The Effect of Negation on Deductive Inferences

Journal of Experimental Psychology: Learning, Memory, and Cognition 2002, Vol. 28, No. 2, 303–317

Copyright 2002 by the American Psychological Association, Inc. 0278-7393/02/$5.00 DOI: 10.1037//0278-7393.28.2.303

The Effect of Negation on Deductive Inferences R. Brooke Lea and Elizabeth J. Mulligan Macalester College Research shows that negation can suppress the activation of propositions presented explicitly in text, but does negation have a similar effect on propositions that can be inferred? That is, does negation inhibit the inference process? Four experiments investigated whether a deductive inference that produces a negated conclusion (therefore not a) is made as readily as a similar inference form that yields an affirmative conclusion (therefore a). A combination of naming latencies, verification times, and reading times indicate that negation does not affect the deductive inference process itself, although it may inhibit the activation of inferred concepts.

Imagine you are having lunch with a friend who tells you that because he is on a diet he will not order both an appetizer and an entre´e. After studying the menu, he decides to have an entre´e. It follows logically that he will not order an appetizer, but do you make this inference? Do you bother to create knowledge about what will not happen? Put a different way, how does negation affect the inference process? Psycholinguistic research has shown that negation can affect various levels of comprehension. Early work (e.g., Clark & Chase, 1972; Gough, 1966; Just & Carpenter, 1971; Trabasso, Rollins, & Shaughnessy, 1971; Wason, 1959, 1963) found that high-level processes were affected by the presence of a negative. For example, fragment completion time and verification latencies were slowed when the sentences participants were completing or verifying contained a negation. Others found comparable effects on the implications of sentences that contained inherently negative words, such as forget (Just & Clark, 1973). The general consensus from much of this work is that extra comprehension processing is required to accommodate the logical entailments of negation. More recently, some theoretical accounts of sentence comprehension include a mechanism that automatically enhances or suppresses the activation of a concept in the reader’s mental representation of a text, and negation has been suggested as one discourse factor that can induce suppression (Gernsbacher & Jescheniak, 1995). Specifically, negated propositions become less active, and therefore they are more difficult to verify as having appeared in the text. In a series of experiments, MacDonald and Just (1989) found that participants were slower and less accurate at verifying that the word bread had appeared in the sentence

(1)

Almost every weekend, Elizabeth bakes no bread but only cookies for the children

compared with a control sentence in which cookies was negated instead of bread: (2)

Almost every weekend, Elizabeth bakes bread but no cookies for the children.

This activation-reduction effect held up in a naming task, although it did not for associates of the negated proposition (e.g., butter). MacDonald and Just also examined reading and verification response times and found that negation produced a progression of effects that appeared to be determined by the level of processing demanded by the task. For example, reading time was not affected by negation, although statement verification times were. MacDonald and Just thus concluded that (a) negation does not affect initial encoding (reading times); (b) negation does produce discourse-level effects on the negated concept (probe times), although not on associated concepts, nor does negation result in any overall processing load effects at the discourse level; and as established in the earlier literature, (c) negation does produce large effects on high-level processes that compute truth value (statement verification times; MacDonald & Just, 1989, p. 642). MacDonald and Just accounted for their probe-time effects by proposing that negation affects activation levels by way of shifts in discourse focus. That is, in much the way that pronominal reference to one name (Gernsbacher, 1989; MacDonald & MacWhinney, 1990) or a nominal reference to one object (Wiley, Mason, & Myers, 2001) has been found to dampen the activation of a nonreferent name, negation also may shift discourse focus from the negated item to one that has not been negated. Thus, if you hear that Mary bakes bread but no cookies, then focus shifts away from cookies. MacDonald and Just (p. 641) proposed that these sorts of changes in focus are realized by changes in activation levels in the discourse representation. But what if you heard that Mary doesn’t have time to bake both bread and cookies, and then you hear that she decides to bake bread? Will you infer that she will not bake cookies? The present research is concerned with the effect that negation might have on the production of these sorts of deductive inferences during reading. Specifically, we examined two logical forms that accomplish focus shifts by means of inference.

R. Brooke Lea and Elizabeth J. Mulligan, Department of Psychology, Macalester College. Elizabeth J. Mulligan is now at the Department of Psychology, University of Colorado at Boulder. We thank Jerome Myers, Chuck Clifton, Wietske Vonk, Keith Rayner, Ira Noveck, Barry Cohen, and Akiva Lieberman for helpful conversations and suggestions regarding this research, and Suzanne Lovett, Al Fuchs, and two anonymous reviewers for trenchant commentary on a draft of this article. We also thank Jennifer Lee, Susan Long, Hilary Mead, and Hannah Jackson for assistance with data collection. Correspondence concerning this article should be addressed to R. Brooke Lea, Department of Psychology, Macalester College, 1600 Grand Avenue, Saint Paul, Minnesota 55105. E-mail: [email protected] 303

LEA AND MULLIGAN

304 Deductive Inference Background

According to the mental logic (M–L) account of logical competence, people are able to draw certain propositional logic inferences easily and accurately (Braine, 1978; Braine & O’Brien, 1998; Rips, 1983, 1994). (Propositional logic inferences are those that depend on sentence particles such as if, and, not, and or.) This claim has held for performance on elementary logic problems, and recently it has been extended to inferences drawn during discourse comprehension (Lea, 1995, 1998; Lea, Lee, & Mulligan, 2001; Lea, O’Brien, Fisch, Noveck, & Braine, 1990). For example, Lea (1995) presented participants with passages like the following: Sniffing Out the Goods (1) (2)

(3) (4) (5)

Special Investigator Evans was trying to crack a Cuban cigar smuggling ring. As he entered the dark warehouse he whispered into his walkietalkie: “If the dog starts to bark then we know there’s tobacco around.” “Okay, I’ll be listening,” replied his partner Newstead, who was monitoring the situation from their unmarked patrol car. There was quiet for a few minutes, then the silence was broken by the sound of Evans’s dog barking. * SMOKE *

Participants read the passages at their own pace and were asked to respond to a semantic associate of an inference that followed from the passage. In the example above, the conditional premise in Line 4, If dog barks then tobacco, can be combined with the information in Line 6, dog barks, to produce the inference that there is tobacco around. Participants were timed making lexical decisions or naming responses to the associate probe word, smoke, and their responses following 28 such passages were reliably faster compared with a control condition in which the inference could not be made. Lea (1995) tested two inference forms from Braine and O’Brien’s M–L model: modus ponens (if p then q; p, therefore q), as in the “Sniffing Out the Goods” example above, and or elimination, ( p or q; not p, therefore q), what we commonly think of as process of elimination. Lea (1995) found that readers reliably made these deductions at the moment both premises were simultaneously available to them. Earlier work had shown that readers so easily made the logical inferences specified in the M–L model while reading that they often falsely recognized their deductions as information that had been presented explicitly in the text (Lea et al., 1990). Other studies using abstract stimuli (e.g., ps and qs on an imaginary blackboard or apples and pears that a Kermit puppet can see in a box) have shown that children and adults are quick and accurate at making the propositional logic inferences described by the M–L model (Braine et al., 1995; Braine & Rumain, 1983). Hence, the evidence is quite good that despite our reputation as flawed reasoners, human adults and children are remarkably competent at integrating information from the environment in a logical fashion. We know relatively less, however, about the effects of discourse factors on deductive inferences (Noveck, Lea, Davidson, & O’Brien, 1991). At present, the M–L model’s predictions for performance in discourse are essentially the same as those for abstract content. There are several discourse-processing situations, however, in which the M–L model’s predictions diverge from expectations derived from the comprehension literature. For ex-

ample, according to the M–L model, the following two inferences should be equivalently easy for participants to make: Or elimination

