Thirty-four Chinese students (13 males and 17 females, M_age = 20.10, range 18–23, SD = 1.35) from a university participated in the current study. All participants were native speakers of Mandarin Chinese with no history of neurological disease or psychiatric disorders. All participants were right-handed as assessed by the Edinburgh Handedness Inventory (Oldfield, 1971) (M = 0.85, range 0.60–1, SD = 13.19) indicating right-handedness. Participants also engaged themselves in a Digit Span Forward task (M = 10.37, range 8–12, SD = 1.07) with a total score of 12 and a Digit Span Backward task (M = 8.07, range 5–10, SD = 1.74) with a total score of 10 (Gong, 1983), indicating high levels of working memory for participants.

A 63 (3 practice and 60 experimental) short passages were chosen from 192 short passages. These short narratives were based on typical daily events. We chose 63 short passages out of the 192 passages by two norming procedures, namely, a cloze procedure for another group of participants to infer what might happen next after reading the first sentences in each passage by writing down two two-character Chinese verbs and a rating procedure by still another group of participants for appraisals in a probability judgment test on a 7-point Likert-type Scale. Details for the norming study of the 63 passages used in the current experiment can be found in Xu et al. (2023). These 63 passages had strong contextual constraints inducing a main strongly activated predictive inference and a weaker alternative through rating. There were five sentences in each passage. Either the contents or the wording of the first four sentences were in exact accordance with those used in Experiment 2 of Xu et al. (2023). The introduction (Sentences 1–3) was kept together with the fourth sentence in the revise condition, which contains conflicting information with the primarily generated predictive inferences and the no revise condition strengthening the possibility of the primary inference as well as a neutral condition continuing merely the story. The first four sentences were followed by an ERP sentence in two versions. The ERP sentence in two versions continued the whole story and differed in the disambiguating words (either Target 1 or Target 2) that ended the ERP sentence. The ERP sentences were modified very carefully to ensure that they could continue the story in a natural way. An example experimental material is shown in Table 1.

The 63 five-sentence long Chinese narratives were followed by comprehension questions encouraging participants to have a good comprehension of the contents. The target words consisted of action verbs of two Chinese characters with no significant difference between the word frequencies of Target 1 (M = 15.16, SD = 22.14) per million and Target 2 (M = 14.33, SD = 23.24) per million (Cai and Brysbaert, 2010), t(59) = −1.20, p = 0.23 or between the number of strokes for Target 1 (M = 18.13, SD = 4.75) and Target 2 (M = 17.07, SD = 4.34), t(59) = 1.34, p = 0.18. In order to make sure that all short narrative texts were globally consistent, a rating task was carried out. Thirty-one participants with similar ages and educational background took part in the rating task judging whether the short narratives were globally coherent or not, on a seven-point scale ranging from 1 (extremely coherent) to 7 (extremely incoherent) with 4 being neutral. Results of t-tests showed that the 5-sentence short narratives were considered as globally coherent both in Target 1 condition (M = 4.67, SD = 0.22) and Target 2 condition (M = 4.65, SD = 0.20), with no significant differences between the two conditions, t(59) = 0.66, p = 0.51.

Stimuli were presented using E-Prime software (Schneider et al., 2002) and were displayed on a 19-inch. CRT video monitor with a refresh rate of 75 HZ on a black background. Each trial began with a cross “+” in the middle of the screen with a duration of 750 ms. The first four sentences in each passage were presented one by one. Participants rested their right thumbs on the line-advance key and pressed the key to erase the present sentence while presenting the next one. The ERP sentence, or the fifth sentence, was presented word by word with a duration of 300 ms. There was also a delay of 700 ms after the disambiguating word to ensure successful recording of the electrophysiological activities for a sufficiently long time-window. Participants were required to try not to blink during the presentation of words in the ERP sentence. Participants were requested to response to a true/false comprehension question after reading the ERP sentence. Participants pressed the designated true (“m”) or false key (“z”) to give a response. Each of the 60 sets of experimental texts (in two conditions) assigned into two versions were counterbanlanced and presented to every participant only once, keeping 10 randomized passages in each condition per block (altogether 6 blocks). Same numbers of participants were engaged in each condition. Participants in the practice section should achieve more than 90% accuracy rate in three trials to make sure that they had comprehended and followed the instructions well. The whole experiment lasted for about 45 min.

