McKoon and Ratcliff (1992) argue for a "minimalist" theory of inference in reading, contrasting it with "constructionist" theories, including theories based on mental or situation models. Minimal inferences are those required for local (not global) coherence and those based on readily available knowledge. This target article argues that minimalism is a hedged and less testable version of an older theory. More important, McKoon and Ratcliff mischaracterize constructionism, failing to notice that local coherence often depends on constructionist processes. When people do not have the knowledge required to establish local coherence, they do not do so during reading. The only unhedged prediction of the minimalist theory is hence incorrect. Although a theory of inference making should be both constructionist and approximately minimalist, a distinction must be made between a computational theory of inference making and a description of the mechanisms underlying our inferential abilities.
1.1. A crucial question for theories of inference during reading concerns how to characterize the inferences people make and the circumstances under which they make them. McKoon and Ratcliff (1992, henceforth M&R) present a minimalist theory of inference and contrast it with constructionist theories. According to minimalism, two kinds of inference are made automatically during reading: those that establish local coherence and those based on readily available information. Other inferences can be made, but only if the reader engages in nonautomatic strategic processing. The minimalist position has much to recommend it; and it is true that most, if not all, constructionist theories have been nonminimalist. I argue, however, that the contrast between minimalism and constructionism one is a false one, since constructionist processes play an essential role in inferences that are made automatically, according to minimalism. M&R's failure to recognize this point results in inadequacies in their account of inference making and renders incorrect the only unequivocal prediction from their theory: that inferences necessary to establish local coherence will be made during reading.
2.1. Minimalism is a version of a theory that is at least fifteen years old: that the only inferences made as a text is read are those that are necessary to establish a coherent interpretation (an idea mooted, in various forms, by, for example, Corbett & Dosher, 1978; Dosher & Corbett, 1982; Singer, 1980; Thorndyke, 1976). According to this theory, inferences that are merely elaborative are made, not when a text is initially processed, but only if they are subsequently required, for example, to answer either a question posed by an experimenter or one arising in the mind of the reader. The minimalist theory places one further restriction on the inferences it claims are made in comprehension, but it hedges the claims of the older theory in two ways. The restriction is that only inferences establishing local cohesive links are made automatically. Even if further inferences are needed to establish global coherence (by connecting "more globally separated pieces of information," M&R: p. 442), they are not made. (The hedges are described in 2.3 - 2.4 below.)
2.2. Although the idea that global coherence is not automatically established strikes many people as counterintuitive, there is considerable evidence that inconsistencies between different parts of a text are not noticed, even by relatively sophisticated readers (e.g., Baker, 1989). Furthermore, M&R present new evidence that global coherence is not automatically established during reading. Nevertheless, some global inferences may be made. For example, Gernsbacher and Robertson (1992) showed that fictional characters' emotional states are readily inferred. Garnham and Oakhill (1992) argue that although such inferences are necessary for a coherent interpretation of a text, they do not establish local links. Thus, to understand a text about Joe, who gets fired from a job because his friend Tom steals money from the till while Joe is in the storeroom, one must infer that Tom is likely to feel guilty. This inference does not link any two adjacent clauses of the text, however. Gernsbacher and Robertson's findings are not necessarily inconsistent with minimalism. Inferences about emotional states could be based on readily available knowledge, but the mundane nature of the story does not guarantee that the information will be readily available. To explain well established findings about instrumental inferences, M&R are forced to conclude that the information that coffee is usually stirred with a spoon is ordinarily not readily available. Unfortunately for M&R, this type of knowledge often underlies inferences that establish local coherence, even when it is not explicit in the text, for example, in:
Mary stirred her coffee. The spoon was dirty and tainted it.
