Thesis Proposal

1
The Role of Background Knowledge in Sentence and
Discourse Processing

• Thesis Proposal
• Raluca Budiu
• February 9, 2000

2
Metaphors

• Time is money.
• People from all cultures use metaphors on an
  every-day basis, irrespective of their level of
  education.
• Language is full of frozen metaphors (Adams
  apple, leg of a table, etc.)
• People understand (most) metaphors easily.

3
Mistakes

• People make mistakes when they speak.
• Often people do not notice mistakes and can
  understand the message communicated
• How many animals of each kind did Moses take
  on the ark?
• Its hard for people not to ignore mistakes.

4
Memory for Text

• People interpret new stories in terms of past
  experiences.
• Doing that helps them remember the new stories
  better.
• Doing than makes them deform the actual facts.

5
Motivation

• Metaphors
• Mistakes
• Memory for text

• Claim all are facets of the same cognitive
  mechanism, which
• accounts for both fallibility and robustness
• uses background knowledge as a heuristic in
  service of the current goal.

6
Thesis Topic Comprehension

• At the semantic level, comprehension works
• bottom-up all the information available is used
  to find an interpretation
• top-down the interpretation is further used to
  help comprehension or recall.
• Proof a unique computational model in ACT-R
  (Anderson Lebiere, 1998)
• explaining and unifying phenomena from various
  domains
• satisfying a number of computational and
  empirical (i.e. fitting actual behavioral data)
  constraints.

7
Behavioral Data

• Metaphor understanding
• Semantic illusions
• Text memory
• Lexical Ambiguities.

8
Overview

• Thesis topic
• A model for sentence comprehension
• Empirical constraints
• Computational constraints
• Summary and work plan.

9
Semantic Interpretation
Understanding a sentence finding a matching
interpretation/context in the background
knowledge.
10
How Does the Model Work?
Incremental From left to right omitting
How many
did
on the
Ark context
Farm context
Ark context
Farm context
Ark context
raise
father
take
Noah
verb
agent
verb
agent
Farm prop
Ark prop
place-oblique
place-oblique
patient
patient
animals
animals
ark
farm
11
Model in the Absence of Context Priming
Read word
Extract Word Meaning
yes
no
Context?
yes
Word matches context?
no
Find context
no
Context found?
no
yes
yes
Old words match?
12
Context Priming
Different processing at the beginning and at the
end of the sentence.
ark?
How many animals did Noah take on the
1. Boat or ship held to resemble that in which
Noah and his family were preserved from the Deluge
Ark story
agent
2. A repository traditionally in or against the
wall of a synagogue for the scrolls of the Torah
Noah
place-oblique
patient
verb
animals
ark(1)
took
13
Model With Context Priming
Read word
Extract Context Role
Context role matches word?
yes
no
Find context
Sentence not comprehended
Context found?
no
yes
no
yes
no
Old words match?
14
Distributed Meaning Assumption
Speak very briefly
Bible char
Navigator
meaning
meaning
word
Noah
Noah
meaning
meaning
Married
Patriarch

• Meaning retrieval extracting word features
• Replace word meaning with feature as unit of
  processing
• Model remains the same.

15
Context Finding With Distributed Meanings
Show It only if you get questions
Noah
took the animals on the ark.
word
Noah
meaning
meaning
meaning
Bible char
Married
Patriarch
Jesus context
Jesus context
Moses context
Moses context
Noah context
16
Summary of the Model

• Incremental
• Trial-and-error strategy
• Mixture of bottom-up and top-down strategies
• Incomplete processing (aka symbolic partial
  matching)
• at the word meaning level (not all features
  extracted)
• at the sentence level
• No syntactic processing thematic roles are
  inputs.

17
Overview

• Thesis Topic
• Model
• Empirical constraints
• Computational constraints
• Summary and work plan.

18
Metaphor-related Phenomena

• Effects of position on metaphor understanding
  (Gerrig Healy, 1983)
• Effects of metaphoric truth on the judgement and
  recall of sentences of the type Some As are Bs
  (Glucksberg, Glidea Bookin, 1982)
• Interferences of literal and metaphoric truth on
  sentence judgements (Keysar, 1989)
• Effects of context length on metaphor
  understanding (Ortony, Schallert, Reynolds
  Antos, 1978)
• Comprehension differences between different types
  of metaphors (Gibbs, 1990 Ortony et al. 1978
  our data).

19
Metaphor Position Effects

• Metaphor-first sentences take longer to
  comprehend than metaphor-second sentences(Gerrig
  Healy, 1983).