Electroencephalogram (EEG) was recorded using Ag-AgCl electrodes mounted on an elastic cap (NeuroScan Inc., Herndon, VA, USA) from 64 active electrodes in a 10/20 system montage. Signals were sampled at 1000 Hz and filtered offline at 0.1–30 Hz. We collected ocular movements and blinks by two pairs of channels. The first pair was the vertical electrooculogram situated in the left eye of the participant, with one electrode supra and another infraorbitally to measure blink artifact. The other pair was the horizontal electrooculogram placed in the external canthi, with one electrode on the left and another on the right side to register eye movements. Impedances were kept below 5 kΩ. EEG preprocessing and ERP analysis were conducted with Curry 8.0 software. Ocular artifacts were eliminated from the data. Epochs containing other artifacts were rejected (1.67%) with potentials exceeding ±100 μv. ERP epoch length was1000 ms and a 200 ms pre-stimulus baseline correction was applied. Individual averages were re-referenced off-line to the average of left and right mastoids. Time windows of interest for ERP analysis were determined by visual inspection and previous studies. Analysis of P3a, P3b, and N400 followed such ROIs based on Pérez et al. (2020) to show the mismatch detection, alternative predictive inference disconfirmation and integration of the alternatives. The six regions of interest (ROI) were chosen out of the 64 electrodes, each containing five electrodes based on electrode sagitality and laterality. The five sites of electrodes were grouped between the two hemispheres (left and right) as the left frontal (LF, averaging among F1, F3, F5, FC3, and FC5); the right frontal (RF: F2, F4, F6, FC4, and FC6); the central (C: C1, C2, CZ, FCZ, and CPZ); the left parietal (LP: P1, P3, P5, CP3, and CP5); the right parietal (RP: P2, P4, P6, CP4, and CP6), and the occipital (O: O1, O2, POZ, PO3, and PO4). Posterior-PNP or P600 mean amplitude was measured across 11 parietal-occipital channels (CP1, CP2, CZ, CP5, CP6, P3, P4, P7, P8, O1, O2) to capture posterior positivity effects (similar to Kuperberg et al., 2006; Rasenberg et al., 2020). For f-PNP, there have been a mixture of electrode choices (see Zirnstein et al., 2018; Kuperberg et al., 2020; Hodapp and Rabovsky, 2021). We chose six frontal electrodes, i.e., F1, F3, F5, F2, F4, and F6 showing the most prominent f-PNP effect similar to Zirnstein et al. (2018) for predictive inference revision analysis.

For answering all the questions above, the mean amplitude data was analyzed to see if participants could detect the mismatch between Target 1 from the information disconfirming the primarily generated predictive inferences and the introduction (Sentences 1–3). If participants could detect the mismatch, there should be a reduction in the mean amplitude data in ROIs associated with P3a while reading Target 2 representing the meaning of alternative predictive inferences compared with Target 1representing the meaning of the primary predictive inferences. We carried out a 2 (target word types: Target 1 and Target 2) × 3 (ROIs: LF, RF, and C) repeated measures ANOVA on the mean amplitude data in the time window of 200–280 ms. We found significant main effect of target word types, F(1, 29) = 11.15, p = 0.002, η²_p = 0.28. The mean amplitude for ROIs related with P3a in Target 1 (M = 4.59, SE = 0.53) were significantly more positive than that in Target 2 (M = 3.89, SE = 0.51). We did not find significant main effect of ROI, F(2, 58) = 2.73, p = 0.07, η²_p = 0.09. We failed to find significant interaction between the contextual conditions and ROIs, F(2, 58) = 2.20, p = 0.12, η²_p = 0.07, either.

According to the hypothesis stated above, readers could further deactivate the revised predictive inference and enhance the activation levels of the alternative predictive inference when supported by more incoming information. We analyzed the mean amplitudes of Target 1 and Target 2 ERPs in the 200–280 ms at posterior and occipital electrodes with a repeated-measures ANOVA. There was a significant main effect of target word types, F(1, 29) = 11.54, p = 0.002, η²_p = 0.29. We found significantly more positivity with Target 1 (M = 2.67, SE = 2.90) than Target 2 (M = 2.07, SE = 3.08). We also detected a significant main effect of ROIs, F(2, 58) = 10.44, p < 0.001, η²_p = 0.27. There was significantly more positivity in RP (M = 3.20, SE = 0.46) than LP (M = 2.33, SE = 0.56) than O (M = 1.57, SE = 0.56). We did not find significant interaction between the contextual conditions and ROIs, F(2, 58) = 1.69, p = 0.19, η²_p = 0.06.

