Transcript Slides (PowerPoint)

Capturing linguistic interaction
in a grammar
A method for empirically evaluating
the grammar of a parsed corpus
Sean Wallis
Survey of English Usage
University College London
[email protected]
Capturing linguistic interaction...
• Parsed corpus linguistics
• Empirical evaluation of grammar
• Experiments
– Attributive AJPs
– Preverbal AVPs
– Embedded postmodifying clauses
• Conclusions
– Comparing grammars or corpora
– Potential applications
Parsed corpus linguistics
• Several million-word parsed corpora exist
• Each sentence analysed in the form of a tree
– different languages have been analysed
– limited amount of spontaneous speech data
• Commitment to a particular grammar required
– different schemes have been applied
– problems: computational completeness + manual
consistency
• Tools support linguistic research in corpora
Parsed corpus linguistics
• An example tree from ICE-GB (spoken)
S1A-006 #23
Parsed corpus linguistics
• 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
• This evidence is necessarily framed within a
particular grammatical scheme
– So… how might we evaluate this grammar?
Empirical evaluation of grammar
• Many theories, frameworks and grammars
– no agreed evaluation method exists
– linguistics is divided into competing camps
– status of parsed corpora ‘suspect’
• Possible method: retrievability of events




circularity: you get out what you put in
redundancy: ‘improvement’ by mere addition
atomic: based on single events, not pattern
specificity: based on particular phenomena
• New method: retrievability of event sequences
Experiment 1: attributive AJPs
• Adjectives before a noun in English
• Simple idea: plot the frequency of NPs with
at least n = 0, 1, 2, 3… attributive AJPs
Experiment 1: attributive AJPs
• Adjectives before a noun in English
• Simple idea: plot the frequency of NPs with
at least n = 0, 1, 2, 3… attributive AJPs
6. 0000
200, 000
Raw frequency
180, 000
Log frequency
5. 0000
160, 000
140, 000
4. 0000
120, 000
3. 0000
100, 000
80, 000
2. 0000
60, 000
40, 000
1. 0000
20, 000
0. 0000
0
0
1
2
3
4
5
6
0
1
2
3
4
5
6
NB: not a straight line
Experiment 1: analysis of results
• If the log-frequency line is straight
– exponential fall in frequency (constant probability)
– no interaction between decisions (cf. coin tossing)
• Sequential probability analysis
– calculate probability of adding each AJP
– error bars (binomial)
– probability falls
• second < first
• third < second
• fourth < second
– decisions interact
Experiment 1: analysis of results
• If the log-frequency line is straight
– exponential fall in frequency (constant probability)
– no interaction between decisions (cf. coin tossing)
• Sequential probability analysis
– calculate probability of adding each AJP
– error bars (binomial)
probability
– probability falls
0. 25
0. 20
• second < first
• third < second
• fourth < second
– decisions interact
0. 1 5
0. 1 0
0. 05
0. 00
0
1
2
3
4
5
Experiment 1: analysis of results
• If the log-frequency line is straight
– exponential fall in frequency (constant probability)
– no interaction between decisions (cf. coin tossing)
• Sequential probability analysis
–
–
–
–
–
calculate probability of adding each AJP
error bars (binomial)
probability
probability falls
decisions interact
fit to a power law
• y = m.x k
• find m and x
-1.2793
0. 25
0. 20
0. 1 5
0. 1 0
0. 05
y = 0.1931x
0. 00
0
1
2
3
4
5
Experiment 1: explanations?
