Rolf A. Zwaan and Arthur C. Graesser (1993) There is no Empirical Evidence That Some Inferences are Automatically or Partially Encoded in Text Comprehension. Psycoloquy: 4(04) Reading Inference (5)

Volume: 4 (next, prev) Issue: 04 (next, prev) Article: 5 (next prev first) Alternate versions: ASCII Summary
PSYCOLOQUY (ISSN 1055-0143) is sponsored by the American Psychological Association (APA).
Psycoloquy 4(04): There is no Empirical Evidence That Some Inferences are Automatically or Partially Encoded in Text Comprehension

Commentary on Garnham on Reading-Inference

Rolf A. Zwaan and Arthur C. Graesser
Department of Psychology
Memphis State University
Memphis TN 38111


In his defense of the constructivist view of language comprehension, Garnham (1992) questions the notion of automatic inferences. We argue that Garnham's arguments could be strengthened by taking a more theoretical perspective on automaticity.


constructionism, inference, mental models, minimalism, reading, text comprehension.


1.1 In his defense of the constructivist view of language comprehension, Garnham (1992) questions the notion of automatic inferences that is so crucial to the minimalist position proposed by McKoon & Ratcliff (1992) (M&R). Garnham provides examples to show that the inferences that M&R assume to be automatically generated may in fact not be automatically generated. In this commentary we argue that Garnham's arguments could be strengthened by taking a more theoretical perspective on automaticity.

1.2 Garnham agrees with M&R on the idea that inference-generation is not an all-or-none process, i.e., inferences are generated at various degrees of strength. We agree that this view may be fruitful and that it may resolve a large number of contradictions in previous research. However, we think it would be unwise to throw out the baby with the bath water: The notion of a continuum in inference-generation should not lead to dispensing with online research into inference generation altogether. We will provide theoretical and methodological arguments to support our position.



2.1 M&R distinguish between inferences that are made automatically and inferences that are not. The automatic inferences are allegedly the only ones that are consistently generated. However, the only criterion M&R use to decide whether an inference is generated automatically is the speed with which the knowledge necessary for the inference becomes available. This is a rather loose criterion for automaticity.

2.2 Speed is not the primary property of automatic processing mentioned in the literature that distinguishes automatic processes from controlled processes (Schneider & Shiffrin 1977; Shiffrin & Dumais 1981; Shiffrin & Schneider 1977). Shiffrin and Dumais (1981) postulated two basic criteria for determining whether a process is automatic or not. First, an automatic process does not use general processing resources or decrease general processing capacity. Second, an automatic process always utilizes general resources and decreases general processing capacity whenever a given set of external stimuli is presented, regardless of a subject's attempt to ignore or bypass the distraction.

2.3 It is unclear whether M&R's automatic inferences satisfy either of these criteria. However, one can design experiments to test whether or not an inference is made automatically. For example, one could have subjects read a text and simultaneously perform a secondary task (e.g., respond to click sounds which occur at instances when the subjects are expected to make an inference). Automatic inference-generation should not be impaired by this secondary task because (a) it does not require working-memory capacity, and (b) once initiated, it cannot be terminated. Unfortunately, M&R never performed these independent tests of whether their inferences were generated automatically. Hence it may be premature to conclude anything about the status of their inferences.

2.4 More should be known about automatic inference generation before the minimalist arguments against the constructionist view can be upheld. For all we know, it may be that no inference is made automatically if a strict criterion for automaticity is used. This issue is particularly important for the minimalist position because of M&R's claim that a minimalist representation forms the database for further processing. However, we do not think the search for automatic inferences should in general be high on the agenda in text comprehension research. In our view, research on inference generation should take comprehension as a threshold for whether or not an inference is generated: Does readers comprehend a text if they haven't made a particular inference? Because text comprehension is strategic, it is relevant to focus on which inferences are generated under which conditions (see our commentary on the Glenberg & Mathew (1992) target article: Zwaan & Graesser 1993).


3.1 Both Garnham and M&R endorse the view that inferences might not be generated in an all-or-none fashion. Instead, they may be partially encoded first and then staggered over time. According to Garnham (paragraph 6.2), this implies that there is no point at which an inference is made -- and this, in his view, renders on-line research into inference-generation superfluous. We think there are serious theoretical and empirical problems with these views.

3.2 The assumption that inferences may be partially encoded does not eliminate the on-line issue: There is still a location in the text at which the inference was first partially encoded. It is important to identify this location and to determine the extent to which the inference was encoded there. In addition, it is important to identify spots further down the line at which the inference was encoded more strongly. It should be noted that even a partially encoded inference may influence further processing. Without information on the precise locus of the initial partial encoding one may not be able to explain aspects of subsequent processing.

3.3 The assumption that there is no point at which inferences are generated makes it difficult to analyze individual differences in processing. For example, it has been demonstrated that domain knowledge affects situation model construction (Tardieu, Ehrlich & Gyselinck 1992). Experts presumably encode particular inferences -- either partially or completely -- earlier in the text than do novices. Novices may need more clues than experts to make some inferences. This issue can be examined properly only by means of a careful analysis of the online evolution of an inference.

3.4 There are methodological problems involving the interpretation of data that support the distributed view on inference generation. M&R conclude that some inferences may be more strongly or completely encoded than others from the fact that different inference words have different error rates and judgment latencies on word recognition tasks. However, these data are aggregated over subjects. Hence they do not provide information about individual subjects. There is accordingly an alternative account for M&R's Ratcliff's (1990) findings: All inferences were encoded completely but some inferences were encoded completely by more subjects than others. This could arise from individual differences among subjects; for example, some inferences were encoded by almost all readers, whereas others were encoded by only the good readers.

3.5 In addition, M&R's inference words may differ in baseline memorability. Differences between words in error rates or response latencies may arise from these baseline variations. Baseline variations therefore have to be taken into account before the partial encoding hypothesis can be properly evaluated.

3.6 To summarize: Whether or not inferences are partially encoded can only be determined on the basis of online within-subject data and with the use of previously normed inference words. The distributed view on inference generation may be theoretically appealing. However, it needs empirical support before any conclusions about its usefulness can be drawn.


Garnham, A. (1992) Minimalism versus constructionism: A false dichotomy in theories of inference during reading. PSYCOLOQUY 3(63) reading-inference-1.1

Glenberg, A. M & Mathew, S. (1992) When minimalism is not enough: Mental models in reading. PSYCOLOQUY 3(64) reading-inference-2.1

McKoon, G & Ratcliff, R. (1992) Inferences during reading. Psychological Review, 99, 440-466.

Schneider, W & Shiffrin, R.M. (1977) Controlled and automatic human information processing: I. Detection, search, and attention. Psychological Review, 84, 1-66.

Shiffrin, R.M. & Dumais, S.T. (1981) The development of automatism. In: J.R. Anderson (Ed.) Cognitive skills and their acquisition. Hillsdale, NJ: Erlbaum.

Shiffrin, R.M & Schneider, W. (1977) Controlled and automatic human information processing: II. Perceptual learning, automatic attending, and a general theory. Psychological Review, 84, 127-190.

Tardieu, H., Ehrlich, M.-F. & Gyselinck, V. (1992) Levels of representation and domain-specific knowledge in comprehension of scientific text. Language and Cognitive Processes, 7, 335-351.

Zwaan, R.A. & Graesser, A.C. (1993) Reading goals and situation models. PSYCOLOQUY 4(3) reading-inference.4

Volume: 4 (next, prev) Issue: 04 (next, prev) Article: 5 (next prev first) Alternate versions: ASCII Summary