Psychol Res (1998) 61: 249±260

Ó Springer-Verlag 1998

ORIGINAL ARTICLE

Manuel G. Calvo á M. Dolores Castillo

Predictive inferences take time to develop

Received: 14 January 1998 / Accepted: 21 April 1998

Abstract In order to determine the time course of inferences about predictable events, predicting or control contexts were presented, followed by a target word (Exps. 1A, B, and C) or a continuation sentence (Exp. 2) that con®rmed or discon®rmed the predicted event. Relative to the control condition, under the predicting condition there was facilitation in naming the con®rming target words 1500 ms after the onset of the last word in the context (Exp. 1C), but not after 500 ms (Exp. 1A), and only a tendency after 1000 ms (Exp. 1B). In addition, there was facilitation in reading the post-target and ®nal regions of the continuation sentence that con®rmed the predicted event, as well as inhibition when the predicted event was discon®rmed, but no e€ect was observed on the target word itself (Exp. 2). It is concluded that, when the predicted event is highly constrained by the context, predictive inferences are likely to be drawn on-line, but they take time to construct.

Introduction Predictive inferences have received considerable attention in reading research (e.g., Fincher-Kiefer, 1996; Keefe & McDaniel, 1993; Lea, 1995; McKoon & Ratcli€, 1986). They occur when we anticipate the consequences of an event, and they convey information about ``what will happen next.'' For example, ``crashed'' might be a predictive inference made when reading: ``With hardly any visibility the plane quickly approached the dangerous mountain and the passengers began to shout in panic.'' The aim of this study is

M. G. Calvo (&) á M. D. Castillo Departmento de Psicololgia Cognitiva, Universidad de La Laguna, 38205 Tenerife, Spain; Fax: 0034-922-317461; e-mail: [email protected]

to extend prior research on the time course of these inferences. Extant models of inferences in reading (see a review in Graesser, Millis, & Zwaan, 1997) ± i.e., the minimalist hypothesis (McKoon & Ratcli€, 1992, 1995) and the constructionist theory (Graesser, Singer, & Trabasso, 1994; Singer, Graesser, & Trabasso, 1994) ± argue that predictive inferences are unlikely to be drawn on-line. The main reason is that these inferences are not necessary to satisfy the coherence and the explanation assumptions. That is, they are not required to make statements in the text locally coherent (e.g., referential assignment: McKoon & Ratcli€, 1992) or globally coherent in the situation model (e.g., superordinate goals: Graesser et al., 1994), nor to explain why actions, events, and states are mentioned in the text (e.g., causal antecedents: Graesser et al., 1994). Nevertheless, both models admit that these inferences can be drawn on-line if they are supported by well-known information readily available in memory (McKoon & Ratcli€, 1992) and if they are highly constrained by the context (Graesser et al. 1994). However, even in this case, both models assume that predictive inferences are not automatic; rather, they involve elaborative construction of meaning, i.e., post-lexical strategic processes. Therefore, they should take time to develop. Some studies have found no evidence that predictive inferences are generated during reading (Du€y, 1986; Fincher-Kiefer, 1993; Magliano, Baggett, Johnson, & Graesser, 1993; Millis & Graesser, 1994; Potts, Keenan, & Golding, 1988; Whitney, Ritchie, & Crane, 1992). However, on-line predictive inferences have been detected in studies that have somehow followed the implications of the minimalist hypothesis and the constructionist theory (Fincher-Kiefer, 1993, 1994, 1995, 1996; Keefe & McDaniel, 1993; Murray, Klin, & Myers, 1993; Whitney et al., 1992). According to these implications, ®rst, the activated inference concept must be kept foregrounded (Whitney et al., 1992) or in attentional focus (Murray et al., 1993) until the time of test in order to have the relevant information

250

readily accessible in memory. This is especially important because predictive inferences are only temporarily held in working memory (Fincher-Kiefer, 1995, 1996; Keefe & McDaniel, 1993). Secondly, the predicting context must strongly imply one main consequence in order to meet the context constraints implication; otherwise, if there are multiple potential consequences, inferences would not be made because of inecient resource expenditure. Once it has been established that predictive inferences can be made under certain restricted conditions ± i.e., when provided with sucient supportive context ± an important step is to analyze their precise time course (see Graesser et al., 1997), which is the main goal of the present study. Most prior experimental research on predictive inferences has not addressed this issue directly, as temporal parameters were not manipulated. In some studies, participants could read the predicting context in a self-paced mode; furthermore, this context was presented either as a whole text (Whitney et al., 1992), one sentence at a time (Fincher-Kiefer, 1993, 1994; Keefe & McDaniel, 1993; Lea, 1995), or one line at a time (Murray et al., 1993). In addition, there was an interval (generally, of 200±500 ms) between the end of the context and the probe or target word. Presumably, if inferences are detected after such a short interval, this implies that they occur on-line with a very short delay. The problem is that, as the participants could decide when to ®nish reading the context, they actually had all the time they needed to make the inference before the context-probe interval. Therefore, this interval becomes meaningless with regard to the time course of the inference. A procedure in which there is free reading time and in which long context segments are visible at a given time represents natural reading, but it does not allow precise assessment of the time course of inferences. In order to overcome this problem, three methodological strategies have been used: (a) presentation of the last word in the inducing context for a limited time, (b) ®xed-pace presentation of the context by means of Rapid Serial Visual Presentation (RSVP) procedures, and (c) manipulation of the interval between the onset of the last word in the context and the onset of the test word by means of the Stimulus Onset Asynchrony (SOA) technique. With the ®rst approach, Millis and Graesser (1994, Exp. 2) used a word-by-word self-paced presentation in which the last word of the context was exposed for either 440 or 940 ms and the test word appeared 100 ms later. With the second strategy, the context was presented word by word with the same pace for all participants (Calvo & Castillo, 1996; Fincher-Kiefer, 1995, 1996; Magliano et al., 1993; Millis & Graesser, Exp. 1). In some of these studies (Fincher-Kiefer, 1995, 1996), the interval between the onset of the last word in the context and the onset of the target word was 1000 ms. As this interval was not varied for di€erent experimental conditions, it does not allow us to determine the time course of inferences. In contrast, with the third