• Feedback loop: for each successive AJP,
it is more difficult to add a further AJP
– Explanation 1: semantic constraints
• tend to say tall green ship
• do not tend to say tall short ship or green tall ship
– Explanation 2: communicative economy
• once speaker said tall green ship, tends to only say ship
– Further investigation required
• General principle:
– significant change (usually, fall) in probability is
evidence of an interaction along grammatical axis
Experiments 2,3: variations
 Restrict head: common and proper nouns
– Common nouns: similar results
– Proper nouns and adjectives are often treated as
compounds (Northern England vs. lower Loire )
 Ignore grammar: adjective + noun strings
– Some misclassifications / miscounting (‘noise’)
• she was [beautiful, people] said; tall very [green ship]
– Similar results
• slightly weaker (third < second ns at p=0.01)
– Insufficient evidence for grammar
• null hypothesis: simple lexical adjacency
Experiment 4: preverbal AVPs
• Consider adverb phrases before a verb
– Results very different
• Probability does not fall significantly between first and
second AVP
• Probability does fall
between third and
second AVP
– Possible constraints
• (weak) communicative
• not (strong) semantic
– Further investigation
needed
Experiment 4: preverbal AVPs
• Consider adverb phrases before a verb
– Results very different
• Probability does not fall significantly between first and
second AVP
• Probability does fall
between third and
second AVP
0.06
0.05
– Possible constraints
• (weak) communicative
• not (strong) semantic
– Further investigation
needed
– Not power law: R2 < 0.24
0.04
0.03
0.02
0.01
probability
0.00
1
2
3
Experiment 5: embedded clauses
• Another way to specify nouns in English
– add clause after noun to explicate it
• the ship [that was tall and green]
• the ship [in the port]
– may be embedded
• the ship [in the port [with the ancient lighthouse]]
– or successively postmodified
• the ship [in the port][with a very old mast]
• Compare successive embedding and
sequential postmodifying clauses
– Axis = embedding depth / sequence length
Experiment 5: method
• Extract examples with FTFs
– at least n levels of embedded postmodification:
Experiment 5: method
• Extract examples with FTFs
– at least n levels of embedded postmodification:
0
1
2
(etc.)
Experiment 5: method
• Extract examples with FTFs
– at least n levels of embedded postmodification:
0
1
2
(etc.)
– problems:
• multiple matching cases (use ICECUP IV to classify)
• overlapping cases (subtract extra case)
• co-ordination of clauses or NPs (use alternative patterns)
Experiment 5: analysis of results
• Probability of adding a further embedded
clause falls with each level
– second < first
– sequential < embedding
• Embedding only:
– third < first
– insufficient data for
third < second
• Conclusion:
– Interaction along embedding and sequential axes
Experiment 5: analysis of results
• Probability of adding a further embedded
clause falls with each level
– second < first
– sequential < embedding
• Embedding only:
– third < first
– insufficient data for
third < second
• Conclusion:
0.06
embedded
0.05
0.04
0.03
sequential
0.02
0.01
probability
0.00
0
1
2
3
– Interaction along embedding and sequential axes
4
Experiment 5: analysis of results
• Probability of adding a further embedded
clause falls with each level
– second < first
– sequential < embedding
• Fitting to f = m.x k
– k < 0 = fall ( f = m/x |k|)
– |k| is high = steep
• Conclusion:
0.06
embedded
y = 0.0539x -1.2206
0.05
0.04
0.03
0.02
sequential
y = 0.0523x -1.6516
0.01
0.00
0
1
– Both match power law: R2 > 0.99
2
3
4
Experiment 5: explanations?
• Lexical adjacency?
– No: 87% of 2-level cases have at least one VP, NP
or clause between upper and lower heads
• Misclassified cases of embedding?
– No: very few (5%) semantically ambiguous cases
• Language production constraints?
– Possibly, could also be communicative economy
• contrast spontaneous speech with other modes
• Positive ‘proof’ of recursive tree grammar
– Established from parsed corpus
– cf. negative ‘proof’ (NLP parsing problems)
Conclusions
• A new method for evaluating interactions
along grammatical axes
– General purpose, robust, structural
– More abstract than ‘linguistic choice’ experiments
– Depends on a concept of grammatical distance
along an axis, based on the chosen grammar
• Method has philosophical implications
– Grammar viewed as structure of linguistic choices
– Linguistics as an evaluable observational science
• Signature (trace) of language production decisions
– A unification of theoretical and corpus linguistics?
Comparing grammars or corpora
• Can we reliably retrieve known interaction
patterns with different grammars?
– Do these patterns differ across corpora?
• Benefits over individual event retrieval
 non-circular: generalisation across local syntax
 not subject to redundancy: arbitrary terms
makes trends more difficult to retrieve
 not atomic: based on patterns of interaction
 general: patterns may have multiple explanations
• Supplements retrieval of events
Potential applications
• 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
Links and further reading
• Survey of English Usage
– www.ucl.ac.uk/english-usage
• Corpora and grammar
– .../projects/ice-gb
• Full paper
– .../staff/sean/resources/analysing-grammaticalinteraction.pdf
• Sequential analysis spreadsheet (Excel)
– .../staff/sean/resources/interaction-trends.xls