Not-both elimination

a or b not a

not both a and b a

therefore b

therefore not b

Braine, Reiser, and Rumain (1984) produced difficulty ratings for all the inferences predicted by the M–L model on the basis of participants’ performance on abstract problems, and these two inference schemas yielded very similar scores: 1.38 for or elimination, and 1.39 for not-both elimination (henceforth referred to as or and not both respectively). Thus, the M–L model predicts—and these data indicate—that when dealing with abstract information, participants make these two inferences equally readily. But do people make them just as readily while reading text? Note that a critical difference between or and not both is that one produces an affirmative conclusion (therefore b), whereas the other produces a negated conclusion (therefore not b). The question therefore arises as to whether the effect that negation has on propositions goes beyond explicit negation; that is, does it inhibit or suppress the production of negative information? If it does, then we might expect participants to more readily draw the or inference “therefore b,” than the not-both inference, “therefore not b.” This outcome would be in the general spirit of minimalist accounts of inference during reading; given that readers do make logical inferences in text, they would be more likely to make deductions about what did or will happen in a passage, for example, than to pursue information about what did not or will not happen. Note that the present approach differs from research that has examined content effects in logical deduction. For example, countless studies have found that specific content facilitates performance on the Wason selection task (Wason, 1966) a deductive reasoning task that in its abstract form is notoriously difficult (e.g., Cheng & Holyoak, 1985; Cosmides, 1989; Griggs & Cox, 1982; Noveck & O’Brien, 1996). Three principal differences distinguish that work with the present effort: (a) The logical inferences under consideration here are drawn easily and accurately when presented in abstract form, that is, real-life content is not needed to improve performance; (b) The content-effect studies on the Wason selection task assume that participants arrive at the correct answer through nonlogical processes, whereas our work is based on a model that is based fundamentally on logical processes; and (c) Real-life content is present in our stimuli because we wish our participants to be reading for comprehension, not solving puzzles. Finally, it is important to note that M–L theories acknowledge that several inference forms, such as modus tollens, are indeed difficult for people not trained in logic. Thus, for example, Braine et al.’s (1995) theory includes not only a direct reasoning routine for inferences that putatively are part of a universal mental logic but also an indirect routine that includes reasoning strategies available only to those who have specific experience with them. The present research is concerned with two inference forms from the M–L model’s direct-reasoning routine. With Experiment 1 we conducted a simple test to see if, in fact, we could produce evidence of the not-both inference during reading. In Experiment 2 we replicated the results of Experiment 1 and added versions of the passages based on the or inference form.

THE EFFECT OF NEGATION

Experiment 3 began an investigation into the role that negation might play in the activation of negated concepts that have been inferred. Experiment 4 ruled out an alternative explanation of the results of Experiment 3 and replicated the main results of Experiment 3.

Experiment 1 We used a reading-time design to investigate whether readers are making the not-both inference. Participants read passages in which certain target sentences would make sense if the inference in question was not made but would appear contradictory if the inference was made. This technique has been exploited to investigate a variety of text-processing phenomena (e.g., Albrecht & Myers, 1995; Lea, Mason, Albrecht, Birch, & Myers, 1998; Myers, O’Brien, Albrecht, & Mason, 1994). For example, in a passage about two brothers making red-hot Texas chili (see Appendix A), the recipe states that they could not use both fresh jalapenos and ground pepper in the chili. When they decide to use fresh jalapenos, the reader can infer that they will not use ground pepper. The target sentence reads, “I’ll get the ground pepper,” said Joseph, which is inconsistent with the inference. The principal difference between the inference version and the no-inference control version of the passage is the omission of the word not in the second sentence (see Appendix A, Story Version 2). Note that the local coherence of the target sentence is identical (and reasonably good) in both the inference and control versions. The target sentence is inconsistent only when the not-both inference is made. Thus, if readers are making the not-both inference, their reading times should be slower on the target sentence compared with the no-inference control.

Method Participants. Thirty-three undergraduate students participated. Most were enrolled in an introductory psychology class and participated to fulfill part of a course requirement; 2 participants were psychology students who volunteered. All participants were native English speakers. Materials. Twelve experimental passages were composed with two versions of each such that 12 passages permitted logical inferences and 12 did not. The no-inference passages differed from their inference pair in Line 2 only, which in the inference version provided a logical premise of the form not both a and b (e.g., According to the recipe, they could not use both fresh jalapenos and ground pepper in the chili.) and in the noinference control presented a neutral sentence that could not be combined with the second logical premise to produce an inference (e.g., According to the recipe, for extra spicy chili they could use both fresh jalapenos and ground pepper.). In both versions Line 4 provided the second premise, a (e.g., “We’ve got plenty of fresh jalapenos—let’s use them!” he said.), and Line 5 stated the target sentence, b (e.g., “I’ll get the ground pepper,” said Joseph.). Note that the target sentence is inconsistent with the logical deduction not b that can be made in the inference version of the passages. If participants have drawn this inference from the passages, the target sentence will be anomalous; otherwise, it will make good sense in a story about people making spicy chili. Each participant read 42 passages. Of these passages, 12 were experimental passages (6 that permitted logical inferences and 6 no-inference controls) and 30 were filler passages. Each passage consisted of five lines of text followed by a one-line comprehension statement about the passage. The 30 filler passages were included to disguise the purpose of the experiment and were composed of 23 passages without any logical premises and 7 passages that contained valid not-both inferences with the not both a and b premise on Line 2, the a premise on Line 4, and the not b conclusion on Line 5. The 7 passages containing

305

inferences were included so that participants could not predict that every passage containing a phrase in the form of not both a and b would end with a sentence that did not make sense. Because only 6 of the 42 passages contained potentially inconsistent target sentences we were confident that participants would not detect a pattern in the stories. The comprehension statements were worded so that for all inference passages the statement referred to information presented in Line 4 (the second premise) and for all no-inference passages the statement referred to information presented in Line 2 (the first premise in the inference passages). Of the 7 filler passages that permitted logical inferences, 6 had comprehension statements that referred to the first premise. This was done to counterbalance the fact that comprehension statements for the 6 experimental inference passages read by each participant always referred to the second premise. For the 42 comprehension statements, the correct answer was true 21 times and false 21 times. For the experimental passages the correct answer was true exactly half of the time, as was the case for the control passages. Procedure. Each participant was seated in front of a computer screen in a room alone. The participant was asked to read a series of passages and respond to a comprehension statement following each passage. Participants were given both oral and written instructions explaining how to move through the passage by pressing the space bar after reading each line and how to respond to the comprehension statements by pressing either the ?/ key for yes, if the statement was consistent with what they had read, or the Z key for no, if the statement was not consistent with what they had read. These keys were labeled with yes and no stickers, respectively. Participants were instructed to keep their fingers on the keys at all times. Participants were then presented with two practice passages.

Results For all of the experiments reported, F1 refers to tests against an error term on the basis of subject variability and F2 refers to tests against an error term on the basis of item variability. Results reported as significant had associated p values of less than .05. Outliers (response times that were more than three standard deviations from the subject means) were discarded, resulting in a loss of less than .05 of the data in each experiment. Fourth-sentence and target-sentence reading times appear in Table 1. Participants read the target sentences 774 ms slower, on average, in the notboth inference versions of the passages than in the no-inference control versions. This difference was significant by subjects and items, F1(1, 31) ⫽ 19.60, MSE ⫽ 534,019.47; F2(1, 10) ⫽ 6.18, MSE ⫽ 370,757.32. The reading slowdown indicates that the participants were making the not-both inference while reading the passages. We inspected fourth-sentence reading times to make sure that the target-sentence effect was not due to processing spillover; if the reading times for the sentence preceding the target sentence was consistently slower in the inference condition, then the parallel slowdown in the target sentence might be due to a processing spillover and not to a contradiction effect, as we would like to claim. However, fourth-sentence reading times were fairly similar

Table 1 Reading Times (in ms) for the Fourth Sentence and Target Sentence in Experiment 1 Condition

Fourth sentence

Target sentence

No inference Not-both inference Difference

2,409 2,251 158

2,826 3,600 ⫺774

306

LEA AND MULLIGAN

and, if anything, faster (by 158 ms on average) in the inference condition, though this difference did not reach significance, F1(1, 31) ⫽ 3.31, MSE ⫽ 106,549.10, p ⫽ .08; F2(1, 10) ⫽ 2.42, MSE ⫽ 50,187.39, p ⫽ .15. Therefore, there is no evidence that the target-sentence slowdown was due to spillover. The latency and accuracy data for the comprehension questions were not analyzed for Experiments 1 and 2 because the contradiction in the inference version would create confusion in the reader’s mind as to the source of the contradiction. That is, it would be unclear what was true in the story and what was false.

