Psychological Research (2001) 65: 158±169

Ó Springer-Verlag 2001

ORIGINAL ARTICLE

Manuel G. Calvo á Enrique Meseguer á Manuel Carreiras

Inferences about predictable events: eye movements during reading

Received: 4 January 2000 / Accepted: 23 October 2000

Abstract Eye ®xations were recorded to assess whether, how, and when readers draw inferences about predictable events. Predicting context sentences, or non-predicting control sentences, were presented, followed by continuation sentences in which a target word referred to a predictable event (inferential word) or an unlikely event (non-predictable word). There were no e€ects on initial target word processing measures, such as launch and landing sites, ®xation probability, ®rst-®xation duration, or ®rst-pass reading time. However, relative to the control condition, the predicting context (1) speeded up reanalysis of the inferential word, as revealed by a reduction in second-pass reading time and regressions, and (2) interfered with processing of the non-predictable word, as shown by an increase in regressions. These results indicate that predictive inferences are active at late text integration processes, rather than at early lexical-access processes. The pattern of ®ndings suggests that these inferences involve initial activation of rather general concepts following the inducing context, and that they are completed or re®ned with delay, after the inferential target word is read.

Introduction This study used an eye-movement methodology to investigate the time course of predictive inferences during reading. Prior studies have employed experimental tasks that involved rather restrictive conditions, regarding either the presentation of the stimuli (®xed-pace rapid serial visual presentation of the context sentences, e.g., FincherKiefer, 1995; or self-paced moving-window procedures, e.g., Calvo & Castillo, 1996), the interval between the inducing context and the probe (e.g., stimulus onset asynchrony manipulations; e.g., Millis & Graesser, 1994), M. G. Calvo (&) á E. Meseguer á M. Carreiras Departamento de PsicologõÂ a Cognitiva, University of La Laguna, 38205 Tenerife, Spain

or the measurement of the responses to assess the inference (e.g., naming isolated target words aloud, while reading silently, e.g., Murray, Klin, & Myers, 1993). While these procedures have been very useful for determining the conditions under which inferences are made on-line, the interpretation of the ®ndings can be bene®ted from the use of procedures that allow more natural, nonobtrusive, reading, such as the eye-movement methodology. Such multiple-task convergent approach has proved fruitful to investigate reading (e.g., Schilling, Rayner, & Chumbley, 1998; see Haberlandt, 1994). Predictive inferences are implicit anticipations of likely outcomes of events, based on the application of our prior world knowledge to explicit information in a message describing such events (e.g., Fincher-Kiefer, 1995; Keefe & McDaniel, 1993; Klin, GuzmaÂn, & Levine, 1999; McKoon & Ratcli€, 1986; Trabasso & Magliano, 1996). For example, we would be making a predictive inference if, when reading that ``the angry man threw the delicate porcelain vase against the wall'', we were to anticipate that the vase ``broke'' (e.g., Potts, Keenan, & Golding, 1988). As forward inferences, predictive inferences facilitate processing of subsequent information when they match it. They are also behaviorally useful, because of their preparatory function to recruit resources and plan adaptive behavior in advance of the actual events. Accordingly, they should be made on-line, as soon as the inducing information is processed. Yet, these elaborations may be restricted under normal reading conditions, because of the overload they could cause during on-line comprehension, and even the risk of wrong anticipations requiring backward corrections. Actually, two extant models of inference processing in reading, such as the minimalist hypothesis (McKoon & Ratcli€, 1992, 1995) and the constructionist theory (Graesser, Singer, & Trabasso, 1994; see Graesser, Millis, & Zwaan, 1997) argue that predictive inferences are unlikely to be drawn on-line. The reason is that these inferences are not required to make statements in the text locally coherent (minimalist hypothesis) or globally coherent (constructionist theory), or to explain

159

why actions, events and states are mentioned in the text (constructionist theory), which would be necessary conditions to make inferences normally. However, both models admit that predictive inferences can be generated if they are supported by well-known information that is readily available in memory (minimalist hypothesis), or if the predicted outcome is highly constrained by the context, with few, if any, alternative consequences (constructionist theory). Experimental research on predictive inferences has obtained mixed results. Some studies have found that these inferences can be made on-line (Calvo & Castillo, 1996, 1998; Calvo, Castillo, & Estevez, 1999; FincherKiefer, 1993, 1995, 1996; Keefe & McDaniel, 1993; Klin, GuzmaÂn et al., 1999; Klin, Murray, Levine, & GuzmaÂn, 1999; Murray et al., 1993; Whitney, Ritchie, & Crane, 1992), whereas others have provided negative evidence (Magliano, Baggett, Johnson, & Graesser, 1993; Millis & Graesser, 1994; Potts et al.,1988), or evidence of only minimal, incomplete (McKoon & Ratcli€, 1986, 1989) or nonspeci®c (Du€y, 1986) inferencing. Some authors (e.g., Calvo & Castillo, 1998; Fincher-Kiefer, 1996; Klin, GuzmaÂn et al., 1999) have argued that the extent to which the passages used to study inferences met the demands of high availability and context constraints, as put forward by the minimalist hypothesis and the constructionist theory, could account for the di€erent ®ndings. Nevertheless, even if predictive inferences can be made on-line in some limited conditions, both the minimalist hypothesis (McKoon & Ratcli€, 1992, 1995) and the constructionist theory (Graesser et al., 1994) assume that these inferences are not automatic. Rather they would involve elaborative construction of meaning, i.e., post-lexical strategic processes. Accordingly, they should take time to develop. The possibility of multiple alternatives to forecast, the limitation of working memory resources, and the interruption of other on-going processes would impose a delay in the on-line generation of these inferences. A number of prior studies have manipulated time parameters, such as the ®xed-pace presentation of the context and the interval between the inducing context and the probe, which allow us to estimate the time course of predictive inferences1. Essentially, these inferences have not been detected within the ®rst 500 ms after the onset of the last word in the context (whereas causal consequence or bridging inferences have; Magliano, Baggett et al., 1993; Millis & Graesser,