approach , several studies have combined the ®xed-pace presentation with SOA manipulations (Calvo & Castillo; Magliano et al.; Millis & Graesser). Using this third approach, neither Magliano et al. (1993), with narrative sentences, nor Millis and Graesser (1994), with scienti®c/expository texts, found evidence for predictive inferencing, regardless of the SOAs (from 250±1200 ms in Magliano et al.; from 540±1040 ms, in Millis & Graesser). However, both these studies might not have met the implications of the minimalist and the constructionist models (i.e., well-known information readily available in memory, or strong implication of one highly salient consequence). Thus, Magliano et al. admitted that only 26% of the participants could generate the inference (target word) in question-answering protocols when they were asked to say ``what happened next'' for each predicting context, before the experiment.1 Millis and Graesser indicated that ``idiosyncratic answers that were generated by only one subject were excluded'' (p. 588). Therefore, the availability and context constraints were probably low in both studies. In a similar approach, Calvo and Castillo (1996) used materials which allowed a much higher .67 production score (i.e., 67% of participants explicitly mentioned the word representing the inference) in the ``what happened next'' protocols. In addition, they used either (a) a wordby-word ®xed-pace RSVP procedure to present the context, followed by a target word representing the inference that the participants were to pronounce (Exp. 1), or (b) a word-by word self-paced presentation of the context, followed by a continuation sentence that represented the inference (Exp. 2). The ®ndings revealed (a) facilitation in naming target words after a 1250-ms SOA, but not after a 500-ms SOA (Exp. 1), and (b) facilitation in reading the target word plus a post-target region (spill-over e€ect) of the continuation sentence, but not in reading the target word itself (Exp. 2). The ®ndings converged to suggest that predictive inferences occur on-line, but with delay. However, some limitations in the experiments conducted by Calvo and Castillo (1996) can be overcome in order to obtain further knowledge on the process of predictive inferencing in the present study.2 First, the use of only two SOAs (500 and 1250 ms) provided a limited range to examine the time course of inferences; with two more SOAs (1000 and 1500 ms) in the present study, the 1

However, in Magliano et al., with a comparable mean production likelihood score of .30, causal antecedent inferences, but not causal consequence (predictive) inferences, showed evidence of on-line encoding. 2 New features in the present study, in comparison with Calvo and Castillo's study, include: (1) slower RSVP rate (100% vs 150%); (2) new SOAs (500±1000±1500 vs 500±1250 ms); (3) enhanced constraints in the predicting contexts (81±82% vs 67%); (4) control of word-based priming (subjective and objective ± i.e., dictionary ± word association vs subjective word association); (5) control contexts (explicit vs no context); and (6) post-target and ®nal regions of the continuation sentences for the con®rming and the discon®rming versions (identical vs di€erent).

251

assessment provides us with greater precision. Secondly, the RSVP rate of context presentation that Calvo and Castillo used might have been sucient to understand explicit information, but not to construct implicit information (i.e., inferences) under the 500-ms SOA condition; a slower rate in the present study is relevant to answer this question. Thirdly, in the inference condition of Calvo and Castillo's study, the target words were preceded by a predicting context; in contrast, in the control condition, there was no prior context. While this no-context condition has some advantages (see Calvo & Castillo, p. 64), the explicit non-inference contexts in the present study seem more appropriate. Finally, the posttarget and the ®nal regions (in the continuation sentences following the predicting/control contexts) were previously di€erent for the version con®rming the inferred event and for the discon®rming version. As the facilitation e€ects occurred when reading the post-target region, it has now become crucial to make this (and the ®nal) region equivalent for both versions (see Appendix), in order to demonstrate delayed inferencing. In the experiments reported in this paper, we presented sentence contexts either predictive (inference) or non-predictive (control) of a highly probable consequence. These contexts were followed by a target word to be named (Exps. 1A, B, C) or a continuation sentence to be read (Exp. 2) that represented a concept which con®rmed or discon®rmed the predicted event. If these inferences are generated on-line, then facilitation in processing (i.e., naming/reading) the con®rming target word/sentence after the predicting context should be observed, relative to when the targets are presented after the control context and in comparison with the discon®rming word/sentence. In order to determine the time course of these inferences, in Exps. 1A, B, C, the interval (SOA) between the last word in the context and the target word was manipulated (500 vs 1000 vs 1500 ms, respectively); and in Exp. 2, reading times for the successive parts (target word, post-target region, and ®nal region) of the continuation sentence were recorded. If predictive inferences are induced immediately and automatically, then the facilitation e€ect should occur with the shorter SOA (Exp. 1A) and on the target word itself (Exp. 2); in contrast, if these inferences are strategically induced with delay, then facilitation should only be observed with the longer SOAs (Exps. 1B or 1C) and in the post-target or ®nal regions (Exp. 2).

Experiments 1A, 1B, and 1C A naming task was used, with either a 500-ms (Exp. 1A), a 1000-ms (Exp. 1B), or a 1500-ms (Exp. 1C) SOA between the onset of the last word in a predicting or a control context and the onset of a target word. Participants were presented with the sentences word by word at a ®xed pace, using an RSVP procedure, and pronounced the target word as soon as it appeared.

The new SOAs allowed us to make a more precise analysis of the time course of inferences around the critical 1250-ms point in which Calvo and Castillo (1996) found a facilitation e€ect. In addition, we increased by 50% the exposure duration of words, thus slowing the RSVP rate. The reason was that, the quicker the RSVP rate, the more dicult it has been to detect inferences (Magliano et al., 1993; Millis & Graesser, 1994, p. 592). Accordingly, we wanted to ascertain whether the absence of e€ects under the 500-ms SOA condition in Calvo and Castillo was due to too a fast rate, and whether inferences could also occur with this short SOA if the readers were provided with more time during the presentation of the context.

General methodology for Exps. 1A, B, and C The three studies in the series3 di€ered only in the SOA. The last word in the context, which was named the pre-target word, was always visible for 450 ms in the three experiments. Accordingly, the di€erence between them consisted of a blank interval between the end of the pre-target word and the onset of the target word. This interval was only 50 ms in Exp. 1A, 550 ms in Exp. 1B, and 1050 ms in Exp. 1C.

Participants. In each experiment, 48 di€erent Spanish psychology undergraduates participated for course credit.

