Text Comprehension Research Paper

Text comprehension involves many cognitive processes that relate the elements in the text to one another and to prior knowledge so that the result is a coherent mental representation. An important issue for models of text comprehension is the contribution of prior knowledge of the text topic and structure to the formation of coherent representations. Texts are seldom completely explicit with respect to the connections among text elements so comprehenders often have to make inferences to bridge these ‘gaps.’ Inferences sometimes are made just from the text information and sometimes involve elaborative processes that rely on prior knowledge.

The outcome of text comprehension is a memory representation that reflects the meaning the comprehender has constructed. The representation may reflect the surface text itself, the meaning constructed from the text elements (e.g., words, phrases, and sentences that were actually in the text), and/or meaning elements derived from prior knowledge. The representation often goes beyond the meaning of the text (the textbase model) to reflect a model of the situation described by the text. Prior knowledge plays a greater role in the situation model than in the textbase.

In general, models of text comprehension seek to describe, explain, and predict various processing activities and their representational consequences. A lower bound on the adequacy of models of text comprehension is that they provide explanatory accounts of the behavioral data. Currently, models of text comprehension provide more adequate descriptions and explanations than predictions of behavioral phenomena. Furthermore, most models do not deal adequately with all of the empirical findings. Most models selectively focus on some subset of empirical findings, although they increasingly attempt to be comprehensive (see, e.g., Britton and Graesser 1996, Gernsbacher 1997, Kintsch 1998, O’Brien and Myers 1999, van Oostendorp and Goldman 1999). The remainder of this research paper discusses current and future directions in the study of text comprehension models.

1. Approaches To Studying Models Of Text Comprehension

Text comprehension research has required the development of methods for systematically analyzing and describing characteristics of the text, the comprehender, and the processing and representation demands of various comprehension tasks. Researchers examine these ‘input’ variables in relation to two types of dependent measures, namely those that assess processing as it is occurring (on-line measures) and those that assess the results of the processing (memory representation measures). Computational models have been formulated to give more precise explanatory and predictive power to models of text comprehension.

1.1 Text Analysis

Texts are often described at the level of the individual ideas (microlevel) as well as the more global organization of these ideas (macrolevel). Origins of modern text analysis systems typically represent the ideas as propositions. Although the specific details of propositional notation differ somewhat across systems, a proposition is a theoretical unit that contains a predicate (e.g., a verb or adjective) and associated arguments (e.g., nouns, embedded propositions). Arguments have a functional role such as agent, object, or location (for discussion, see Graesser et al. 1997 a, Goldman 1997).

Consider the first sentence in a hypothetical paragraph: Mike drove to the gas station and got out of the car. There are two ideas linked together by an and. Mike is the agent of both driving and getting out of the car and the location of both actions is the gas station. The ideas in subsequent sentences would be similarly described and attempts made to link them. Suppose the next several sentences relate details of Mike’s attempt to rob the gas station, including the fact that the attempt was foiled when the car stalled as Mike was making his getaway. The global organization would reflect an attempted robbery theme, provided that the comprehender had prior knowledge of robbery scenarios. Ideas more central to the robbery theme would have greater importance, be more central, and be at a more superordinate level in the representation.

Empirical research has shown that propositions are a useful way to characterize text and prior knowledge. Recent challenges to propositional representations contend that there is more to meaning than the symbolic form, i.e., the word (Glenberg 1997). Clearly this is the case, but it does not seem to deny the usefulness of propositional representations for text research.

Although prior knowledge can guide the organization of ideas in a text, the surface text itself contains cues to the underlying semantic organization. Relationships among ideas may be signaled by repetition of specific concepts, either in exactly the same form (e.g., car) or through synonym (e.g., jalopy for the stalled car in the Mike example). Rhetorical devices (e.g., The key idea is, The first point is) and graphic cues (e.g., paragraph indentations, boldface font) can help readers recognize and create connections among ideas (Goldman and Rakestraw 2000, Zwaan and Radvansky 1998).