1994). Rather, it has been shown that predictive inferences take at least 750 ms or more after the end of the last word in the context (Fincher-Kiefer, 1995, 1996; Calvo & Castillo, 1996, 1998; Calvo et al., 1999). However, the results obtained with the aforementioned approach might have some limitations, because the tasks employed did not allow participants to read the way they do in normal circumstances. The eye-movement method assesses reading processes under more natural conditions (see Rayner & Sereno, 1994). We used this methodology to validate, re®ne and extend prior ®ndings regarding the time course of inferences about predictable events. As far as we know, eye-®xation measures have not been previously used to investigate these inferences (see Rayner, 1998)2. A main purpose of the present study is to determine to what extent predictive inferences involve early or late processing stages. In our study, participants were presented with inducing or control context sentences, followed by continuation sentences that included a pretarget region, a target word, a posttarget region, and a ®nal region. Evidence for on-line predictive inferencing will involve facilitation in reading a target word that represents the predicted event following the predicting context, relative to when that word follows the non-predicting control context. Furthermore, relevant evidence for the time course of these inferences, and the processes involved, are provided by two groups of eye-®xation measures (see Liversedge, Paterson, & Pickering, 1998; Rayner, 1998): (1) initial processing measures, such as probability of ®rst-pass ®xation (i.e., skipping rate) or ®rst-®xation duration, and (2) reanalysis measures, such as secondpass reading time and regressions (see description of all measures in the Method section). If predictive inferences a€ect early lexical-access processes, facilitation should occur on the inferential target word in the initial processing measures. These measures have proved to be sensitive to predictability of words within a context sentence, as a function of lexical or semantic association. Thus, Altarriba, Kroll, Sholl, and Rayner (1996), Ehrlich and Rayner (1981), Rayner and Well (1996), or Schustack, Ehrlich, and Rayner (1987) have found that high-predictable words are skipped (i.e., not ®xated) more frequently, and ®xated more shortly than low-predictable words (see Brysbaert & Vitu, 1998, for a detailed discussion). Accordingly, if predictive inferences involve lexical predictability, they

1 Other studies, though demonstrating the on-line occurrence of predictive inferences, could not determine their time course with precision, because the procedures allowed self-paced presentation of the inducing contexts and/or the probe (e.g., Fincher-Kiefer, 1993; Keefe & McDaniel, 1993; Klin, GuzmaÂn et al., 1999; Murray et al., 1993; Whitney et al., 1992). On-line, as opposed to ``o€line'', means that the cognitive processes are performed during comprehension of the stimuli, as opposed to being induced by a test at later retrieval. Presumably, there is a continuum from automatic to elaborative on-line processes, ranging from less than 500 ms to a little over 1 s following the stimulus (e.g., Till, Mross, & Kintsch, 1988).

2 Nevertheless, also using eye-movement measures, O'Brien, Shank, Myers, and Rayner (1988) obtained evidence for a di€erent type of semantic inferences, involving elaboration of contextually appropriate meanings or instantiation of noun category (e.g., inferring `knife' from the phrase `the mugger stabbed the woman with his weapon'). There was reduced gaze duration when reading the word that represented the implied concept (e.g., knife) following the implying context phrase. However, this occurred only when the phrase strongly suggested (e.g., stabbed vs assaulted) the implied concept, and when there was an anaphoric relationship between the implying and the implied concepts (Garrod, O'Brien, Morris, & Rayner, 1990).

160

will also reveal a similar pattern regarding initial processing measures. In contrast, if, as expected, predictive inferences are higher level processes which involve representation of a situation model (prediction of events) rather than lexical or semantic association (prediction of words) (FincherKiefer, 1996), then the eye-movement patterns will be di€erent. Thus, if predictive inferences a€ect messageintegration processes, there would be facilitation in the late processing measures on the inferential target word, and interference on an alternative, non-predictable target word, representing an unlikely event. Thus far, based on prior research, we have shown preference for the hypothesis that predictive inferences involve late rather than early processes. However, to adequately test (and, eventually, rule out) the earlyprocesses hypothesis, several low-level oculomotor factors associated with saccadic programming must be controlled. Some early eye-movement measures, such as word skipping and landing site, are particularly sensitive to these factors, even more than to semantic context factors (see Brysbaert & Vitu, 1998). Thus, if no di€erence is found for our inferential target words between the predicting and the control context, this does not necessarily imply that early processing is not facilitated in the former relative to the latter condition. Rather, this lack of e€ect might be due to contextual facilitation not being strong enough to make processing of the target word in the facilitated condition faster than the saccadic programming time. Accordingly, to maximize the semantic context e€ects on early processing measures, the contribution of oculomotor variables should be made comparable for the di€erent context conditions (Lavigne, Vitu, & d'Ydewalle, 2000). With this aim in mind, we have made the pretarget region, the target word, and the posttarget/®nal regions identical for both context conditions (thus controlling the e€ect of powerful linguistic variables impinging upon oculomotor and non-contextual linguistic factors, such as word length and frequency). Moreover, we have used several alternative eye-movement measures, such as launch site and duration of the last ®xation before ®rst landing on the target word, to ensure that the target word was pre-processed in parafoveal vision in the same way under the di€erent context conditions.

