames imaginable: jokes, prayers, legal declarations. We will mostly focus here on referencegames, description games and action games as they are the most common and exploringthem already gives us plenty of very difficult challenges.

Semantic indeterminacy

The concepts expressed by language can be organised in terms of semantic domains. Forexample, there is the domain of color categories, the domain of spatial relations, the domainof bodily actions, and so on. Each domain organises some aspect of reality, for example theexperience of hue in the case of color or the location in space with respect to other objectsin the case of spatial relations. Knowing that the meaning of a particular word falls within

17

1. LANGUAGE GAMES 18

a certain semantic domain is enormously helpful to a language learner and this has lead toa distinction between a Naming Game and a Guessing Game.

1. The semantic domain of a word or grammatical construction in the Naming Game issufficiently restricted so that the possible meaning can be inferred directly. This is forexample the case if B says “What is the color of this pen?” and C replies “Brown”.B knows in this case that the meaning associated with the word is within the colordomain and so, if he did not know the word “brown” yet, he can try to infer therelevant perceptual category from the context and the object of joint attention.

2. In the Guessing Game, the possible semantic domains are open-ended. For example, ifthe speaker points to a pen and says “brown”, the hearer cannot know whether “brown”is a color, or the name of a writing instrument, or anything else that could possibly betrue about this pen. Consequently, if the hearer does not know this particular word itis much harder to infer its meaning.

The Guessing Game brings in Quine’s Gavagai problem.20 The philosopher Quine tells thestory of the anthropologist who encounters a rabbit in the forest. The native who is with himshouts “Gavagai” and points to the rabbit. But does this word now mean rabbit, or “Thereis an animal running away”, or “This is what we are going to have for dinner”? It is notreally possible to know just from this single interaction. The issue is not one of referentialuncertainty. The anthropologist knows what object “Gavagai” refers to. It is rather semanticindeterminacy, i.e. how constrained the semantic domain is. Narrowing down the semanticdomain is not necessarily a matter of language. The turn-taking patterns and cooperativeinteractions expected in a particular language game as well as the shared context contributejust as much. If the builder shouts a word to his assistant, then the assistant knows he isasking for someting that is relevant to the building process, is already present in the currentcontext, and is not yet named by an existing word.

1.3.3 Ways of studying language games

There are various approaches for studying language games. The first approach is empirical.Real human dialogs are recorded and then analysed in order to find their structure or theconceptualisations that are used. This kind of research is primarily done by pragmaticsresearchers, psycholinguists and discourse analysts21

More recently, experimental psychologists have started to use a second, experimental ap-proach which relies on laboratory experiments in emergent communication. They give hu-man subjects a specific task and then observe the communication system that emerges whenthe subjects play language games within this context. In some experiments, humans can use

20 Quine, W. (1960) Word and Object. The MIT Press, Cambridge Ma.21 Clark, H. H. (1996). Using language. Cambridge University Press, Cambridge.

18

1.3. LANGUAGE GAMES 19

1-----B: Tell me where you are?

2-----A: Ehm : Oh God (laughs)

3-----B: (laughs)

4-----A: Right : two along from the bottom one up:

5-----B: Two along from the bottom, which side?

6-----A: The left : going from left to right in the second box.

7-----B: You're in the second box.

8-----A: One up (1 sec.) I take it we've got identical mazes?

9-----B: Yeah well : right, starting from the left, you're one along:

10----A: Uh-huh:

11----B: and one up?

12----A: Yeah, and I'm trying to get to ...

Figure 1.1: Example of a human dialog in which one partner (A) has to describe to the otherpartner (B) which square in the maze he has chosen as topic (here indicated with an arrow). Theinitial exchanges establish the communicative goal and create a feeling of complicity. Then thefirst information exchanges take place, which quickly lead to greater rather than less confusion sothat B even starts to doubt whether they are looking at the same maze. Note how none of theutterances here is a well-formed grammatical sentence and how A thrusts upon B a particular wayof conceptualising locations in terms of paths from shared initial points.

their normal natural language (see figure 1.122) and one sees the hesitations, the incompleteand ungrammatical utterances, as well as the inferential coding and alignment phenomenadiscussed earlier.