Discussion The results of Experiment 1 are clear. Target-sentence reading times were more than 700 ms slower when they appeared in the inference versions of the passages compared with their almost identical no-inference controls. Although local coherence was no different between conditions, the target sentence would appear inconsistent with a discourse representation that included the notboth inference. This robust target-sentence reading-time difference supports the hypothesis that readers were making the not-both inference; pretarget reading times did not invite a spillover account of the data. A possible alternative explanation of the reading slowdown in Experiment 1 is that it reflects, in part, the making of the logical inference. By this account, readers do not make the not-both inference when the second premise is first available (at the fourth sentence), but do so only when they encounter the target sentence. Thus, in the inference version of the passage, participants read We’ve got plenty of fresh jalapenos—let’s use them! without integrating that information with can’t use both ground pepper and jalapenos stated two lines earlier. Instead, they read on, and when they encounter “I’ll get the ground pepper,” said Joseph they then consider how this information fits with what came earlier. In the no-inference version the fit is made easily thanks to the could use both ground pepper and jalapenos premise; in the inference condition the fit is more complex because it involves integrating Lines 2 and 4. Consequently, according to the late-integration account, the reading slowdown we found may have been due to extra processing prompted by the target sentence (in the inference version only) and not because readers were noticing a contradiction in the passage. This alternative explanation is tenable only if one is willing to assume that readers were, for some reason, systematically integrating at Line 5 but not at Line 4 over the 12 experimental passages. Both Line 4 and Line 5 are equally relevant to the first premise, therefore in the absence of some principled reason why readers would systematically choose to integrate after Line 5 but not after Line 4, it seems unlikely to us that the effects are due to inference making at Line 5 and not at Line 4. Furthermore, in writing these passages we were careful to choose propositions that pragmatically work well together; if it were odd that people spicing their red-hot Texas chili would put both fresh jalapenos and ground pepper in the pot, then it might sound strange to readers that Joseph is getting the ground pepper after Eric finds the fresh jalapenos, and the alternative explanation above would be more plausible. But we purposefully chose disjuncts with pragmatics that would permit both inclusive and exclusive use. Thus, there is nothing inherently odd about someone spicing chili with jalapenos and then spicing it

with ground pepper— unless one’s situation model includes information that prohibits it (i.e., the inference in question). In Experiment 2 we tried to replicate the contradiction effect of Experiment 1, and we added the or inference for comparison. The latter inference form has been tested in numerous studies with a variety of dependent measures and has consistently produced an inference effect (Lea, 1995; Lea, Charnley, & Mead, 2001; Lea, Lee, & Mulligan, 2001; Lea et al., 1990). Because the first four lines of the or passages used in Experiment 2 were essentially identical to those used in previous experiments in which probes were primed when they appeared immediately after Line 4, we were confident that our participants would make the inference at Line 4 and not at Line 5, as predicted by the late-integration explanation. Thus, Experiment 2 allowed us to observe reading behavior when we were certain that a logical inference was being made.

Experiment 2 Method Participants. Forty-five introductory psychology undergraduate students participated for course credit. All were native English speakers. Design, materials, and procedure. Experiment 2 consisted of a 2 (schema: or vs. not-both) ⫻ 2 (inference: inference version vs. noinference version) design. Both factors were within-subjects. We assembled four versions of each of the 12 experimental passages used in Experiment 1 and created four stimulus lists to which participants were randomly assigned. As in Experiment 1, each passage contained a target sentence that was locally coherent and consistent with both inference and no-inference versions of the texts unless the logical inference had been made. If the inference was made, then the target sentence would appear anomalous. A sample or passage appears in Appendix B. All other aspects of the materials and procedure were identical to that in Experiment 1.

Results Fourth-sentence and target-sentence reading times are presented in Table 2. Overall, the pattern of results was comparable with that in Experiment 1. Participants read the target sentences more than 700 ms slower, on average, in the inference versions of the passages than in the no-inference control versions. This difference was significant by subjects and by items, F1(1, 41) ⫽ 38.7, MSE ⫽ 639,727.75; F2(1, 8) ⫽ 37.49, MSE ⫽ 121,239.96. The reading slowdown implies that across the two inference-schema types participants were noticing the contradiction created by the logical inferences. Follow-up tests showed that the inference effect was

Table 2 Reading Times (in ms) for the Fourth Sentence and Target Sentence in Experiment 2 Condition

Fourth sentence

Target sentence

Not both No inference Inference Difference

1,973 1,981 ⫺8

2,400 3,118 ⫺718

Or No inference Inference Difference

2,005 2,068 ⫺63

2,279 3,005 ⫺726

THE EFFECT OF NEGATION

significant by subjects and items for the not-both inference, F1(1, 41) ⫽ 17.19, MSE ⫽ 656,670.46; F2(1, 8) ⫽ 21.49, MSE ⫽ 72,221.58, and for the or inference, F1(1, 41) ⫽ 36.36, MSE ⫽ 371,775.69; F2(1, 8) ⫽ 23.44, MSE ⫽ 133,556.37. Thus, there was a strong inference effect for both inference forms. Neither the main effect for inference type (not-both vs. or) was significant (both ps ⬎ .23) nor was the Inference Type ⫻ Inference Effect interaction significant, F1 ⬍ 1; F2(1, 8) ⫽ 1.622, MSE ⫽ 84,537.99, p ⫽ .24. Therefore, as the means indicate, both inference forms produced large and similar target-sentence reading slowdowns. Reading times for the fourth sentence were relatively uniform across the four conditions. (Comparisons between the inference types are not meaningful because the sentences were different.) Indeed, there were no significant main effects or interactions in the subjects analysis of fourth sentence reading times (all ps ⬎ .20). The items analysis also revealed no main effects, although the Inference Type ⫻ Inference Effect interaction was marginal, F2(1, 8) ⫽ 4.42, MSE ⫽ 34,029.58, p ⫽ .07. Overall, then, the fourthsentence reading times offer no support for a spillover explanation for the target-sentence effects.

Discussion Experiment 2 was designed to exploit our knowledge of the or inference to clarify the not-both inference. Every text-processing experiment involving or that we know of has produced strong inference effects. Therefore we are certain that Experiment 2 participants were making the or inference at Line 4 and no later. The effect was a 726-ms slowdown on a target sentence that contradicted the inference. Turning to the not-both inference we saw a very similar pattern. Readers in Experiment 1 and Experiment 2 slowed down 774 and 718 ms, respectively, when reading a target sentence that was inconsistent with the not-both inference. We found no evidence whatsoever that readers were waiting until the target sentence to integrate Lines 4 and 2 (i.e., that they made the inference) and then finding the contradiction. Indeed, if readers were doing that extra processing at the target sentence (keeping three sentences plus an inference in working memory simultaneously) we might expect a larger slowdown than we got (cf. Frazier & Rayner, 1982; Myers et al., 1994; Myers, Shinjo, & Duffy, 1987)—and certainly a larger slowdown than that produced by the or condition, in which we are certain the inference had been made before the target sentence. Given these results, we feel confident in concluding that the target sentence slowdowns in Experiments 1 and 2 were due to the difficulty readers had in integrating the sentence with their situation model of the text that already included the not-both inference.

Experiment 3 Thus far in our study we had found evidence that negation does not affect the inference process, at least with respect to the notboth inference form; participants seem just as likely to go beyond the information given to make the therefore not b inference as the therefore b inference. But what is the fate of the negated inferred proposition? Earlier work found that explicitly negated propositions are less available compared with ones that are not negated (MacDonald & Just, 1989). Does the same effect hold for inferred negated propositions? We pursued this question in Experiments 3 and 4 by using a combination of naming and verification times.

307

Consider a passage about a student, Vikki, deciding on what courses to take. Suppose that she discovers that she must take either history or English and that after some thought she decides not to take English. One can infer that she will take history (see Appendix C, Story Version 1). Alternatively, suppose the situation is that Vikki cannot take both history and English. Eventually she decides to take English; you can infer that she did not take history. Appendix C presents four versions of this example passage; Story Versions 1 and 2 represent the or and not-both examples, respectively. The other versions provided controls for our dependent measures. Following the MacDonald and Just (1989) study, we presented inference-related probe words that participants were asked to name aloud as quickly as possible. Naming latencies were chosen because they are thought to be less affected by postlexical processes than are binary decision measures such as probe recognition and lexical decision response times (see Keenan, Golding, Potts, Jennings, & Aman, 1990, for a detailed discussion of this concern). If participants made the inference that Vikki decided to take history in the or version of the passage, then naming times for “history” should be reliably faster than in a baseline, no-inference control condition in which the inference could not be made. In Story Version 3 (No inferenceA), the logical particle in the first premise has been altered so that Vikki has decided to take history and English, whereas the second premise (that she is taking English) is essentially the same as in the two inference versions—the important difference is that no (logical) inference can be made after the second premise in Story Version 3. Thus, the or inference (therefore history) in Story Version 1 should make the concept “history” highly active and naming times ought to be speeded relative to the control Story Version 3, in which no inference about “history” could be made. The outcome for the not-both inference (therefore not history) in Story Version 2 is less clear. On the one hand, the naming times might be speeded compared with the control because the inference had just been made. On the other hand, however, a possible outcome of using an activation measure such as naming time in the not-both case is that the inference not history might fail to produce fast responses to the probe history. Recall that MacDonald and Just (1989) found that negating a noun (e.g., not bread) inhibited naming and lexical decision responses to that noun. Thus, unspeeded probe responses to negated propositions in the present study could mean that although the inference was made, its activation was reduced at the discourse level, as MacDonald and Just found. MacDonald and Just (1989) found that negation produced measurable effects both on activation at the discourse level and on tasks that require the assessment of truth value, such as statement verification. Therefore we also measured participants’ response times to a comprehension question that followed each text. For a subset of the passages they read, the comprehension question concerned the inference in question and thus constituted a verification. That is, to determine whether the comprehension statement was true or false, participants would have to have drawn the inference. For example, in the not-both inference passage presented in Appendix C, the comprehension statement is Vikki did not take history that semester. A participant who has drawn the inference (therefore not history) while reading the story will be able to judge the verification statements quickly and accurately. We compared the verification latencies associated with inference

