Graesser, A. C., Singer, M., & Trabasso, T. (1994). Constructing inferences during narrative text comprehension. Psychological Review, 101, 371-395.
Chapters 14 through chapter 22 in Handbook of Psycholinguistics
(1994) Edited by M. A. Gernsbacher. San Diego: Academic Press.
Discourse = multiple sentences or conversational
turns (short stories, text books, conversations, etc.)
Psycholinguistic emphasis on discourse comprehension (rather
than production) and written discourse (rather than oral discourse)
Importance of Discourse:
Discourse > collection of sentences/turns
discourse comprehension involves more than comprehending each sentence
discourse production involves more
than producing each turn in isolation
Language and cognition intertwined in discourse (nonmodular)
comprehension/production requires coordination
of reasoning, memory, perceptual
processes in addition to linguistic processes
Discourse can mirror events (e.g., goals, conflicts, etc.) In our and hence manner in which represented (can be used as model for studying human cognition)
Methodology: On-line (priming, reading times)
vs. Off-line (memory, question answering)
Difficult to determine if inference occurred as material was comprehended (on-line) or later (off-line).
Major debates about how to determine if inferences are
made on-line. For example: (from Holtgraves & Ashley)
"I had the highest GPA on my floor last year"
Reaction probe: Brag
Faster Rts if probe characterizes speech act performed.
Is speech activated or does this reflect context checking?
Coherence necessary for set of sentences/turns to be considered
But what is coherence?
"I went to the doctor yesterday. Doctors make too much
money. I'd have a lot of money if I won the lottery. The government shouldn't
run lotteries. "
Sentences are related but it is not coherent.
Example has local coherence but not global coherence.
Local coherence: connections can be established
between closely adjacent sentences/propositions (e.g., argument overlap;
as in above example)
Global coherence: Construction/recognition
of a discourse theme (what the discourse is about; e.g., what is above
Debate: Is coherence in the text or in the reader's mind?
Cohesion is partly a result of text-based cohesion markers
(e.g., and, because, therefore, anaphoric and cataphoric reference, etc.).
But text-based markers alone are insufficient:
Example: At dinner last night John burnt his mouth. The
soup was hot.
Example not coherent unless reader infers that John burnt
his mouth with the soup; this is not explicitly stated.
So, inferential processing is required.
But inferential processing must be constrained; Not all
possible inferences are made when we read a discourse (e.g., that John
is less than 8 ft. tall).
Much debate regarding amount, type, and nature of on-line inferential processing.
Positions on this issue can be ordered as follows:
No inferences (text-base only); completely bottom-up (Ideal
Minimalist hypothesis (McKoon & Ratcliff); only those
inferences required for local coherence are automatically generated on-line
Constructionist theory (Graesser et al): inferences to
achieve local coherence plus inferences necessary to understand why actions,
events, etc. are mentioned in text (search after meaning)
Promiscuous inference generation (all classes of inferences
are generated) (Ideal type)
NOTE: inferences generated will depend on reader's goal,
processing capacity, etc.
Anaphoric reference - a term's (e.g., pronoun) meaning/identity is based on an earlier expression (antecedent)
Cataphoric reference - foregrounds a future expression (e.g., this dog of mine....)
Deicitic expressions - references the world outside
of the text
Anaphoric has been studied; cataphoric & deictic ignored
Anaphoric production - (note common ground interpretation)
First mention of concept use indefinite reference (e.g., a dog); it may
not be in common ground. Subsequently, can use pronominal reference (e.g.,
he) if concept is in discourse focus, or definite reference (e.g., the
dog) if concept not in focus.
Noun phrase reference
Cohesion achieved by locating an anaphoric antecedent
Bridging inferences (Haviland & Clark, 1974);
Reading times slowed if search necessary:
Herb unpacked the picnic supplies (beer)
The beer was warm
Antecedents are reactivated (Dell et al., 1983);
A burglar surveyed the garage set back from the street.
Several milk bottles were piled at the curb. The banker and her husband
were on vacation. The criminal (cat) slipped away from the streetlamp.
Participants faster at verifying burglar was in text when
tested after criminal (reactivates burglar concept) than after cat
Alternative view (Garrod and Sanford); readers build situation models such that they need not reactivate certain concepts; those concepts are currently activated.
Simon drove to the city/ Simon took his car to the city
The car kept overheating
No differences in reading second sentence; car is part
of reader's situation model and need not be activated.
Some support for both views:
O'Brien et al. (1995) - distance effect; sentence reading
times slowed as a function of antecedent distance (supports search and
reactivation) but distance effects did not occur if concept was important
to the model being described.
So, beer is not an important feature of the picnic model,
but car is an important feature of driving model (can reconcile Haviland
& Clark and Garrod & Sanford).
Sometimes more difficult to locate antecedent because no semantic similarity. So, use syntax (agree in gender and number), foregrounding (prominent concepts), general knowledge (e.g., Al helped Bob because he.......; he refers to Al; implicit causality)
Local coherence established via causal connections
- Jordan attempted a shot at the buzzer. I closed my eyes.
Note: Some causal connections are required for causal
coherence; others are elaborative and may not be required, although they're
compatible with the text.