In other experiments, human subjects have to invent a communication system from scratchand these experiments therefore recreate a situation similar to the origins of human language.This is for example the case in an experiment set up by Bruno Galantucci.23. Players movearound in a computer game world consisting of a set of rooms located on a grid and markedwith figures. Players have to move to the same room, but they only have a local view andso they cannot see where the other player is located. As they need to know this in order todecide on their next move, players are encouraged to develop ways for describing their ownpositions, where they intend to move next, or what they suggest the other player should do.Galantucci introduced an unusual graphical medium by which players can communicate: ascratchpad that moves vertically as one draws on it. Thus drawing a horizontal line results ina diagonal line with a slant that reflects the velocity profile of the drawing motion. Becauseof this novel medium, players are forced to invent a new communication system. There isno prior inventory, not even a prior set of signs to build from. Remarkably, most pairs of

22 From: Pickering, M., and S. Garrod (2004) Toward a mechanistic psychology of dialogue. Behavioral andBrain Sciences, 27, 169-225.

23 Galantucci, B. (2005) An Experimental Study of the Emergence of Human Communication Systems.Cognitive Science (29) 737-767

19

1. LANGUAGE GAMES 20

players manage, although some clearly do not, and we observe the same findings as seen innatural dialogue: innovations, alignment, and variation among individuals. We also observeclearly how the emergent communication system is tightly embedded in the coordination ofthe behavioral processes between dialogue partners.

Language games have also been studied by logicians, and particularly by Jaako Hintikka.24

Some of the basic ideas of Wittgenstein are retained, in particular that language is interwovenwith activities in the world and meaning comes through these activities, but logicians thenfocus on verifying the truth of (quantified) propositions through an adversary game, whereone (imaginary) opponent brings arguments for the truth of the proposition and the otheropponent tries to disqualify or falsify them. So this is quite a different notion of a languagegame compared to the one used here.

Finally there is the modeling approach, which is the one most relevant in the present context.It proposes to build explicit models of all or part of the cognitive structures and processesthat go into a language game. This is obviously enormously challenging and so complex thatmodels can only be tested with computer simulations even for very basic and simple languagegames. Fortunately advances in computer science, robotics, and Artificial Intelligence havegiven us incredibly powerful platforms to build and test language game models. It haseven become possible to program robots that play language games in a shared environment(figure 1.2). These robots have perceptual systems to identify, recognise, and track objectsin the world, including other robots, they have motor systems to engage in cooperativeactions including joint attention, and ways to carry out all the cognitive steps needed toplay a language game. Fluid Construction Grammar has grown out of this kind of detailedmodeling efforts and has already been used in a wide variety of embodied language gameexperiments.25

1.4 THE SEMIOTIC CYCLE

It only takes a few minutes of reflection to realise that extraordinary complex cognitiveactivities are going on in a language game, both from the side of the speaker and fromthe side of the hearer. Even before language starts playing a role, there is the ongoingperception of the scene and the planning and execution of actions. There is the complicatedsubtle process of turn-taking between speaker and hearer guided by a loose, tacit scriptspecific to each language game. Joint attention needs to be established by tracking eyegaze or by pointing gestures. And a specific goal must be conceived and shared by speakerand hearer that fits with their joint objectives. The cognitive processes specifically relevantfor language can be decomposed and organised into a series of steps which we will call thesemiotic cycle.

24 Hintikka, J. (1973) Logic, Language Games, and Information. Kantian Themes in the Philosophy of Logic.Oxford University Press, Oxford

25 Steels, L. (2003a) Evolving grounded communication for robots. Trends in Cognitive Science, 7(7):308-312.

20

1.4. THE SEMIOTIC CYCLE 21

Figure 1.2: Today it is possible to construct complete models of the cognitive processes andactions that are required to play language games about objects in the world and embed thesemodels in highly complex fully autonomous humanoid robots. In the scene that is shown here, onerobot asks the other one to perform an action and the other robot carries it out.