LEA AND MULLIGAN

308

versions of the stories with identical statements of the same truth value based on information that was explicitly presented in a control version of the story. For example, Story Version 4 presented in Appendix C states explicitly that Vikki has decided not to take history; the verification statement for this version is Vikki did not take history that semester— exactly the same statement as the one that follows the not-both inference passage. Indeed one might even expect a response-time advantage for verification based on inferences because the target concept would have just been generated and therefore be highly activated. Thus, responses to the verification task should be approximately the same as— or even faster than—those associated with a passage in which the information had been presented explicitly. However, if negation affects processes that compute truth value, as MacDonald and Just (1989) and others have found, then verification times for not-both inferences may show negation effects. Specifically, we might expect that verifications for negated inferences would not only be slower than their affirmative counterparts but also less speeded compared with the control. The latter prediction is based on the expectation that the activation boost enjoyed by the affirmative inferences produced in the or condition might be attenuated in the case of the negated inference concepts in the not-both passages. Thus, we may well find an Inference Type ⫻ Verification Response Time interaction.

Method Participants. Seventy undergraduate students participated for course credit. All were native speakers of English. Materials and design. Thirty texts were written in four versions that manipulated whether a specific logical inference could be drawn and whether the inference produced an affirmative or negative conclusion (or vs. not both). Table 3 presents a summary of the four experimental conditions in Experiment 3, and Appendix C presents all four versions of an example passage in which Vikki is thinking about taking courses in English and history; the target concept is “history.” Stimulus lists were constructed such that one of the four versions of each experimental passage appeared on a list. In addition to the 30 passages contained in each stimulus list that contained logical premises (10 inference plus 20 no-inference), 24 filler texts were designed around the same form but contained no logical particles. Sixteen of the filler passages contained naming probes that appeared after the first sentence of the passage. Probes appeared after the third sentence in the remaining 8 fillers. These balanced the probe positions in the experimental passages, in which 10 probes appeared after the second sentence and 20 appeared after the fourth sentence. Thus, over all 54 passages, the probe was presented in the first half of the passage 26 times and in the second half of the passage 28 times. All of the probe words appeared in the passage in which the probe was presented. The probe word was taken from the first sentence of the passage 24 times and in the second sentence 30 times (in some of the filler passages the word appeared in more

Table 3 Summary of Experimental Conditions (Story Versions) in Experiment 3 Condition

Function

Or (1) Not both (2) No inferenceA (3) No inferenceB (4)

Logical inference Logical inference Naming control and verification control for or Verification control for not both

than one sentence of the story). Exactly half of the comprehension statements were true and half were false. The comprehension statements were verifications of a logical inference in 10 of the 54 passages. The remaining 81% of the comprehension statements were either true of information stated explicitly in the passage, or they were false. Participants were randomly assigned to read one of the six lists; this resulted in six participant groups. Thus, the within-subjects factor (story version) and between-subjects factor (participant group) yielded a 4 ⫻ 6 split-plot design. Participant-group effects were not predicted. To summarize, participants read 54 stories, 10 of which permitted a logical inference and 44 of which did not. The naming probes could appear after any of the four sentences, and their presentation position was approximately balanced between the first and second half of the passages. Participants were seated before a computer screen in sound-insulated rooms. A response box with three keys (yes, no, continue) was placed within easy reach. Each participant adjusted the position of a microphone, which was attached to a boom stand, so that it was comfortably close to his or her mouth. Procedure. The experimental session began with 30 practice naming trials to help orient the students to the microphone and response box. They were given detailed instructions about how not to accidentally trip the voice key. The stories were presented to participants in a random order. Participants read two practice stories before they began the experiment. A story trial began with the title presented in the center of the screen for 2 s, followed by the first sentence of the passage. Participants were asked to read the stories (sentence by sentence) at their own pace; when they read and understood the sentence on the screen they pressed the continue key and moved on to the next sentence. Sentences replaced each other on the screen; no more than one sentence was present on the screen at a time. When participants read the sentence immediately preceding the probe, their keypress cleared the screen and activated the naming probe, which was presented in all capital letters with an asterisk at either side. Participants understood that they should pronounce the word as quickly and accurately as possible. One key on the response box functioned as the yes key, and an equally accessible key served as the no key. A third key on the response box was used to advance the text line by line. Participants were instructed to keep a finger on these three keys at all times. At the end of each story, a sentence appeared on the screen and participants were asked to indicate whether the sentence follows from, or is true of, the story they just read. Participants’ response to this comprehension statement advanced them to the next story.

Results Some conditions served as controls for only one dependent measure (e.g., Story Version 4 is a control for the verification task but not for the naming task). Therefore, only the relevant measures will be reported. Mean naming times for Experiment 3 are presented in Table 4. Only naming latencies associated with correct responses are reported. Naming time. Both the M–L model and the focus-shift hypothesis1 predict that naming times in the or condition should be faster than those in the no-inference control. Whether the negated inference in the not-both case (e.g., therefore not history) leads to faster naming times, however, depends on the effect that negation has on inferred propositions. The focus-shift hypothesis predicts that naming latencies in the not-both condition will be similar to those in the no-inference control because they are identical in terms of 1 By focus-shift hypothesis we mean an account of shifting focus that is not achieved through inference (e.g., by negation as MacDonald & Just, 1989, found). One can think of the logical inferences under investigation as one of a class of linguistic devices that can shift focus.

THE EFFECT OF NEGATION

Table 4 Naming Times (in ms) for Conditions (Story Versions) in Experiment 3 Condition

Naming time

Or (1) Not both (2) No inferenceA (3)

496 519 525

discourse focus and that both will be slower than those in the or condition because the latter shifts focus (implicitly, by means of inference) toward the target. The means depicted in Table 4 conform well with the latter prediction. On average, participants named the word probes significantly faster in the or condition compared with the No-inferenceA control, F1(1, 64) ⫽ 7.71, MSE ⫽ 3,603.80; F2(1, 24) ⫽ 5.25, MSE ⫽ 1,010.46, whereas the not-both response times were not different from the control (both Fs ⬍ 1). Thus, the naming-time data indicate that the inference concepts were not highly activated in the not-both passages. Verification latency. As mentioned earlier, we expected that verification latencies in the inference conditions would be no different than, or possibly faster than, their controls. Indeed, to the extent that negation affects responses that evaluate truth value, we thought it possible that an Inference Type ⫻ Response Effect interaction would result. The verification latencies for Experiment 3 presented in Table 5 show such an effect. Participants did not take significantly more time to complete the not-both verifications compared with the control (both Fs ⬍ 1), but verification times for or passages were speeded compared with the control, F1(1, 64) ⫽ 7.45, MSE ⫽ 563,568.16; F2(1, 24) ⫽ 5.57, MSE ⫽ 288,609.12. The interaction proved to be significant by subjects, F1(1, 64) ⫽ 4.75, MSE ⫽ 604,019.63, and marginal by items, F2(1, 24) ⫽ 3.12, MSE ⫽ 279,008.98, p ⫽ .09. The verification accuracy data followed a similar pattern; the or inferences were verified more accurately than the control, F1(1, 64) ⫽ 15.30, MSE ⫽ .04; F2(1, 24) ⫽ 7.82, MSE ⫽ .03, and the not-both inferences were not significantly different from control, F1(1, 64) ⫽ 2.08, MSE ⫽ .02, p ⫽ .15; F2(1, 24) ⫽ 3.17, MSE ⫽ .01, p ⫽ .09. This difference led to an interaction that was significant by subjects and items, F1(1, 64) ⫽ 14.17, MSE ⫽ .03; F2(1, 24) ⫽ 8.72, MSE ⫽ .02.

