Levels

[ Psy 5054 ]

Levels of Representation in Memory for Discourse

• How are these sentences similar? Different?
• Levels of Representation
• Singer’s (1990) Principles
• Surface Memory
• The Propositional Textbase
• The Situation Model
• Simultaneous Investigations of All Three Levels
• Unanswered Questions

How are these sentences similar? Different?

• "The frog ate the bug."
• "The bug was eaten by the frog."
• "The frog had the bug for breakfast."
• "The bug had the frog for breakfast."

Levels of Representation

• Van Dijk & Kintsch (1983) proposed that comprehension results in three separate levels of representation in episodic memory.

• • The most superficial and short lived captures the surface form of each sentence; a syntactically, semantically, and pragmatically interpreted string of words.
  • The meaning of a discourse is captured by a richly interconnected network of ideas called the propositional textbase.
  • The most enduring level of representation, the situation model, is similar to the representation that results from direct experience with the situation described by a discourse.

Van Dijk and Kintsch (1983)

Singer’s (1990) Principles

• Principle 1: Sentence recognition should vary with the number of representations with which a test sentence is consistent.
• Principle 2: It should be possible to identify task, understander, and discourse factors that influence some of the levels but not others.
• Empirical support for the surface form, propositional textbase, and situation model distinction is based on these principles.

Surface Memory

• Jarvella (1971)
• Surface Memory in Natural Settings
• Variables that Influence Long-Term Surface Memory

Jarvella (1971)

• Interrupted a story and tested surface memory for the last three clauses under two conditions:

• • "The confidence of Kofach was not unfounded. To stack the meeting for McDonald the union had even brought in outsiders."
  • "Kofach had been persuaded by the international to stack the meeting for McDonald. The union had even brought in outsiders."

Jarvella (1971)

• These results suggest that:

• • Surface form is held in STM until a sentence boundary is reached, then purged to make room for the next sentence.
  • Only meaning is stored in LTM.

Surface Memory in Natural Settings

• Overview
• Memory versus Reproduction
• Converging Support

Overview

• Did Jarvella and others really show that there is no long-term memory for surface form?
• Research done in natural settings during the late 1970s showed that people can discriminate between sentences they have heard (or read) and meaning preserving paraphrases.

• • Classroom Lectures (Kintsch & Bates, 1978)
  • Lunch Room Conversations (Keenan, MacWhinney & Mayhew, 1977)
  • Soap Opera’s (Bates, Masling & Kintsch, 1978)

Memory versus Reproduction

• An Example

• • "Mary told me she saw John/him at the basketball game."
  • What if you remember that this was the first (or only) time that John was mentioned?

• "In Context" Recognition Tests

• • Embed the Recognition Test in the Discourse
  • Any difference between participants with and without prior exposure to the discourse can be attributed to "real" surface memory.
  • Studies using this procedure (e.g., Bates, Kintsch, Fletcher & Giuliani, 1980) provide strong support for long-term surface memory.

Converging Support

• Re-reading times are slower when the surface form is changed (Tardif & Craik, 1989)
• Sentence verification times are slower when the surface form is changed (Anderson, 1974).

Variables that Influence Long-Term Surface Memory

• Distinctiveness (Kintsch & Bates, 1977)

• • Sentences that are different in some way (e.g., jokes and announcements) produce stronger surface memory.

• Interactive Content (Keenan, MacWhinney & Mayhew, 1977)

• • Sentences that tell us about the relationship between the speaker and listener produce more surface memory.

• Cognitive Effort (Bates, Kintsch, Fletcher & Giuliani, 1977)

• • Sentences that require more processing effort produce more surface memory.

The Propositional Textbase

• Meaning versus Form
• Propositions
• Predicates and Arguments
• The Psychological Status of Propositions
• The Organization of Propositions in the Textbase

Meaning versus Form

• Memory for meaning exceeds memory for form (Bransford & Franks, 1971)
• Test Sentences

• • "The ants ate the sweet jelly which was on the table." (3 ideas)
  • "The ants in the kitchen ate the jelly." (2 ideas)
  • "The ants in the kitchen ate the sweet jelly which was on the table." (4 ideas)
  • "The jelly was sweet." (1 idea)

• The Meaning Conveyed by these Sentences

• • (EAT AGENT:ANTS OBJECT:JELLY)
  • (LOCATION:IN ANTS KITCHEN)
  • (PROPERTY:SWEET JELLY)
  • (LOCATION:ON JELLY TABLE)

• What gets remembered?

