Comprehension

Abstract

Comprehension is the cognitive process by which a reader or listener builds a coherent mental representation of a text or utterance, going beyond the words to grasp what they describe. This article treats comprehension as the construction of layered representations: a surface record of the exact words, a textbase of the propositions they express, and a situation model of the state of affairs the text is about. It follows the construction-integration model through which knowledge and context settle meaning, the inferences that knit sentences into a coherent whole, and the working-memory and word-level skills that make one comprehender better than another. Three demonstrations exercise the resolution of an ambiguous word by context, the updating of a situation model across a narrative, and the multiplicative simple view of reading.

Keywords: comprehension, situation model, inference, reading

To comprehend a text is to end up knowing what it is about, not merely what it says. A reader who finishes a paragraph about a camping trip can report who went, where they were, and what went wrong, often without recalling a single sentence in its original wording. That everyday achievement conceals a demanding construction: the mind must identify words, combine them into propositions, link each sentence to what came before, draw on stored knowledge to fill the gaps the writer left implicit, and assemble the result into a model of the situation the text depicts. Cognitive psychology has spent half a century specifying these layers and the processes that build them, turning comprehension from a synonym for understanding into one of the most precisely modeled achievements of the mind. The sections below build the construct from its levels of representation, work through the construction-integration model and the situation model it yields, examine the inferences comprehension requires, and turn to the memory and word-level skills that explain individual differences.

What Comprehension Is

Comprehension is the process of constructing a mental representation of the meaning of a message from language, written or spoken, such that the comprehender can use, question, and remember what the message conveys rather than merely its form. Its defining feature is that understanding is not a copy of the input but a construction: the reader supplies knowledge, resolves ambiguity, and infers connections the text leaves unstated, so the representation built is richer and more integrated than the words on the page. A classic demonstration of the shift from form to meaning comes from Sachs, who had people listen to a passage and then tested their memory for individual sentences. Recognition of the exact wording and syntax decayed within seconds of continued reading, while memory for the meaning of a sentence remained accurate, showing that comprehenders rapidly discard the surface form once its meaning has been extracted (Sachs, 1967).

This observation motivates the central idea of modern comprehension research, that a text is represented at several levels at once. Kintsch and van Dijk proposed that comprehension yields a hierarchy rather than a single representation: a surface memory of the exact words, a textbase of the propositions the text asserts together with their connections, and a higher-level structure that integrates those propositions with the reader's knowledge into a coherent whole (Kintsch & van Dijk, 1978). Figure 1 sets out the three levels this article assumes, running from the transient surface form through the propositional textbase to the situation model that represents what the text is about.

Levels of Representation

The middle level, the textbase, is the network of propositions the text explicitly states. A proposition is the smallest idea that can be judged true or false, typically a predicate with its arguments, and the sentences of a text map onto a set of such propositions linked by shared arguments. Kintsch and van Dijk distinguished the microstructure of the textbase, the local web of propositions and the coherence relations that connect adjacent sentences, from its macrostructure, the gist or hierarchical summary that macro-rules derive by deleting detail and generalizing across propositions (Kintsch & van Dijk, 1978). The macrostructure is why a reader can report the main point of a passage while forgetting its particulars: the summarizing operations that build it discard the very details that made up the microstructure.

The highest level, the situation model, cannot be built from the text alone; it requires the reader to supply knowledge the writer assumed. Bransford and Johnson made the dependence vivid with passages that were grammatical and fully explicit yet nearly impossible to understand or recall without an organizing context. A paragraph describing an everyday procedure in deliberately vague terms was rated incomprehensible and poorly remembered, but the same words became clear and memorable when a title or picture supplied the topic in advance, because only then could readers activate the schema needed to relate the sentences to one another (Bransford & Johnson, 1972). Comprehension, on this evidence, is not decoding followed by storage but the integration of the text with what the reader already knows.

