Knowledge of Grammar, Knowledge of Usage

1
Knowledge of Grammar, Knowledge of Usage

• Syntactic Probabilities Affect Pronunciation
  Variation (Gahl and Garnsey)

2
Topics

• The Dialogue
• Experimental design
• Results and Discussion

3
The Dialogue

• Papers arguing that grammar is probabilistic
• Bybee 2001, Jurafsky 2001, 2003a, Henry 2002
• Example claim the way language is used affects
  the way it is represented cognitively
• This is an obvious critique of UG
• But - Newmeyer 2003 Grammar is grammar and usage
  is usage

4
Newmeyer 2003 abstract

• A number of disparate approaches to language,
  ranging from cognitive linguistics to stochastic
  implementations of optimality theory, have
  challenged the classical distinction between
  knowledge of language use of language.
  Supporters of such approaches point to...
  language users' sensitivity to the frequency of
  occurrence of grammatical elements.... In this
  article I defend the classical position provide
  evidence.... The article also questions the
  relevance of most corpus-based frequency
  probability studies to models of individual
  grammatical competence.

5
Some particular criticisms

• Probabilities are a function of meaning,not
  syntax i.e. one interpretation is much more
  likely in the world, but this shows nothing about
  Grammar
• Much previous research could be seen as looking
  at probabilities of word adjacency, but grammar
  is about syntactic categories (verb, noun).
• The current paper is Gahl and Garnseys response

6
GG Lit Review

• This is the literature review, not their own
  research!
• Frequent words tend to be shorter. High-frequency
  multi-word expressions tend to shorten.
• Repeated words and words presented earlier in
  discourse tend to shorten. GG note this makes
  no reference to syntax.

7
Lit Review 2

• Reduction can be overall shortening or segment
  deletion.
• Previous work shows that lexical frequency is a
  significant predictor of /t,d/ segment deletion
  (just, fast, went)
• Certain word-to-word associations cause
  shortening (Supreme Court)

8
Lit review 3

• Bybee 2002 showed that final segments of words
  that tend to appear before consonants, where
  deletion or lenition is likely, are more prone to
  delete word-finally. This is a phonological
  category frequency effect.
• GG so maybe other levels of abstraction, i.e.
  syntax, would be similar?
• Active verbs more prone to segment deletion than
  passive verbs

9
Criteria for Experiment Design

• Probabilities must be based on phrase types
  rather than specific lexical items (sexy and
  Hunter while highly frequent together are not
  phrase types)
• They should make reference to explicitly
  syntactic relationships (that a generativist
  would accept like subcategorization and not
  simple adjacency)
• Not based on plausibility or meaning

10
Verb Bias does the Trick

• Verb bias is the probability that a verb will
  appear in a particular syntactic structure
• Some verbs biased towards appearing with a direct
  object. We confirmed the date of our visit.
• Some verbs biased towards appearing with
  sentential complements We believed Hunter makes
  the best presentations always.
• Bias is determined based on corpus and norming
  studies.

11
Bias-conforming and Bias-Violating

• Sentences in which a DO-Biased Verb actually has
  a DO are bias-conforming
• Sentence in which a DO-Biased Verb occurs with an
  SC are bias-violating.
• Previous studies show that bias-conforming
  sentences are processed more rapidly and with
  greater accuracy than bias-violating ones.

12
Verb-bias meets all 3 criteria

• Verb bias applies not to single words but to all
  verb complements of a given type (DO or SC)
• Verb bias is about a syntactic relationship the
  sort of complement taken not simple adjacency
  or co-occurrence
• The difference is not based on meaning. SC and
  DO complements can be equally plausible, and yet
  the verb is still DO-biased.

13
Verb Bias and Phonetics

• GG will use verb bias and phonetic modifications
  to study probability in syntactic representations
• Focus on the verb itself and the NP following the
  verb (and pauses). The NP following the verb
  will be a Subject if the verb complement is a
  clause and a DO if it is not a clause.