Table 1 Example of passages, including the predicting and the control context, and the continuation sentence with its four regions. Italics indicate the target words (the inferential word studied: outside the brackets; the non-predictable word slept: inside the

Method Participants Forty native Spanish speaking psychology undergraduates at La Laguna University participated for course credit. They had normal uncorrected vision. Materials Forty short Spanish passages were used (see Table 1 and Appendix), each of which was composed of (a) one predicting context sentence, (b) one non-predicting, control context sentence, (c) one continuation sentence including a target word that represented the predicted event (inferential target word) in the predicting context, and (d) one continuation sentence including a target word that represented an unlikely event (non-predictable target word) for both contexts. Besides the target word, there were three relevant regions in the continuation sentence: pretarget (1 or 2 words long), posttarget (2 or 3 words), and ®nal (1 or 3 words); these three regions were identical for both versions (i.e., c and d) of the continuation sentence (only the target word was di€erent). Four versions of each passage were constructed (see Design). Each participant was presented with one of these versions (a di€erent version for each passage). Thus, on each trial participants read either a predicting or a control context sentence, followed by a continuation sentence including either an inferential or an nonpredictable target word. Norming data Prior to the experiment, we conducted several norming studies and analyses to validate the materials (see details in Calvo et al., 1999). Essentially, ®rst, we controlled word-based priming (Keenan, Golding, Potts, Jennings, & Aman, 1990) by including 42% of the same content words in the predicting and the control contexts. These were words that could be semantically associated with the inferential target words (see words preceded by an asterisk in Table 1 and Appendix). They were rearranged in the control contexts so that they did not suggest the inference. Second, in a sentencecompletion study, we con®rmed that the predicting contexts did induce the presumed inferences, and that these could be expressed linguistically. The participants in this study wrote one-word continuations about ``what might happen next'' following the event that was described in each context. These one-word predictions were subsequently used as target words. The inferential target words were mentioned by 82% of the participants after the predicting contexts, and by 8% after the control contexts; the respective ®gures for the non-predictable target words were 2% and 3%. Finally, we checked that the predicting and the control contexts were equivalent in number of words (M ˆ 20.6, for both contexts), and that the inferential and the non-predictable target words were equivalent (all ts < 0.5) in number of characters

brackets). Words in parentheses and the asterisks did not appear in the stimuli. Asterisks indicate content words shared by the predicting and the control contexts to control for word-based priming. See original Spanish sentences in the Appendix

Predicting context: Three days before the *examination the *pupil went to the *library, looked for a separate *table and opened his *notebook Continuation sentence: (Pretarget) The pupil (Target) studied [slept] (Posttarget) for an hour (Final) approximately Control context: The *pupil, who was a little tired after ®nishing his *examination, forgot his *notebook and left it on a *table in the *library Continuation sentence: (Pretarget) The pupil (Target) studied [slept] (Posttarget) for an hour (Final) approximately

161 (M ˆ 6.6 vs 6.7) and lexical frequency (M ˆ 33.4 vs 35.4 per million; Alameda & Cuetos, 1995). The range of target word length was four to ten letters, for both inferential and non-predictable target words. The range of for lexical frequency was 1±314 for inferential words; and 1±325 for non-predictable words. Design A 2 (Priming context: predicting vs control) ´ 2 (Target type: inferential vs non-predictable) within-subjects factorial design was used. We constructed four versions of each passage, and then these versions were combined in four lists of stimuli to address the design requirements (i.e., orthogonal and counterbalanced combination of the context and target factors). Thus, each list consisted of: ten predicting contexts followed by inferential target continuation sentences; ten control contexts followed by inferential target continuation sentences; ten predicting contexts followed by non-predictable target continuation sentences; and ten control contexts followed by non-predictable target continuation sentences. The assignment of target words 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 list, with 40 trials in random order. It should be noted that, for each participant, only 25% of the trials (predicting + inferential) induced inferences, whereas 50% were neutral (control), and 25% (predicting + non-predictable) clearly discouraged such inferential activity. In these conditions, it is not likely that the participants consciously made inferences as a strategy to meet the demands of the experiment. Apparatus Eye movements were recorded by a Forward Technologies Dual Purkinje 5.5 Eyetracker, which has a resolution of less than 10 min of arc. The eyetracker was interfaced with an IBM compatible PC that controlled stimulus display and data storage. The passages were each presented in three lines up to 80 characters per line. The pretarget region, the target word, and the posttarget region were located on the third line, which began with the last preceding context word (so the pretarget, target, or posttarget word were never located at the beginning or end of a line). All of the characters, except the ®rst letter of the context sentence and of the continuation sentence, were presented in lowercase. The position of the eye was sampled every millisecond, and the computer stored data on the duration and location of each ®xation for later analysis. The computer was also interfaced with a super-VGA visual display unit on which the stimulus passages were presented. The display was 60 cm from the participant's eye, and four characters equaled 1° of visual angle. Viewing was binocular, though eye movements were monitored from the participant's right eye. Participants viewed the screen with their heads positioned in a deep chin rest and a forehead rest, with a strip around the head, to minimize movements during the experiment. Procedure Before the experiment started, participants were informed that the study was about reading comprehension of short passages, which would be displayed on a screen. They were told to read at their normal rate and to comprehend the sentences as well as they could. They were also told that they would periodically be asked to answer comprehension questions about the passages. Then, the participant sat in front of the eyetracker and calibration was performed. Following the calibration period (5±10 min), participants read six practice passages before reading the 40 experimental passages. When readers ®nished each passage, they pressed a key and the computer either displayed a statement (e.g., `The pupil was in a bookshop', for the example passage presented in Table 1) on half of the trials, or proceeded to the next passage. Half of these

statements had ``yes'' answers, and half ``no''. Readers responded to the statements by pressing one of two buttons and received no feedback on their answers. After readers completed half of the experiment, the experimenter recalibrated the equipment, and readers had a short break (in addition, recalibration was performed whenever necessary). Readers normally completed the experiment in about 30 min. Measures Three groups of eye-movement measures were examined for each of the four regions of the continuation sentence, with particular interest in the target word. First, two typical measures of initial or `early processing' were computed (given that the region at issue was ®xated during ®rst-pass reading): (a) ®rst-®xation duration, which is the duration of the ®rst ®xation on the region independent of the number of ®xations; (b) ®rst-pass reading time or gaze duration, which is the sum of all ®xations durations on a region prior to moving to another region, including within-zone regressions prior to an exit right or left. In an attempt to further explore early processes, several additional measures were obtained: (c) probability that the target word was ®xated when it was initially encountered, (d) landing position of ®rst ®xation, and (e) number of ®xations received during ®rst-pass reading. Second, measures of reanalysis or `late processing': (f) second-pass reading time, which is the time spent ®xating a region after the reader has ®xated at least once away from it; (g) probability of regressions in ®rst-pass reading, i.e., backward eye movements that begin at the right-most region the reader has ®xated in a ®rst pass, and leaves the currently ®xated region to the left. And, third, measures to `control' for `parafoveal' pre-processing of the target word: (h) launch site, or the last location of the eyes before ®rst landing on the target word, and (i) the duration of the ®xation at the launch site.