2.3. The first of M&R's two hedges is that their theory describes only inferences that are made "automatically" as a text is read. This hedge generalizes the widely accepted idea that elaborative inferences can be made subsequent to the initial reading of a text. According to M&R, nonminimal inferences may also be made during reading if readers engage in nonautomatic strategic processing, either for reasons of their own, or in following experimental instructions. These inferences, however, should be distinguishable from minimal inferences using the normal criteria for differentiating automatic and strategic processes. Such nonautomatic inferences are sometimes needed to establish local coherence (M&R: p. 441), and M&R strongly imply, though they do not state, that such inferences are made. Furthermore, their Experiment 2 suggests that such inferences resolve local "inconsistencies" in texts, such as buying yoghurt and grapefruit as an alternative to riding a bike. Thus, if local coherence is "automatically" established (M&R: Abstract), a different sense of "automatic" (roughly, "as a matter of course") must be intended, rather than the one in which automatic contrasts with strategic. Indeed, despite their claims, the notion of automaticity does no real work for M&R. The crucial determinants of inference making are, on the one hand, the establishment of local coherence and, on the other hand, ready availability of knowledge, which can underlie elaboration. Furthermore, for inferences that establish local coherence, M&R fail to consider the possibility that the relevant knowledge is not only not readily available, but not available at all.
2.4. M&R's second hedge: An inference that would otherwise be nonminimal can be made if it depends only on "information that is quickly and easily available" (1992: p. 440). This is not a new idea, but it is problematic. In some instances, independent considerations may suggest what knowledge is readily available to a particular reader. However, if such independent evidence is not available, it is all too easy to "determine" whether knowledge is available by discovering whether it permits an inference to be made automatically. Thus, this second hedge reduces the testability of M&R's theory.
2.5. M&R could justifiably argue that the aspects of their theory that make it less testable are those that make it more realistic. The idea that inference making is (at least partly) a strategically controlled process and depends on what knowledge is readily available, which may vary from person to person, helps make sense of conflicting results in the literature. What is more damaging to M&R's position is their incorrect characterization of constructionist processes and their failure to recognize that such processes may be implicated in inferences that, according to the theory, are minimal.
3.1. According to M&R, constructionist theories, particularly those based on mental or situational models, hold that "the mental representation of a text automatically specifies, in some complete way, the real-life situation described by the text" (p. 458). However, the notion of a "life-like" representation is too vague to be useful, and the notion of a "complete" representation is incoherent. A complete representation implies elaborative (nonminimalist) inferences, because texts do not describe situations completely. However, the idea that a representation of a situation can be both complete and finite (and, hence, a viable candidate for a mental entity) cannot be sustained. Situations can be described at increasingly fine levels of details and in increasingly bizarre and irrelevant ways. There is no end to the ways in which a representation can be elaborated, and no sense in which it is ever complete. Mental models of situations are intended to be incomplete, as is apparent from the formal semantic theories that have influenced our own version of mental models theory (Garnham, 1987; Johnson-Laird, 1983). In both discourse representation theory (Kamp, 1981) and situation semantics (Barwise & Perry, 1983), text representations are partial submodels of an overall model of the world -- submodels because they do not represent all of the world, only part of it, and partial because they contain only partial information about the bit of the world they do represent: in the first instance, those aspects that are explicitly mentioned.
3.2. In contrast to M&R, Bransford, Barclay and Franks (1972; henceforth BB&F), with whom M&R concede that the constructionist position originates, identified the two crucial aspects of that position. The first is that the representation of the content of a (descriptive) text is a representation of the situation in the world that the text is about ("a wholistic description of the overall situation linguistically communicated", BB&F: p. 202). The second is that, in constructing this representation, information that is explicit in the text (almost always) has to be combined with relevant knowledge about the world from long-term memory, either specific knowledge about particular people, places and things, or general knowledge about kinds of people, places and things ("one is assumed to use linguistic information in conjunction with previous knowledge to construct semantic descriptions" BB&F: p. 207).
3.3. The first aspect of the constructionist position originally distinguished it from the view that sentence comprehension results in mental representations of the linguistic form of sentences. Constructionism claims that the elements of text representations stand for things in the world (people, objects, places and so on), not for linguistic entities. "Wholistic" representations are created by integrative processes that link information from different parts of the text. More specifically, information from the clause currently being read must be incorporated into the model of the text built so far. The information initially extracted from a clause is what is explicit in that clause. When this information is integrated into the developing mental model of the text, local links are usually established (but see below). Because making local links is a constructive process, it may add to the representation information that is not explicit in the text. M&R claim (p. 442) that "a constructed representation of the situation described by a text would not necessarily include aspects of the situation that were mentioned in close proximity." On one reading, this claim is inconsistent with M&R's idea that constructed representations are complete. More important, it misses the point that links between adjacent clauses are a primary factor in determining whether a text is well-written (see e.g. Garnham, Oakhill & Johnson-Laird, 1982).