Materials. Forty short Spanish passages were used (see Appendix). Each passage was composed of (a) one predicting sentence context, (b) one non-predicting, control sentence context, (c) one target word (a verb) that con®rmed the predicted event and represented the inference concept, and (d) one target word (a verb) that discon®rmed the predicted event and was unrelated to the inference concept. At the end of the predicting/control contexts, a pre-target word was added (a noun ± e.g., the lift, in the ®rst example of the Appendix ± which was the subject of the action ± e.g., went up ± represented by the target word). On each trial, each participant was presented with either a predicting or a control context of each passage, followed by either a con®rming or a discon®rming target word. Several preliminary studies and analyses were conducted to validate the materials before the experiments. First, we controlled word-based priming, i.e., activation of the target word due to semantic associations with individual words in the context (Keenan, Golding, Potts, Jennings, & Aman, 1990, p. 296). Both by means of a comprehensive dictionary of synonyms and associates in Spanish (Ortega, 1990) and by means of subjective ratings made by independent judges, we identi®ed those words in the predicting contexts that could have any, though remote, relationship with the target words. Subsequently, we included those words in the respective control contexts. With this procedure, the predicting and the control contexts shared 42.2% of content-words (see Appendix). These words were rearranged in the control contexts, so that they did not suggest the inference. 3

We did not manipulate SOA as a repeated-measures variable because it could have been potentially disruptive to comprehension: It might have increased uncertainties in the reader of not knowing about the timing of the test word. Accordingly, these experiments will be treated as a series, rather than a single one or as three independent studies. The SOA conditions were run in a ®xed order.

252 Subsequently, in a sentence completion study (see Magliano et al., 1993), we ensured that the predicting context could actually induce the presumed inferences, and that these could be expressible in language. Altogether, 104 undergraduates were presented with the contexts and asked to write the ®rst word that came to mind in order to complete the contexts. Thus, participants were expected to indicate ``what happens next,'' which characterizes predictive inferences. The one-word predictions (or close synonyms) served as targets. Two targets were selected for each context: con®rming (which represented highly likely consequences) and discon®rming (plausible, but unlikely consequences). The former were mentioned by 82.4% of participants after the predicting contexts and by 8% after the control contexts; for the latter, the respective scores were 2% and 3%. Finally, we checked that the predicting and the control contexts were equivalent in length (both contexts, M ˆ 20.6, number of words). In addition, the con®rming target words did not di€er from the discon®rming words either in number of characters (M ˆ 5.97 vs 6.00, respectively) or word frequency (M ˆ 33.2 vs 35.4; Alameda & Cuetos, 1995). Design. A 2 (Context: predicting vs control) ´ 2 (Target: con®rming vs discon®rming) within-subjects factorial design was used. In the predicting conditions (both con®rming and discon®rming), the target word was preceded by a sentence suggesting a highly likely consequence. In the control conditions, the target word was preceded by a sentence non-predictive of any particular consequence but compatible with both the con®rming and the discon®rming target words. Four lists of materials were constructed, each consisting of ten predicting contexts followed by con®rming targets, ten control contexts followed by con®rming targets, ten predicting contexts followed by discon®rming targets, and ten control contexts followed by discon®rming targets. The assignment of targets to the predicting or the control context was reversed across the lists so that a given participant saw a particular context and target only once. Each participant received one of the lists, with 40 trials in random order. Apparatus and procedure. The materials were displayed on a computer screen. Stimulus presentation and response collection were controlled by ALR-486 computers. Sentences were presented word by word with a ®xed-pace RSVP procedure. In Calvo and Castillo (1996), each word was presented for 300 ms, plus 25 ms per letter, and there was a 50-ms blank interval between successive words. In the present experiments we increased the word exposure time by 50%: Each word was shown for 450 ms, plus 37.5 ms per letter ± e.g., a six-letter word such as target would appear on the screen for 450 + (37.5 ´ 6) ˆ 675 ms ± with the blank interval between words remaining at 50 ms. A trial included one sentence context and one target word. Each trial began when the participant pressed the space-bar; 500 ms later, the words began to appear (and disappear) on the center of the screen, one at a time, according to the temporal parameters men-

Table 1 Mean naming latencies (in ms) and SD for target words in the con®rming and discon®rming versions, as a function of stimulus onset asynchrony (SOA) and context

SOA and context

500-ms SOA (Exp. 1A) Predicting Control 1000-ms SOA (Exp. 1B) Predicting Control 1500-ms SOA (Exp. 1C) Predicting Control

tioned above. After the SOA at the end of the context, the target word appeared ¯anked by asterisks (e.g., ** went up **). Participants had been warned that the asterisks were the cue to pronounce the word in between as quickly as they could, trying not to make errors. A microphone attached to the participant's chin and connected to the computer registered the naming responses. The target word remained on the screen until the participant made a vocal response. Latencies were timed in milliseconds from the onset of the target word to the onset of the participant's response. Then a comprehension question was presented on the screen. It consisted of a recognition test which reworded (by means of synonyms or paraphrases) the content of the context. Participants responded by pressing one of two keys (Yes or No). After that, the instruction to begin a new trial appeared on the screen. There were 12 practice trials.

Results For this whole series of studies, pronunciation errors occurred less than 2% of the time and did not vary signi®cantly as a function of experimental conditions. Comprehension accuracy was 88.0% (SD ˆ 7.42), and it was equivalent for all conditions. Mean latencies for correctly pronounced target words are shown in Table 1. They were analyzed in a 2 (Context) ´ 2 (Target) ANOVA, both by subjects (F1) and by items (F2). We used two criteria to deal with outliers: (a) First, reaction times that were above 2000 ms or below 300 ms were replaced by scores of 2000 or 300, respectively, which represented 0.89% of total scores; (b) second, reaction times that were still above or below 2.5 standard deviations (SD) from the mean were replaced by the participant's mean score plus or minus 2.5 SD in that experimental condition; this procedure a€ected 1.2% of scores. Exp. 1A. Naming times were faster in the predicting context conditions than in the control conditions, F1 < 1.0; F2(1, 78) ˆ 4.67, MSE ˆ 2,933, p < .05, and faster for the con®rming than for the discon®rming target words, F1(1, 47) ˆ 35.53, MSE ˆ 1,278, p < .0001; F2(1, 78) ˆ 6.44, MSE ˆ 4,883, p < .025. Exp. 1B. Naming times were faster for the con®rming than for the discon®rming target words, F1(1, 47) ˆ 41.85, MSE ˆ 2,001, p < .0001; F2(1, 78) ˆ 16.23, MSE ˆ 4,323, p < .001. The context ´ target interac-