1.2 Comprehender Characteristics

Several characteristics of comprehenders provide constraints on text comprehension. The most frequently examined are prior knowledge relevant to the topic of the text and discourse forms, working memory, and interest.

1.2.1 Prior Knowledge. Researchers have developed various techniques to assess the impact of prior knowledge on text comprehension and vice versa. Techniques for assessing domain-relevant concept knowledge include concept relatedness ratings, question answering, sorting, and cued associative responding. Assessments of discourse genre knowledge, and rhetorical structures include judgment tasks, their use in generated text, and patterns of on-line processing measures such as reading time and verbal reports. By comparing groups that differ in prior knowledge measures, researchers can assess the impact of knowledge on text comprehension. Researchers infer the impact of the text on knowledge from changes in measures of prior knowledge before and after comprehension.

1.2.2 Working Memory. Although all humans have limited attentional resources, there are individual differences in working memory span for verbal material and these are related to text processing and memory. Daneman and colleagues (see Daneman and Tardif 1987) developed a technique for assessing working memory span. Working memory limitations mean that, in most cases, the whole text cannot be processed at once but must be processed in cycles. Ericsson and Kintsch (1995) proposed that one way of dealing with resource constraints is to construct retrieval structures in long-term working memory.

1.2.3 Interest. Personal interest in particular topic areas is assessed through self-report measures typically. Text comprehension is positively related to personal interest in the topic of the text. Situational, or text-based interest, is associated with specific information in a specific text and often is measured through ratings of each sentence in a text. Information with high situational interest, sometimes called ‘seductive details,’ is well recalled regardless of its importance to the overall theme of the text (see review by Alexander et al. 1994).

1.3 Analysis Of Task Demands

Task demands are characterized in terms of the memory representation needed to perform on postcomprehension tasks and whether the tasks require that the comprehender just retrieve information or retrieve and reason with information. For example, comprehenders can do an excellent job of retelling a story so long as they have a good representation of the surface elements and textbase of the story. However, if they need to decide whether the main character acted ethically, they need to reason about the events in the story. The reasoning involves the situation representation.

Other analyses of task demands examine indices of processing difficulty (e.g., unfamiliar vocabulary or complex syntax) in order to compare different texts or task situations. Under conditions of high task demands, comprehenders should either take longer to process the text or end up with less coherent representations. The effects of task demands will be minor if comprehenders do not interpret them appropriately and strategically respond to them (Goldman 1997).

1.4 Assessing Online Processing Of Text

Online techniques allow text processing to be assessed independently of the resulting representation and its subsequent use in memory or problem solving tasks. Online measures indicate what is active in working memory at the same time. This is important because limits on attentional resources constrain the amount of text and prior knowledge that can be active simultaneously. Simultaneous activation is necessary for forming links among text or prior knowledge elements, according to virtually all current models of comprehension (Britton and Graesser 1996, Kintsch 1998, van Oostendorp and Goldman 1999).

Some online techniques (e.g., eye movements, reading times, and inspection patterns) provide estimates of where readers look, for how long, and the order in which they look at specific information. Other techniques provide estimates of what information is active in working memory at specific points during the reading process (e.g., probe memory, priming, and think-aloud techniques). Several models and mechanisms have been proposed to account for observed patterns of processing and activation of prior knowledge and previously processed portions of text, including the structure building framework (Gernsbacher 1997), associative priming (Kintsch 1998), resonance (O’Brien and Myers 1999), constructionist theory (Graesser et al. 1994), and the memory- based processing model (O’Brien and Myers 1999). In addition, think-aloud protocols indicate that there are age and skill-related differences in strategies that comprehenders use to form coherent representations of text, with more effective comprehenders using explanation-based strategies to a greater degree than less effective comprehenders (Chi in press, Cote et al. 1998, Trabasso and Magliano 1996).