3.4. The focus, in the first instance, on information that is explicitly presented, is a crucial part of the mental-models theory. It is also important in the mental-models account of reasoning. It forms the basis of an explanation of many biases and errors in reasoning that arise because people focus on what is stated in the premises and fail to consider possibilities that are not explicitly mentioned.
4.1. Both the original discussion by BB&F, and many subsequent discussions, exemplify constructive processes by elaborative inference making. Thus, inferences about the relative positions of the turtles, fish and log in the famous BB&F sentences:
Three turtles rested on/beside a floating log and a fish swam beneath it/them.
are not necessary to establish local coherence, and they are not necessarily based on readily available knowledge. (The information that would have to be readily available, however, is not "that the fish swam under the log" [M&R: p. 441]. M&R confuse the inferred information with the information on which the inference is based, in this case the explicit information in the sentence, together with general knowledge about spatial relations and about the relative sizes of turtles, fish, and logs.) Inferences to default case fillers, or highly probable ones, such as instruments (e.g. knives for cutting steaks, Paris & Lindauer, 1976), which are paradigm cases of constructive inferences, are not usually necessary to establish coherence, either.
4.2 Nevertheless, many of the most straightforward inferences that do establish local coherence are constructive. For example, the most famous bridging inference in the psycholinguistic literature is one in which the background knowledge that picnic supplies in California often include beer licenses involves an otherwise unexpectedly definite reference to beer ("The beer was warm.") following a reference to picnic supplies ("We checked the picnic supplies.") (Haviland & Clark, 1974). Similarly, in establishing the referent of the pronoun "he" in:
John blamed Bill because he spilled the milk.
it is necessary to use the information that spilling milk is usually undesirable, and knowledge about the circumstances under which one person blames another. Whether this information is used during reading is, of course, an empirical question. But if it is not, M&R's claims about local coherence are incorrect. In any case, inferences that are minimal according to M&R's definition, depend on constructive processing. Hence their contrast between a minimalist and a constructionist theory is an untenable one. Ironically, M&R discuss both bridging inferences and anaphoric processing, but the bridging inference they present (1992: p. 444) is a much more difficult one that may well not be made; and their discussion of anaphoric processing (1992: pp. 443-444) fails to consider either the different kinds of information that may be needed to resolve an anaphor or where that information comes from (text or long-term memory). M&R's presentation suggests that most local links are established automatically (i.e. nonstrategically). However, more careful consideration suggests that almost all local links depend on background knowledge that is not readily available. If the information that spoons are used to stir coffee is not readily available, surely the information that picnic supplies can contain beer is not.
5.1. The need for constructive processes in making "minimal" inferences emphasizes the complexity of inference making. In particular, like most previous authors, M&R fail to distinguish between a textual signal that an inference is needed to make a (local) link, and the use of background knowledge (and hence constructive processes) to make the link. Garnham and Oakhill (1992) pointed out this distinction in discussing the findings of Noordman and Vonk (1992). In one experiment, Noordman and Vonk included sentences with two clauses linked by the subordinating conjunction "because" in short texts. This conjunction indicates a causal, inferential or reason-based relation between the information in the two clauses. Recognizing the existence of this relation is a part of establishing the local coherence of the text, but one part only. To form a coherent representation it is also necessary to determine the particular connection between the two pieces of information. For example, in one of Noordman and Vonk's texts the crucial sentence was:
Chlorine compounds are frequently used as propellants, because they do not react with other substances.
The link between the two clauses depends on knowledge about propellants and why they ought to be inert. Noordman and Vonk's texts were chosen so that ordinary subjects were unlikely to have the background knowledge that lay behind the inferences. By comparing texts in which this knowledge was made explicit before the crucial sentence with those in which it was not, Noordman and Vonk showed that their subjects did not carry out inferential work as they read that sentence. However, if they were later asked to verify the information underlying the inference, it was at that point that they made the inference. These findings falsify the one unequivocal prediction from M&R's minimalist theory: that inferences necessary to establish local coherence are made as a matter of course. The way this prediction is falsified further emphasizes the contribution of constructive processes to minimal inferences.