Con®rming version

Discon®rming version

M

SD

M

SD

688 696

100 103

721 724

111 120

648 665

109 101

699 698

107 124

645 687

99 106

722 716

133 126

253 Fig. 1 Estimated activation of the target concepts representing the predicted consequence (con®rming) or an alternative nonpredicted consequence (discon®rming), as a function of stimulus onset asynchrony (SOA), with naming time difference between the primed and the unprimed conditions, i.e., predicting ) control

tion was not signi®cant, in spite of a trend, F1(1, 47) ˆ 1.86, MSE ˆ 2,025, p ˆ .17; F2(1, 78) ˆ 2.16, MSE ˆ 1,458, p ˆ .14. Exp. 1C. Naming times were faster following the predicting context than following the control context, F1(1, 47) ˆ 7.18, MSE ˆ 2,159, p < .01; F2(1, 78) ˆ 6.82, MSE ˆ 2,531, p < .01, and faster for the con®rming than for the discon®rming words, F1(1, 47) ˆ 34.91, MSE ˆ 3,846, p < .0001; F2(1, 78) ˆ 17.77, MSE ˆ 3,611, p < .0001. However, there was a context ´ target interaction, F1(1, 47) ˆ 9.14, MSE ˆ 2,948, p < .01; F2(1, 78) ˆ 9.19, MSE ˆ 2,531, p < .01. Tests for simple e€ects indicated that, for the con®rming words, naming latencies were faster after the predicting context than after the control context, F1(1, 47) ˆ 30.15, MSE ˆ 1,382, p < .0001; F2(1, 78) ˆ 15.93, MSE ˆ 2,531, p < .001; in contrast, for the discon®rming words, there was no di€erence as a function of context (both Fs < 1.0). This suggests that inferences were made in the predicting-and-con®rming condition. Combined analysis of Exp. 1A, B, and C. Data from the three experiments were combined in a 3 (SOA) ´ 2 (Context) ´ 2 (Target) ANOVA, in order to see the effects across the SOA factor. There were main e€ects of context, F1(1, 141) ˆ 8.65, MSE ˆ 1,846, p < .01; F2(1, 234) ˆ 12.89, MSE ˆ 2,308, p < .001, and target, F1(1, 141) ˆ 106.00. MSE ˆ 2,379, p < .0001; F2(1, 234) ˆ 37.75, MSE ˆ 4,272, p < .0001, with the same meaning as in the single experiments. They were quali®ed by a context ´ target interaction, F1(1, 141) ˆ 8.53, MSE ˆ 2,372, p < .01; F2(1, 234) ˆ 7.42, MSE ˆ 2,308, p < .01. The lack of a clear threeway interaction, F1(2, 141) ˆ 2.30, MSE ˆ 2,371, p ˆ .10; F2(2, 234) ˆ 2.08, MSE ˆ 2,308, p ˆ .12, is probably due to the fact that in Exp. 1B there was a trend in the same direction as in Exp. 1C. Accordingly, we compared Exps. 1A and 1C, leaving out the intermediate values of 1B. A three-way interaction emerged, F1(1,

94) ˆ 4.04, MSE ˆ 2,546, p < .05; F2(1, 156) ˆ 4.64, MSE ˆ 2,689, p < .05, which corroborated that there was facilitation in naming con®rming target words following the predicting context in the longer (1500-ms) SOA, but not in the shorter (500-ms) SOA. This di€erential contribution of SOA can be illustrated in terms of activation scores (i.e., naming times in the control condition subtracted from those for the same target word in the predicting condition; see Fig. 1). These scores show the degree of facilitation in naming the target word and are assumed to re¯ect activation of the inference concept4 (see Graesser et al., 1994; Till, Mross, & Kintsch, 1988). Figure 1 shows that the activation of concepts representing the inference is close to the baseline in the 500-ms SOA, increases in the 1000-ms SOA (but does not reach statistical signi®cance), and becomes greatest in the 1500-ms SOA condition. Discussion Facilitation (i.e., shorter naming latencies) for the words that con®rmed the predicted event after the inducing context relative to the control context (a) in the 1500-ms SOA, (b) as a similar tendency in the 1000-ms SOA, (c) but not in the 500-ms SOA, implies that predictive inferences (a) were generally drawn by 1s after the end of the context, (b) were probably initiated, but not completed, 550 ms after the context, and (c) were not generated immediately (50 ms) after the context. Previous research using the naming task has reported discrepant ®ndings. Potts et al. (1988) and Whitney et al. (1992) found no evidence for predictive inferences, 4

The di€erence score (predicting ) control) indicates to what extent the target concept is activated after reading the predicting context. Positive scores (longer naming latencies) reveal inhibition; negative scores (shorter latencies) show facilitation. If a predictive inference is drawn from the predicting context, then quicker naming latencies will be observed in this condition than in the control condition.

254

whereas Calvo and Castillo (1996), Keefe and McDaniel (1993), Lea (1995), and Murray et al. (1993) did. As naming is not contaminated by reconstructive processes at the time of test (Keenan et al., 1990), a backward context-checking explanation can be ruled out. Nevertheless, most previous studies could not determine the time course with precision because (a) the predicting context was presented with a self-paced procedure, either one sentence (Keefe & McDaniel; Lea), or one line (Murray et al.) at a time, and (b) the participants could decide when to ®nish reading the context and make the probe appear. With such large text units and free reading time, it was not possible to determine when the inference occurred. In contrast, in the present study, the RSVP procedure and the manipulation of the SOA provided greater control over the unit and the temporal presentation. Using di€erent materials, Calvo and Castillo detected predictive inferences with a 1250-ms SOA, but not with a 500-ms SOA, which is consistent with the present ®ndings. In addition, the present study provides further knowledge regarding two temporal factors, i.e., the possible threshold for these inferences after the context (1000-ms SOA) and the role of the time available during context processing (slower RSVP rate) (see General discussion). Nevertheless, both the RSVP procedure and the naming task have been criticized. The former may truncate the natural self-paced reading comprehension process (see Millis & Graesser, 1994). The latter may lack the sensitivity to detect processes at a level of sentence analysis, higher than the word-access level (Fincher-Kiefer, 1993, p. 120; Whitney et al., 1992). Therefore, it is possible that (a) predictive inferences could be drawn immediately or earlier if the reader had control over the pace of presentation, or that (b) they could have been generated, but were not observed in the shorter SOA conditions because naming within these intervals could be sensitive to perceptual rather than conceptual priming. Accordingly, in the next experiment we used relatively natural reading conditions, and the task to measure inferences involved semantic processing.