Discussion The purpose of Experiment 3 was to investigate the effect that negation has on inferred propositions. Experiments 1 and 2 establish that the not-both inference is made on-line. The question addressed by Experiment 3 was the following: Would these inferred (and negated) propositions (e.g., not history) experience the sort of inhibition effects that MacDonald and Just (1989) found for explicitly negated concepts? Using the passages from the first two experiments, we found that naming latencies for inference concepts were significantly speeded when they were affirmative but not when they were negated. This result, though indirect, is at least consistent with the hypothesis that negation inhibits inferred propositions. The verification latencies also showed negation effects. Affirmative statements (e.g., Vikki took history that semester) were verified more rapidly compared with a control in which the prop-

309

osition to be verified was presented explicitly three sentences earlier. Because the answer in both conditions was yes, and the statements to be verified were identical, the response advantage in the inference condition is due, presumably, to the fact that the verification concept had just been computed and was still in working memory. Negated statements (e.g., Vikki did not take history that semester), on the other hand, did not show this availability boost, perhaps because the availability advantage enjoyed by inferences made in the or condition was offset or attenuated when the concept to be verified was negated. The significant interaction between inference type and verification-sentence response time supports research noted previously that has repeatedly found that negation affects processes that compute truth value, such as those involved in a verification task. Although Experiment 3 was designed to pursue the effects of negation on inferred concepts, it is important to note that the focus-shift hypothesis also accounts neatly for the naming results in this experiment. Probes were not speeded in the not-both passages because there was no shift of focus to “history” at the time of test, whereas probes were speeded in the or passages because the second premise (I’m not going to take English) naturally shifts focus to “history.” Of course, such a shift can take place only if the inference is drawn. Still, the focus-shift hypothesis can be tested more directly by separating the explicit shift in focus from the presentation of the second premise in the text. In other words, we needed to unconfound the licensing conditions for the inference (simultaneous availability of the two premises) with an explicit shift in focus in the text. Experiment 4 was undertaken to accomplish this goal.

Experiment 4 We reversed the order of the stories’ premises in Experiment 4 so that no change in discourse focus occurred at the time of test. For example, in the or passage presented in Appendix D (Story Version 1), the reader first learns that Vikki cannot take English and then later finds out that she must take either history or English to satisfy graduation requirements. It follows logically that she must take history. At the time of the naming probe, however, there is no change in discourse focus as there had been in Experiment 3. Thus, if increased activation of the targets in the or condition in Experiment 3 was due to a change in discourse focus—and not to inference making—then that effect should disappear in Experiment 4. One might argue that stating that Vikki cannot take English in the second sentence would constitute a shift in focus to “history.” However, the sentence preceding the probe reinstates English as

Table 5 Verification Times (in ms) and Percentage of Verification Error for Conditions (Story Versions) in Experiment 3 Condition Not both Inference (2) No inferenceB (4) Or Inference (1) No inferenceA (3)

Verification time

% Error

3,137 3,101

.14 .10

2,383 2,727

.13 .25

LEA AND MULLIGAN

310

part of the graduation requirement, and it is not until one integrates the two premises that the focus can shift to “history.” That integration, of course, is the or inference. The focus-shift hypothesis predicts that naming targets in the or condition will not be speeded because the discourse focus is not shifting toward those targets at the time of test (other than the shift accomplished by the inference). Thus, targets in both inference conditions should be as slow or slower than the control, according to focus shift. As in Experiment 3, we anticipated that verification responses would be at least as fast as the controls, with more of a facilitation effect for the affirmative verifications. We also measured reading times on the stories’ fourth sentence; this sentence contained the second premise in inference versions of the passages and immediately preceded the probe. Even though the sentences could not be identical across conditions, we wanted to be reassured that differential reading times could not lead to an alternative (spillover) interpretation of the naming results.

Method Participants. Fifty-four undergraduate students participated for course credit. All were native speakers of English. Materials and procedure. The passages used in Experiment 3 were rewritten so that the order of the premises was reversed. A summary of the experimental conditions in Experiment 4 is presented in Table 6, and an example passage is presented in Appendix D. Four conditions were required to test the hypotheses in Experiment 2. Participants read 20 experimental passages and 24 fillers. All other aspects of the materials, design, and procedure were the same as Experiment 3.

Results Mean fourth-sentence reading times and naming latencies for Experiment 4 are presented in Table 7, and the verification results appear in Table 8. Naming times. Overall, the pattern of results was the same as Experiment 3. On average, participants named the word probes faster in the or condition compared with the no-inference control. This effect was significant by subjects, F1(1, 50) ⫽ 6.36, MSE ⫽ 3,883.33, and marginal by items, F2(1, 16) ⫽ 3.45, MSE ⫽ 1,546.71, p ⫽ .08. As in Experiment 3, there was no difference in naming time between the not-both condition and the control (both Fs ⬍ 1). Verification latency. The verification latencies for Experiment 4 are presented in Table 8. As in Experiment 3, participants did not take significantly more time to complete the not-both verifications compared with the no-inference control in which the inference was stated in the story, F1(1, 50) ⬍ 1; F2(1, 16) ⫽ 1.52, MSE ⫽ 100,973.86, p ⫽ .24. Verifications for the or inference also were no slower than the control; indeed they were faster, as they

Table 6 Summary of Experimental Conditions (Story Versions) in Experiment 4 Condition

Function

Or (1) Not both (2) No inferenceA (3) No inferenceB (4)

Logical inference Logical inference Verification control for not-both Naming control and verification control for or

Table 7 Naming and Fourth-Sentence Reading Times (in ms) for Conditions (Story Versions) in Experiment 4 Condition

Fourth sentence

Naming

Or (1) Not both (2) No inferenceA (4)

4,237 4,657 4,068

497 531 525

had been in Experiment 3, F1(1, 50) ⫽ 19.57, MSE ⫽ 235,310.20; F2(1, 16) ⫽ 6.22, MSE ⫽ 83,277.61. This difference in the two inference forms produced an interaction that was significant by subjects, F1(1, 50) ⫽ 11.01, MSE ⫽ 238,940.39, and marginal by items, F2(1, 16) ⫽ 3.69, MSE ⫽ 14,124.26, p ⫽ .074. Again, verification accuracy followed the latency response pattern. The or inferences were verified more accurately than the control, F1(1, 50) ⫽ 11.81, MSE ⫽ .02; F2(1, 16) ⫽ 1.53, MSE ⫽ .05, p ⫽ .23, whereas the not-both inferences were no different than the control (both Fs ⬍ 1). The interaction was significant by subjects, F1(1, 50) ⫽ 8.07, MSE ⫽ .02, but not by items, F2(1, 16) ⫽ 1.46, MSE ⫽ .03, p ⫽ .25. Reading times. We looked at the reading times on the passages’ fourth sentence to reassure ourselves that response differences in the naming task were not an artifact of differences in reading times on that sentence. The fourth sentence was the last in each story, the one that contained the second premise in inference conditions, and the one that immediately preceded the naming probe in all conditions. If fourth-sentence reading times were significantly slower in the naming control condition, then one might worry that carryover effects might account for the difference between the control and or conditions on the naming task. However, as is evident in the reading times reported in Table 8, fourth-sentence reading times for the control condition were faster than any other condition. Indeed, they were 167 ms faster on average than in the or condition, a difference that approached significance by subjects, F1(1, 50) ⫽ 2.11, MSE ⫽ 276,226, p ⫽ .15; F2(1, 16) ⬍ 1. Only the fourth-sentence reading times for the not-both condition were significantly slower than the other conditions (all ps ⬍ .01 by subjects and ⬍ .05 by items). Thus, there is no reason to believe that spillover was affording any speed advantage to the inference probes; if anything, the naming responses in the or and not-both conditions may have been inhibited by slower reading of the immediately preceding sentence.

Discussion Experiment 4 was undertaken to separate the effects of focus shift and inference. By reversing the order of premises, we created Table 8 Verification Times (in ms) and Percentage of Verification Error for Conditions (Story Versions) in Experiment 4 Condition Not both Inference (2) No inferenceA (3) Or Inference (1) No inferenceB (4)

Verification time

% Error

2,847 2,800

.12 .10

2,296 2,689

.10 .19

THE EFFECT OF NEGATION

passages in which the licensing conditions for the logical inferences occurred without an explicit focus shift toward or away from the naming targets. If the significant difference in naming response time in Experiment 3 had been due to a focus shift at the time of test, then that difference should have disappeared without the focus shift. However, the or naming probes still were significantly faster than in the control condition, thus indicating that the parallel priming effect in Experiment 3 was likely due to inference and not just to focus change. Readers’ verifications were faster when they inferred the proposition through the or inference than when the proposition had appeared explicitly in the story. As we found in Experiment 3, producing information by means of inference may give those propositions privileged status with regard to salience or memory (cf. the generation effect). If this is true, then this result points to an interesting way in which negation cancels or at least works against the inference boost.