Propositions

• Psychologists borrowed the notion of a proposition from philosophy and linguistics.
• Each of Bransford & Franks’ "ideas" is a proposition.
• Propositions are the smallest units of meaning to which we can assign a truth value (extensional definition).
• Each propositions includes a predicate and one or more arguments (intensional definition).

Predicates and Arguments

• Predicates are abstract states and events.
• Predicates are realized in the surface structure as verbs, adjectives, adverbs and prepositions.
• Each predicate has argument slots that must be filled before we can assess its truth value.
• Argument slots can be filled by concepts (nouns in the surface structure) or by other propositions.

• • "Mary saw John."
  • "Mary saw John break the window."

The Psychological Status of Propositions

• Cued Recall (Lesgold, 1972)
• All-or-None Recall (Kintsch & Glass, 1974)
• Reading Time (Kintsch & Keenan, 1973)
• Priming (Ratcliff & McKoon, 1978)

Cued Recall (Lesgold, 1972)

• Participants study a list of unrelated sentences.

• • "The large, black dog chased the frightened cat."

• DV: What is the probability of recalling a specific target word ("cat") in a cued recall task?
• IV: Recall cues can be words from the same proposition ("chased") or from a different proposition ("large").
• Participants are more likely to recall the target when the cue is from the same proposition.

All-or-None Recall (Kintsch & Glass, 1974)

• Participants studied sentences consisting of one or three propositions.

• • "The policeman issued the driver a summons."

  • • (ISSUE AGENT:POLICEMAN RECIPIENT:DRIVER OBJECT:SUMMONS)

  • "The crowded passengers squirmed uncomfortably."

  • • (CROWDED PASSENGERS)
    • (SQUIRMED PASSENGERS)
    • (MANNER (SQUIRMED PASSENGERS) UNCOMFORTABLY)

• Partial recall was more likely with the three proposition sentences.

Reading Time (Kintsch & Keenan, 1973)

• With number of words held constant, reading time increases with the number of propositions.

Priming (Ratcliff & McKoon, 1978)

• Participants are presented with sentences to study.

• • "The mausoleum that enshrined the czar overlooked the square."

• How long does it take to verify that "square" was in one of the sentences?

• • Baseline: 671 msec
  • When the preceding test word is "czar" (different proposition): 571 msec
  • When the preceeding test word is "mausoleum" (same proposition): 551 msec

The Organization of Propositions in the Textbase

• The Textbase
• The Role of Co-Occurrence
• The Role of Co-Reference

The Textbase

• The textbase is a richly interconnected set of propositions.
• Two propositions are connected in the textbase if they share an argument (or one is an argument to the other) and they co-occur in STM.

The Role of Co-Occurrence

• Models of human learning and memory suggest that connections are formed only while items are jointly active in STM.
• Related propositions prime each other only if they co-occur (McKoon & Ratcliff, 1992).
• Propositions influence one another’s free recall only if they co-occur in STM (Fletcher & Bloom, 1988).

The Role of Co-Reference

• Cued-Recall (Fletcher, Chrysler, van den Broek, Deaton & Bloom, 1992)

• • "The receptionist named the ship and the cow kicked over the stool." (34%)
  • "The receptionist named the cow and the cow kicked over the stool." (62%)

• Priming (McKoon & Ratcliff, 1980)

• • The number of referential steps separating two propositions predicts the degree of priming between them.

The Situation Model

• Why isn’t the propositional textbase enough?
• Spatial Models
• Causal Models
• Why isn’t the situation model enough?

Why isn’t the propositional textbase enough?

• Individual Differences in Comprehension

• • Two people can agree on what was said but disagree on its meaning.

• Crossmodality Integration

• • How do we combine the information from verbal directions and a map?

• Updating

• • Where is information from related texts combined?

• Learning

• • Learning from a text is not the same as learning a text.

• Perceptual and Sensory Qualities

• • Some texts give rise to perceptual and sensory properties that are not well captured by propositions.

Spatial Models

• The Central Role of Spatial Models
• Mani & Johnson-Laird (1982)
• Morrow, Greenspan & Bower (1987)

The Central Role of Spatial Models

• Spatial models have dominated research on situation models.
• The properties of spatial models are well understood.
• Space is difficult to represent with propositions.
• There is less linkage between the textbase and situation model when the situation isn’t represented with propositions.
• It is easier to separate the textbase and the situation model when the text describes a spatial layout.

Mani & Johnson-Laird (1982)

• Determinant spatial descriptions are easier to remember than indeterminant descriptions.

• • "The bookshelf is to the right of the chair. The chair is in front of the table. The bed is behind the table."
  • "The bookshelf is to the right of the chair. The chair is in front of the table. The bed is behind the chair."