Experiment 2 A self-paced reading, moving-window (e.g., de Vega, Carreiras, Calvo, & Alonso, 1990; Just, Carpenter, & Woolley, 1982) paradigm was used which allowed the detection of concept activation at several points during processing. In contrast to Exp. 1, in which only a target word was presented after the predicting/control contexts, in Exp. 2 a whole continuation sentence was included. Reading times were collected for a pre-target region, the target word, the post-target region, and the last region in this sentence. If inferences were drawn when reading the predicting context, then reading subsequent words or sentences

con®rming, or congruent with, the inference concept should be speeded up, whereas reading time should be slowed down for discon®rming, incongruent information. Furthermore, reading times at di€erent points in the sentence following the context can reveal the time course of the inferences. Presumably, facilitation in reading the target word that represents the inference concept would indicate that predictive inferences involve immediate/automatic processes. In contrast, if the e€ects occurred in a post-target region, or the last word of the sentence, they could be attributed to post-lexical processes involving elaboration or integration and could reveal delayed, strategic inference construction (see Klin, 1995; O'Brien, Shank, Myers, & Rayner, 1988). Method Participants. Sixty-four psychology undergraduates participated for course credit. Materials. Sixty-four passages were presented in Spanish (40 of them were equivalent to those used in Exp. 1, and there were 16 additional passages of the same type: length, shared content words, etc.). Each passage was composed of (a) one predicting sentence context, (b) one non-predicting, control context, (c) one continuation sentence in which a target word con®rmed the predictable event, and (d) one continuation sentence in which a target word discon®rmed the predictable event. Besides this target word, the continuation sentences included a pre-target region (1±2 words), a post-target region (2±3 words) and a ®nal region (1±2 words). These regions were the same for both the con®rming and the discon®rming versions; the only di€erence was the target word (see Appendix). Design. A 2 (Context sentence: predicting vs control) ´ 2 (Continuation target sentence: con®rming vs discon®rming) within-subjects factorial design was used. Each participant received either a predicting (trials: N ˆ 32) or a control (N ˆ 32) context of each passage, followed by either a con®rming (N ˆ 16 after a predicting context + 16 after a control context) or a discon®rming version (N ˆ 16 + 16) of the continuation sentence (see Exp. 1 for further details). Apparatus and procedure. Sentences were presented on a computer screen by means of the self-paced moving-window technique (de Vega et al., 1990; Just et al., 1982). The length of the moving window varied from 1±4 words. The target word was always displayed alone. The rest of the words were grouped according to linguistic criteria (i.e., syntactic relatedness: e.g., article-noun-verb, noun-adjective, verb-adverb, etc.). Stimulus presentation and response collection were controlled by ALR±486 computers. A trial included one context sentence and one continuation sentence. Each trial began when the participant pressed the spacebar; then the screen was ®lled with masks corresponding to the letters of the sentences. With each keypress, 1±4 words were typed on the screen, replacing the masks corresponding to those words, and the previous word or words were replaced with a mask, and so on until the end of both sentences. As in natural self-paced reading on a printed page, the eye moved from left to right and made return sweeps when advancing to the next line, though regressive ®xations to previous sentence segments were not possible. The reading time for a word (or string of words) was de®ned as the interval between successive keypresses. After having

255 read the two sentences in each trial, a comprehension (recognition test) question was presented.

Results Accuracy on comprehension questions was 85.0% (SD ˆ 8.6), and it was equivalent under all the experimental conditions. Reading times that were above or below 2.5 SD from the mean in each experimental condition were replaced by the participant's mean score plus or minus 2.5 SD (1.1% of scores were a€ected). Several ANOVAs were conducted on reading times, both by subjects (F1) and items (F2). Continuation sentence: Total reading times Initially, a 2 (Context) ´ 2 (Target) ANOVA was performed on the total reading times of the continuation sentences. Reading times were faster for the con®rming than for the discon®rming sentences, F1(1, 63) ˆ 59.15, MSE ˆ 56,567, p < .0001; F2(1, 126) ˆ 12.85, MSE ˆ 234,700, p < .001. However, this di€erence depended on which type of context was presented, F1(1, 63) ˆ 17.88, MSE ˆ 44,448, p < .0001; F2(1, 126) ˆ 3.88, MSE ˆ 224,099, p ˆ .05. Simple e€ects tests indicated that sentences involving con®rmation of the predicted consequence were read faster after the predicting context than after the control context, F1(1, 63) ˆ 28.27, MSE ˆ 38,649, p < .0001; F2(1, 126) ˆ 4.95, MSE ˆ 224,099, p < .05; in contrast, for the discon®rming sentences, there was no di€erence as a function of context (all Fs < 1.0). (See mean scores in Table 2.)

189) ˆ 187.4, MSE ˆ 67,620, p < .0001; F2(3, 378) ˆ 392.2, MSE ˆ 39,068, p < .0001, which is not theoretically relevant, because of di€erent region length. A target e€ect and a context ´ target interaction emerged again, with the same meaning as in the previous (total reading times) analysis. In order to decompose a three-way interaction, F1(3, 189) ˆ 11.16, MSE ˆ 5,131, p < .0001; F2(3, 378) ˆ 3.79, MSE ˆ 16,731, p ˆ .01, separate ANOVAs were conducted for each region. (a) Pre-target region. No reliable e€ects were found. (b) Target word. Target words were read faster following the predicting context than following the control context, F1(1, 63) ˆ 8.54, MSE ˆ 2,177, p < .01; F2(1, 126) ˆ 4.52, MSE ˆ 6,027, p < .05, and con®rming targets were read faster than discon®rming targets, F1(1, 63) ˆ 31.87, MSE ˆ 3,331, p < .0001; F2(1, 126) ˆ 7.94, MSE ˆ 12,706, p < .01). There was no interaction (all Fs < 1.0).

Continuation sentence: Partial reading times

(c) Post-target region. Reading times were faster for the con®rming than for the discon®rming version, F1(1, 63) ˆ 33.41, MSE ˆ 4,681, p < .0001; F2(1, 126) ˆ 3.51, MSE ˆ 44,908, p ˆ .06. A context ´ target interaction, F1(1, 63) ˆ 11.54, MSE ˆ 4,748, p < .01; F2(1, 126) ˆ 5.09, MSE ˆ 12,105, p ˆ .025, revealed that the participants took less time to read the post-target region in the con®rming version after the predicting context than after the control context, F1(1, 63) ˆ 15.68, MSE ˆ 3,892, p < .001; F2(1, 126) ˆ 3.79, MSE ˆ 12,106, p < .05; in contrast, in the discon®rming version, though di€erences showed an opposite trend, they were not statistically signi®cant, F1(1, 63) ˆ 1.03, MSE ˆ 6,860, p > .10; F2(1, 126) ˆ 1.55, MSE ˆ 12,106, p > .10.