Results Approximately 6.3% of the ®xation data were eliminated because of track losses or because ®xations were shorter than 70 ms (see Rayner, Sereno, Morris, Schmauder, & Clifton, 1989, for justi®cations for eliminating data), with equal distribution across conditions. Mean comprehension performance was 87.1%, with no participant performing at less than 75% accuracy on the comprehension questions. For each dependent measure and region, we carried out 2 (context) ´ 2 (target) ANOVAs, using both subjects (F1) and items (F2) as random e€ects. Interactions were analyzed by means of planned comparisons (between the predicting and the control condition). All of the reported analyses are signi®cant at the 0.05 level, using the Bonferroni procedure, unless otherwise indicated. Initial processing First-®xation duration, and ®rst-pass reading time There were no signi®cant di€erences between the predicting condition and the control condition, nor Context ´ Target interactions, in either ®rst-®xation duration or ®rst-pass reading time on any region separately (see mean scores across the four regions of the

162 Fig. 1 Mean ®rst-®xation durations (in ms; SD in parenthesis) in reading the pretarget (PRE), target (TAR), posttarget (POS), and ®nal (FIN) regions of the continuation sentence, as a function of context sentence (predicting vs control) and target word (inferential vs nonpredictable)

Fig. 2 Mean ®xation durations (in ms; SD in parenthesis) in ®rst-pass reading of the pretarget (PRE), target (TAR), posttarget (POS), and ®nal (FIN) regions of the continuation sentence, as a function of context sentence (predicting vs control) and target word (inferential vs non-predictable)

continuation sentence in Figs. 1 and 2)3. Only a marginally signi®cant e€ect of Target appeared in ®rst ®xation duration on the posttarget region, F1(1, 39) ˆ 7.50, MSE ˆ 404, F2(1, 39) ˆ 3.50, MSE ˆ 786, P ˆ 0.069, revealing an increased ®xation time on the posttarget region when it followed a non-predictable target word (M ˆ 257 ms), relative to when it followed an inferential target word (M ˆ 247 ms). In additional Region (pretarget vs target) ´ Context ´ Target ANOVAs, we took 3

Our main analyses excluded 0-ms ®xations that occurred when readers skipped a region. We subsequently performed a second set of analyses that included 0-ms ®xations (see Pickering & Traxler, 1998, p. 945). The results of these analyses matched the results of the main analyses, so we do not report them.

the pretarget region as a baseline to examine possible changes from the pretarget to the target word. No e€ect was signi®cant, including the three-way interaction, which further con®rms the absence of e€ects on both ®rst-®xation and ®rst-pass times. Additional initial-processing and parafoveal measures (see mean scores in Table 2) There were no signi®cant di€erences between the predicting and the control condition, nor Context ´ Target e€ects, either in the probability or the number of ®xations on the target word, nor in landing position on this word. No signi®cant e€ects appeared on launch site

163 Table 2 Mean (a) ®rst-pass ®xation probability on the targetword, (b) landing site on the target word (position from the beginning of the region, in number of characters), (c) number of ®rst-pass ®xations on the target word, (d) launch site prior to target word ®xation (position in number of characters backwards from

the beginning of the target word), and (e) duration (in ms) of the last ®xation before landing on the target word, as a function of context sentence (predicting vs control) and target word (inferential vs non-predictable)

Inferential target word Context Predicting Additional measures of early processing (a) Probability of ®xation (b) Landing site (c) Number of ®xations Control measures of parafoveal processing (d) Launch site (e) Duration of last ®xation

0.91 3.0 1.3 )3.2 250

control 0.91 3.1 1.3 )3.2 252

Non-predictable target word Context Predicting 0.93 3.0 1.3 )3.0 256

control 0.90 2.8 1.3 )2.8 250

Fig. 3 Mean ®xation durations (in ms; SD in parenthesis) in second-pass reading of the pretarget (PRE), target (TAR), posttarget (POS), and ®nal (FIN) regions of the continuation sentence, after having abandoned the region in a forward direction, as a function of context sentence (predicting vs control) and target word (inferential vs non-predictable). Asterisks indicate signi®cant di€erences between the predicting condition and the control condition

prior to the target word either, nor in duration of the ®xation prior to the target word. This indicates that the target word was pre-processed in parafoveal vision in the same way under the di€erent conditions, and suggests that the (lack of) contribution of early processes is not masked by parafoveal preview e€ects. Reanalysis and text integration Second-pass reading time (after an exit either right or left; see mean scores in Fig. 3) Signi®cant e€ects emerged only on the target word4. Main e€ects of Target, F1(1, 39) ˆ 33.16, MSE ˆ 7,873, 4 Thirty-nine participants (out of 40) spent more than 70 ms rereading the target word in one or more of the four experimental conditions. For all 40 items, second-pass reading time was more than 70 ms in one or more conditions.

F2(1, 39) ˆ 15.57, MSE ˆ 16,494, were quali®ed by a Context ´ Target interaction, F1(1, 39) ˆ 6.54, MSE ˆ 5,404, F2(1, 39) ˆ 4.79, MSE ˆ 7,434. Planned contrasts indicated that reinspection times were shorter for inferential target words following predicting contexts than when these words followed control contexts, F1(1, 39) ˆ 12.69, MSE ˆ 3,253, F2(1, 39) ˆ 5.71, MSE ˆ 7,302; in contrast, there was no signi®cant di€erence between the two context conditions for non-predictable target words (both Fs < 1.0). There was also a similar, though nonsigni®cant (both Ps > 0.10), interactive trend for the posttarget region. An entirely consistent pattern appeared when the pretarget region was included as a baseline in a Region (pretarget vs target) ´ Context ´ Target ANOVA. The three-way interaction was signi®cant, F1(1, 39) ˆ 7.07, MSE ˆ 3,375, F2(1, 39) ˆ 8.07, MSE ˆ 3,011, but the Context ´ Target interaction was not (both Ps > 0.10). This corroborates that the di€erences occurred speci®cally on the target word, even when baseline di€erences were controlled.