1.4.1 Steps in the semiotic cycle

An overview of the steps in the semiotic cycle is given in figure 1.3. The semiotic cycle canstart if both partners in communication are situated in a shared world, either physically ormentally, and if they have a way to perceive and act upon this world. Telepathy or any otherform of direct mind-reading or control is excluded. The only interactions that are possiblebetween speaker and hearer go through the real world itself.

The semiotic cycle begins when the speaker has formulated a communicative goal or acollection of goals that fits within the ongoing cooperative actions that make sense in thespecific context. Next the following steps take place, first in language production (the leftside of the cycle) and then language comprehension (the right side of the cycle).

• Conceptualisation: As already explained earlier, natural languages express concep-tualisations of the world, i.e. ways to divide up the world into different objects, waysto categorise the objects and their relations, and ways to introduce a perspective onthe scene and give it emotional and social significance. So the first challenge for thespeaker is to come up with an appropriate conceptualisation that makes sense in thepresent context and is relevant for the communicative goal.

• Production: The next challenge for the speaker is to turn this conceptualisation into

21

1. LANGUAGE GAMES 22

Figure 1.3: The semiotic cycle decomposes the different cognitive activities that are requiredfor the linguistic subtasks of a language game. On the left we see the activities carried out by thespeaker and on the right the activities carried out by the hearer.

a verbal form using the lexical and grammatical inventory available in his languagesystem, a process we will call production. It has also been called formulation, ver-balisation, or grammatical encoding. Production results in an abstract description ofwhat the utterance should look like. The mapping from conceptualisation to languageis not a simple translation process. It involves choices because there are usually manyways to express the same thing, and it may even require that the speaker goes backand re-conceptualises what he wants to say because there is no straightforward way toexpress the chosen meaning.

• Rendering. Next, the constraints on the utterance have to be translated into themotor control commands that trigger the articulatory movements of the speech organs.We call this process rendering. Rendering through speech is in itself extra-ordinarycomplex and will not be studied in any detail here.

Then the hearer goes through his own steps in the semiotic cycle:

• De-Rendering: The speech signal must first be translated back into a representationof the relevant properties of the utterance. This means that individual words, intona-tion, stress patterns, word order, or other aspects of the utterance must be identifiedand categorised. Again, this is a very complicated process which works remarkablywell, considering all the noise that comes with real speech and the sloppiness of human

22

1.4. THE SEMIOTIC CYCLE 23

articulation.

• Parsing: Parsing operates over an abstract description of the form of the utteranceas provided by the de-rendering processes. Parsing is the process of reconstructing thevarious deeper structures that underly the sentence with the objective of reconstructingthe meaning, i.e. the conceptualisation that might have been intended by the speaker.Parsing involves a set of subprocesses that can to some extent be further untangled,like processes for morphology, syntax, constituent structure, etc.

• Interpretation. The reconstructed meaning must then be interpreted in terms ofthe experiences and categorisations that the hearer was able to derive from his ownperception and action in the world, and the communicative goal must be inferredfrom the language game, the shared context, and the recent history of cooperativeinteractions between speaker and hearer. We call this process interpretation.

Once an interpretation is available, the hearer can chose an action that is compatible withthe script of the language game. Execution of this action may lead to some sort of reactionfrom the speaker. He may show happiness in getting the coffee he wanted, or wait to seesome nodding by the hearer as a sign that the sentence was understood. In any case, boththe speaker and the hearer learn something from the interaction. The speaker learns whetherthe utterance had the required impact and the hearer can see whether he understood theutterance correctly and whether he made the right inferences about what had to be done asa consequence.