According to our hypothesis, readers could integrate the alternative predictive inferences when they read Target 2 at the end of the ERP sentence. We conducted a 2 (target word types: Target 1 and Target 2) × 6 (ROIs: LF, RF, C, LP, RP, and O) repeated measures ANOVA on the mean amplitude data for N400 in the 300–410 ms time window. This analysis showed a significant main effect of target word types, F(1, 29) = 5.50, p = 0.03, η²_p = 0.16, with less negativity for Target 2 (M = 1.13, SE = 2.29) compared with Target 1 (M = 0.60, SE = 0.32). We found no significant main effect of ROIs, F(5, 145) = 2.30, p = 0.11, η²_p = 0.07. No significant two-way interaction effect between target word conditions and ROIs was found, F(5, 145) = 1.83, p = 0.16, η²_p = 0.06, either.

In accordance with the above hypothesis, readers could carry out re-analyses when primarily generated predictive inferences were in conflicts with the new input when reading Target 2 at the end of the ERP sentence. We analyzed the mean amplitudes of target words with a 2 (target word types: Target 1 and Target 2) × 11 (electrodes: CP1, CP2, CPZ, CP5, CP6, P3, P4, P7, P8, O1, and O2) repeated measures ANOVA for P600 component in the time window of 500–700 ms. We found a significant main effect of target word types, F(1, 29) = 7.44, p = 0.030, η²_p = 0.15. There was significantly more positivity tendency with Target 1 (M = 0.99, SE = 0.18) than Target 2 (M = 0.78, SE = 0.15). We also found a significant main effect of electrodes, F(10, 290) = 10.22, p < 0.001, η²_p = 0.26, with CP1 (M = 1.37, SE = 0.23), CP2 (M = 1.14, SE = 0.21), CPZ (M = 1.18, SE = 0.22), and CP5 (M = 1.22, SE = 0.17) generally having larger amplitude than other electrodes. We also found a significant two-way interaction effect between target word types and electrodes, F(10, 290) = 18.71, p < 0.001, η²_p = 0.39. Simple effects analysis showed that there were significant effects for the 11 electrodes in both target word conditions and there were significant effects for target words at electrodes such as CP2, CPZ, P4, and CP6.

To detect whether readers could revise the disconfirmed predictive inferences when they read Target 2 at the end of the ERP sentence, we conducted a 2 (target word types: Target 1 and Target 2) × 6 (electrodes: F1, F3, F5, F2, F4, and F6) repeated measures ANOVA for f-PNP in the six electrodes in the 500–700 ms. A significant main effect of target word types, F(1, 29) = 6.42, p = 0.02, η²_p = 0.18, with significantly more positivity tendency with Target 1 (M = 1.25, SE = 0.21) than with Target 2 (M = 0.84, SE = 0.21) was found. There was a significant main effect of electrodes, F(5, 145) = 3.74, p = 0.003, η²_p = 0.11. The frontal areas including F2 (M = 1.23, SE = 0.23), F1 (M = 1.22, SE = 0.21), F3 (M = 1.06, SE = 0.18), and F4 (M = 1.05, SE = 0.25) had bigger amplitudes than F5 (M = 0.84, SE = 0.15) and F6 (M = 0.89, SE = 0.23). There was no significant two-way interaction effect between contextual conditions and electrodes, F(5, 145) = 0.43, p = 0.83, η²_p = 0.01.

Reduction in P3a mean amplitude data when reading Target 2 suggests that the word representing the once weakly activated predictive inference alternative is more consistent with the current situation model than the primarily strongly activated but later revised predictive inference. Readers are keen on detecting the mismatching information during reading comprehension. This is in accordance with the two-component model of evaluative comprehension (Isberner and Richter, 2014) that the initial detection of inconsistencies is a routine part of comprehension. And for the RI-Val model, the detection processes are memory-based and carried out routinely and efficiently, requiring few demands on cognitive resources (Richter et al., 2009). Therefore, the detection of mismatch is almost automatic.

The significant reduction in P3b mean amplitude data for Target 2 than Target 1 shows that readers could suppress or disconfirm inconsistent predictive inferences when mismatch information follows (e.g., Iseki, 2006; Pérez et al., 2015). According to the Structure Building Framework (Gernsbacher, 1990; Gernsbacher, 1997), activation levels of the activated memory nodes are subject to the mechanisms of suppression. Through suppression, inappropriate information is made unavailable while with enhancement, appropriate information becomes more easily available (Gernsbacher, 1997). The enhancement in P3b mean amplitude data of Target 1 showed readers’ efforts to update the once-activated but no-longer-relevant information, or in other words, the suppression procedure.

The significant differences of N400 mean amplitude data for Target 1 and Target 2 suggest easier integration of the latter, which represents the alternative predictive inferences. This result is consistent with findings of Pérez et al. (2015, 2019) which found more negativity in the revise condition than the no revise and neutral conditions for participants with high working memory capacity. These findings indicate that readers could integrate the alternative predictive inferences upon encountering the inconsistent information input. Yet, the results are different from those of Xu et al. (2023), which failed to find either successful integration of the alternative predictive inferences while reading the conflicting sentences. Major differences between the present experiment and others lie in the specific times at which integration was examined. By contrasting the results in the current study and those of previous research, we infer that the alternative inferences could be integrated within readers’ situation model, but with a delay.