164 Fig. 4 Mean probability of regressions (i.e., percentage of trials in which there was at least one or more regressions; SD in parenthesis) landing on the pretarget (PRE), target (TAR), posttarget (POS), and ®nal (FIN) regions of the continuation sentence, as a function of context sentence (predicting vs control) and target word (inferential vs non-predictable). Asterisks indicate signi®cant di€erences between the predicting condition and the control condition

A Context ´ Target ANOVA was conducted on second-pass reading times of the target word only after the readers had moved forward away to the right of the region (thus excluding second-pass ®xations after an exit left). The results were totally convergent with the previous ones. There was a Context ´ Target interaction, F1(1, 39) ˆ 4.39, MSE ˆ 5,949, F2(1, 39) ˆ 3.99, MSE ˆ 6,863, P ˆ 0.05, with a signi®cant di€erence between the predicting (M ˆ 71 ms) and the control condition (M ˆ 113 ms) only for inferential target words, F1(1, 39) ˆ 13.94, MSE ˆ 2,598, F2(1, 39) ˆ 5.53, MSE ˆ 6,128, but not for non-predictable target words (M ˆ 177 vs 168 ms; both Fs < 1.0). Regressions Two aspects of regressive eye movements were examined: (a) the probability that regressions landed on a region (i.e., inward regressions); and (b) the probability that (one or more) regressive saccades were launched from one particular region to another particular region. The analysis of `inward' regressions (see mean scores in Fig. 4) showed only main e€ects of Target on the target word, F1(1, 39) ˆ 30.04, MSE ˆ 201, F2(1, 39) ˆ 20.83, MSE ˆ 276, which were quali®ed by a Context ´ Target interaction, F1(1, 39) ˆ 14.71, MSE ˆ 123, F2(1, 39) ˆ 15.61, MSE ˆ 121. Planned contrasts revealed that fewer regressions landed on inferential target words in the predicting condition than in the control condition, F1(1, 39) ˆ 7.24, MSE ˆ 106, F2(1, 39) ˆ 6.89, MSE ˆ 107; furthermore, the opposite applied to non-predictable target words, F1(1, 39) ˆ 6.25, MSE ˆ 168, F2(1, 39) ˆ 7.59, MSE ˆ 155. A consistent pattern emerged when the pretarget region was included as a baseline in a Region (pretarget vs target) ´ Context ´ Target ANOVA. The

three-way interaction was signi®cant, F1(1, 39) ˆ 9.67, MSE ˆ 108, F2(1, 39) ˆ 7.02, MSE ˆ 144, with the same meaning as for second-pass reading time. Of all possible `regressive saccades' between two regions (Fig. 5 shows four of them having special relevance), signi®cant Context ´ Target interactive e€ects occurred only on regressions from (a) the posttarget region to the target word, F1(1, 39) ˆ 5.87, MSE ˆ 958 (only data by subjects are available for regressive saccades), and (b) from the ®nal region to the target word, F1(1, 39) ˆ 8.35, MSE ˆ 396. Planned contrasts indicated that there were fewer regressions from the posttarget region to the inferential target word following the predicting context than following the control context, F1(1, 39) ˆ 6.52, MSE ˆ 0.85, whereas the di€erence (in the opposite direction) was non-signi®cant for the nonpredictable target words, F1(1, 39) ˆ 1.22, ns. In contrast, there were more regressions from the ®nal region to the non-predictable target word following the predicting context than following the control context, F1(1, 39) ˆ 10.24, MSE ˆ 0.54, whereas regressions from the ®nal region to the inferential target word were not a€ected by the context (F < 1.0)5.

5

Number of participants (p; out of 40) that made each number of regressive saccades (rs; out of 10 trials) for each of the critical conditions. From Posttarget region to Target word: Predicting + Inferential: 29p ˆ 0rs; 8p ˆ 1rs; 2p ˆ 2rs; 1p ˆ 3rs. Control + Inferential: 18p ˆ 0rs; 16p ˆ 1rs; 2p ˆ 2rs; 4p ˆ 4rs. Predicting + Non-predictable: 10p ˆ 0rs; 20p ˆ 1rs; 7p ˆ 2rs; 2p ˆ 3rs; 1p ˆ 4rs. Control + Non-Predictable: 16p ˆ 0rs; 14p ˆ 1rs; 9p ˆ 2rs; 1p ˆ 3rs. From Final region to Target word: Predicting + Inferential: 29p ˆ 0rs; 7p ˆ 1rs; 4p ˆ 2rs. Control + Inferential: 25p ˆ 0rs; 13p ˆ 1rs; 2p ˆ 2rs. Predicting + Non-predictable: 19p ˆ 0rs; 8p ˆ 1rs; 6p ˆ 2rs; 7p ˆ 3rs. Control + Non-Predictable: 25p ˆ 0rs; 11p ˆ 1rs; 3p ˆ 2rs; 1p ˆ 3rs.

