Upload
elvin-barnett
View
212
Download
0
Embed Size (px)
Citation preview
Workshop: Corpus (1)Workshop: Corpus (1)What might a corpus of spoken data tell What might a corpus of spoken data tell
us about language?us about language?
OLINCO 2014
Olomouc, Czech Republic, June 7
Sean WallisSurvey of English Usage
University College London
OutlineOutline
• What can a corpus tell us?
• The 3A cycle
• What can a parsed corpus tell us?
• ICE-GB and DCPSE
• Diachronic changes– Modal shall/will over time
• Intra-structural priming– NP premodification
• The value of interaction evidence
What can a corpus tell us?What can a corpus tell us?
• Three kinds of evidence may be obtained from a corpus Frequency (distribution) evidence of a particular
known linguistic event Coverage (discovery) evidence of new events Interaction evidence of the relationship between
events
• But if these ‘events’ are lexical, this evidence can only really tell us about lexis– So corpus linguistics has always involved
annotation
The 3A cycleThe 3A cycle
• Plain text corpora– evidence of lexical phenomena
Text
The 3A cycleThe 3A cycle
• Plain text corpora– evidence of lexical phenomena
• Need to annotate– add knowledge of frameworks– classify and relate phenomena– general annotation scheme
• not focused on particular research goalsAnnotation
Corpus
Text
The 3A cycleThe 3A cycle
• Plain text corpora– evidence of lexical phenomena
• Need to annotate– add knowledge of frameworks– classify and relate phenomena– general annotation scheme
• not focused on particular research goals
• Corpus research = the ‘3A’ cycle– Annotation
Annotation
Corpus
Text
The 3A cycleThe 3A cycle
• Plain text corpora– evidence of lexical phenomena
• Need to annotate– add knowledge of frameworks– classify and relate phenomena– general annotation scheme
• not focused on particular research goals
• Corpus research = the ‘3A’ cycle– Annotation Abstraction
Annotation
Abstraction
Corpus
Text
Dataset
data transformation (“operationalisation”)
The 3A cycleThe 3A cycle
• Plain text corpora– evidence of lexical phenomena
• Need to annotate– add knowledge of frameworks– classify and relate phenomena– general annotation scheme
• not focused on particular research goals
• Corpus research = the ‘3A’ cycle– Annotation Abstraction
Analysis
Annotation
Abstraction
Analysis
Corpus
Text
Dataset
Hypotheses
data transformation (“operationalisation”)
Annotation Annotation Abstraction Abstraction
• Abstraction– selects data from annotated corpus– maps it to a regular dataset for statistical
analysis– bi-directional (“concretisation”)
• allows us to interpret statistically significant results
Annotation Annotation Abstraction Abstraction
• Abstraction– selects data from annotated corpus– maps it to a regular dataset for statistical
analysis– bi-directional (“concretisation”)
• allows us to interpret statistically significant results
• Even ‘lexical’ questions need annotation:– 1st person declarative modal verb shall/willabstraction relies on annotation
What can a What can a parsedparsed corpus tell corpus tell us?us?• Three kinds of evidence may be obtained
from a parsed corpus Frequency evidence of a particular known rule,
structure or linguistic event Coverage evidence of new rules, etc. Interaction evidence of the relationship
between rules, structures and events
• BUT evidence is necessarily framed within a particular grammatical scheme– So… (an obvious question) how might we
evaluate this grammar?
What can a parsed corpus tell What can a parsed corpus tell us?us?• Parsed corpora contain (lots of) trees
– Use Fuzzy Tree Fragment queries to get data
– An FTF
What can a parsed corpus tell What can a parsed corpus tell us?us?• Parsed corpora contain (lots of) trees
– Use Fuzzy Tree Fragment queries to get data
– An FTF
– A matchingcase in a tree
– UsingICECUP(Nelson et al, 2002)
What can a parsed corpus tell What can a parsed corpus tell us?us?• Trees as handle on data
– make useful distinctions– retrieve cases reliably– not necessary to “agree” to framework used
• provided distinctions are meaningful
What can a parsed corpus tell What can a parsed corpus tell us?us?• Trees as handle on data
– make useful distinctions– retrieve cases reliably– not necessary to “agree” to framework used
• provided distinctions are meaningful
• Trees as trace of language production process– interaction between decisions leave a probabilistic
effect on overall performance• not simple to distinguish between source
– depends on the framework • but may also validate it
Why spoken corpora?Why spoken corpora?