A subsequent 4 (Region) ´ 2 (Context) ´ 2 (Target) ANOVA explored e€ects across the regions (see Table 2). Reading times varied for each region, F1(3,

(d) Last word. Reading times were faster for the con®rming than for the discon®rming version, F1(1, 63) ˆ 40.00, MSE ˆ 23,229, p < .0001; F2(1,

Table 2 Mean reading times (in ms) and SD for the pre-target region, the target word, the post-target region, and the ®nal region of the continuation sentence, in the con®rming and discon®rming versions as a function of context

Region and context

Pre-target Predicting Control Target Predicting Control Post-target Predicting Control Final Predicting Control Total Predicting Control

Con®rming version

Discon®rming version

M

M

SD

SD

728 755

128 127

751 768

151 147

593 617

106 107

641 651

133 135

718 762

152 181

797 782

191 182

1033 1121

283 309

1223 1172

359 346

3071 3255

587 612

3412 3373

709 704

256 Fig. 2 Estimated activation of the target concepts in the con®rming and discon®rming versions across the di€erent regions of the continuation sentence with reading time di€erence between the primed and the unprimed conditions, i.e., predicting ) control

126) ˆ 8.90, MSE ˆ 91,747, p < .01. However, this di€erence depended on the context, F1(1, 63) ˆ 19.67, MSE ˆ 15,786, p < .0001; F2(1, 126) ˆ 4.35, MSE ˆ 77,970, p < .05. Simple e€ects tests indicated that, for the con®rming version, the last word was read faster after the predicting context than after the control context, F1(1, 63) ˆ 18.42, MSE ˆ 13,645, p < .0001; F2(1, 126) ˆ 3.51, MSE ˆ 77,971, p ˆ .06; in contrast, for the discon®rming version, the di€erences were in the opposite direction, with longer reading times following the predicting context than following the control context, F1(1, 63) ˆ 4.04, MSE ˆ 20,336, p < .05; F2(1, 126) ˆ 1.16, MSE ˆ 77,971, p > .10. Activation scores across the regions. As in Exp. 1, activation scores were computed (see Fig. 2). As they were obtained from the raw scores (i.e., predicting minus control reading times; see Table 2), the statistical e€ects were the same as those described in (a) to (d). Discussion As a summary of ®ndings, Fig. 2 indicates that, whereas (a) the pre-target region and the target word were not a€ected, (b) the predicting context, relative to the control context, facilitated (speeded up) reading of the posttarget and ®nal regions of the continuation sentences that con®rmed the predicted event, but (c) it inhibited (slowed down) processing of the ®nal region in the sentences that discon®rmed the predicted event. These results provide converging, supportive evidence for those reported in Exps. 1A, B, C. First, predictive inferences occur on-line, as revealed by the signi®cant reduction in the amount of time to read the continuation sentences speci®cally when they matched the information induced by the context (i.e., con®rmation of the predicted events). Secondly, predictive inferences are generated or completed with some delay, as the e€ects did not occur on the target word, but spilled over to the

post-target and ®nal regions. Millis, Morgan, and Graesser (1990) also found facilitation in reading sentences that matched previous expectations. However, as reading times were measured for a sentence as a whole, Millis et al. (1990) could not ascertain whether there was facilitation at any particular point within the sentence. Our multiple and sequential measurement of individual parts of the sentence allowed a more precise assessment of the time course. Particularly interesting are the inhibitory e€ects, as they have not been investigated previously in predictive inferencing. Inhibition has been observed with related phenomena, either explicit descriptions (e.g., Albrecht & O'Brien, 1993) or causal/bridging inferences (Klin, 1995). Thus, reading times for a continuation sentence were longer when an earlier description was inconsistent with it (Albrecht & O'Brien), and participants were slow to read a sentence following a causal coherence break that contradicted the intended inference (Klin). Presumably, this reveals that alternative, noncompatible concepts are inhibited, regardless of whether the source is explicit information or an inference. In the present study, these e€ects represent additional knowledge in comaprison with Calvo and Castillo's (1996) study, in which inhibition was not observed. By making the post-target/®nal regions identical for the con®rming and the discon®rming version in the present study, unlike Calvo and Castillo, we have probably gained in sensitivity to inhibition. Moreover, our ®ndings coincide with those of Klin's, in that inhibition was detected with the reading time paradigm, but not with the naming technique. Nonetheless, the reason for this di€erence is not clear. Inhibitory mechanisms should be the focus of further research on inferences. The delayed nature of facilitation and inhibition has several interpretations. It is consistent with one of the central concepts of the minimalist hypothesis (McKoon & Ratcli€, 1990, 1992), according to which predictive inferences are gradually constructed. At the ®rst en-

257

counter with the inducing information, inferences would not be encoded in an all-or-none fashion, but tenuously or minimally. Readers would not make precise and de®nite interpretations of the predicting sentences at the very moment they read them. Relatively general inference concepts would be activated instead, and the complete inference would be delayed until clear evidence was found later. Only when the participant has read the con®rming target word would he/she complete the inference. As a consequence, when the con®rming word appeared, readers would have to complete and/or make speci®c their previous partial or general inference. This completion and/or re®nement process would require some additional time when processing the con®rming target word. This would explain why there was no facilitation when reading the target word. However, once the inference was completed (i.e., after having read the target word), processing of subsequent consistent information would be facilitated, whereas processing of inconsistent information would be inhibited. This is precisely what happened in the post-target and ®nal regions of the continuation sentence. Alternatively, the delay in facilitation and inhibition could be accounted for in terms of integration processes (see Haberlandt & Graesser, 1990) or even straight priming from the target words. Thus, it can be argued that the post-target and ®nal regions are more semantically coherent with the predicting context and the con®rming target word than with the control context and the discon®rming target word. Accordingly, facilitation and inhibition in these regions would re¯ect easiness of integration with the prior context/target. However, the source of this cohesiveness can be either the explicit information in the prior context/target or, as we assume, the implicit information (i.e., inference) that the predicting context suggests and that the con®rming target word completes.5

5

If the main source were the explicit information, then integration processes would have been a€ected similarly for the con®rming and the discon®rming versions of the post-target/®nal regions after the predicting context (which was not the case). The reason is that the context was the same for both versions, which were also identical themselves. Only the target word di€ered. Nevertheless, it could be further argued that it is precisely the di€erent target word that was responsible, on its own, for a priming e€ect on the post-target/®nal regions. However, if this had been the case, then integration should have been comparable regardless of whether the target word came after the predicting or the control contexts. The reason is that the target word was the same in both conditions. Accordingly, our position is that the possible easiness (or diculty) to perform integration processes on the post-target/®nal regions was actually dependent on the prior implicit information source ± i.e., the inference process initiated when reading the context and completed when reading the target word. In this process, rather than having a straight priming e€ect, the function of the target word would be to complete (con®rming) or to unmake (discon®rming) the prior partial inference. Because of this, the integration processes on the post-target/®nal regions would be performed more or less readily.