14
Hypothesis 1

• Verbs in bias-matching contexts will be more
  likely to undergo /t,d/ deletion than verbs in
  bias-violating contexts
• So a DO-biased verb will delete in a DO context,
  and an SC-biased verb will delete in an SC
  context
• Motivated by previous studies showing high
  lexical frequency and high word-to-word
  probability cause deletion

15
Hypothesis 2

• Bias-violating prosodic boundaries will affect
  the duration of words and pauses greater than
  bias-matching boundaries
• DO-verbs will be longer in SC contexts than when
  in DO contexts
• SC-verbs will be longer in DO contexts than in SC
  contexts.
• Clauses have prosodic boundaries this will be
  compensated for in bias-matching contexts by the
  reductions, but there will be no such
  compensation in bias-violating ones, making them
  longer

16
Table 1 Examples

• The CIA director confirmed (Short) the rumor once
  it had spread widely.
• The CIA director confirmed the rumor (Long)
  should have been stopped sooner.
• The job applicant believed the interviewer (Long)
  when she discussed things with her.
• The job applicant believed (Short) the
  interviewer had been dishonest with her.

17
Methods

• 10 verbs of each bias type
• Divided up into groups and randomized
• 20 students (10 M, 10 F)
• Each verb appeared once in bias-conforming
  context and once in bias-violating context
• Participant were asked to read the sentences
  first and then when comfortable with them, read
  them aloud.
• Controls for equal plausibility from earlier
  norming and corpus studies.
• Used Praat to measure durations.

18
Results

• Everything they thought was right and everything
  those silly UG people thought was wrong

19
The End

• The End
• Thank you

20
Results 2

• /t,d/ deletion rates were higher in
  bias-conforming contexts than in bias-violating
  ones. (Figure 1)
• The rates did not deviate based on verb-type (SC
  vs DO verbs) but on whether or not the bias
  matched the context. This is critical.

21
Logistic Regression

• Logistic regression is a statistical model that
  relates one or more predictors and looks for
  significance.
• Eight possible predictors (GG 760-1), some were
  significant, some not.
• Importantly, the match between verb bias and
  syntactic structure was one of them.
• They are not arguing that Bias-matching is the
  only predictor of all phonetic reduction, just
  one.

22
Strength of Bias

• This is big the strength of the bias predicted
  the likelihood of deletion.
• So a strongly biased DO-verb was much more likely
  to cause deletion than a weakly biased SC-verb.

23
Further results Hypothesis 2

• Verbs, post verbal pauses, and following NPs all
  followed hypothesis 2s predictions. Duration
  was longer in bias-violating conditions.
• The duration of everything after the ambiguous
  NP, however, was not predicted by bias-violating.
• Repetition had no effect on durations.

24
Discussion Articulatory Practice

• Could the shortening be due to articulatory
  practice instead of probabilistic grammar?
• Unlikely 1 some of the sentences were high
  probability and yet might not have ever been
  encountered before. Remember they are testing
  grammatical-probability, not lexical.
• Unlikely 2 Their repetition data showed no
  repetition effects on the verbs.

25
Discussion - Retrieval

• Could the lengthening be due to difficulties in
  retrieval?
• Unlikely 1 Retrieval lengthening is typically
  associated with overall speaking rate. These
  results showed no duration changes after the
  ambiguous NP. Duration modifications were
  specific.
• Unlikely 2 The participants were asked to read
  the sentences first before recording, so all
  lexical items were already primed.

26
Discussion Speaker Control

• Speaker-controlled variation Speakers choose to
  exert more or less articulatory effort taking
  into account factors that might affect
  intelligibility.
• So highly-predictable forms are produced with
  less articulatory effort than low-probability
  forms.
• GG think this matches their data

27
Conclusion 1

• This experiment overcame previous objections
  about meaning and syntactic classes and provided
  evidence that grammar is probabilistic.
  (Remember that the strength of bias correlated
  with the frequency of deletion. It is
  stochastic.)

28
Conclusion 2

• What is the representation of these probabilistic
  form variations? Their experiment does not show.
• They argue that the simplest theory is to have
  the grammar itself be probabilistic.
• Results are consistent with the idea that
  grammatical competence is dependent upon language
  exposure i.e. not genetics.