The Construction-Integration Model

The most influential account of how meaning is assembled is Kintsch's construction-integration model, which explains comprehension as the settling of a network in two phases. In the construction phase the text and the reader's associations generate a large, unruly set of candidate meanings and propositions, produced by weak, context-insensitive rules that fire promiscuously, so that irrelevant word senses and unwarranted inferences are activated alongside appropriate ones. In the integration phase the network relaxes into a stable state through spreading activation: elements that are mutually consistent reinforce one another, while those that fit poorly with the emerging representation lose activation and drop away (Kintsch, 1988). Coherent interpretation is thus an emergent product of constraint satisfaction rather than a set of carefully applied rules, which lets the same architecture handle knowledge integration and error correction without a separate intelligent controller.

The model is at its most concrete with lexical ambiguity. When a reader meets a word with several meanings, the construction phase briefly activates all of them regardless of context, and only integration with the surrounding sentence suppresses the meanings that do not fit. This predicts the momentary activation of contextually inappropriate senses that comprehension research has repeatedly observed, followed by their rapid decay once context is brought to bear. The demonstration below runs this settling process on a single ambiguous word, letting the reader choose the sentence context and watch the network suppress the meaning that does not cohere with it.

Situation Models

If the textbase is what the text says, the situation model is what the text is about: a representation of the people, objects, and events described, structured more like a model of a real or imagined world than like a string of sentences. Zwaan and Radvansky synthesized the evidence that comprehenders build and continually update such models, and that memory for a text is largely memory for its situation model rather than for its wording or even its propositions (Zwaan & Radvansky, 1998). The strongest support comes from findings that readers are sensitive to the described situation in ways the text does not explicitly mark. Glenberg and colleagues showed that an object described as spatially associated with the protagonist, carried rather than set down, remained more accessible in memory as the narrative continued, evidence that readers track spatial relations in a model of the situation and not merely the sentences that reported them (Glenberg, Meyer, & Lindem, 1987).

How the model is kept current is specified by the event-indexing model, which holds that readers monitor several dimensions of the unfolding situation simultaneously. Zwaan, Langston, and Graesser proposed that each new event is indexed on five dimensions, namely time, space, causation, the protagonist or object involved, and intentionality or goals, and that reading slows whenever an incoming event breaks continuity on one or more of these dimensions, because the model must be updated to absorb the shift (Zwaan, Langston, & Graesser, 1995). A change of scene, a time jump signaled by a phrase such as an hour later, or a switch of protagonist each impose a measurable processing cost. The demonstration below advances through a short narrative and shows which of the five indices change at each step, with a processing-load reading that rises with the number of simultaneous shifts.

Inference and Coherence

Because writers leave much implicit, comprehension depends on inference to establish coherence between sentences. The most basic kind is the bridging inference required by the given-new structure of language. Haviland and Clark showed that readers treat the definite article and similar devices as an instruction to find an antecedent already in the discourse: a sentence whose given information matches something stated earlier is read quickly, whereas one that presupposes information not yet established forces the reader to infer a bridge, which takes measurably longer (Haviland & Clark, 1974). Reading the beer was warm after a first sentence mentioning some beer is fast; reading it after a sentence mentioning a picnic basket is slower, because the reader must infer that the basket contained beer before the sentence can cohere.

How far inference extends beyond such necessary bridges is contested, and Table 1 sets out the two positions. McKoon and Ratcliff advanced a minimalist hypothesis: in the absence of specific goals, readers automatically encode only the inferences needed for local coherence and those that follow readily from information quickly available, and they do not routinely compute the many elaborative or predictive inferences a text could support (McKoon & Ratcliff, 1992). Graesser, Singer, and Trabasso countered with a constructionist theory, on which readers engaged in normal comprehension are driven by a search after meaning: they try to explain why actions, events, and states are mentioned, and so generate the goal, causal, and thematic inferences needed for a globally coherent explanation of the text (Graesser, Singer, & Trabasso, 1994). The debate is not whether readers infer but which inferences are routine, and the answer depends on the reader's purpose and the demands of the material.