• Speech predates writing– historically – literacy growth and spread– child development – internal speech
during writing
Why spoken corpora?Why spoken corpora?
• Speech predates writing– historically – literacy growth and spread– child development – internal speech
during writing
• Scale– professional authors recommend 1,000
words/day– 1 hour of speech 8,000 words (DCPSE)
Why spoken corpora?Why spoken corpora?
• Speech predates writing– historically – literacy growth and spread– child development – internal speech during
writing
• Scale– professional authors recommend 1,000 words/day– 1 hour of speech 8,000 words (DCPSE)
• Spontaneity– production process lost: many written sources
edited
Why spoken corpora?Why spoken corpora?
• Speech predates writing– historically – literacy growth and spread– child development – internal speech during writing
• Scale– professional authors recommend 1,000 words/day– 1 hour of speech 8,000 words (DCPSE)
• Spontaneity– production process lost: many written sources edited
• Dialogue– interaction between speakers
ICE-GB and DCPSEICE-GB and DCPSE
• British Component of the International Corpus of English (1990-92)– 1 million words (nominal)– 60% spoken, 40% written– speech component is orthographically transcribed – fully parsed
• marked up, POS-tagged, parsed, hand-corrected
• Diachronic Corpus of Present-day Spoken English– 800,000 words (nominal)– orthographically transcribed and fully parsed– created from subsamples of LLC and ICE-GB
• Matching numbers of texts in text categories• Not sampled over equal duration
– LLC (1958-1977) – ICE-GB (1990-1992)
0.0
0.2
0.4
0.6
0.8
1.0
1955 1960 1965 1970 1975 1980 1985 1990 1995
p(shall | {shall, will})
Modal Modal shallshall vs. vs. willwill over time over time
• Plotting modal shall/will over time (DCPSE)
• Small amounts of data / year
Modal Modal shallshall vs. vs. willwill over time over time
• Plotting modal shall/will over time (DCPSE)
0.0
0.2
0.4
0.6
0.8
1.0
1955 1960 1965 1970 1975 1980 1985 1990 1995
p(shall | {shall, will})• Small amounts
of data / year
• Confidence intervals identify the degree of certainty in our results
Modal Modal shallshall vs. vs. willwill over time over time
• Plotting modal shall/will over time (DCPSE)
0.0
0.2
0.4
0.6
0.8
1.0
1955 1960 1965 1970 1975 1980 1985 1990 1995
p(shall | {shall, will})
• Small amounts of data / year
• Confidence intervals identify the degree of certainty in our results
• Highly skewed p in some cases
– p = 0 or 1 (circled)
Modal Modal shallshall vs. vs. willwill over time over time
• Plotting modal shall/will over time (DCPSE)
0.0
0.2
0.4
0.6
0.8
1.0
1955 1960 1965 1970 1975 1980 1985 1990 1995
p(shall | {shall, will})
• Small amounts of data / year
• Confidence intervals identify the degree of certainty in our results
• We can now estimate an approximate downwards curve
(Aarts et al., 2013)
Intra-structural primingIntra-structural priming
• Priming effects within a structure – Study repeating an additive step in
structures
• Consider– a phrase or clause that may (in principle)
be extended ad infinitum• e.g. an NP with a noun head
N
Intra-structural primingIntra-structural priming
• Priming effects within a structure – Study repeating an additive step in
structures
• Consider– a phrase or clause that may (in principle)
be extended ad infinitum• e.g. an NP with a noun head
– a single additive step applied to this structure
• e.g. add an attributive AJP before the head
N
AJP
Intra-structural primingIntra-structural priming
• Priming effects within a structure – Study repeating an additive step in structures
• Consider– a phrase or clause that may (in principle) be
extended ad infinitum• e.g. an NP with a noun head
– a single additive step applied to this structure• e.g. add an attributive AJP before the head
– Q. What is the effect of repeatedly applying this operation to the structure?
shipN
N
AJP
Intra-structural primingIntra-structural priming
• Priming effects within a structure – Study repeating an additive step in structures
• Consider– a phrase or clause that may (in principle) be
extended ad infinitum• e.g. an NP with a noun head
– a single additive step applied to this structure• e.g. add an attributive AJP before the head
– Q. What is the effect of repeatedly applying this operation to the structure?