General discussion We have found support for the hypothesis that predictive inferences are drawn on-line, but with delay, during reading. A word (Exp. 1) or a sentence (Exp. 2) representing an event predicted by a previous context was named (Exp. 1) or read (Exp. 2) more quickly following this inducing context than following a control context. In contrast, when the word or sentence represented alternative events, then (a) naming latencies were not affected by the context (Exp. 1), and (b) reading times were slower following the predicting context (Exp. 2). These e€ects occurred only 1050 ms after the end of the context (1500-ms SOA); there was a similar (but nonsigni®cant) trend 550 ms after the context (1000-ms SOA); and they were not observed immediately (50 ms) after the context (500-ms SOA) (Exp. 1). Furthermore, they occurred when the post-target and the ®nal regions of the continuation sentence were read, but not when the target word itself was read (Exp. 2). Predictive inferences are likely to be drawn when two types of criteria are met by the text materials and their presentation: (a) if the predicted inference is highly constrained by the context, as speci®ed by the constructionist theory (Graesser et al., 1994), and (b) if the relevant information is readily accessible in memory, as proposed by the minimalist hypothesis (McKoon & Ratcli€, 1992), and the inference concept is kept activated until the time of test, as speci®ed by the temporary activation conceptualization (Fincher-Kiefer, 1995, 1996; Keefe & McDaniel, 1993). With (a) validation of materials in the preliminary, norming studies, and (b) the presentation of the target word as a continuation from the context, we have presumably met these criteria. Predictive inferences take time to develop or to complete. They are not drawn automatically or immediately, but need some hundred milliseconds after the reader has comprehended the information that is explicit in the context. Unlike bridging or backward inferences, which can be made in less than 500 ms after the context (Magliano et al., 1993; Millis & Graesser, 1994), predictive inferences are comparable to thematic inferences (i.e., contextually relevant meanings of ambiguous words: Till et al., 1988) in that both require more than 500 ms. Furthermore, though predictive inferences can be detected 800 ms after the end of the inducing context (Calvo & Castillo, 1996, using a 1250-ms SOA; see also Fincher-Kiefer, 1995, 1996), they are not detected a little earlier (550 ms after the context, i.e., 1000-ms SOA). Therefore, a critical point in the time course to draw predictive inferences seems to be located between 550 and 800 ms after the context. The spill-over e€ects found in Exp. 2, in which facilitation occurred when reading the post-target and ®nal regions, but not in the target word, reinforce this conceptualization. However, in this case the time course cannot be determined with the same precision. Furthermore, the delay ± though convergent ± may not have strictly the same meaning in both

258

paradigms. The reason is that, when reading the continuation sentence (Exp. 2), time is ``®lled'' with additional information (post-target and ®nal regions), whereas time is ``empty'' when naming the target word (Exp. 1). The additional information could have induced readers to wait and see before risking an inference earlier. Nevertheless, there are a number of quali®cations regarding the time course of predictive inferences. First, the length of the required delay could depend to some extent on individual di€erences (e.g., in working memory capacity or in prior knowledge). Secondly, the delay could also vary as a function of the degree of context constraints; thus, some predictive inferences could be made faster than others. Both these issues are worthy of further research. Thirdly, the time required after the context might depend on the time available during context processing. Experiment 1 in the present study provides a negative answer to this question. Thus, even though the rate of context presentation was 50% slower in the present experiments than in Calvo and Castillo (1996), no inferences were drawn in the 500-ms SOA condition. This suggests that predictive inferences need additional time after the end of the context, regardless of the additional time available during the presentation of the context. This during-context time does not seem to compensate for the lack of time after the context, nor to accelerate the time course of these inferences. A rigorous test of this hypothesis would imply orthogonal manipulation of RSVP rate and SOA in future research. In general, the time-course ®ndings are consistent with the idea that predictive inferences involve post-access elaboration of information implicit in the context, which requires additional time beyond the processing of the explicit information (Kintsch, 1988). They are also consistent with a graded view on the construction of inferences, as proposed by the minimalist hypothesis (McKoon & Ratcli€, 1990, 1992). While predictive inferences could be initiated earlier in a minimal form, only later would they reach a strength level or a degree of re®nement that could be expressible and detected in a linguistic form.6 Acknowledgements This research was supported by Grant PS940079 from the DGICYT, Spanish Ministry of Education and Science. We are grateful to Arthur C.Graesser and Manuel Carreiras for their comments on an earlier version of this article, to Manuel GonzaÂlez Mauricio for the elaboration of the software used for the present study, and to Alejandro JimeÂnez for his assistance in data collection.

Appendix Examples of experimental passages (as translated into English). Predicting context

On arriving at the &entrance of the &building where he had his &apartment, Francisco entered the &lift, pressed the button, and it began to &move. The lift/ Control context

Francisco bought an &apartment in the &building with the luxurious &entrance, but he hardly used the &lift to &move from one ¯oor to another. The lift/ Con®rming target word and continuation sentence

Exps. 1A, B, and C (henceforth: Exp. 1): went up (*) Exp. 2: went up (*)/without/any noise./ Discon®rming target word and continuation sentence

Exp. 1: went down (*) Exp. 2: went down (*)/without/any noise./ Predicting context

Lola was eager to &know the end of the &novel, so she lay down comfortably and opened it at the &page she had reached the last time. Lola/ Control context

Lola &knew the author of the &novel whose photo appeared on the ®rst &page of the newspaper, so she phoned to congratulate her on her success. Lola/ Con®rming target word and continuation sentence

Exp. 1: read Exp. 2: read/the pages/that remained./ Discon®rming target word and continuation sentence

Exp. 1: counted Exp. 2: counted/the pages/that remained./ Predicting context

Three days before the &examination the &student went to the &library, looked for a separate table and opened his ¬ebook. The student/

6

Nevertheless, a problem with the notion that the generation of a predictive inference might be graded is that we could confuse the gradation of one inference with two or more di€erent inferences. Thus, in the typical example of ``the actress who fell from a 14th story window,'' then ``something bad happens to her,'' ``she is hurt,'' and ``she dies'' may be completely independent inferences rather than di€erent grades of the same inference.