Note. The two accounts agree that bridging inferences needed for coherence are made during reading and differ over whether readers routinely generate the wider set of elaborative and explanatory inferences. Adapted from McKoon and Ratcliff (1992) and Graesser, Singer, and Trabasso (1994).

Working Memory and Individual Differences

Comprehension unfolds under a tight capacity limit, because the propositions of earlier sentences must be held available while later ones are processed and integrated. Daneman and Carpenter captured this with the reading span task, which requires people to read a series of sentences aloud while holding the final word of each for later recall, so that the same system must process and store at once. Reading span measured this way correlated strongly with reading comprehension, and far better than a simple word span that involves storage alone, because it taxes the coordination of processing and storage that comprehension itself demands (Daneman & Carpenter, 1980). The finding established working-memory capacity as one of the most reliable predictors of comprehension skill.

Just and Carpenter built this into a capacity theory of comprehension, in which a single pool of activation supports both the computations of comprehension and the temporary storage of intermediate products. When a sentence is syntactically complex or an integration spans many words, demand can exceed the available capacity, and readers with a smaller pool slow down or fail on exactly the constructions that require holding more material in an active state (Just & Carpenter, 1992). A complementary account locates individual differences in a specific process rather than a general resource. Gernsbacher's structure building framework holds that comprehenders lay a foundation and then map coherent incoming information onto it, and that less-skilled comprehenders are worse at suppression, the mechanism that dampens irrelevant or no-longer-relevant information such as the inappropriate meaning of an ambiguous word (Gernsbacher, Varner, & Faust, 1990). On this view a poor comprehender is not simply short of capacity but slow to shut down what no longer fits.

Skilled and Developing Comprehension

Reading comprehension also depends on the machinery that turns print into language in the first place. Gough and Tunmer proposed the simple view of reading, which holds that reading comprehension is the product of two separable components, decoding and linguistic comprehension, and that both are necessary and neither is sufficient. Expressed as a product rather than a sum, the model implies that a reader who cannot decode will not comprehend text however strong the underlying language, and that a fluent decoder with weak language comprehension will not comprehend either (Gough & Tunmer, 1986). The two classic patterns of reading difficulty, the dyslexic reader with intact oral language but poor decoding and the poor comprehender with fluent decoding but weak understanding, fall out of the model as failures of different factors.

At the word level, Perfetti's lexical quality hypothesis specifies what a strong decoder brings to comprehension. High-quality lexical representations, in which a word's spelling, sound, and meaning are precise, redundant, and tightly bound, can be retrieved rapidly and reliably, freeing resources for integration; low-quality representations retrieve slowly or ambiguously and create bottlenecks that ripple up into comprehension (Perfetti, 2007). In development, the component skills that support comprehension can be measured before comprehension itself matures. Cain, Oakhill, and Bryant found that children's inference-making, comprehension monitoring, and knowledge of story structure each predicted later reading comprehension even after word reading and verbal ability were controlled, evidence that comprehension draws on higher-level skills that are partly independent of decoding and vocabulary (Cain, Oakhill, & Bryant, 2004). The demonstration below makes the multiplicative logic of the simple view concrete.

Worked Example

The simple view of reading states that reading comprehension is the product of decoding and linguistic comprehension, R equals D times C, where each component is expressed as a proportion of full skill between 0 and 1. The multiplicative form is the substantive claim, and its consequences are easiest to see by computing a few cases. Take a typical competent reader with decoding at 0.90 and language comprehension at 0.80. Their predicted reading comprehension is 0.90 times 0.80, which equals 0.72. Now weaken one factor at a time and hold the other at 0.90.