1.5 Assessing Mental Representations

Obviously, mental representations are not directly observable. Researchers make inferences about mental representations on the basis of observable indices of those representations, such as the information comprehenders produce when asked for the ‘most important’ information in a text or when asked to recall a text. Assessments of text representations are taken after the comprehender has read or heard the text, and thus differ from on-line measures.

Methods for studying comprehenders’ representations are typical of those used to study broader issues of memory. A primary distinction is between recognition and recall or free response, techniques. Recognition memory is often used to determine the content and form of elements in the representation (e.g., surface, text, or situation). Recall or free response measures can be more or less open-ended. Comprehenders may be asked to recall the text, recall just the main ideas from the text, summarize the text, create a diagram of the important information from the text, or answer questions about presented and potentially inferred information.

Assessments of mental representations have established a number of important findings, summarized here. Ideas that are more important, more causally central, or more superordinate are remembered better and more frequently included in summaries than are subordinate ideas. Generally, this is the same information that receives more extensive processing. In addition, the more coherent the text, the better the recall, especially for comprehenders who are ‘low knowledge’ with respect to the topic of the text. Finally, interest affects what is remembered. When interest value is pitted against importance, interest wins.

2. Computational Models

Computational models simulate text comprehension on a computer. Computer models must be extremely specific and precise with respect to the hypothesized processes, representations, and interactions. The fit of the simulated output can be compared with the output of humans for a comparable task to provide a basis for validating and updating models of text comprehension. Likewise, mathematical models quantify precise patterns of processing times, memory scores, ratings, and other psychological data. Failures of fit can be as informative as good fits because these discrepancies can lead to new insights about the limitations of existing models (Goldman et al. 1996, Graesser et al. 1997 a).

Computational models of text processing have been influenced by symbolic and connectionist cognitive theories. In symbolic theories (Anderson 1983), actions are linked to specific conditions under which they might occur. These are called production rules and take the form of If <condition> Then <action>. When the condition side of a production is satisfied, the production ‘fires’ and the action occurs. When new input enters working memory, all production rules are evaluated to determine which ones have conditions that are satisfied, and one is selected to fire, completing a cycle. On successive cycles, the input to working memory dynamically changes, and different productions are satisfied.

Connectionist theories (Rumelhart and McClelland 1986) distribute representations and processes among a large set of simple units analogous to neural units. Each word, concept, or proposition has a corresponding ensemble of neural units. The units are connected with one another through weights that are excitatory, inhibitory, or neutral and form a neural network. During comprehension, a particular set of units that matches the input is initially activated and activation spreads throughout the network consistent with the excitatory and inhibitory links. Activation continues spreading until there are relatively minimal changes in the activation levels of the units and a stable pattern of activation is achieved. This pattern of activation captures the meaning of the input.

Most models of text comprehension are hybrids of symbolic and connectionist theories. Currently, the two most influential models are those of Just and Carpenter (1992) and Kintsch (1998).

2.1 Just And Carpenter’s Collaborative Activation-Based Production System (CAPS) Model

The CAPS architecture (Just and Carpenter 1992) is a production system architecture augmented by two characteristics that give it a connectionist flavor. First, the collaborative aspect of CAPS results from removing the restriction that only one production fires on each cycle. Rather, all matched productions are simultaneously fired. Second, the activation process is not all or none. The condition sides of productions accrue levels of activation. The collaborative and changing activation characteristics of CAPS combine to generate a process in the connectionist spirit: productions fire in parallel, gradually changing the activations of working memory elements by an amount proportional to the activations of other elements. When the activation of a particular element falls below a threshold activation level, that element is no longer noticed by the system and is unable to participate in further processing. Just and Carpenter constructed a comprehension-simulation model within the CAPS architecture, READER. It was designed to process a short, expository passage on flywheels. Lexical, syntactic, and semantic representations were built simultaneously due to the parallel firing of all productions that matched on a given cycle.