Control context

The &student, who was very tired after ®nishing his &examination, forgot his ¬ebook and left it in the &library. The student/

259 Con®rming target word and continuation sentence

Exp. 1: studied Exp. 2: studied/for an hour,/approximately./

terisks (*) indicate that there was only one word in Spanish. The 40 (Exp. 1) and 64 (Exp. 2) passages, each with its four (context by resolution) versions, are available in Spanish

Discon®rming target word and continuation sentence

Exp. 1: slept Exp. 2: slept/for an hour,/approximately./ Predicting context

While Maria &walked &barefoot over the rocks, she &put her foot &down, without realizing, on a piece of &glass which had been left on the &¯oor. Maria/ Control context

In order to avoid &putting her dirty shoes &down on the &¯oor, Maria &walked &barefoot to the &glass display cabinet to place the present in it. Maria/ Con®rming target word and continuation sentence

Exp. 1: cut herself (*) Exp. 2: cut herself (*)/with pain/ at that moment./ Discon®rming target word and continuation sentence

Exp. 1: slipped Exp. 2: slipped/ with pain/ at that moment./ Predicting context

When Aida saw her &father in the &airport, she run up to him, and he &bent down over the &child. The father/ Control context

Before the trip with Aida, the &father &bent down to show a scale model of the &airport to his &child. The father/ Con®rming target word and continuation sentence

Exp. 1: embraced Exp. 2: embraced/ the child/ as usual./ Discon®rming target word and continuation sentence

Exp. 1: spoke Exp. 2: spoke to (*)/ the child/ as usual./ Note: Target words are in bold letters. Ampersands (&) and slashes did not appear in the stimuli. Ampersands indicate the content words shared by the predicting and the control context, to control for word-based priming. Slashes show the regions (pre-target, target, post-target, and ®nal) in the continuation sentence. As-

References Alameda, J. R., & Cuetos, F. (1995). Diccionario de frecuencias de las unidades linguÈõÂsticas del castellano [Dictionary of lexical frequencies of Spanish linguistic units]. Oviedo, Spain: S.P. Universidad de Oviedo. Albrecht, J. E., & O'Brien, E. J. (1993). Updating a mental model: Maintaining both local and global coherence. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 1061±1070. Calvo, M. G., & Castillo, M. D. (1996). Predictive inferences occur on-line, but with delay: Convergence of naming and reading times. Discourse Processes, 22, 57±78. Du€y, S. A. (1986). Role of expectations in sentence integration. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 208±219. Fincher-Kiefer, R. (1993). The role of predictive inferences in situation model construction. Discourse Processes, 16, 99±124. Fincher-Kiefer, R. (1994). The e€ect of inferential processes on perceptual indenti®cation. Discourse Processes, 18, 1±17. Fincher-Kiefer, R. (1995). Relative inhibition following the encoding of bridging and predictive inferences. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 981±995. Fincher-Kiefer, R. (1996). Encoding di€erences between bridging and predictive inferences. Discourse Processes, 22, 225±246. Graesser, A. C., Millis, K. K., & Zwaan, R. (1997). Discourse comprehension. Annual Review of Psychology, 48, 163±189. Graesser, A. C., Singer, M., & Trabasso, T. (1994). Constructing inferences during narrative text comprehension. Psychological Review, 101, 371±395. Haberlandt, K., & Graesser, A. C. (1990). Integration and bu€ering of new information. In A. C. Graessser & G. H. Bower (Eds.), Inferences and text comprehension (pp. 71-87). San Diego, CA: Academic Press. Just, M. A., Carpenter, P. A., & Woolley, J. D. (1982). Paradigms and processes in reading comprehension. Journal of Experimental Psychology: General, 111, 228±238. Keefe, D. E., & McDaniel, M. A. (1993). The time course and durability of predictive inferences. Journal of Memory and Language, 32, 446±463. Keenan, J. M., Golding, J. M., Potts, G. R., Jennings, T. M., & Aman, C. J. (1990). Methodological issues in evaluating the occurrence of inferences. In A. C. Graesser & G. H. Bower (Eds.), Inferences and text comprehension (pp. 295±312). San Diego, CA: Academic Press. Kintsch, W. (1988). The role of knowledge in discourse comprehension: A construction-integration model. Psychological Review, 95, 163±182. Klin, C. M. (1995). Causal inferences in reading: From immediate activation to long-term memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1483±1494. Lea, R. B. (1995). On-line evidence of elaborative logical inferences in text processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 1469±1482. Magliano, J. P., Baggett, W. B., Johnson, B. K., & Graesser, A. C. (1993). The time course of generating causal antecedent and causal consequence inferences. Discourse Processes, 16, 35±53. McKoon, G., & Ratcli€, R. (1986). Inferences about predictable events. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 82±91. McKoon, G., & Ratcli€, R. (1990). Dimensions of inference. In A. C. Graesser & G. H. Bower (Eds.), Inferences and text comprehension (pp. 313±328). San Diego, CA: Academic Press.

260 McKoon, G., & Ratcli€, R. (1992). Inference during reading. Psychological Review, 99, 440±466. McKoon, G., & Ratcli€, R. (1995). The minimalist hypothesis: Directions for research. In C. A. Weaver, S. Mannes, & C. R. Fletcher (Eds.), Discourse comprehension (pp. 97±116). Hillsdale, NJ: Erlbaum. Millis, K. K., & Graesser, A. C. (1994). The time-course of constructing knowledge-based inferences for scienti®c texts. Journal of Memory and Language, 33, 583±599. Millis, K. K., Morgan, D., & Graesser, A. C. (1990). The in¯uence of knowledge-based inferences on the reading time of expository text. In A. C. Graesser & G. H. Bower (Eds.), Inferences and text comprehension (pp. 197±212). San Diego, CA: Academic Press. Murray, J. D., Klin, C. M., & Myers, J. (1993). Forward inferences in narrative text. Journal of Memory and Language, 32, 464± 473. O'Brien, E., Shank, D. M., Myers, J. L., & Rayner, K. (1988). Elaborative inferences during reading: Do they occur on-line? Journal of Experimental Psychology: Learning, Memory, and Cognition, 14, 410±420.

Ortega, D. (1990). Thesaurus de sinoÂnimos, antoÂnimos v asociacioÂn de ideas [Thesaurus of synonyms, antonyms, and associates]. Barcelona, Spain: Sopena. Potts, G. R., Keenan, J. M., & Golding, J. M. (1988). Assessing the occurrence of elaborative inference: Lexical decision versus naming. Journal of Memory and language, 27, 399±415. Singer, M., Graesser, A. C., & Trabasso, T. (1994). Minimal or global inference during reading. Journal of Memory and Language, 33, 421±441. Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory and Cognition, 16, 283±298. de Vega, M., Carreiras, M., Calvo, M. G., & Alonso, M. (1990). Lectura v comprensioÂn: Una perspectiva cognitiva [Reading and comprehension: A cognitive perspective]. Madrid: Alianza. Whitney, P., Ritchie, B. G., & Crane, R. S. (1992). The e€ect of foregrounding on readers' use of predictive inferences. Memory and Cognition, 20, 424±432.