Sentence Comprehension
- TGG (Chomsky, 1957)
- The Psychological Reality of Phrase Structure Rules
- The Derivational Theory of Complexity
- A Computational Analysis of Sentence Comprehension
- Bottom-Up versus Interactive Models
- Comprehension Strategies
- A PDP Model of Sentence Understanding
- Sentence Comprehension and the Brain
TGG (Chomsky, 1957)
- The goal of TGG was to create a grammar of symbolic rules capable of generating all the
grammatical sentences of a language without also generating any ungrammatical sentences.
- The Significance of TGG
- TGG changed the focus of linguistic research.
- TGG offered an exciting alternative to behaviorism.
- Phrase Structure Rules and Transformations
Phrase Structure Rules
- Phrase structure rules generate simple active sentences.
- S --> NP + VP
- NP --> Det + MN
- MN --> Mod + N
- VP --> V + NP
- Det --> "the" or "a"
- Mod --> "tall" or "dying" or . . .
- N --> "bay" or "woman" or . . .
- V --> "saved" or . . .
Transformations
- Transformations map phrase structure trees for simple active sentences into phrase
structure trees for other kinds of sentences (e.g., passive, question, negative, preposed
indirect object).
- Add Constituents
- Delete Constituents
- Move Constituents
The Psychological Reality of Phrase Structure Rules
- Distance Judgements (Levelt, 1970)
- Chunking (Graf & Torrey, 1966)
- Production Latency (Ammon, 1968)
- Errors in Free Recall (Johnson, 1965)
- Click Detection
Distance Judgements (Levelt, 1970)
- Participants are shown a sentence.
- They are then presented with 3 consecutive words and asked to choose the 2 that seem
most related.
- Hierarchical clustering reproduces the phrase structure.
Chunking (Graf & Torrey, 1966)
- Reading times are faster when line breaks correspond to major phrase structure
boundaries,
- Breaks at Major Boundaries
- "During World War II,
- even fantastic schemes
- received consideration
- if they gave promise
- of shortening the conflict."
- Breaks Not at Major Boundaries
- "During World War
- II, even fantastic
- schemes received
- consideration if they gave
- promise of shortening the conflict."
Production Latency (Ammon, 1968)
- Participants listen to a sentence.
- Then they hear a probe word and are asked to produce the following word in the sentence
as quickly as possible.
- hear "boy" --> say "saved"
- The further apart the words are in the phrase structure, the longer it takes to respond.
Errors in Free Recall (Johnson, 1965)
- Participants listen to several unrelated sentences then try to recall them.
- The probability of recalling word N+1 given that word N has been recalled increases as
their distance apart in the phrase structure decreases.
Click Detection
- Clicks that are embedded in a sentence tend to "migrate" toward major phrase
structure boundaries.
- "In her hope of marrying An(click)na was surely impractical."
- "Your hope of marrying An(click)na was surely impractical."
- This was originally interpreted as a perceptual phenomenon.
- "Phantom" click studies (Reber & Anderson, 1970) showed that it is the
result of a response bias.
The Derivational Theory of Complexity
- What is it?
- Early Evidence
- Counterexamples
What is it?
- According to the Derivational Theory of Complexity (Miller, 1962):
- Sentences have to be "de-transformed" during comprehension.
- Sentences are represented in memory as a simple active "kernal" plus a list of
transformations.
- "The frog ate the bug." --> (ATE FROG BUG)
- "The bug was eaten by the frog." --> (ATE FROG BUG) + PASSIVE
- "Was the bug eaten by the frog?" --> (ATE FROG BUG) + PASSIVE + QUESTION
- This suggests that as the number of transformations increases, comprehension difficulty
and memory load should increase.
Early Evidence
- Transform and Locate (Miller & McKean, 1964)
- Recognition Confusions (Clifton & Odom, 1966)
- STM Capacity (Savin & Perchonock, 1965)
Transform and Locate (Miller & McKean, 1964)
- Participants were presented with a sentence (e.g., "The frog ate the bug.")
and a sentence type (e.g., PASSIVE QUESTION).
- Their task was to point at the named variant of the sentence (e.g., "Was the bug
eaten by the frog?") on a display.
- Transformations had variable but additive effects on response time.
Recognition Confusions (Clifton & Odom, 1966)
- Participants were presented with a list of unrelated sentences.
- The DV was recognition memory.
- The probability of a false alarm increased as the transformational distance between the
target sentence and the foil decreased.
STM Capacity (Savin & Perchonock, 1965)
- Present Sentence --> Present Digits --> Recall Sentence --> Recall Digits
- As the number of transformations increases, the number of digits recalled decreases.
Counterexamples
- Subsequent Research produced so many counterexamples that the Derivational Theory of
Complexity was abandoned.
- "Fred phoned the girl up." = "Fred phoned up the girl."
- "Fred runs faster than the girl runs." > "Fred runs faster than the
girl."
A Computational Analysis of Sentence Comprehension
- What information is available?
- A sequence of words.
- Context?
- What is the goal of the computation?