165 Fig. 5 Mean probability (SD in parenthesis) of regressive movements (i.e., number of regressive movements divided by ten trials per condition per subject) (a) from the pretarget region to the preceding context sentence (PRE-CON), (b) from the target word to the pretarget region (TAR-PRE), (c) from the posttarget region to the target word (POS-TAR), and (d) from the ®nal region to the target word (FIN-TAR), as a function of context sentence (predicting vs control) and target word (inferential vs non-predictable). Asterisks indicate signi®cant di€erences between the predicting condition and the control condition

Discussion These results support two main conclusions. First, readers generate inferences about predictable events during reading. This is demonstrated by (1) facilitation (i.e., shorter second-pass time, and fewer regressions) in reading target words that represented predictable events following the predicting context, in comparison with when these words followed the non-inducing, control context; and by (2) interference (i.e., more regressions) in processing target words that represented unlikely events following the predicting context, in comparison with the control context. Secondly, predictive inferences are not drawn, or completed, immediately after having read the inducing context, but they take some time to develop. This is demonstrated by the fact that neither facilitation nor interference a€ected early processing measures6. Rather, these e€ects were observed only on late processing measures, involving reanalysis of the target word (i.e., second-pass time and regressions). In general, our results are consistent with those of prior studies using a variety of tasks (mainly, naming) other than eye movements, which have indicated that predictive inferences can be drawn on-line if the context highly constrains the predictability of a main consequence (Calvo & Castillo, 1996; Calvo et al., 1999; Fincher-Kiefer, 1993, 1995, 1996; Keefe & McDaniel, 1993; Klin, GuzmaÂn et al., 1999; Klin, Murray et al., 1999; Murray et al., 1993; Whitney et al., 1992). More6 If predictive inferences involved early cognitive processes, then facilitation in processing the inferential target word following the predicting context would have revealed itself in one or more of the following indices: a shorter ®rst-®xation or ®rst-pass reading time on the target word, a higher probability of skipping the target word, a farther ®rst landing position from the beginning of the target word, or fewer ®xations on it when initially encountered. The fact that none of these indices was a€ected is contrary to the early processing hypothesis.

over, regarding the time course of predictive inferences, our ®ndings are also consistent with those of studies that have manipulated the rate of presentation of the context one word at a time and the interval between the context and the probe. These studies have shown that predictive inferences are not made within an interval of less than 750 ms after the end of the context (Calvo & Castillo, 1996, 1998; Magliano, Baggett et al., 1993; Millis & Graesser, 1994), even when the ®xed pace presentation of the context is signi®cantly slowed down (Calvo & Castillo, 1998; Calvo et al., 1999). Rather, these inferences take additional post-context time (Calvo & Castillo, 1996; Calvo et al., 1999; Fincher-Kiefer, 1995, 1996), which indicates that they are constructed with delay. However, the delay is not totally comparable in these ®xed-pace studies (in which there was a post-context ``blank'' interval before the target word) and in the self-paced eye-movement study (in which the delay was ``®lled'' with text in the continuation sentence). In a closer approach to the eye-movement method, Calvo and Castillo (1998, Experiment 2) used a selfpaced moving-window procedure, in which the participants read the same sentences as in the present study. These authors found spillover e€ects, revealing facilitation in reading the posttarget and the ®nal regions of the continuation sentence that con®rmed a predicted event, and interference in the ®nal region when the target word discon®rmed a predicted event; in contrast, there were no e€ects on the target word itself. These moving-window results can now be interpreted in the light of the eye-®xation data. It has been argued that the movingwindow method may induce an arti®cial bu€ering strategy, and produce ``lag'' e€ects (e.g., Magliano, Graesser, Eymard, Haberlandt, & Gholson, 1993). Taking this into account, Calvo and Castillo's (1998) ®ndings are coincident with the eye-®xation data. Both show delayed processing of predictive inferences, though this is revealed by di€erent measures in each study.

166

Thus, in both cases, (a) facilitation occurred after the target word had already been read (i.e., on second-pass reading and regressions to the target word ± eye ®xations study ± , or on the regions following the target word when it could not be reread ± moving window study); and (b) interference occurred with the longest possible delay (i.e., regressions from the ®nal region ± eye ®xations ± , or a slow down in reading the ®nal region when regressions were not possible ± moving window). Both the facilitation and the interference e€ects converged to indicate that the concept representing the predictable event was active when reading the sentence following the inducing context. Presumably, activation of this concept enabled readers to comprehend the corresponding words more readily. Furthermore, it is reasonable that activation of the inferential concept (e.g., studied) caused diculty in processing inconsistent words that represented unlikely events (e.g., slept; i.e., the more activated the inferential concept is, the more likely it is expected to interfere on the processing of an inconsistent concept). Nevertheless, if the inferential concept had been completely and speci®cally activated by the time the target word was initially encountered, facilitation and inhibition would have occurred on that word at that time (e.g., ®rst ®xation). The fact that there were no such early e€ects suggests that the inferential concept might have been only vaguely or generically activated (e.g., the pupil read or worked, etc.) by the time the target word was read ®rst. Only after the participant had read the inferential target word would the inference be completed, which would then reduce the need of reanalysis behavior (i.e., as re¯ected in less second reading and regressions). This interpretation is consistent with the claims that predictive inferences are drawn only partially or minimally (McKoon & Ratcli€, 1986, 1989), or that they do not take the form of highly speci®c predictions (Du€y, 1986). The lack of early complete and speci®c activation of the inferential concept would be insucient to a€ect initial access to the target word meaning. However, once this concept is completed and re®ned when the target word is processed, it becomes capable of a€ecting later text integration processes, as revealed by the facilitation and inhibition e€ects that emerged after the target word was read ®rst. Nevertheless, though both the facilitation and the interference e€ects take place with delay, their time course seems to be di€erent. Thus, facilitation occurred immediately after the target word had been read, as revealed by the fact that regressions to the inferential target word mainly came from the posttarget region (in the control condition, relative to the predicting condition). In contrast, interference revealed itself later, in the form of regressions from the ®nal region to the non-predictable target word, when this word was inconsistent with the expectation induced by the predicting context. Prior studies have also demonstrated the existence of such interference e€ects. Thus, in Du€y's (1986) study, participants were slower to read an expectation-violating sentence embedded in a high-expectation text; and in