A child with strong oral language but a decoding deficit, the dyslexic profile, might have decoding 0.40 and comprehension 0.90, giving 0.40 times 0.90, or 0.36. A child who decodes fluently but understands language poorly, the poor-comprehender profile, might have decoding 0.90 and comprehension 0.40, giving 0.90 times 0.40, again 0.36. Both land at the same predicted reading comprehension, and both are far below the 0.81 that 0.90 on both factors would yield, because the product is capped by whichever factor is lower. An additive model would treat the two profiles as merely average and would miss the central clinical fact that a severe weakness in either component, uncompensated by strength in the other, depresses comprehension sharply (Gough & Tunmer, 1986). The demonstration above lets both components be varied so the multiplicative cap can be traced across the full range.

Discussion

Comprehension has become one of the best-specified constructs in cognitive psychology because researchers stopped treating it as a single act and decomposed it into levels and processes. The levels-of-representation framework separates what a reader keeps, the durable situation model, from what a reader quickly loses, the surface form, and thereby explains why understanding and verbatim memory dissociate (Kintsch & van Dijk, 1978). The construction-integration model supplies a mechanism, constraint satisfaction over a noisy network, that builds coherence and resolves ambiguity without a homunculus directing the work (Kintsch, 1988). The situation-model tradition shows that the endpoint of comprehension is a model of a world, tracked along several dimensions at once (Zwaan & Radvansky, 1998), and the inference literature specifies which of the connections that model requires are drawn as a matter of course and which wait upon the reader's goals (Graesser, Singer, & Trabasso, 1994). Running beneath all of this is a capacity constraint that makes working memory a reliable index of who comprehends well (Daneman & Carpenter, 1980), and a word-level foundation whose quality determines how much capacity is left for the higher-level work (Perfetti, 2007). The strands converge on a single picture: comprehension is the incremental construction of a coherent model of a described situation, built from limited-capacity processing over representations of varying durability, and its failures can be traced to a shortfall at any level, from a low-quality lexical entry to an unmade inference to a situation model that was never grounded in the right knowledge.

Glossary

Bridging inference.

An inference a reader must draw to connect a new sentence to earlier information, as when a definite reference presupposes an antecedent that was only implied.

Construction-integration model.

Kintsch's account in which comprehension first constructs a noisy network of candidate meanings and then integrates it by spreading activation into a coherent interpretation.

Decoding.

The process of translating printed words into their spoken forms and meanings; one of the two components of reading in the simple view.

Elaborative inference.

An inference that adds information a text implies but does not require for coherence, such as a likely consequence or an instrument; not always generated during reading.

Event-indexing model.

A model of situation-model updating in which readers monitor time, space, causation, protagonist, and goals, with reading slowing when an event breaks continuity on a dimension.

Given-new contract.

The convention that a sentence marks some information as given, to be matched to an existing antecedent, and some as new, guiding the bridging inferences a reader makes.

Inference.

Information a comprehender adds to a text, drawing on knowledge or context, to establish coherence or to explain why events are described.

Lexical quality hypothesis.

Perfetti's proposal that precise, well-bound representations of a word's spelling, sound, and meaning support fast retrieval and free resources for comprehension.

Macrostructure.

The gist or hierarchical summary of a text, derived from the microstructure by rules that delete detail and generalize across propositions.

Microstructure.

The local network of propositions in a textbase and the coherence relations that connect adjacent sentences.

Proposition.

The smallest unit of meaning that can be judged true or false, typically a predicate with its arguments; the building block of the textbase.

Reading span.

A measure of working memory in which a person processes sentences while retaining their final words for recall; a strong predictor of comprehension skill.

Simple view of reading.

The model that reading comprehension is the product of decoding and linguistic comprehension, so a severe weakness in either factor caps the outcome.

Situation model.

A mental representation of the state of affairs a text describes, structured like a model of a world and constituting what a reader chiefly remembers.

Structure building framework.

Gernsbacher's account in which comprehenders lay a foundation and map coherent information onto it, with suppression damping what no longer fits.

Suppression.