Speaker and hearer take turns in language games, so that each individual needs to havecompetence in both and presumably uses the same representation of linguistic knowledge inboth tasks. Often, a child or a second language learner appears to be better in comprehensionthan speaking, but that is because comprehension can ignore many details of language,such as morphology or syntax, that cannot be ignored in language production. A sentencecan often be understood even if most of the grammar is thrown out and some words areunknown, simply because the context and the language game help to delinate the set ofpossible meanings drastically.

The semiotic cycle shown in figure 1.3 seems to suggest that the different activities take placeone after another, but in reality they are of course entirely intermingled. The hearer does notwait for the speaker to start inferring what could be intended and may interrupt the hearerin midcourse, possibly helping to complete the sentence. Moreover the flow of informationis not just in the direction of the arrows in figure 1.3. There is as much information going inthe other direction, both for the speaker and the hearer. For example, the speaker may usecertain words that force the hearer to pay attention to particular visual details of the scenewhich he would otherwise ignore. Or by using a particular grammatical construction, thespeaker may force the hearer to conceptualise the scene in a particular way, for example takeinto account specific participant roles in the event. So language can actually influence thevision system and the conceptualisation system and it certainly influences what actions thehearer is performing. Conversely, the speaker may change course in mid-stream if he cannot

23

1. LANGUAGE GAMES 24

find ’the right words’ to express the original conceptualisation or if he sees puzzlement onthe face of the hearer. So in a more realistic model, all these processes are much moreintermingled and every process influences every other process. However, for reasons ofanalysis and scientific investigation, it is useful, indeed necessary, to distinguish differentactivities and investigate each one separately. It is also useful to assume a linear flow, beforestudying more tightly coupled and tangled interactions between all these processes.

A detailed analysis of a speech signal shows that there is a huge gap between the physicalsound and the abstract categories that constitute the interface to lexical and grammaticalprocesses. Speech production is entirely comparable with motor control in other domains.It is carried out by a set of articulators (the vocal tract, the tongue, the lips) which movearound extremely rapidly with very tight motor control. Speech perception is similar tosensory perception in other domains, relying on complex signal processing, segmentation,pattern recognition and structural analysis. Both speech production and speech perceptionare hugely complex subsystems in their own right and a lot of research has been carried outon modeling and operationalsing the relevant processes.

Speech is a big domain of study on its own and we will not get into it here. In order toshortcut all the issues related to real speech, we will postulate an abstract representation oflinguistic forms that can then act as the input to the articulatory system on the one handor be reconstructed by the speech recognition system on the other, even though it must beemphasised that the division of the semiotic cycle into different processes is artificial. Inreal human language all steps in the cycle have an effect on each other and also in languageevolution there may be processes at the speech level (for example the conflation of speechsounds due to the rapid movement of articulators or the difficulty of hearing certain soundsat the end of a word) that may then impact the lexicon and even the grammar, driving thegrammaticalisation process.

1.4.2 The meta-level

The semiotic cycle as discussed in the previous subsection is a reasonable model when thingsgo well. But there is clear evidence, from neurobiological observations, psychological exper-iments, and from information processing requirements, that more must be going on. Thereare many occasions when smooth processing is not possible, because routine solutions donot work or because the words and structures being detected in the utterance of the speakerare not conform to current expectations.

This suggests that we need a two-leveled system (illustrated in figure 1.4). At the firstlevel, routine processing is going on. Conceptualisation, production, and rendering for thespeaker and de-rendering, parsing, and interpretation for the hearer. The second level isa meta-level. When things start getting difficult or going wrong, this level becomes moreand more active. It is able to catch failures, execute diagnostics, and even repair failures,after which processing at the routine layer may continue, possibly with a changed lexical

24

1.4. THE SEMIOTIC CYCLE 25

and grammatical inventory.26 For example, if the speaker does not find a word in his lexiconto express certain parts of the meaning, he may retrieve a word that comes close and simplyexpand its usage by analogy, thus extending his own lexicon. Or if the hearer encountersa word that he never heard before, he may wait to gather more evidence about possiblemeanings and then add a new entry into his lexicon to capture the hypothesised meaning ofthe unknown word.