General Discussion The present study was designed to investigate the role of negation in the inference process. MacDonald and Just (1989) found that explicitly negated concepts (e.g., not cookies) were less accessible from readers’ mental representations of the text compared with concepts that were not negated. MacDonald and Just explained their effects in terms of discourse focus; when one hears that Elizabeth baked bread but not cookies, one’s focus turns away from “cookies” and onto “bread.” The question posed here is the following: if one hears that Mary doesn’t have time to bake both bread and cookies, and therefore she decides to bake bread, will one then infer that she will not bake cookies? Though MacDonald and Just did not address the question directly, a reasonable extension of their focus-shift hypothesis is, in this case, that once we know that Elizabeth will make bread, our focus will shift away from “cookies.” But does that mean (a) that the inference will not be made, or (b) that once made, it will be inhibited? When, if at all, will negation affect these inferences? Experiments 2, 3, and 4 compared two logical inferences that are fundamentally similar: (a) Both inference forms begin with two possible outcomes, and given additional information, you are left with only one; and (b) previous research shows that adults find the two inference forms equally easy to make (when working with abstract materials; Braine et al., 1984). Indeed, according to M–L theory (e.g., Braine & O’Brien, 1998) the or and not-both inferences are predicted to be made equally readily whether a person is engaged in explicit problem solving or discourse processing. Researchers in text processing who recently have devoted a considerable effort to the question of which sorts of inferences are made during reading and under what conditions they are made, now believe that readers do not necessarily make all of the inferences that psychologists previously believed they did (e.g., Graesser & Bower, 1990; McKoon & Ratcliff, 1992; Singer, Graesser, & Trabasso, 1994). This message has resulted in an effort (in the text-processing literature) to better specify the inference process during reading (e.g., Graesser, Singer, & Trabasso, 1994; McKoon, Gerrig, & Greene, 1996; Myers & O’Brien, 1998). The current work aims to contribute to this discussion by challenging theories of psychological deduction with scopes that include discourse processing to consider the constraints that reading might impose on inference making. In other words, just because people easily make these logical inferences when they are trying to solve

311

problems about ps and qs on an imaginary blackboard, does not mean that they do so routinely during reading when their cognitive goals are less focused and their resources more heavily tapped. Specifically, negation is a discourse factor that long has been understood to affect text processing in a variety of ways (e.g., Carpenter & Just, 1975; Just & Clark, 1973; MacDonald & Just, 1989), and the present work constitutes, in part, a test of M–L theory with regard to the effect of negation on inference making; although the logic model predicts that people make the or and the not-both inferences equally readily while reading, we were skeptical. We suspected that readers would be less likely to make inferences about negative information—what did not or will not happen—in a text than they would be to make inferences about positive information—what has happened or will happen—in the text. Experiment 1 provided direct evidence that readers were making the not-both inference. Participants were much slower to read a sentence when it contradicted the not-both inference compared with when the sentence appeared in a control version of the story that did not sanction the inference. Experiment 2 replicated the reading-slowdown effect and addressed the possibility that the Experiment 1 reading slowdown was due to readers making inferences at the target sentence rather than with the presentation of the second premise at Line 4. In Experiment 2 we added the or inference and found an identical pattern of reading slowdowns between the inference forms. If participants were waiting until the target sentence to integrate Lines 2 and 4 and then comparing the inference with Line 5 (the target sentence), we would have expected to see a slowdown larger than in the or condition for which we are certain that the inference was made before the target sentence. However, we found no evidence for an Inference Type ⫻ Inference Effect interaction; if anything, the slowdown was slightly smaller in the not-both case. Thus, these two experiments present a clear case that the not-both inference is made on-line. Experiments 3 and 4 produced naming-time results that make the story more interesting. The or probes were named significantly faster than the control, but the not-both probes were not. Given the results from Experiments 1 and 2, together with the fact that we used exactly the same passages in Experiment 3 as in the first two experiments, we are left to conclude that inferring negative information does not afford an activation boost detectable with naming. Indeed, we found a similar pattern with the verification data. For both inference types, the verification statement was true of the inference that followed from the fourth line of the passage. In their control versions, the statement was true of information explicitly presented earlier in the story. For the affirmative or inference (therefore b), we found significant facilitation compared with the control. For the negated not-both inference (therefore not b), we found no difference from the control. Therefore it appears that negation negated, if you will, the availability boost afforded to affirmative inferences. An alternative possibility is that readers were not making the not-both inference in Experiments 3 and 4. Militating against this view are (a) the results of Experiments 1 and 2 that show a clear inference effect for not-both, (b) the fact that the passages used in Experiment 3 were exactly the same as those used in Experiments 1 and 2, and (c) the lack of any inference effect detected in the verification results in Experiments 3 and 4. Therefore, the evidence converges on the conclusion that readers do indeed make the not-both inference on-line and that the inference fails to activate the negated concept. To-

312

LEA AND MULLIGAN

gether, then, these four experiments present a complex and potentially interesting picture of the role that negation might play in inference making. First, negation does not seem to inhibit the inference making process—at least not in a way relevant to the current inferences; Experiments 1 and 2 established that readers were in fact making the negated not-both inference. Second, it appears that negation does affect the inference after it is made. Specifically, the present data indicate that inference concepts that are negated are less accessible than comparable inferences that are not negated. This result is parallel to MacDonald and Just’s (1989) finding that explicitly negated concepts (as opposed to inferred negated concepts) were less accessible from the reader’s mental representation of the text. As discussed earlier, MacDonald and Just explained their effects in terms of discourse focus. However, there is more going on in the present passages than focus shifts alone. By separating focus shift from presentation of the second premise in Experiment 4 we established that the priming of the or inference targets was not due to a shift in discourse focus. In addition, Experiment 1 and 2’s demonstration that readers were in fact making the not-both inference shows that the inference mechanism takes precedence (at least chronologically) over whatever mechanism drives focus shift. That is, the inference is made first and then the focus shift, or suppression mechanism, is activated. Thus, the inference is made even though it yields information that is not particularly relevant to the discourse focus. Gernsbacher’s structure building framework specifies a suppression mechanism that accounts for similar phenomena at a different level of comprehension (e.g., Gernsbacher, 1990; Gernsbacher & Faust, 1991). For example, although both meanings of an ambiguous word are immediately activated even when the sentence context clearly indicates one meaning, within 1 s only the appropriate meaning is still active (e.g., Swinney, 1979; Till, Mross, & Kintsch, 1988). According to the structure building framework, the inappropriate meanings are actively suppressed by the reader; that is, the activation of information that is no longer necessary for the (discourse) structure being built is decreased. Thus, the meaning “hidden microphone” is activated and then quickly suppressed when the word bug is encountered in a text about insects. Other theories of discourse processing model these sorts of results in slightly different ways. For example, according to Kintsch’s (1988) construction integration (C–I) model, discourse comprehension proceeds in a two-stage fashion. In the first stage, a text base is constructed from the linguistic input as well as from the comprehender’s knowledge base; many elements are activated in this phase, some of which are redundant or contradictory. An integration process is then used to strengthen the contextually appropriate elements and inhibit the inappropriate ones. Both the C–I model and structure building framework provide a good theoretical context in which to view the current data. According to the M–L model, the or and not-both inferences are drawn at the moment that both premises are simultaneously available to the reader—regardless of how the inference relates to the currently read text. This assumption is consistent with the construction phase of Kintsch’s model and is consistent with Gernsbacher’s assumption that suppression and enhancement occur after “memory cells” are activated by incoming stimuli. After the inference is drawn, comprehension mechanisms operate on the information available and the relevant, useful information is kept (enhanced or integrated according to Gernsbacher, 1990, and Kintsch, 1988, respectively),

while irrelevant and redundant information is suppressed or not integrated. For present purposes, an inference about what did not or will not happen in a text would be seen as less relevant than an inference about what has happened or will happen in the text. In other words, both inferences are made immediately, and then the output is examined for relevance to the discourse model of the text. According to this view, we should find evidence that both the or and the not-both inference are being made but only evidence for higher activation of the or-inference concept. The results of the present experiments fit well with this prediction. Does the present work inform the debate between competing theories of deduction? M–L theories are often compared with the Mental Models (M–M) approach to deductive reasoning. Briefly, M–M theory assumes that reasoning is accomplished by a threestep semantic process in which people (a) build an initial model on the basis of the meaning of the premises and containing only true contingencies; (b) derive a conclusion based on these premises; and (c) try to build further models in which that conclusion could be false. If no such models are discovered, then the initial conclusion is endorsed (e.g., Johnson-Laird, Byrne, & Schaeken, 1992). Both M–L and M–M theory predict that readers will make the or and not-both inferences in texts such as those used in our study. However, different predictions can emerge from the two theories depending on the order in which the premises arrive. For example, Girotto, Mazzocco, and Tasso (1997) found that reversing the premise order of certain conditional forms led to a divergence in predictions that allowed them to compare the two theories. In the present work we reversed the order of premises in Experiment 4 to rule out a focus-shift explanation of the results of Experiment 3; according to M–M, the not-both inference should be easier to make when the categorical premise is presented first (a; not both a and b), as opposed to when the categorical premise is presented second (not both a and b; a; P. N. Johnson-Laird, personal communication, July 26, 2001). M–L predicts no difference because of premise order. Our results showed that readers routinely made the inference in both orders, which might be interpreted as support for the M–L approach. However, our experiments were not designed to detect subtle differences such as those predicted by M–M theory. Therefore, our data should not be seen as diagnostic with respect to these competing psychological theories of deduction.2 Note that in many discourse situations a premise can be a fact stored in long-term memory, rather than information that was presented earlier in the text or conversation. Consider the following example in which the or and not-both premises might be pragmatically induced from world knowledge: Minnesota Version (1) (2)

Sam was asking Pat about her new car. They lived in Minnesota, and he had been trying to talk her into getting the four-wheel drive transmission. (3) Sam: “What did you decide about the transmission for your new car?” (4) Pat (sighing): “It was a hard choice, but in the end I went with the regular two-wheel drive.” (5) Inference: Pat did not get the four-wheel drive transmission.