5.2. In Noordman and Vonk's passages the information underlying the inferences can in principle be recovered during reading, on the assumption that the texts are both coherent and true. If chlorine compounds are used as propellants because they do not react with other compounds, inertness must be a desirable property of propellants. However, although this information is potentially available to readers, they do not appear to derive it. In a subsequent experiment, Noordman and Vonk used an expert-novice paradigm (and a different set of passages, about economic matters). They found that experts, for whom the knowledge underlying the inferences should have been readily available, did carry out inferential work as they read sentences with linguistically signaled links. This finding is consistent with M&R's idea that inferences based on readily available information will be made. Nevertheless, Noordman and Vonk's findings considerably complicate a theory of inference making. Competent readers certainly recognize linguistic signals of how the parts of a text are linked. They know from knowledge of their language that "because" signals causal, inferential and reason-based relations, "but" contrastive relations, and the use of a definite noun phrase that its referent should be given information. But recognizing that there is a link is not the same as knowing what underlies it. Fleshing out such links typically requires the use of background knowledge and, hence, constructive processes.
6.1. Mental-models theorists must be constructionists, but they may also be reconstructed minimalists. Indeed, they might hold a stronger view than M&R: that inferences are only made if they are both necessary to establish local coherence and based on readily available information. If the information is not readily available, the inference is not made. Nevertheless, many mental models theorists are nonminimalists, and one reason they appear to favor on-line elaboration is because of their focus on spatial inference. The spatial domain is ideal for distinguishing between a representation of what is described (a spatial layout of some kind) and the language used to describe it. However, this feature of the spatial domain further complicates an analysis of inference making. M&R argue that some of the results suggesting automatic spatial inferences have alternative explanations. Nevertheless, the results concerning one-dimensional spatial arrays (e.g. Potts, 1974) are firmly established, so a few comments on spatial inference are in order.
6.2. Since a representation of a spatial array is far removed structurally from the sentences used to describe the array, it is not obvious what inferences are encoded in it. And although M&R, with their notion of partially encoded inferences (see p. 458 for discussion), are among the few authors to have broken away from an oversimplistic view of inference making, they still characterize an inference as an encoded proposition (p. 463). In the representation of a spatial array, however, propositions that might be used to describe the array are not directly represented. To express something that might be characterized as "an inference encoded in the array" it is necessary to generate a sentence (expressing a proposition) from the array, but this process may itself contribute to the making of the inference. Such complications are partly what I had in mind when I argued (Garnham, 1989: pp. 168-169) that inferences might be thought of as implicitly encoded into representations of spatial arrays, and that processes at encoding, storage and retrieval might all contribute to a single inference. The notion of "making an inference" suggests that there is some moment at which an inference is made, but this suggestion is misleading. Furthermore, if there is no moment at which an inference is made, the question of whether it is made automatically needs to be reformulated.
6.3. Given the complexity of spatial inferences, it is not surprising that M&R's analysis applies more directly to simple cases, such as inferences about case fillers, where the difference between information encoded into a mental model and a linguistic expression that can be used to express that information is less apparent. Nevertheless, even "simple" case filling inferences may depend on processes with complex temporal properties. For example, in Garnham (1982) I presented an "omission theory" of case filling inferences in which highly probable or default case fillers are omitted from a representation of content both when they are implicit and when they are explicit. In either case, they can be reconstructed from an adequate representation of the content of a text, if required. In the omission theory, unlike in the theory that highly plausible case fillers are immediately inferred, there is no single moment at which an inference is made. These theories are difficult to distinguish, however, because they make identical predictions in many cases (Garnham, 1982). In the omission theory, the question about what inferences are made "during reading" is beside the point.
6.4. I have argued that minimalist and constructionist theories of inference making are not only compatible with one another, but that a viable theory of inference making will definitely be constructionist and is likely to be approximately minimalist. Indeed, I have previously argued (Garnham, 1982, 1989) for an account of inference making within the mental-models framework (said by M&R to be constructionist and hence antiminimalist), one that might be described as a modified minimalist theory. The idea that a theory of inference making should be an amalgam of minimalist and constructionist ideas is misleading, however. In their version of minimalism, M&R draw on notions such as automatic versus strategic processing. These notions describe properties of the processes that underlie inference making. They are quite inappropriate for analyzing what it is to make an inference. Yet it is just such analytic questions, both in the domain of inference making and in other domains such as anaphoric processing, that the mental-models theory is intended to answer. Indeed, the older notion of minimalism, defined in terms of a distinction between inferences necessary to establish a coherent interpretation of a text and those that are merely elaborative, is part of an analytic description of inference.