Klin, GuzmaÂn et al.'s (1999) study there was a slow down in reading times for a contradicting sentence following the predicting context. These authors measured reading times for the sentence as a whole. By examining reading behavior at di€erent points in the continuation sentence that represented an unlikely event, we have obtained more precise information about the time course of the interference process. It appears that, when readers draw an inference in the predicting condition, and then ®nd an inconsistent word, they are not able to reanalyze (and change) their inference completely on-line (e.g., by slowing reading of the non-predictable target word). Rather, readers wait until the end of the sentence; then they go back and reread the target word to recover its meaning and perform the text integration process. Our ®ndings can be compared with those on lexical or semantic predictability of words as a function of context constraints, also using the eye-tracking method. When words are highly constrained by a preceding sentence, they are ®xated less frequently and for less time in ®rst®xation and ®rst-pass reading than unconstrained words (Altarriba et al., 1996; Ehrlich & Rayner, 1981; Rayner and Well, 1996; Schustack et al., 1987). This type of lexical context constraints have been determined via a cloze task in which the participant predicts the word (normally, a noun) that could ®t in an incomplete sentence fragment. In contrast, in our norming studies, we used a procedure in which the participants predicted the event (typically, a verb) that could result from a described situation. The high-constraint target words used in the lexical-predictability studies are typically produced between 80% and 90% of the time, and the lowconstraint words are mentioned less than 10% of the time. Similarly, for our own inferential target words, the respective predictability scores were 82% (predicting context) and 8% (control context). Yet, our inferential words were initially ®xated for equivalent time and skipped similarly both in the predicting and the control conditions. Accordingly, we did not ®nd any e€ects on the initial processing measures, unlike those studies on lexical predictability, but, rather, on late processing measures. Moreover, this occurred even though we had presumably made the predicting and the control contexts equivalent in word-based priming on the target words (i.e., intralexical associability). This suggests that the processes involved in the predictability of words within a context sentence are different from those involved in the predictability of events (see Du€y, 1986). Although anticipation of a speci®c word can be made mainly by lexical-access processes, anticipation of events would involve the construction of situational models (Fincher-Kiefer, 1996), which do not have a speci®c linguistic representation. Rather general or nonspeci®c concepts would be initially activated to anticipate events, whose information could be conveyed in a number of ways (i.e., by di€erent words). This is why initial lexical access to a speci®c target word representing the predictable event is not facilitated. For lexical access to be facilitated, a closer lexical or

167

semantic correspondence would be needed between the target word and the inferential representation. In contrast, later text integration of the target word can be bene®ted if the meaning of the inferential representation is generally consistent, though not speci®cally coincident, with the target word. Accordingly, the eye-®xation measures, though essentially concordant with prior measures obtained under less natural reading conditions, have revealed that predictive inferences are active at text integration rather than at initial lexical access. Furthermore, the eye-®xation measures have depicted the pattern of reanalysis behavior involved in late inference construction. Acknowledgements This research was supported by Grants PB971481 to Manuel G. Calvo, and PB96-1048 to Manuel Carreiras, from the DGES, Spanish Ministry of Education and Science.

Appendix Example of experimental passages, as translated into English, and original Spanish sentences (in brackets). PREDICTING CONTEXT: Three days before the *examination the *pupil went to the *library, looked for a separate *table and opened his *notebook. {Tres dõ as antes del *examen el *alumno fue a la *biblioteca, busco una *mesa apartada y abrio su *cuaderno de apuntes} CONTROL CONTEXT: The *pupil, who was a little tired after ®nishing his *examination, forgot his *notebook and left it on a *table in the *library. {El *alumno se olvido el *cuaderno de apuntes en una *mesa de la *biblioteca despueÂs de haber terminado el *examen cansado} CONTINUATION SENTENCE: The pupil / studied [slept] / for an hour / approximately. {El alumno / estudio [durmioÂ] / una hora / aproximadamente} PREDICTING CONTEXT: When Eva saw her *father in the *airport, she ran up to him, and he *bent down over his *child. {Cuando Eva vio a su *padre en el *aeropuerto, corrio hacia eÂl, y el padre se *inclino sobre la *ninÄa.} CONTROL CONTEXT: Before the trip with Eva, the *father *bent down to show a scale model of the *airport to his *child. {Antes de irse de viaje con Eva, el *padre se *inclino para ensenÄarle una maqueta del *aeropuerto a la *ninÄa} CONTINUATION SENTENCE: The father / embraced [spoke to] / the child / as usual. {El padre / abrazo [habloÂ] / a la ninÄa / como siempre.}

PREDICTING CONTEXT: The woman went into the *church, spoke with the *priest for a few minutes and afterwards *knelt down in front of the *altar {La mujer entro en la *iglesia, hablo durante unos minutos con el *cura y luego fue a *arrodillarse frente al *altar.} CONTROL CONTEXT: After having spoken with the *priest for a few minutes, in front of the *church's *altar, the woman *knelt down to do her shoe up {DespueÂs de haber hablado con el *cura unos minutos frente al *altar de la *iglesia, la mujer se *arrodillo para anudarse el zapato.} CONTINUATION SENTENCE: The woman / prayed [wrote] / a prayer / with devotion. {La mujer / rezo [escribioÂ] / una plegaria / con devocioÂn.} PREDICTING CONTEXT: The *angry *bees *¯ew out of the *hive and some of them *threw themselves at Julia's arm. {Las *furiosas *abejas *volaron fuera de la *colmena y varias de ellas se *lanzaron sobre el brazo de Julia.} CONTROL CONTEXT: Julia was *angry because someone had *thrown rubbish near the new *hive and the *bees had *¯own out. {Julia estaba *furiosa porque alguien habõ a *lanzado basura cerca de la *colmena de las *abejas, que habõ an *volado.} CONTINUATION SENTENCE: The bees / stung [soiled] / Julia's / arm. {Las abejas / picaron [mancharon] / a Julia / en el brazo} PREDICTING CONTEXT: When the *party was over, there were *bags and *papers all over the *¯oor, so Susana picked up the *broom. {Al terminar la *®esta, habõ a bolsas y *papeles por todo el *suelo, asõ que Susana cogio la *escoba.} CONTROL CONTEXT: In order to decorate the *party, Susana hung up the colored *papers with the *broom that was on the *¯oor. {Para adornar la *®esta, Susana colgo *papeles de color con ayuda de la *escoba que estaba en el *suelo.} CONTINUATION SENTENCE: Susana / swept [mopped] / the ¯oor / thoroughly. {Susana / barrio [fregoÂ] / el suelo / con esmero.} PREDICTING CONTEXT: Early in the morning the *gardener took the *hose, connected it to the *water *tap and started his work.