Figure 1.4: Language processing requires a two-leveled system with one level dealing with routinesolutions, even if they involve some search, and the second, meta-level, coming into action whencertain problems become manifest, which could then be diagnosed and repaired so that routineprocessing can continue.

This kind of two-level architecture is known in computer science as computational reflection.Computational reflection means that when a computational system gets into an error stateit is able to jump up to a meta-level and repair the damage before moving back to object-level processing. For example, if a LISP system encounters an undefined function, it willjump into an error state where the programmer can inspect how the system got there, seebindings of variables or examine definitions of functions, and then repair the error-state, forexample by defining the missing function, and then continue object-level computation as ifnothing happened. It is possible to envision that this debugging activity is done by anotherprogram, in which case that program needs to have access to the complete error state andmust be able to change anything, including the program code that caused the error. Moderncomputer systems, and particularly operating systems, have facilities that perform this kindof computational reflection, otherwise computers would crash without mercy all the time, asthey used to do before all this was invented. In some sophisticated computational reflectionarchitectures, there is even the possibility of further regress, in the sense that if the meta-levelitself gets into trouble, there could still be a meta-meta level that becomes active and tries todeal with problems at the first meta-level. In principle we can keep going like this, climbingthe infinite tower of meta-level and coming down with solutions, but, quite amazingly, ithas been shown also that we can close the loop and build a reflective system that is able torecurse on itself by using the same information structures and language constructs at each

26 Levelt, W. (1989) Speaking: From Intention to Articulation. The MIT Press, Cambridge Ma.

25

1. LANGUAGE GAMES 26

computational level.27

In Fluid Construction Grammar , the meta-level is considered to be as important as the’object-level’ and the knowledge and skills needed at the meta-level are as much an integralpart of the knowledge of a language as knowledge of the rules of the lexicon and grammarthemselves. Meta-level language processing must be quite sophisticated. For example, themeta-level of the speaker must include mechanisms for self-monitoring. Because speakersalways have the competence to be listeners as well, they can in principle simulate the effectof an utterance on the hearer by using themselves, i.e. their own inventory of rules and theirown world model as a model of the hearer. Consequently the speaker can detect whetherthere is too much ambiguity in what he is going to say and repair it by selecting other wordsor by adding more grammatical structure to the utterance. Conversely, a human hearer isin principle able to simulate himself how he would express the meaning that he believesthe speaker wants to convey, and he can use the outcome of such a simulation to performtop-down predictions and thus figure out the meaning of unknown words, or understandwhy certain grammatical markers are there, or even correct the utterances produced by thespeaker.

Meta-level capacities are absolutely essential to deal with actual language use, simply becausereal world utterance are full of incomplete and ungrammatical phrases, errors, words whosemeanings have been stretched almost beyond recognition, the coercion of words or phrasesinto new syntactic categories, and so on. They are also essential to explain why languagecan be an open creative system that expands itself to express meanings that have not beenexpressed before or to make sense of inventions by the speaker which have never been usedbefore by anyone.

1.5 CONCLUSIONS

This chapter has broadly set the scene for the more technical treatment that will follow.We looked at the notion of a linguistic formalism and why it is needed, at the motivationbehind the development of Fluid Construction Grammar, and at language games, proposedas more appropriate units for linguistic investigation than the isolated sentence. Then thedifferent steps in the semiotic cycle of language production and language comprehensionwere outlined. The coming chapters study progressively more complex language, startingfrom single-word sentences, and each time introduce the necessary mechanisms to handlethat level of complexity using FCG. On the way, we will pause now and then to study alsothe semantic aspects of language in more detail.

27 Research on computational reflection started with the work of Brian Smith in the early eighties and hassince flourished into a research field on its own. The historical starting point of the literature is: Smith,B. (1982) Reflection and semantics in a procedural language. Technical Report MIT-LCS-TR-272, MIT,Cambridge Ma.

26