A Paradigm for the missing middle

ABSTRACT

This commentary highlights what I see as the most important of Walter Kintsch’s many contributions to cognitive psychology and especially text comprehension. His work filled the great void between the top and the bottom in research on comprehension. The state of the field at the time was (a) strong research foundation on word reading, (b) detail-focused research on syntactic processes by a vibrant sentence processing research community, and (c) demonstration research on the importance of higher-level top down influences. The void between (b) and (c) meant there was no possibility of a comprehension science that was both rigorous, like (a) and (b), and reflective of the role of outside-the-text knowledge, like (c). Beyond the specific empirical contributions of Kintsch’s work is that he showed that it was possible to create a paradigm for the missing middle, a characterization of how word-driven and knowledge-driven processes yield a coherent meaning-based understanding of a text a few pieces at a time. This allowed the flourishing of work on the incremental processes that result in a reader’s (often) coherent representation of text meaning.

Introduction

Walter Kintsch’s contributions to the scientific study of the human mind are both varied and profound, not only to the field of text comprehension but more generally to the multidisciplinary program of the cognitive sciences, extending even to how cognitive science can capture the mental representations of beauty (Kintsch, 2012). In this commentary I focus on a multifaceted contribution of high impact, one that has enabled the emergence of text comprehension as a genuine scientific subfield: Kintsch provided a paradigm to build-in the missing middle that had existed between the top-down guidance of comprehension and the bottom-up word level and syntactic processes that build meaning incrementally.

Filling this missing middle allowed the following description of how reading and comprehending a text unfolds to become widely accepted: Text comprehension results from word-by-word, phrase-by-phrase, and sentence-by-sentence process guided by multiple knowledge sources accessible to the reader. How this happens in detail requires the study of incremental, localized processes across microlevels of language. Perhaps because these processes are challenging to study, research on comprehension started at the high end, where global structures might influence local word and sentence processes.

Comprehension from the top and disconnected word processing from the bottom

The study of text comprehension was not a focus of research until the cognitive revolution was in full swing in the 1970s. Early artificial intelligence systems demonstrated the power of global organizers in guiding comprehension, for example, creating “scripts” for specific situational comprehension (Schank & Abelson, 1977). Thus, knowledge about restaurant scripts—key actors and typical sequences of actions—guided the reader’s understanding of simple stories about restaurant experiences. Similarly, approaches in psychology and education emphasized situated conceptual structures or schemata (Anderson & Pearson, 1984). Experimental evidence for global top-down guidance came from studies of vaguely written texts that became comprehensible when readers were given a helpful title (Bransford & Johnson, 1972). Moreover, this top-down approach was able to demonstrate individual differences in comprehension that were due to background knowledge. Thus, Anderson et al. (1977) showed that a text that lacked referential specificity could be understood as being about either music or card playing depending on whether the reader was a student in music education or physical education. Other top-down approaches hypothesized roles in narrative comprehension for generalized mental structures based on everyday experiences, for example, story grammars (Mandler & Johnson, 1977; Stein & Glenn, 1979). Trabasso and colleagues (Trabasso et al., 1984) hypothesized causality as a fundamental mental structure and showed that causal expectations activated during reading predict how readers understand sentences (Trabasso & Suh, 1993).

These studies were groundbreaking in establishing comprehension as an object of study and in demonstrating the importance of knowledge beyond the text in guiding comprehension. This was accomplished while ignoring the processes that identified words, retrieved their meanings, integrated word meanings and syntactic structures, and established coherence across sentences.

Meanwhile, there was substantial progress from other research communities focusing on the word-level of the reading system. Experimental work and modeling of word identification was flourishing and research on word meaning processes was addressing issues relevant for comprehension. A major issue was whether, in the context of a sentence, all meanings of a word were activated or only the single meaning that fit the context. Thus, the problem was not that the bottom-up part of reading was ignored but that it was disconnected from comprehension in both research goals and research communities. Both the top and bottom were targets of research. It was the middle that was missing: an approach to comprehension that included a prominent role for word meanings.