6.5 In Marr's (1982) terminology, the mental-models theory would aim to provide a computational theory of the making of inferences: a theory of what constitutes inference making and why people make the inferences they do. Notions such as automatic versus strategic processing are appropriate for describing mechanisms at the abstract level of algorithm and representation. As for implementation (Marr's third level), we know virtually nothing about the neural mechanisms that underlie inference making. It is true that many mental models theorists have emphasized not just constructive processing, which is an essential part of text comprehension, but on-line elaborative inference making. However, on-line elaboration is not an essential part of a mental-models theory of text comprehension (as M&R themselves admit at one point, p. 458). The essential aspects of such a theory are (1) the idea that the representation of the content of a text is a representation of the part of the world that the text is about, not of the linguistic expressions used to describe it, (2) a conceptually coherent account of the processes that establish the reference of linguistic expressions, and (3) an emphasis on the incremental nature of the construction of text representations.
Baker, L. (1989). Metacognition, comprehension monitoring, and the adult reader. Educational Psychology Review, 1, 3-38.
Barwise, J. & Perry, J. (1983). Situations and attitudes. Cambridge, MA: MIT Press/Bradford Books.
Bransford, J.D., Barclay, J.R. & Franks, J.J. (1972). Sentence memory: A constructive versus interpretive approach. Cognitive Psychology, 3, 193-209.
Corbett, A.T. & Dosher, B.A. (1978). Instrument inferences in sentence encoding. Journal of Verbal Learning and Verbal Behavior, 17, 479-491
Dosher, B.A. & Corbett, A.T. (1982). Instrument inferences and verb schemata. Memory and Cognition, 10, 531-539.
Garnham, A. (1982). Testing psychological theories about inference making. Memory and Cognition, 10, 341-349.
Garnham, A. (1987). Mental models as representations of discourse and text. Chichester: Ellis Horwood.
Garnham, A. (1989). Inference in language understanding: What, when, why and how. In R. Dietrich & C.F. Graumann (Eds.), Language processing in social context. (pp.153-172). Amsterdam: North Holland.
Garnham, A & Oakhill, J.V. (1992). Discourse representation and text processing from a "mental models" perspective. Language and Cognitive Processes, 7, 193-204.
Garnham, A., Oakhill, J.V., & Johnson-Laird, P.N. (1982). Referential continuity and the coherence of discourse. Cognition, 11, 29-46.
Gernsbacher, M.A. & Robertson, R.R.W. (1992). Knowledge activation versus sentence mapping when representing fictional characters' emotional states. Language and Cognitive Processes, 7, 353-371.
Haviland, S.E & Clark, H.H. (1974). What's new: Acquiring new information as a process in comprehension. Journal of Verbal Learning and Verbal Behavior, 13, 512-521.
Johnson-Laird, P.N. (1983). Mental models: Towards a cognitive science of language, inference, and consciousness. Cambridge: Cambridge University Press.
Kamp, H. (1981). A theory of truth and semantic representation. In J. Groenendijk, T. Janssen & M. Stokof (Eds.), Formal methods in the study of language (pp. 255-278). Amsterdam: Mathematical Centre Tracts.
Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. San Francisco: Freeman
McKoon, G. & Ratcliff, R. (1992). Inference during reading. Psychological Review, 99, 440-466.
Noordman, L.G.M. & Vonk, W. (1992). Reader's knowledge and the control of inference in reading. Language and Cognitive Processes, 7, 373-391.
Paris, S.G. & Lindauer, B,K. (1976). The role of inferences in children's comprehension and memory for sentences. Cognitive Psychology, 8, 217-227.
Potts, G.R. (1974). Storing and retrieving information about ordered relationships. Journal of Experimental Psychology, 103, 431-439.
Singer, M. (1980). The role of case-filling inferences in the comprehension of brief passages. Discourse Processes, 3, 185-201.
Thorndyke, P.W. (1976). The role of inferences in discourse comprehension. Journal of Verbal Learning and Verbal Behavior, 15, 437-446.
AUTHOR NOTE: My work on mental models has been supported by ESRC grant C 0023 2439 "Mental models and the interpretation of anaphora". Thanks to Jane Oakhill for comments on an earlier draft of this paper.