2

We thank an anonymous reviewer for suggesting this comparison.

THE EFFECT OF NEGATION San Francisco Version (6) (7)

Sam was asking Pat about her new car. They lived in San Francisco, and he had been trying to talk her into getting an automatic transmission. (8) Sam: “What did you decide about the transmission for your new car?” (9) Pat (sighing): “It was a hard choice, but in the end I decided against the automatic transmission.” (10) Inference: Pat got the manual transmission.

In these examples, the major premise is world knowledge stored in both Sam’s and Pat’s long-term memory about car transmissions (that they come in one of two forms but not both). In both versions Pat can take this mutual knowledge about cars for granted, assume that Sam will retrieve the appropriate premise along with geographic information that San Francisco has many steep hills and that Minnesota is snowy, and be assured that she has provided information relevant to his query (cf. Sperber & Wilson, 1986). Thus, premise information can be retrieved from the participants’ world knowledge, and local pragmatics will dictate what the speaker will say explicitly and what will be left to inference. More generally, the pragmatic foundations of these two inference forms concern contexts in which two options are narrowed to one by the process of asserting choice affirmatively (not-both) versus asserting choice negatively (or). These examples also illustrate how inference mechanisms permit some economy in discourse; Pat can disregard two important elements of their exchange (the major premise and conclusion) and is free to move the conversation in whatever direction she wishes (cf. Singer & Halldorson, 1996). Finally, the present work further establishes logical deductions as part of the common repertory of inferences that people make in everyday cognitive processing. Evidence that readers make the 12 propositional logic inferences specified in Braine and colleagues’ M–L model (e.g., Braine et al., 1984; Braine & O’Brien, 1998) was first reported by Lea et al. (1990) in an off-line memory task in which participants often thought logical inferences that they had made while reading were instead information presented explicitly in the texts. Lea (1995) found that two of the model’s inference forms, or and modus ponens, were made on-line in locally coherent texts. The present work tested an unlikely prediction of the M–L model: that readers will bother to make inferences about negative information—about what has not or will not happen in a passage. Despite our expectations to the contrary, it appears possible that these simple logical deductions are among the elementary and mandatory linguistic processes that precede more deliberate and constructive cognitive operations (cf. Bonatti, 1998; Gentzen, 1964; Macnamara, 1986; Noveck & Politzer, 1998; O’Brien, 1993; Osherson, 1975; Pinker, 1989; Politzer, 1986; Rips, 1994; Sperber & Wilson, 1986). Of course, more work is needed to integrate models of reasoning with those of discourse processing before definitive conclusions can be made. In the meantime, however, the present research and the mental logic model tested herein provide a basis from which we can make advances toward a more complete understanding of deduction and comprehension.

References Albrecht, J. E., & Myers, J. L. (1995). The role of context in accessing distant information during reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1459 –1468.

313

Bonatti, L. (1998). Why it took so long to bake the mental-logic cake: Historical analysis of the recipe and its ingredients. In M. D. S. Braine & D. P. O’Brien (Eds.), Mental Logic (pp. 7–22). Hillsdale, NJ: Erlbaum. Braine, M. D. S. (1978). On the relation between the natural logic of reasoning and standard logic. Psychological Review, 85, 1–21. Braine, M. D. S., & O’Brien, D. P. (1998). Mental Logic. Hillsdale, NJ: Erlbaum. Braine, M. D. S., O’Brien, D. P., Noveck, I. A., Samuels, M. C., Lea, R. B., Fisch, S. M., & Yang, Y. (1995). Predicting intermediate and multiple conclusions in propositional logic inference problems: Further evidence for a mental logic. Journal of Experimental Psychology: General, 124, 263–292. Braine, M. D. S., Reiser, B. J., & Rumain, B. (1984). Some empirical justification for a theory of natural propositional logic. In G. H. Bower (Ed.), The psychology of learning and motivation: Advances in research and theory (Vol. 18; pp. 313–371). New York: Academic Press. Braine, M. D. S., & Rumain, B. (1983). Logical reasoning. In J. H. Flavell & E. M. Markman (Eds.), Handbook of child psychology: Vol. 3 Cognitive Development (pp. 263–339). New York: Wiley. Carpenter, P. A., & Just, M. A. (1975). Sentence comprehension: A psycholinguistic processing model of verification. Psychological Review, 82, 45–73. Cheng, P. W., & Holyoak, K. (1985). Pragmatic reasoning schemas. Cognitive Psychology, 17, 391– 416. Clark, H. H., & Chase, W. G. (1972). On the process of comparing sentences against pictures. Cognitive Psychology, 3, 472–517. Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped how humans reason? Studies with the Wason selection task. Cognition, 31, 187–276. Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178 –210. Gentzen, G. (1964). Investigations into logical deduction. American Philosophical Quarterly, 1, 288 –306. Gernsbacher, M. A. (1989). Mechanisms that improve referential access. Cognition, 32, 99 –156. Gernsbacher, M. A. (1990). Language comprehension as structure building. Hillsdale, NJ: Erlbaum. Gernsbacher, M. A., & Faust, M. E. (1991). The mechanism of suppression: A component of general comprehension skill. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 245–262. Gernsbacher, M. A., & Jescheniak, J. D. (1995). Cataphoric devices in spoken discourse. Cognitive Psychology, 29, 24 –58. Girotto, V., Mazzocco, A., & Tasso, A. (1997). The effect of premise order in conditional reasoning: A test of the mental model theory. Cognition, 63, 1–28. Gough, P. B. (1966). The verification of sentences: The effects of delay of evidence and sentence length. Journal of Verbal Learning and Verbal Behavior, 5, 492– 496. Graesser, A. C., & Bower, G. H. (1990). Inferences and text comprehension. New York: Academic Press. Graesser, A. C., Singer, M., & Trabasso, T. (1994). Constructing inferences during narrative text comprehension. Psychological Review, 101, 371–395. Griggs, R. A., & Cox, J. R. (1982). The elusive thematic-materials effect in Wason’s selection task. British Journal of Psychology, 73, 407– 420. Johnson-Laird, P. N., Byrne, R. M. J., & Schaeken, W. (1992). Propositional reasoning by models. Psychological Review, 99, 418 – 439. Just, M. A., & Carpenter, P. A. (1971). Comprehension of negation with quantification. Journal of Verbal Learning and Verbal Behavior, 10, 244 –253. Just, M. A., & Clark, H. H. (1973). Drawing inferences from the presupposition and implications of affirmative and negative sentences. Journal of Verbal Learning and Verbal Behavior, 12, 21–31.

314

LEA AND MULLIGAN

Keenan, J. M., Golding, J. M., Potts, G. R., Jennings, T. M., & Aman, C. R. (1990). Methodological issues in evaluating the occurrence of inferences. In A. C. Graesser & G. H. Bower (Eds.), Inferences and text comprehension (pp. 295–312). New York: Academic Press. Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integration model. Psychological Review, 95, 163–182. Lea, R. B. (1995). On-line evidence for elaborative logical inferences in text. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1469 –1482. Lea, R. B. (1998). Logical inference and comprehension: How mental logic and text processing theories need each other. In M. D. S. Braine & D. P. O’Brien (Eds.), Mental Logic (pp. 63–78). Hillsdale, NJ: Erlbaum. Lea, R. B., Charnley, J., & Mead, H. (2001). Working memory size affects pragmatic and logical inferences differently. Manuscript in preparation. Lea, R. B., Lee, J. & Mulligan, E. J. (2001). Accessing distant premise information: How memory feeds reasoning. Manuscript submitted for publication. Lea, R. B., Mason, R. A., Albrecht, J. E., Birch, S., & Myers, J. L. (1998). Who knows what about whom: What role does common ground play in accessing distant information? Journal of Memory and Language, 39, 70 – 84. Lea, R. B., O’Brien, D. P., Fisch, S. M., Noveck, I. A., & Braine, M. D. S. (1990). Predicting propositional logic inferences in text comprehension. Journal of Memory and Language, 29, 361–387. MacDonald, M. C., & Just, M. A. (1989). Changes in activation levels with negation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 633– 642. MacDonald, M. C., & MacWhinney, B. (1990). Measuring inhibition and facilitation from pronouns. Journal of Memory and Language, 29, 469 – 492. Macnamara, J. (1986). A border dispute: The place of logic in psychology. Cambridge, MA: MIT Press. McKoon, G., Gerrig, R. J., & Greene, S. B. (1996). Pronoun resolution without pronouns: Some consequences of memory-based text processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 919 –932. McKoon, G., & Ratcliff, R. (1992). Inference during reading. Psychological Review, 99, 440 – 466. Myers, J. L., & O’Brien, E. J. (1998). Accessing the discourse representation during reading. Discourse Processes, 26, 131–157. Myers, J. L., O’Brien, E. J., Albrecht, J. E., & Mason, R. A. (1994). Maintaining global coherence during reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 876 – 886. Myers, J., Shinjo, M., & Duffy, S. (1987). Degree of causal relatedness and memory. Journal of Memory and Language, 26, 453– 465. Noveck, I. A., Lea, R. B., Davidson, G. M., & O’Brien, D. P. (1991). Human reasoning is both logical and pragmatic. Intellectica, 11, 81–109.