168

{Por la manÄana temprano el *jardinero cogio la *manguera, la enrosco al *grifo del *agua e inicio su tarea.} CONTROL CONTEXT: With *water, the *gardener cleaned the *hose and the *tap that the painters had carelessly soiled the day before. {El *jardinero limpio con *agua la *manguera y el *grifo que los pintores habõ an ensuciado el dõ a anterior} CONTINUATION SENTENCE: The gardener / irrigated [pruned] / the plants / of the garden. {El jardinero / rego [podoÂ] / las plantas / de la urbanizacioÂn.} Note: Target words in italics (inferential targets: outside the brackets; non-predictable targets: inside the brackets). Asterisks, slashes, and brackets did not appear in the stimuli. Asterisks indicate the content words shared by the predicting and the control context, to control for word-based priming. Slashes (/) in the continuation sentence separate each of the four regions (i.e., pretarget, target, posttarget, and ®nal).

References Alameda, J. R., & Cuetos, F. (1995). Diccionario de frecuencia de las unidades linguÈõÂsticas del castellano [Dictionary of lexical frequencies of Spanish linguistic units]. Oviedo, Spain: SEP, Universidad de Oviedo Altarriba, J., Kroll, J. F., Sholl, A., & Rayner, K. (1996). The in¯uence of lexical and conceptual constraints on reading mixed-language sentences: evidence from eye ®xations and naming times. Memory and Cognition, 24, 477±492 Brysbaert, M., & Vitu, F. (1998). Word skipping: implications for theories of eye movement control in reading. In G. Underwood (Ed.), Eye guidance in reading and scene perception (pp. 125± 147). Oxford: Elsevier 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 Calvo, M. G., & Castillo, M. D. (1998). Predictive inferences take time to develop. Psychological Research, 61, 249±260 Calvo, M. G., Castillo, M. D., & Estevez, A. (1999). On-line predictive inferences: processing time during vs after the priming context. Memory and Cognition, 27, 834±843 Du€y, S. A. (1986). Role of expectations in sentence integration. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 208±219 Ehrlich, S. F., & Rayner, K. (1981). Contextual e€ects on word perception and eye movements in reading. Journal of Verbal Learning and Verbal Behavior, 20, 641±655 Fincher-Kiefer, R. (1993). The role of predictive inferences in situation model construction. Discourse Processes, 16, 99±124 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 Garrod, S., O'Brien, E. J., Morris, R. K., & Rayner, K. (1990). Elaborative inferencing as an active or passive process. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 250±257 Graesser, A. C., Millis, K. K., & Zwaan, R. A. (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. (1994). Methods in reading research. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 1±31). San Diego, CA: Academic Press 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 Klin, C. M., GuzmaÂn, A. E., & Levine, W. H. (1999). Prevalence and persistence of predictive inferences. Journal of Memory and Language, 40, 593±604 Klin, C. M., Murray, J. D., Levine, W. H., & GuzmaÂn, A. E. (1999). Forward inferences: from activation to long-term memory. Discourse Processes, 27, 241±260 Lavigne, F., Vitu, F., & d'Ydewalle, G. (2000). The in¯uence of semantic context on initial eye landing sites in words. Acta Psychologica, 104, 191±214 Liversedge, S. P., Paterson, K. B., & Pickering, M. J. (1998). Eye movements and measures of reading time. In G. Underwood (Ed.), Eye guidance in reading and scene perception (pp. 55±75). Oxford: Elsevier 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 Magliano, J., Graesser, A. C., Eymard, L. A., Haberlandt, K., & Gholson, B. (1993). Locus of inference and inference processes during text comprehension: a comparison of gaze durations and word reading times. Journal of Experimental Psychology: Learning, Memory, and Cognition 19, 704±709 McKoon, G., & Ratcli€, R. (1986). Inferences about predictable events. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 82±91 McKoon, G., & Ratcli€, R. (1989). Semantic association and elaborative inference. Journal of Experimental Psychology: Learning, Memory, and Cognition, 15, 326±338 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 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. J., 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 Pickering, M. J., & Traxler, M. J. (1998). Plausibility and recovery from garden paths: an eye-tracking study. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24, 940± 961 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 Rayner, K. (1998). Eye movements in reading and information processing: 20 years of research. Psychological Bulletin, 124, 372±422 Rayner, K., & Sereno, S. C. (1994). Eye movements in reading. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 57±81). San Diego, CA: Academic Press Rayner, K., Sereno, S. C., Morris, R. K., Schmauder, A. R., & Clifton, C. (1989). Eye-movements and on-line comprehension

169 processes. Language and Cognitive Processes, 4 (special issue), 21±49 Rayner, K., & Well, A. D. (1996). E€ects of contextual constraint on eye movements in reading: a further examination. Psychonomic Bulletin and Review, 3, 504±509 Schilling, H. E., Rayner, K., & Chumbley, J. I. (1998). Comparing naming, lexical decision, and eye ®xation times: word frequency e€ects and individual di€erences. Memory and Cognition, 26, 1270±1281 Schustack, M., Ehrlich, S. F., & Rayner, K. (1987). Local and global sources of contextual facilitation in reading. Journal of Memory and Language, 26, 322±340

Till, R. E., Mross, E. F., & Kintsch, W. (1988). Time course for priming and inference words in a discourse context. Memory and Cognition, 16, 283±298 Trabasso, T., & Magliano, J. P. (1996). Conscious understanding during comprehension. Discourse Processes, 21, 255±287 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