Note: Causal connections can be backward or forward (similar
to reference); forward generally not required for coherence
Local coherence may be established by activating a previously
mentioned causal antecedent (recent - short term memory; or distant - long
term memory) or by activating relevant background knowledge (LTM - semantic
Background knowledge activation (Singer
et al., 1992)
1a Dorothy poured the bucket of water on the bonfire.
1b The fire went out.
2a Dorothy placed the bucket of water by the bonfire.
2b The fire went out.
Question: Does water extinguish fire? Yes - No
Faster for 1b than 2b; For 1, knowledge that water extinguishes
fire is activated to establish coherence (1a caused 1b), but this is not
necessary for 2.
Reinstatement (van den Broek & Thurlow, 1990)
1 Joanne was working late.
2 Joanne was working late.
It was raining hard. It was raining hard.
She left the office building. She left the office building.
The wind blew open her folder. She dropped the papers in a puddle.
The papers got all wet. The papers got all wet.
Need to reinstate from LTM that it was raining in 1 but not in 2.
Reading time for last sentence longer in 1 than 2.
Lexical decision for rain faster in 1 (reactivate) than 2.
3. Case Role Assignment
NP assigned specific case structure roles (e.g., agent, recipient, object, location, time, etc.)
I gave the money to my bookie.
Agent = I; object = money; recipient = bookie
4. Argument overlap
Propositional representations (see textbase below).
Many systems exist (e.g., Kintsch, Graesser, Kieras, etc.)
Systems developed in 1970s when psychologists move away from word lists and began examining memory for meaningful texts.
Proposition = argument (verb, adjective, adverb, connective) plus one or more arguments (nouns, embedded propositions).
Proposition analysis preserves (literal) meaning; it's the gist.
Coherence achieved via argument overlap (inferences not required)
I bet on the game. bet (I, game)
I lost my bet. lost (I, lost)
Note: in some systems, case role assignment is activated (agent = I, etc.)
Note: argument overlap is often insufficient; inferences required:
I bet on the game
I paid my bookie the money I owed.
According to Graesser et al's constructivist theory, the following inferences are generated on-line (contra the minimalist view).
In general, these inferences are a result of a comprehender's
(readers) search after meaning; an attempt to understand 'why' various
actions, events, etc. are occurring in the text.
1. Superordinate goal
Goal statements are reactivated in order to achieve coherence
Trabasso & Suh (1993) (adapted)
1.There was a girl named Betty.
2.There was a girl named Betty.
Her mother's birthday was soon. Her mother's birthday was soon.
She wanted to giver her a present. She wanted to give her a present.
Everything was too expensive. She bought her a purse.
She decided to knit a sweater. She decided to knit a sweater.
She finished the sweater. She finished the sweater.
She folded the sweater carefully. She folded the sweater carefully.
Recognition of goal statement (wanted to give her a present) when probed later in discourse was faster in 1 than 2.
2. Emotional Reactions of actors
Gernsbacher et al (1992):
Mike stole money from a store where his best friend worked. His friend was subsequently fired.
Probe: guilty < pride
Main point or moral of text
- little empirical research (theme often same as or quite similar to superordinate goal; See Trabasso study above).
Compatible with text, but not necessary for local/global coherence.
Readers sometimes infer instruments used to bring about actions. Likelihood depends on contextual constraints and necessity for achieving coherence.
The worker drove the nail.
The tool was too small for the task.
Name: hammer; speed equal to naming hammer when it is explicitly mentioned.
2. Predictive inferences (causal consequences)
- variability among readers; so, difficult to establish
- may be activated if it aids discourse and text constrains.
Klim et al (1999):
Consequence: Steven had been married for years, and his
resentment had been building up. One day, no longer able to control his
anger, he threw a delicate porcelain vase against the wall.
Motivation: Brad had not money but he just had to have
the beautiful ruby ring for his wife. Seeing no salespeople around, he
quietly made his was closer to the counter.
Probe recognition faster relative to a control.
- a more specific or concrete concept activated
O'Brien et al (1988):
All the mugger wanted was to steal the woman's money. But when she screamed, he stabbed her with his weapon/knife in an attempt to quiet her. He threw the knife into the bushes and ran away.
Gaze time for knife preceded by weapon = gaze time for knife preceded by knife.
4. Speech acts
Rarely studied; conversational processing may be different than text processing.
"I had the highest GPA on my floor"
1. Surface code = exact wording and syntax; preserved only for most recent clause (unless specific wording is informative in some way)
2. Textbase = propositions that preserve meaning, but does not include specific wording or syntax (see above for example). Sometimes textbase will include inferences made so as to achieve coherence. Some theorists (e.g., Gernsbacher) argue that textbase is bypassed; go from surface to situation
3. Situational model = microworld or model constructed via the interaction between the text and the reader's world knowledge; a mental image of what the text is about.
Less well explored:
4. Communication model = writer's purpose in constructing text; to persuade, etc.
5. Text genre; newspaper article, web article,
encyclopedia article, JEP paper, novel, etc.
Memory: Surface code decays very rapidly (held in WM only)
unless wording is significant in some sense (e.g., if reader believes text
is literature, surface memory is enhance; if wording conveys politeness
or lack thereof, surface memory enhanced)
Discrepancies can exist; e.g., between the textbase (literal
meaning) and communication model (or speaker meaning).