The Interaction of Lexical and Syntactic Ambiguity

1
The Interaction of Lexical and Syntactic Ambiguity

• by
• Maryellen C. MacDonald
• presented by
• Joshua Johanson

2
What do you mean by lexical and syntactic
ambiguity?

• Lexical Ambiguity
• financial bank/river bank
• Syntactic Ambiguity
• I saw the spy with the binoculars
• Lexical and Syntactic Ambiguity
• to watch/a watch

3
Approach 1 Unified Approach

• Both syntax and lexical ambiguities resolved via
  probabilistic constraints
• Very complex interaction
• All lexical and syntactic interpretations are
  activated until one is favored
• The alternative view uses garden sentences to
  argue that only one syntactic interpretation is
  activated
• I told the girl that the boy met the story
• While Mary was mending the sock fell on the floor

4
Approach 2 Different mechanism for each type of
ambiguity

• Lexical Ambiguity
• Representations are created and stored
• All senses are activated in parallel until most
  probable one is chosen
• Sytactic Ambiguity
• Parses are constructed, not stored
• There is only one serial parse
• Lexical and Semantic Ambiguity
• The Delay Model

5
Delay Model

• If there is an interaction between lexical and
  syntactic ambiguity, resolution of the ambiguity
  waits until it is no longer ambiguous
• I know that the desert trains could resupply the
  camp.
• I know that the desert trains soldiers to be tough

6
Frazier and Rayners experiment

• Measured eye fixation for reading times
• Compared ambiguous readings with unambiguous
  readings
• Used this/these to disambiguate
• This desert trains
• These desert trains

7
Results
8
Criticism

• End region is not consistent (3-10 words)
• Delay may come from another source besides the
  ambiguity
• This and These are usually used as anaphoric
  references, but there is no mention of which
  desert or which trains these refer to.
• This could be awkward and may be the cause in the
  delay, not the disambiguation

9
MacDonalds Experiment

• 53 MIT undergraduates tested
• Saw one word at a time on the screen
• Pressed space when they were done
• Asked comprehension question at the end (No
  feedback)
• Threw out tests if the comprehension question
  were wrong
• Threw out 5 participants who got more than 20 of
  the comprehension questions wrong

10
Sentences

• Modified Fraziers questions to have exactly four
  words after the ambiguous phrase
• Instead of using the modifier to dis-ambiguate,
  they used the part of speech
• the deserted trains
• the desert trained
• Kept the this/these distinction

11
Sample Question Set
12
Results from Ambiguous Region

• No difference between the desert trains and
  the deserted trains
• Using anaphoric modifiers increases reading time
  across the board
• Disambiguation increases reading time in the NV
  interpretation

13
Why does disambiguation increase reading time

• Inherent NN bias affected the reading times
• Readers might have confused desert trained,
  which is NV, with desert-trained which is an
  adjective
• The desert trains soldiers doesnt necessarily
  imply NV
• The desert trains (that) soldiers attacked were
  destroyed

14
Results from the End Region

• Obvious benefits from disambiguating the
• Still some residual this/these confusion,
  except when it disambiguates these desert trains

15
Experiment 2

• Hypothesis A strong semantic bias affects
  ambiguity resolution
• This would disprove the delay model, which
  suggests all disambiguation waits until the
  disambiguation is resolved.
• Avoids anaphoric determiners
• corporation fires (ambiguous)
• corporations fire (unambiguous)

16
Setup

• 44 (out of 46) MIT undergrads
• Same basic setup as before with keyboard and
  comprehension questions
• 16 questions, each with and without and ambiguity
  with and without supportive bias
• Uses different NV combinations to bias the
  interpretation
• corporation fires NV bias
• warehouse fires NN bias

17
(No Transcript)
18
Results
19
Observations

• Very little difference in the Supportive Bias
  times
• Still seems to be a reverse ambiguity effects
• There might be something to do with it taking
  longer to process NVs that NNs.
• Using a word that is still ambiguous but has a
  stronger bias affects the results
• This is not supported by the delay model

20
Can we predict how strong the effect will be?

• How often is the word the head of the phrase
  rather than the modifier
• How often is the word a verb rather than a noun
• (to warehouse prisoners)
• How often do the words appear together
• How plausible is the situation
• (Is a corporation more likely to have a fire or a
  warehouse?)

21
How they collected it

• Wall Street Journal corpus (yea!!!)
• Counted the number occurrences for the how
  often questions.
• Plausibility is subjective, and could be
  influenced by other factors
• They tested 96 native English speakers on
  sentence completion and counted number of times
  they completed it as a NN vs. NV
• The warehouse fires

22
Results

• Percentage Heads
• Supportive bias (corporation fires) 85.5
• Unsupportive bias (warehouse fires) 58.8
• Noun/Verb interpretation
• 6.5 verb usage (to warehouse prisoners)
• Co-occurrence
• Supportive bias .1
• Unsupportive bias 2.1(exact) and 42.(combined)
• Still very sparse, went with a Boolean exist or
  not exist
• Sentence Completion Norms
• Supportive bias 50.9 NV interpretation
• Unsupportive bias 9.6 NV interpretation

23
How do these biases affect reading times?

• Stepwise regression function
• Stepwise regression only used the head measure
  results
• r2 .19, F(1,30), p .01
• Simple regression
• As supportiveness for NV interpretation increase,
  reading time increases
• Co-occurrence had the reverse effect, since
  co-occurrence seems to promote NN interpretation

24
Do these affect unambiguous sentences

• Yes. People tend to read plausible sentences
  more quickly than less plausible sentences. This
  may indicate that plausibility might help
  disambiguate more than modifiers.
• the corporations fire was read more quickly
  than the warehouses fire.

25
Lets do another step-wise regression!

• This time we use the difference between the
  ambiguous and unambiguous readings
• The partial correlation between reading time
  differences and the percentage head measure was
  .47
• If the word is more likely to be a head, the
  reading time increased with the ambiguity.
• The partial correlation with the sentence
  completion norms was -.37
• If the word is more plausible, the reading time
  decreases with the ambiguity

26
Shouldnt it be the other way around?

• Apparently not. We know that NVs take longer to
  read than NNs. Maybe an incorrect reading of the
  word as a NN actually decreases the reading time
  more than having the correct reading of a NV.
• the warehouse fires
• the warehouses fire

27
I wonder how that affect carries out to the rest
of the sentence?
28
(No Transcript)
29
Conclusions

• Both lexical and combinatorial factors influence
  the ambiguity resolution process
• These factors accounted for a significant portion
  of variance in reading times
• More constraints promoted NV, the smaller the
  effect of ambiguity
• This data does not support the delay model
• Psycholinguistics has underestimated the
  influence of probabilistic information

30

• Do we really keep track of the
• number of times that a noun is used
• as the head of a noun phrase?
• We might need it to disambiguate between head
  nouns and noun modifiers
• Maybe it just reflects other factors
• Animacy?
• Morphological analysis

31
Questions?