- A propositional representation?
- What strategy is used to achieve the goal with the available information?
- Bottom-Up/Modular
- Interactive/Non-Modular
Bottom-Up versus Interactive Models
- The Modularity Hypothesis
- Forster’s Model versus HEARSAY
- Plausibility Effects
- How should we interpret plausibility effects?
- Forster’s Explanation
The Modularity Hypothesis
- Fodor’s (1983) Modularity hypothesis is the major source of disagreement among
researchers who study sentence comprehension.
- According to this proposal, sentence comprehension is a strictly bottom-up process
carried out by informationally encapsulated language "modules" that are not
influenced by our other cognitive abilities.
- Theories based on TGG also assumed that comprehension is a bottom-up process that is not
influenced by non-linguistic abilities.
Forster’s Model versus HEARSAY
- Forster’s Arguments Against the Interactive Model
- Too Powerful
- Lacks Parsimony
- Intractable
- As a result, we should assume that the bottom-up model is correct until it is proven
wrong.
Plausibility Effects
- Does meaning influence the time required for syntactic analysis?
- Evidence showing that it does?
- Slobin (1966)
- Steedman & Johnson-Laird (1967)
Slobin (1966)
- Slobin investigated the effects of "reversibility" on the time required to
process active and passive sentences.
- Reversible Active: "The boy chased the girl."
- Reversible Passive: "The girl was chased by the boy."
- Non-Reversible Active: "The boy ate the bagel."
- Non-Reversible Passive: "The bagel was eaten by the boy."
- Subjects judged the congruence of a sentence and a picture.
- Reversible Sentence: Active < Passive
- Non-Reversible Sentences: Active = Passive
Steedman & Johnson-Laird (1967)
- Steedman & Johnson-Laird (1967) investigated the effects of "plausibility"
on the processing of preposed indirect object sentences.
- Unconstrained: "The man took the girl the boy."
- Constrained: "The man took the girl the coat."
- Participants read a sentence, pressed a button as soon as they understood it, then
answered a question about it.
- Constrained < Unconstrained
How should we interpret plausibility effects?
- Proponents of interactive models view plausibility effects as evidence for interactive
processing.
- Forster offers 2 explanations for plausibility effects that preserve the bottom-up
modular organization of his model:
- Differential Access to the GPS
- Two Language Processing Systems
Forster’s Explanation
Comprehension Strategies
- What’s a strategy?
- How do strategies work?
- Ambiguous Sentences
What’s a strategy?
- Heuristics versus Algorithms
- Most contemporary models of sentence comprehension (especially interactive models) are
based on strategies.
- Comprehension strategies can be either semantic or syntactic.
- Examples:
- Look for the first N-V-N sequence to be agent-action-object. (syn)
- After encountering a verb, look for the number and kind of arguments appropriate to that
verb. (syn)
- Look for constituents that fit the semantic requirements of each verb. (sem)
How do strategies work?
- MacWhinney, Bates & Kliegl (1984)
- Word Order
- Agreement with the Verb
- Animacy
- Conclusions
MacWhinney, Bates & Kliegl (1984)
- MacWhinney, et al. asked native speakers of English, German and Italian to identify the
agent in some simple sentences:
- "The eraser the pig chases."
- "Licks the cow the goat."
- "The dog grabs the pencil."
- These sentences were designed to de-confound cues based on:
- Word Order
- Agreement with the Verb
- Animacy
Word Order
- Native speakers of English seem to rely primarily on word order.
Agreement with the Verb
- Italian speakers rely primarily on agreement with the verb, but are also influenced by
animacy.
Animacy
- Native speakers of German are influenced by animacy and, to a lesser extent, by
agreement with the verb.
- What about case markers?
Conclusions
- Strategies interact and people make use of whatever information they can to understand a
sentence.
- As the frequency and reliability of a strategy increase, people rely on it more.
- Just looking at English give us an incomplete picture of how language comprehension
works.
Ambiguous Sentences
- Garden Path Sentences
- "The horse raced past the barn fell."
- "The old man the boats."
- Strategies
- Minimal Attachment
- Late Closure
- What about frequency and reliability?
A PDP Model of Sentence Understanding
- Pollack and Waltz (1985) have developed an interactive activation model of sentence
understanding.
- How would you modify their model to implement the comprehension strategies described by
MacWhinney, Bates & Kliegl (1984)?
Sentence Comprehension and the Brain
- Evidence for the Modular Position
- Evidence for the Interactive Position
Evidence for the Modular Position
- Aphasia is caused by damage to specific areas of the left cerebral cortex.
- Functional imaging studies identify these same areas as involved in language
comprehension and production.
- ERP studies of language typically show stronger effects in the left cerebral cortex.
Evidence for the Interactive Position
- Neural Placticity
- Symmetry of Neural Connections
- In sentence contexts, semantic anomalies produce an N400 while syntactic anomalies
produce a P600 (Osterhout, 1994).
- "He spread the warm bread with socks."
- "The broker persuaded to sell the stock."
The End!
|