Building the missing middle

Word meanings are central in reading, the output of the word identification system and the input to the comprehension system, as emphasized in the Reading Systems Framework (Perfetti & Stafura, 2014). Identifying words and integrating them into comprehension structures (semantic and syntactic) are the main recurring events in reading. Kintsch and his colleagues brought word meanings into focus. Kintsch and van Dijk (1978) approached text comprehension as a cyclical process in which every text element activated meanings on the way to building a coherent representation of the text. The Construction Integration (CI) Model (Kintsch, 1988) proposed two rapidly occurring phases of comprehension: An initial short-lived construction phase, prompted by word meaning, spreads activation across memory of both text elements and general knowledge in a passive, automatic process. (In the CI Model, “construction” contrasts sharply with its use in other comprehension accounts, where it entails an active role for the reader who “constructs” understanding, e.g., Graesser et al., 1994.) A rapidly occurring integration phase uses meaning overlap from the preceding cycle and additional knowledge to constrain the information remaining for the next cycle, while meaning overlap across multiple cycles leads to a coherent representation of the text. The CI Model moved text comprehension research toward a processing approach, incorporating memory-based, word-meaning, and sentence level components.

It is hard to overestimate the significance of the CI Model—not just the model itself but the approach it represents—for research on reading comprehension. Beyond a number of direct extensions of the model (e.g., Goldman & Varma, 1995; Schmalhofer, 1995) is the impact of the CI approach on the kinds of questions that emerged when words, phrases, and sentences became important in accounting for text comprehension. For example, the structure building framework (Gernsbacher, 1990, 1997) emphasized the complementary processes of memory-based meaning mapping and structure building. Later models retained a focus on bottom-up, memory-based processes, including the Resonance Model (Myers & O’Brien, 1998) and the Resonance, Integration, and Validation Model (Cook & O’Brien, 2014).

The text comprehension field did not retreat from the study of global, top-down influences on comprehension. Such influences continued to be emphasized in constructivist theories (Graesser et al., 1994) and theories that postulated specific mental structures (dimensions of events) that guide the reader’s construction of narrative coherence (Zwaan et al., 1995). The Landscape Model combined the automatic bottom-up processes of memory-based models with top-down influences (Van den Broek et al., 1999, 2005) from the reader’s standard for coherence and reading goals (Van den Broek & Helder, 2017; Van den Broek et al., 1995).

Kintsch also influenced how the research field treated knowledge as integral with text-driven processes. In the situation model (Van Dijk & Kintsch, 1983), comprehension occurs at multiple levels, one close to the explicit meaning of the text and one enriched through knowledge-based inferencing that produces a richer understanding of a text. Successful situational comprehension thus includes meanings that are referentially specific but also well aligned with the text meaning.

An interesting aspect of the CI Model is its omission of syntactic influences, an omission that Kintsch acknowledged. Indeed, this is true for nearly all models of text comprehension, which ignore the parsing processes (influenced by both syntactic knowledge and semantic context) that configure words into hierarchical structures associated with meaning relations (propositions). This omission is largely due to the specialized scientific communities across levels of reading: Word processing, sentence processing, and text processing are areas of research with little overlap either in research goals or researchers. This may not be a problem for a theory of text comprehension. The propositions that are the units of meaning can be assumed rather than computed from sentences, that is, we assume that cognitive operations (parsing) on word strings have resulted in the meaning units that are input to a text processing model. On another view, one can ask where propositions come from (Perfetti & Britt, 1995), with the answer to this question being part of a comprehensive theory of comprehension.

A thought about the future as now

The future is always one step ahead of our attempts to characterize it. For text researchers, the future, which is already present, includes the promise of new tools to study comprehension and production of language. The most compelling of these are the opportunities and challenges associated with the power of large language models and generative artificial intelligence. These models provide an opportunity for applications to research on basic issues of incremental comprehension. Their applicability reflects the research perspectives provided by Walter Kintsch’s work on text comprehension. For example, one consequence of treating comprehension as incrementally constructed from reading words and phrases is that research methods that can assess the reading of every word in a text become important. At the experimental level, eye tracking and event-related potentials have long provided the kind of word-by word data needed for tracking incremental processes. Now these experimental tools (and others) can be enhanced by knowledge-related data provided by large language models.

Large language models come into this picture in several ways. First, they provide statistical data based on large corpora, so that the surprisal of a word (the inverse of predictability) in a given text is measured reliably and probably more meaningfully than experimental norming. Because they also allow the measurement of meaning similarity among words in a sentence, the models provide statistical data that are differentially related to the construction phase (word meaning associations) and the integration phase (surprisal) of the CI Model. Second, they allow researchers to compare different large language models on behavioral, eye tracking or event-related potential data. The models have different architectures and different inferable links to cognitive processes. The third follows from the second and is the most intriguing. Because the models have different architectures and different links to cognitive processes, there exists the potential to use the relative success of different models to inform ideas and stimulate research about how humans comprehend texts.

Disclosure statement

No potential conflict of interest was reported by the author(s).