Drops of molten silver filled
the sky
4.21s (4.23s)

Container context
Container context
Stars context
The sky was filled with
drops of molten silver
3.53s (2.84s)

Stars context
Stars context

Predictions
20
Effects of Metaphoric Truth
hide

• Some roads are snakes gt Some flutes are jails
  (Glucksberg et al., 1982)
• snakes needs to be processed more deeply in
  order for Some roads are snakes to be judged as
  false.
• Congruent sentences lt incongruent sentences
  (Keysar, 1989)
• All features are equally informative in the
  congruent conditions.

RT
RT
21
Types of Metaphors
hide

• Literal sentences are as fast to understand as
  metaphorical sentences (Ortony et al., 1978)
• The hens clucked noisily.
• Metaphoric anaphoras are sometimes harder to
  understand than equivalent literals (Gibbs,
  1990)
• The creampuff did not show up for the box match.
• Does the literality of a metaphoric sentence make
  a difference?
• The hens/women clucked/talked noisily.

22
What Are Semantic Illusions?

• How many animals of each kind did Moses take on
  the ark?
• Semantic illusions are very robust (Reder
  Kusbit, 1991) however, not anything can make an
  illusion.
• Good vs. bad illusions
• How many animals did Adam take on the ark?

23
Semantic Illusion Datasets

• Illusion rates for good and bad distortions
  (Ayers, Reder Anderson, 1996)
• Percent correct for good and bad distortions in
  the gist task (Ayers et al., 1996)
• Latencies in the literal and gist task (Reder
  Kusbit, 1991)
• Processing of semantic anomalies and
  contradictions (Barton Sanford, 1993)
• When an aircraft crashes, where should the
  survivors be buried? vs. When a bicycle accident
  occurs where should the survivors be buried?

24
Good vs. Bad Illusions
All levels of distortion are significantly
different from one another.
25
Gist Task
Hide this
Undistorted gt Bad

• People are faster and very good at performing the
  gist task (Reder Kusbit, 1991)

26
Meaning Overlap
hide
Patriarch
Navigator
Noah
Moses
Moses
Noah
Egyptian
Married
Bible char
Adam
Adam
First man
Eve
Eden born
27
Modeling Semantic Illusions
Moses

• Model says Distorted if it finds no
  interpretation
• Key idea meaning overlap (supported by van
  Oostendorp Mul, 1990 van Oostendorp Kok,
  1990)
• Model predicts an effect of position of
  distortion in the sentence late distortions are
  harder to detect.

Noah
take
Adam
verb
agent
Ark prop
place-oblique
patient
animals
ark
28
Memory for Text

• Prior schemas can influence text memory
  (Bartlett, 1932 Bransford Johnson, 1972
  etc.)
• If a text is consistent with a pre-existent
  script (paradigmatic situation/previous
  experience)
• subjects recall more propositions from the text,
• but also make more script-consistent intrusions
• (Owens, Bower Black, 1979).

29
Text Memory Datasets

• Recall and recognition of sentences from multiple
  episodes related or not by a common setting
  (Owens et al., 1979)
• Interferences from related stories on recall and
  recognition of text (Bower, Black Turner,
  1979)
• Text recall in the presence or absence of a topic
  (Bransford Johnson, 1972)
• Recall of single, related and unrelated facts
  (Bradshaw and Anderson, 1982).

30
Interferences Among Related Stories

• The number of intrusions can increase if
  subjects study more variants of the same script
  (Bower, Black Turner, 1979)
• At the Dentists --- about Bill
• At the Doctors --- about Tom

31
Modeling Script Effects
Visiting-healthcare-professional script
Script Propositions
Studied Propositions
32
Elaborations
mihaib hide

• recall improved when subjects were shown the
  topic of a passage before studying the passage
  (Bransford Johnson, 1972)
• recall improved when subjects studied a number of
  related sentences about one historical figure,
  compared with the conditions in which they
  studied unrelated sentences about that figure or
  a single fact (Bradshaw Anderson, 1979).

33
Difficulties With Modeling Script Effects

• Parsing the discourse into a unitary and coherent
  representation (solve the problem of binding)
• Text representation that allows recursive
  schemas
• Modeling different types of intrusions,
  especially abstract intrusions

• Studied
  Intruded
• Bill paid the bill.
  Tom paid the bill.
• The nurse x-rayed Bills
  The nurse checked Toms
• teeth.
  blood pressure.

34
Lexical Ambiguity Resolution

• Although not designed for data from this domain,
  our model makes strong predictions about
  ambiguity resolution.
• Does context influence meaning access for an
  ambiguous word?
• Possible answer both meanings are activated, but
  activation depends additively on both context
  and individual meaning frequency (Tabossi, 1988
  Duffy, Morris Rayner, 1988 Rayner Duffy,
  1986 Rayner Frazier, 1989 Lucas, 1999).