Just and Carpenter (1992) embellished the CAPS model with the assumption of an activation limit (‘cap’) on working memory. The capacity-constrained architecture (3CAPS) means that productions compete for available activation. When fired production rules request more activation than is available, the ‘cap’ has been reached. Processing at the ‘cap’ results in a graceful loss from working memory of productions that are not participating in processing, and an overall system slowdown.

2.2 Kintsch’s Construction–Integration Model

The Construction–Integration (CI) model (Kintsch 1998) represents the input as propositional nodes in a network of linked nodes. The node is a symbolic representation of the content that has a weighted link in the connectionist manner to other nodes in the network. The model simulates the construction of a text-base representation and the incorporation of prior knowledge to form a model of the situation described by the text. The model proceeds in two phases. During Construction, propositions are linked based on argument (concept) overlap to create an initial representation. Additional links can be specified depending on assumptions about causality, temporality, and logical relations. Prior knowledge is also activated ‘indiscriminately’ and some irrelevant information may be included as part of the network representation. Kintsch (1998) uses the Latent Semantic Analysis (LSA) model of knowledge representation (Landauer and Dumais 1997) to simulate prior knowledge activation. In LSA, knowledge is represented as vectors in a high-dimensional ‘semantic space.’ Concepts in the input text are located in the semantic space and neighbors are activated.

During Integration, activation spreads through the network of nodes. Activation flows out of nodes with few or weak connections and to nodes that are richly connected, until the network reaches a stable state. During Integration, relevant knowledge associations tend to maintain activation and become part of the representation, while irrelevant associations tend to lose activation and drop out of the representation. At the completion of a CI cycle, each text node has a particular strength value and link strengths for each node to which it is connected. At the conclusion of comprehension of the entire text, each node and link in the network has a cumulative strength value, reflecting its cumulative activity during all cycles of comprehension. These values are used to create long-term memory strength values that can be used to predict the relative likelihood of remembering a particular node or relationship.

The CI model takes into account working memory limitations in that input propositions are processed in cycles and only the most active are held over for processing in subsequent cycles. The number that are ‘held over’ is a parameter in the model. Propositions not held over can be reinstated in subsequent comprehension cycles. Kintsch (1998) uses the theory of Long-term Working Memory (Ericsson and Kintsch 1995) to elaborate the relationship between working memory and long-term memory. This theory postulates that expert performance for tasks such as reading is accomplished by having highly organized and easily accessible knowledge structures in long-term memory. During processing, working memory utilizes ‘pointers’ to these structures, which makes them quickly and easily accessible, thus accounting for the fast and accurate application of prior knowledge under conditions of normal comprehension.

2.3 Other Activation-Based, Dynamic Models

The CI and 3CAPS models have spawned a ‘family’ of models that share assumptions about fluctuating activation of elements and dynamic behavior of the representations. One model, Capacity Constrained Construction Integration (3CI), combines CI and 3CAPS by replacing the fixed-buffer working memory of the CI model with the dynamic, fixed-capacity working memory system of 3CAPS (Goldman et al. 1996). 3CI allows all propositions that are currently active after a CI processing cycle to be carried over to the next processing cycle. Loss of activation is a natural result of processing when the ‘cap’ is reached. 3CI results in an enhanced ability to establish coherence across sentences by having more propositions potentially participate in the same processing cycle, thereby increasing the link strength between them. Simulation results show that for texts that have hierarchical organizations (such as a premise and examples), the 3CI model is able to assess the global discourse structure and relate specific information to that global structure more effectively than the CI model.

Other models in the activation-based family focus on the nature of the links among text elements (Britton and Graesser 1996, van Oostendorp and Goldman 1999). For example, Langston and Trabasso (1999) implemented a model in which causal relatedness drives the creation of links between clauses and is responsible for text coherence. The model accounts for a variety of narrative processing phenomena such as reading time, importance ratings, and the contents of think-aloud protocols.