Noveck, I. A., & O’Brien, D. P. (1996). To what extent do pragmatic reasoning schemas affect performance on Wason’s selection task? Quarterly Journal of Experimental Psychology, 49(A), 463– 489. Noveck, I. A., & Politzer, G. (1998). Leveling the playing field: Investigating competing claims concerning relative inference difficulty. In M. D. S. Braine & D. P. O’Brien (Eds.), Mental Logic (pp. 367–384). Hillsdale, NJ: Erlbaum. O’Brien, D. P. (1993). Mental logic and irrationality: We can put a man on the moon, so why can’t we solve those logical reasoning problems? In K. I. Manktelow and D. E. Over (Eds.), Rationality (pp. 110 –135). Chichester, UK: Routledge. Osherson, D. N. (1975). Models of logical thinking. In R. Falmagne (Ed.), Reasoning: Representation and process in children and adults (pp. 81–91). Hillsdale, NJ: Erlbaum. Pinker, S. (1989). Learnability and cognition: The acquisition of argument structure. Cambridge, MA: MIT Press. Politzer, G. (1986). Laws of language use and formal logic. Journal of Psycholinguistic Research, 15, 47–92. Rips, L. J. (1983). Cognitive processes in propositional reasoning. Psychological Review, 90, 38 –71. Rips, L. J. (1994). The psychology of proof: Deductive reasoning in human thinking. Cambridge, MA: MIT Press. Singer, M., Graesser, A. C., & Trabasso, T. (1994). Minimal or global inference during reading. Journal of Memory and Language, 33, 421– 441. Singer, M., & Halldorson, M. (1996). Constructing and validating motive bridging inferences. Cognitive Psychology, 30, 1–38. Sperber, D., & Wilson, D. (1986). Relevance: Communication and cognition. Cambridge, MA: Harvard University Press. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645– 659. Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory & Cognition, 16, 283–298. Trabasso, T., Rollins, H., & Shaughnessy, E. (1971). Storage and verification stages in processing concepts. Cognitive Psychology, 2, 239 –289. Wason, P. C. (1959). The processing of positive and negative information. Quarterly Journal of Experimental Psychology, 11, 92–107. Wason, P. C. (1963). The contexts of plausible denial. Journal of Verbal Learning and Verbal Behavior, 4, 7–11. Wason, P. C. (1966). Reasoning. In B. Foss (Ed.), New horizons in psychology. Harmondsworth, Middlesex, England: Penguin. Wiley, J., Mason, R. A., & Myers, J. L. (2001). Accessibility of potential referents following categorical anaphors. Journal of Experimental Psychology: Learning, Memory, and Cognition, 27, 1238 –1249.

THE EFFECT OF NEGATION

Appendix A Experiment 1 Example Passages With *Target Sentences* and ?Comprehension Statements? Story version 1: Not-both inference Joseph and Eric were spicing their famous red-hot Texas chili for their guests one evening. According to the recipe, they could not use both fresh jalapenos and ground pepper in the chili. “Well,” said Eric looking in the fridge, “let’s see what we have handy.” “We’ve got plenty of fresh jalapenos—let’s use them!” he said. *“I’ll get the ground pepper,” said Joseph.* ?They used jalapenos in the chili.? Story version 2: No inference (target sentence control) Joseph and Eric were spicing their famous red-hot Texas chili for their guests one evening. According to the recipe, for extra spicy chili they could use both fresh jalapenos and ground pepper. “Well,” said Eric looking in the fridge, “let’s see what we have handy.” “We’ve got plenty of fresh jalapenos—let’s use them!” he said. *“I’ll get the ground pepper,” said Joseph.* ?According to the recipe, they could use both jalapenos and ground pepper to spice the chili.?

Appendix B Experiment 2 Or Example Passages With *Target Sentences* and ?Comprehension Statements? Story version 1: Or inference Joseph and Eric were spicing their famous red-hot Texas chili for their guests one evening. According to the recipe, they could add either fresh jalapenos or ground pepper. ‘Well,’ said Eric looking in the fridge, ‘let’s see what we have handy.’ ‘We’ve got no fresh jalapenos—so we won’t use them,’ he said. The chili was great without ground pepper. ?They didn’t use jalapenos to spice the chili.? Story version 2: No inference (target sentence control) Joseph and Eric were spicing their famous red-hot Texas chili for their guests one evening. According to the recipe, they should use fresh jalapenos instead of the usual ground pepper. ‘Well,’ said Eric looking in the fridge, ‘let’s see what we have handy.’ ‘We’ve got fresh jalapenos—so let’s use them,’ he said. The chili was great without ground pepper. ?They used jalapenos instead of ground pepper to spice the chili.?

(Appendixes follow)

315

316

LEA AND MULLIGAN

Appendix C Experiment 3 Example Passages, *TARGETS*, ?Verification Statements?, and Correct Response to the Verification Task (True/False) Story version 1: Or inference Vikki had several slots left in her course schedule. She had already decided to take either History or English. Choosing between the alternatives was depressing, however. ‘Well,’ Vikki sighed to herself, ‘I’m not going to take English,’ she finally decided. *HISTORY* ?Vikki took History that semester.? (True) Story version 2: Not-both inference Vikki had several slots left in her course schedule. She had already decided that she could not take both History and English. Choosing between the alternatives was depressing, however. ‘Well,’ Vikki sighed to herself, ‘I’m going to take English,’ she finally decided. *HISTORY* ?Vikki did not take History that semester.? (True) Story version 3: No inferenceA (Naming control; Verification control for or) Vikki had several slots left in her course schedule. She had already decided to take History and English. Looking at the alternatives was depressing, however. ‘Well,’ Vikki sighed to herself, ‘at least I’ve decided to take English—I know I’ll like that.’ *HISTORY* ?Vikki took History that semester.? (True) Story version 4: No inferenceB (Verification control for not-both) Vikki had several slots left in her course schedule. She had already decided not to take History, but to take English instead. Looking at the alternatives was depressing, however. ‘Well,’ Vikki sighed to herself, ‘at least I’ve decided to take English—I know I’ll like that.’ *HISTORY* ?Vikki did not take History that semester.? (True)

THE EFFECT OF NEGATION

317

Appendix D Experiment 4 Example Passages, *TARGETS*, ?Verification Statements?, and Correct Response to the Verification Task (True/False) Story version 1: Or inference Vikki had one slot left in her course schedule. She could not take English because she had already taken it. Before registering she consulted the college handbook to check the graduation requirements. She had to take either History or English that semester in order to satisfy the requirements. *HISTORY* ?Vikki took History that semester.? (True) Story version 2: Not-both inference Vikki had one slot left in her course schedule. She had to take English because she hadn’t already taken it. Before registering she consulted the college handbook to check the graduation requirements. She didn’t need to take both History and English that semester in order to satisfy the requirements. *HISTORY* ?Vikki didn’t take History that semester.? (True) Story version 3: No inferenceA (Verification control for not-both) Vikki had one slot left in her course schedule. She had to take English because she hadn’t already taken it. Before registering she consulted the college handbook to check the graduation requirements. She could take no History but only English that semester to satisfy the requirements. *HISTORY* ?Vikki didn’t take History that semester.? (True) Story version 4: No inferenceB (Naming control; Verification control for or) Vikki had several slots left in her course schedule. She had to take English because she hadn’t already taken it. Before registering she consulted the college handbook to check the graduation requirements. She had to take both History and English that semester to satisfy the requirements. *HISTORY* ?Vikki took History that semester.? (True)

Received February 23, 2001 Revision received August 3, 2001 Accepted August 27, 2001 䡲