35
Lexical Ambiguity Datasets

• Gaze duration on balanced and unbalanced
  homophones (Duffy et al., 1988)
• Mean reading time per character in the
  disambiguation region (Duffy et al., 1988)

36
Lexical Ambiguity An Eye Movement Study (Duffy
et al., 1988)
Mention controls hide
Disambiguation-before
Disambiguation-after

• Because it was kept on the back of a high shelf,
  the pitcher (whiskey) was often forgotten.

Of course the pitcher (whiskey) was often
forgotten because it was kept on the back of a
high shelf.
Balanced
When she finally served it to her guests, the
port (soup) was a great success.
Last night the port (soup) was a great success,
when she finally served it to her guests.
Unbalanced
Context always supports subordinate meaning for
unbalanced words.
37
Gaze Durations on Homophones

• Duffy et al. (1988) manipulated position of
  disambiguating region and relative frequency of
  the homophones meanings
• Disambiguating region before/after the homophone
• Homophone could be balanced (pitcher) or
  unbalanced (port)

38
Gaze Duration on Homophones

• Times longer than controls reflect multiple
  access.
• Times equal with controls reflect selective
  access.

39
Time Spent on Disambiguating Region
mihaib hide
40
Fitting the Data

• Disambiguation-after
• no context priming
• individual meaning activation is proportional
  with meaning frequency (ACT-R assumption)
• ACT-R is serial (no multiple access), but close
  competitors can slow down retrieval (tentative
  ACT-R assumption).
• Disambiguation-before
• context priming context is an extra source of
  activation
• If the wrong meaning is more frequent, context
  priming may not be enough.

41
Overview

• Thesis Topic
• Model
• Empirical constraints
• Computational constraints
• Summary and work plan.

42
Computational Constraints

• Realistic reaction times
• Integration with background knowledge
• Allowing for errors of the syntactic processor
  (i.e. wrong thematic roles).

43
Syntactic Ambiguity As a Computational Constraint

• Garden path effects have been largely
  documented in the literature
• The horse raced past the barn fell
• The cop arrested by the detective was guilty of
  taking bribes.

Solution thematic roles as meaning features
later omitted.
44
Summary

• Language comprehension theory to be embodied in
  a unique ACT-R model
• Semantic rather than syntactic level of
  processing (no parser)
• The theory should satisfy
• Computational constraints
• Realistic reaction times
• Integration with background knowledge
• Syntactic ambiguity.
• Empirical constraints
• Metaphor understanding
• Semantic illusions
• Lexical ambiguity
• Memory for text script effects and elaborations.

45
Empirical Constraints

• Metaphor understanding
• Effects of position on metaphor understanding
  (Gerrig Healy, 1983)
• Effects of metaphoric truth on the judgement and
  recall of sentences of the type Some As are Bs
  (Glucksberg et al., 1982)
• Interferences of literal and metaphoric truth on
  sentence judgements (Keysar, 1989)
• Effects of context length on metaphor
  understanding (Ortony et al., 1978)
• Comprehension differences between different types
  of metaphors (Gibbs, 1990 Ortony et al. 1979
  our data).

46
Empirical Constraints (contd.)

• Semantic illusions
• Illusion rates for good and bad distortions in
  the literal and gist tasks (Ayers et al., 1996)
• Latencies in the literal and gist task (Reder
  Kusbit, 1991)
• Processing of semantic anomalies and
  contradictions (Barton Sanford, 1993).
• Lexical ambiguity
• Gaze duration on balanced and unbalanced
  homophones (Duffy et al., 1988)
• Mean reading time per character in the
  disambiguation region (Duffy et al., 1988)

47
Empirical Constraints (contd.)

• Memory for text (script effects and
  elaborations)
• Recall and recognition of sentences from multiple
  episodes related or not by a common setting
  (Owens et al., 1979)
• Interferences from related stories on recall and
  recognition of text (Bower et al., 1979)
• Text recall in the presence or absence of a topic
  (Bransford Johnson, 1972)
• Recall of single, related and unrelated facts
  (Bradshaw and Anderson, 1982).

48
Model Validation

• Collect new empirical data to validate side
  effects or other predictions of the model, not
  covered by the previous list of empirical
  phenomena
• E.g. position effects for Moses illusion.
• Test it on other sets of data (for the same
  phenomena) than the ones it has been built for
  in order to avoid overfitting.

49
Work Plan
Garden path
Lexical ambiguity
Text memory
Semantic illusions
Metaphor
20
10
15
30
25

• Modeling and parameter fitting
• Data collection metaphors and semantic
  illusions
• The model still has to solve the more difficult
  problems of discourse representation.