Van den Broek et al. (1999) focus on different kinds of information and connections among nodes. They proposed a ‘Landscape’ model in which a text representation is constructed from information that is activated from the text in the current reading cycle, the text in the preceding reading cycle, information earlier in the text, and from prior knowledge. Different types of connections carry different weights, and a limited capacity working memory provides constraints on the amount of information that can be processed in any particular cycle. The Landscape model can account for both the order and the content of free recall.

3. Issues For The Future

Conceptual advances in models of text comprehension require that they continue to focus on knowledge activation and inference making, especially for learning from text in content and technical domains. Optimal learning from text occurs when individuals who vary in their prior knowledge of the text topic are ‘matched’ to texts that vary in conceptual difficulty. While this observation is not novel, Wolfe et al. (1998) provided evidence that LSA provides a reasonably convenient way to establish a range of texts on the same topic that differ in conceptual complexity.

Models of text comprehension also need to expand their focus in a number of ways. First, they need to include the processing of the surface text, including parsing, knowledge of discourse structure and genre, and the impact these have on the creation of propositional representations. Advances in techniques for processing natural language make it more feasible for computational models to take surface text input rather than a researcher-parsed list of propositions. This expansion of computational models will permit more detailed understanding of interactions between surface structure characteristics of the discourse and interpretations that comprehenders construct (Goldman et al. 1996).

A second expanded focus concerns the impact of new input on existing knowledge structures (O’Brien and Myers 1999). This is important for learning in content domains and is important in the everyday conduct of human activity. Adults are resistant to updating information they have read in newspapers and magazines when contradictory information is published subsequently. Whether updating occurs appears to depend on the combined impact of strength and explicitness of the contradiction (see chapters in van Oostendorp and Goldman 1999). Future research on updating needs to examine domain-relevant prior beliefs and knowledge because these are likely to affect readers’ judgments about the plausibility of the original and updating information.

Models of text comprehension also need to expand the nature of the texts they examine. In response to the ubiquitous nature of new multimedia technologies and the information resources they make available, individuals are comprehending an expanding array of texts. Research is beginning to examine models of text comprehension in some of these new venues, such as hypertext and hypermedia environments, particularly with respect to scanning and inspection strategies. Comprehenders are accessing multiple sources of information quickly and easily, sometimes in response to self-assessment of comprehension. We know a little about the processing, representation, and integration of information from different sources (Perfetti et al. 1999). We need to know more.

Finally, models of text comprehension need to move beyond the single reader and deal with the social construction of meaning. Text comprehension, like other aspects of cognition, is situated in particular social, historical, and cultural contexts that impact the interpretation process (Bloome and Egan-Robertson 1993, Goldman 1997). What individuals regard as important and meaningful in texts is governed by their interactions with others, as well as by their interactions with the text. Models of text comprehension need to examine the impact of such contexts on the psychological processing and representation of meaning.

Bibliography:

1. Alexander P A, Kulikowich J M, Jetton T L 1994 The role of subject-matter knowledge and interest in the processing of linear and non-linear texts. Review of Educational Research 64: 201–52
2. Anderson J R 1983 The Architecture of Cognition. Harvard University Press, Cambridge, MA
3. Bloome D, Egan-Robertson A 1993 The social construction of intertextuality and classroom reading and writing lessons. Reading Research Quarterly 28: 305–33
4. Britton B K, Graesser A C (eds.) 1996 Models of Text Comprehension. Erlbaum, Hillsdale, NJ
5. Chi M T H in press Self explaining: the dual process of generating inferences and repairing mental models. In: Glaser R (ed.) Advances in Instructional Psychology. Erlbaum, Mahwah, NJ, pp. 161–238
6. Cote N, Goldman S R, Saul E U 1998 Students making sense of informational text: relations between processing and representation. Discourse Processes 25: 1–53
7. Daneman M, Tardif T 1987 Working memory and reading skill re-examined. In: Coltheart M (ed.) Attention and Performance 12: The Psychology of Reading. Erlbaum, Hillsdale, NJ, pp. 491–508
8. Ericsson K A, Kintsch W 1995 Long-term working memory. Psychological Review 102: 211–45
9. Gernsbacher M A 1997 Two decades of structure building. Discourse Processes 23: 265–304
10. Glenberg A M 1997 What memory is for. Behavioral and Brain Sciences 20: 1–55
11. Goldman S R 1997 Learning from text: reflections on the past and suggestions for the future. Discourse Processes 23: 357–98
12. Goldman S R, Rakestraw J A Jr. 2000 Structural aspects of constructing meaning from text. In: Kamil M L, Mosenthal P B, Pearson P D, Barr R (eds.) Handbook of Reading Research. Erlbaum, Mahwah, NJ, Vol. 3, pp. 311–35
13. Goldman S R, Varma S, Cote N 1996 Extending capacityconstrained construction integration: toward ‘smarter’ and flexible models of text comprehension. In: Britton B K, Graesser A C (eds.) Models of Text Comprehension. Erlbaum, Hillsdale, NJ, pp. 73–113
14. Graesser A C, Gernsbacher M A, Goldman S R 1997a Cognition. In: van Dijk T A (ed.) Discourse as Structure and Process. Discourse Studies: A Multidisciplinary Introduction. Sage, Thousand Oaks, CA, Vol. 1, pp. 292–319
15. Graesser A C, Millis K K, Zwaan R A 1997b Discourse comprehension. Annual Review of Psychology 48: 163–89
16. Graesser A C, Singer M, Trabasso T 1994 Constructing inferences during narrative text comprehension. Psychological Review 101: 371–95
17. Just M A, Carpenter P A 1992 A capacity theory of comprehension: individual differences in working memory. Psychological Review 99: 122–49
18. Kintsch W 1998 Comprehension: A Paradigm for Cognition. Cambridge University Press, New York
19. Langston M, Trabasso T 1999 Modeling causal integration and availability of information during comprehension of narrative texts. In: van Oostendorp H, Goldman S R (eds.) The Construction of Mental Representations During Reading. Erlbaum, Mahwah, NJ, pp. 29–69
20. Landauer T K, Dumais S T 1997 A solution to Plato’s problem: the latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological Review 104: 211–40
21. O’Brien E J, Myers J L 1999 Text comprehension: a view from the bottom up. In: Goldman S R, Graesser A C, van den Broek P (eds.) Narrative Comprehension, Causality, and Coherence: Essays in Honor of Tom Trabasso. Erlbaum, Mahwah, NJ, pp. 35–53
22. Perfetti C A, Rouet J-F, Britt M A 1999 Toward a theory of documents representation. In: van Oostendorp H, Goldman S R (eds.) The Construction of Mental Representations During Reading. Erlbaum, Mahwah, NJ, pp. 99–122
23. Rumelhart D E, McClelland J L, P&P Research Group 1986 Parallel Distributed Processing: Explorations in the Micro-structure of Cognition. MIT Press, Cambridge, MA
24. Trabasso T, Magliano J P 1996 Conscious understanding during comprehension. Discourse Processes 21: 255–87
25. Van den Broek P, Yang M, Tzeng Y, Linderholm T 1999 The landscape model of reading: Inferences and the online construction of memory representation. In: van Oostendorp H, Goldman S R (eds.) The Construction of Mental Representations During Reading. Erlbaum, Mahwah, NJ, pp. 71–97
26. van Oostendorp H, Goldman S R (eds.) 1999 The Construction of Mental Representations During Reading. Erlbaum, Mahwah, NJ
27. Weaver III C A, Mannes S, Fletcher C R (eds.) 1995 Discourse Comprehension: Essays in Honor of Walter Kintsch. Erlbaum, Hillsdale, NJ
28. Wolfe M B W, Schreiner M E, Rehder B, Laham D, Foltz P W, Kintsch W, Landauer T K 1998 Learning from text: matching readers and texts by latent semantic analysis. Discourse Processes 25: 309–36
29. Zwaan R A, Radvansky G A 1998 Situation models in language comprehension and memory. Psychological Bulletin 123: 162–85