The mechanism that reduces the activation of irrelevant or no-longer-relevant information, such as the inappropriate meaning of an ambiguous word; weaker in poor comprehenders.

Surface form.

The verbatim record of the exact words and syntax of a message, which decays within seconds as its meaning is extracted.

Textbase.

The level of representation consisting of the propositions a text explicitly asserts and their connections, intermediate between surface form and situation model.

Key Researchers

Morton Ann Gernsbacher. Developed the structure building framework and identified weak suppression as a source of individual differences in comprehension skill; Vilas Research Professor of Psychology at the University of Wisconsin-Madison, ORCID 0000-0003-0397-3329. ORCID · Google Scholar · Faculty page · Wikipedia

Arthur C. Graesser. Developed the constructionist search-after-meaning theory of inference generation and built intelligent tutoring systems grounded in discourse comprehension; Professor Emeritus of Psychology at the University of Memphis, ORCID 0000-0003-0345-6866. ORCID · Google Scholar · Faculty page · Wikipedia

Walter Kintsch (1932-2023). With Teun van Dijk, formulated the levels-of-representation account of text comprehension, and alone developed the construction-integration model that remains the field's dominant architecture; Professor Emeritus of Psychology and former Director of the Institute of Cognitive Science at the University of Colorado Boulder. Faculty page · Wikipedia

Charles A. Perfetti. Proposed the lexical quality hypothesis linking the precision of word representations to reading comprehension; University Professor and Senior Scientist at the Learning Research and Development Center, University of Pittsburgh, ORCID 0000-0002-0211-8518. ORCID · Google Scholar · Faculty page · Wikipedia

Rolf A. Zwaan. Co-developed the event-indexing model and synthesized the evidence that comprehension culminates in a situation model that memory preserves; Professor of Biological and Cognitive Psychology at Erasmus University Rotterdam, ORCID 0000-0001-9967-7879. ORCID · Google Scholar · Faculty page

Frequently Asked Questions

What is comprehension in cognitive psychology?
Comprehension is the process of building a coherent mental representation of the meaning of a spoken or written message, in which the comprehender integrates the words with prior knowledge rather than storing them verbatim. Its product is a model of what the message is about, not a copy of its form (Kintsch & van Dijk, 1978).

What is a situation model?
A situation model is the mental representation of the state of affairs a text describes, structured like a model of a real or imagined world rather than a string of sentences, and it is largely what a reader remembers of a text once the wording has faded (Zwaan & Radvansky, 1998).

What is the construction-integration model?
It is Kintsch's account of comprehension as a two-phase process: a construction phase activates a noisy set of candidate meanings, and an integration phase settles the network by spreading activation so that mutually consistent elements survive and poorly fitting ones fade (Kintsch, 1988).

Why does memory for exact wording fade so quickly?
Because comprehension extracts meaning and discards form: recognition of the surface wording and syntax of a sentence decays within seconds of continued reading, while memory for its meaning remains accurate (Sachs, 1967).

Do readers infer everything a text implies?
Not according to the minimalist hypothesis, which holds that during ordinary reading people automatically encode only the inferences needed for local coherence or those readily available, rather than the full set of elaborative and predictive inferences a text could support (McKoon & Ratcliff, 1992).

Why do some people comprehend better than others?
Working-memory capacity is a strong predictor: the reading span task, which requires processing sentences while holding their final words, correlates closely with comprehension because comprehension itself demands coordinating processing and storage (Daneman & Carpenter, 1980).

What is the simple view of reading?
The simple view holds that reading comprehension is the product of decoding and linguistic comprehension, so that both are necessary and a severe weakness in either component caps the outcome regardless of strength in the other (Gough & Tunmer, 1986).

How does prior knowledge affect comprehension?
Prior knowledge supplies the schema needed to relate a text's sentences to one another; a passage that is grammatical and explicit can be nearly incomprehensible until a title or context activates the relevant knowledge, after which the same words become clear and memorable (Bransford & Johnson, 1972).

References