Morrow, Greenspan & Bower (1987)

• Participants memorized the layout of a hypothetical building.
• They read about a character doing things in the building.
• Pairs of test words were presented and participants were asked to decide if they named objects from the same or different rooms.
• RT increased as the number of rooms separating the objects from the protagonist increased.

Causal Models

• Causal Models of Narrative Texts
• Effects of Number of Causal Connections
• Effects of Causal Chain Status
• Textbase or Situation Model?

Causal Models of Narrative Texts

• When we read a narrative we attempt to construct a causal model.

• • We feel that we understand a statement when we know what caused it and what it caused.
  • We feel that we understand the text as a whole when we can identify a continuous chain of causally connected events that connect its opening to its final outcome.

Causal Models of Narrative Texts

• (1) Danny wanted to have the red bike (2) that he saw in the window of the neighborhood bike shop. (3) Danny knew that first he had to have $50 (4) to buy the bike. (5) He asked his parents if they would give him the money. (6) His parents denied his request. (7) They suggested that Danny earn the money himself (8) by getting a job. (9) Danny was mad at his parents (10) for not giving him the money, (11) but he was determined to get the $50 somehow. (12) He knew that he would have to find a job, (13) so he called the newspaper (14) and asked for a paper route. (15) He started delivering papers in his neighborhood the next week (16) and earned $10 a week. (17) With this job, (18) Danny had $50 within a few weeks. (19) He took his hard-earned money to the shop, (20) bought the bike, (21) and happily rode home.

Causal Models of Narrative Texts

Effects of Number of Causal Connections

• Statements with many causal connections are recalled better than otherwise similar statements with fewer causal connections (Trabasso & van den Broek, 1985)
• They are also rated as more important (Trabasso & Sperry, 1985).

Effects of Causal Chair Status

• Statements with antecedents reaching to the opening of a narrative and consequences reaching to its end are recalled better then causal "dead ends" (Trabasso & van den Broek, 1985).
• They are also rated as more important (Trabasso & Sperry, 1985).

Textbase or Situation Model?

• Causal connections typically connect clusters of propositions corresponding syntactically to clauses and semantically to individual states or events.

• • "The 750 pound orange and black cat walked into the bedroom and the burglar left quickly."

• Causal connections only improve cued recall performance (relative to referential connections) when they are embedded in a text or when participants are explicitly instructed to pay attention to them (Fletcher, et al., 1992).

• • "The receptionist startled the cow and the cow kicked over the stool." (89%)

Why isn’t the situation model enough?

• Negation

• • "My car is red."
  • "My car isn’t red."

• Any others?

Simultaneous Investigations of All Three Levels

• Recognition Memory (Fletcher & Chrysler, 1990)
• The Time Course of Retrieval (Schmalhoffer & Glavanov, 1986)
• Decay Rates (Kintsch, Welsh, Schmalhoffer & Zimny, 1990)

Recognition Memory (Fletcher & Chrysler, 1990)

• The Situation Model: Carpet/Rug < Painting < Necklace < Vase < Statue
• Recognition Tests

• • Surface Test: "George says that his wife was angry when she found out that the necklace cost more than the carpet/rug." (71%)
  • Textbase Test: "George says that his wife was angry when she found out that the necklace cost more than the carpet/painting." (87%)
  • Model Test: "George says that his wife was angry when she found out that the necklace cost more than the carpet/vase." (98%)

Recognition Memory (Fletcher & Chrysler, 1990)

• These results support the distinction between the surface form, textbase and situation model iff:

• • All three distractors differ from the surface from of the original by the same amount.
  • The textbase and situation model distractors differ from the propositional representation of the original sentence by the same amount.
  • Reconstructive inferences can be ruled out.

• How can we be sure these conditions are met?

The Time Course of Retrieval (Schmalhoffer & Glavanov, 1986)

• Participants were required to respond to probe sentences every 2 seconds.
• The three levels of representation are retrieved from LTM at different rates.

Decay Rates (Kintsch, Welsh, Schmalhoffer & Zimny, 1990)

• Decay rates for the three levels of representation differ.

Unanswered Questions

• Do situation models always incorporate spatial and sensory qualities?
• Are causal connections an important feature of situation models for non-narrative discourse?
• Under what conditions are the textbase and situation model tightly linked (as with narratives) and under what conditions are they relatively independent (as with spatial descriptions)?
• Do any surface features of a discourse have a direct impact on the situation model or are all such effects mediated by the propositional textbase?

The End!

[ Psy 5054 ]

The views and opinions expressed in this page are strictly those of the page author. The contents of this page have not been reviewed or approved by the University of Minnesota

This page was last updated on 11/11/99.