shipNAJP
tall
N
AJP
Intra-structural primingIntra-structural priming
• Priming effects within a structure – Study repeating an additive step in structures
• Consider– a phrase or clause that may (in principle) be
extended ad infinitum• e.g. an NP with a noun head
– a single additive step applied to this structure• e.g. add an attributive AJP before the head
– Q. What is the effect of repeatedly applying this operation to the structure?
shipNAJP
very greentallAJP
N
AJP
Intra-structural primingIntra-structural priming
• Priming effects within a structure – Study repeating an additive step in structures
• Consider– a phrase or clause that may (in principle) be
extended ad infinitum• e.g. an NP with a noun head
– a single additive step applied to this structure• e.g. add an attributive AJP before the head
– Q. What is the effect of repeatedly applying this operation to the structure?
shipNAJP
very greentallAJP
N
AJP
AJP
old
NP premodificationNP premodification
• Sequential probability analysis– calculate probability of adding each AJP– error bars: Wilson intervals– probability falls
• second < first• third < second
– decisions interact
– Every AJP addedmakes it harderto add another
0.00
0.05
0.10
0.15
0.20
0 1 2 3 4 5
probability
NP premodification: NP premodification: explanations?explanations?• Feedback loop: for each successive AJP,
it is more difficult to add a further AJP
• Possible explanations include: logical and semantic constraints
• tend to say the tall green ship • do not tend to say tall short ship or green tall ship
communicative economy• once speaker said tall green ship, tends to only say ship
memory/processing constraints• unlikely: this is a small structure, as are AJPs
NP premod’n: speech vs. NP premod’n: speech vs. writingwriting• Spoken vs. written subcorpora
– Same overall pattern– Spoken data tends to have fewer attributive AJPs
• Support for communicative economy or memory/processing hypotheses?
– Significance tests• Paired 2x1 Wilson tests
(Wallis 2011)• first and second
observed spoken probabilities are significantly smallerthan written
0.00
0.05
0.10
0.15
0.20
0.25
0 1 2 3 4 5
probability
written
spoken
Potential sources of Potential sources of interactioninteraction• shared context
– topic or ‘content words’ (Noriega)
• idiomatic conventions– semantic ordering of attributive adjectives (tall green ship)
• logical-semantic constraints– exclusion of incompatible adjectives (?tall short ship)
• communicative constraints– brevity on repetition (just say ship next time)
• psycholinguistic processing constraints– attention and memory of speakers
What use is interaction What use is interaction evidence?evidence?• Corpus linguistics
– Optimising existing grammar• e.g. co-ordination, compound nouns
• Theoretical linguistics– Comparing different grammars, same language– Comparing different languages or periods
• Psycholinguistics– Search for evidence of language production
constraints in spontaneous speech corpora• speech and language therapy• language acquisition and development
What can a parsed corpus tell What can a parsed corpus tell us?us?• Trees as handle on data
– make useful distinctions– retrieve cases reliably– not necessary to “agree” to framework used
• provided distinctions are meaningful
• Trees as trace of language production process– interaction between decisions leave a probabilistic
effect on overall performance• not simple to distinguish between source
– results enabled by the framework • but may also validate it
The importance of annotationThe importance of annotation
• Key element of a ‘3A cycle’– Annotation Abstraction Analysis
• Richer annotation – more effective abstraction– deeper research questions?
• Multiple layers of annotation– new research questions– studying interaction between layers
• Algorithmic vs. human annotation
More informationMore information
• ReferencesAarts, B. Close, J. and Wallis, S.A. (2013) Choices over time:
methodological issues in current change. In Aarts, Close, Leech and Wallis (eds)The Verb Phrase in English. Cambridge University Press.
Nelson, G., Wallis, S.A. and Aarts, B. (2002) Exploring Natural Language. Amsterdam: John Benjamins.
Wallis, S.A. (2011) Comparing χ2 tests for separability. London: Survey of English Usage.
• Useful links– Survey of English Usage
• www.ucl.ac.uk/english-usage– Fuzzy Tree Fragments
• www.ucl.ac.uk/english-usage/resources/ftfs– Statistics and methodology research blog
• http://corplingstats.wordpress.com