There was a significant difference for P600 data in the two target word conditions, showing costs of disconfirming the originally dominant predictive inferences. Previous studies showed that syntactic reanalysis in “garden-path sentences” happened immediately when comprehenders encountered information in conflicts with their initial analysis (Frazier and Rayner, 1982). The reduction in the amplitude data of P600 for Target 2 in our experiment suggests that the costs of readers in their semantic/syntactic adaptation or updating when incoming words are not explained away by the predictions that readers have primarily made. This suggests that the disconfirmed predictive inferences remain to be active in readers’ memory representation though disconfirmed by conflicting information input, incurring cognitive processing costs. It is possible that the already-encoded primary predictive inferences still lingers on in readers’ working memory representation, although their activation levels have been reduced after the primary inferences have been disconfirmed.

The reduction in the mean amplitude data of f-PNP showed that readers had successfully revised the disconfirmed predictive inferences after encountering conflicting information with the primary inferences. Unlike previous investigation into predictive inference revision, which found immediate revision of inconsistent inferences upon accessing conflicting information inputs (e.g., Pérez et al., 2015, 2019), the current study detected the successful revision procedure, even with a delay. Studies of prediction at the sentence level have found that in natural sentence processing, predictions are instantaneously updated whenever new information becomes available (e.g., Szewczyk, 2018). However, information that disconfirms a strong prediction may have a disruptive effect on processing and may prevent comprehenders from immediately revising their predictions (Federmeier et al., 2007; van Petten and Luka, 2012).

In sum, results of the current study corresponds to those of other studies (e.g., Nahatame, 2014; Pérez et al., 2015), which shows that readers could detect the mismatch, disconfirm the inconsistent predictive inferences, integrate the predictive inference alternatives, and revise the primarily highly activated predictive inferences. The results of our experiment suggest that readers could successfully integrate an initially weakly activated alternative predictive inference after receiving enough new supporting information. This might also indicate that the very strong initial contextual constraint could hardly be disconfirmed by insufficient local disconfirmation. In addition, the primarily generated but later disconfirmed predictive inferences still exert influences at later reading comprehension stages This is in accordance with such theories as the KReC framework (Kendeou and O’Brien, 2014) holding that even though the revised information has lost activation and becomes less accessible for the comprehension of subsequent texts, it still exists in the long-term memory representation, and it can still be reactivated and disrupt comprehension. This could explain why there is delayed revision of the disconfirmed predictive inferences.

With all these indications, it is possible that the revised predictive inference has really been deactivated and the activation levels of the originally low-competitive alternative inferences have been enhanced with more supporting information. This might well explain the failure of integration or revision in some studies. For example, Xu et al. (2023) found that readers failed to integrate the low-competitive alternative predictive inferences into their long-term representation in Experiment 1 in which little attention was paid to the activation levels of an alternative predictive inference in the first place. Experiment 2 managed to ensure a high contextual constraint for the preliminary predictive inference while providing low probability of eliciting other strongly constrained alternative predictive inferences, thus eliciting only weakly constrained alternatives that might be derived from the introduction at the same time. Integration of the alternative predictive inferences, though supported by the disconfirming information, failed to happen immediately after the disconfirmation of the conflicting information. By using the P600 component as an index of predictive inference re-analysis or revision procedure in the experiments, the research did not find evidence of successful revision by readers upon the disconfirming information, either. Failure to integrate the alternative predictive differences or revise the disconfirmed predictive inferences immediately upon encountering conflicting new inputs can happen for many reasons.

There are three possible explanations for the delayed integration of the alternative predictive inferences and revision of the disconfirmed predictive inferences. Firstly, for materials used in Xu et al. (2023), the activation of a newly revised predictive inference came mainly from a single sentence containing a contradictory context. There was a lack of enough resonance of the alternative predictive inferences from the contextual information and the disconfirming information in the fourth sentence. Therefore, it might be contextually insufficient for the activation and encoding of the alternative predictive inferences. The alternative inferences that were initially weakly activated were not probable enough to affect the processing of the target inference (see Cranford and Moss, 2023). According to the minimal encoding hypothesis (McKoon and Ratcliff, 1992), when there is another possible consequence, only some of the semantic features of the inferred proposition are encoded into readers’ working memory. With more supporting information, however, one of the alternatives can become more specific and highly activated. Secondly, the activation levels of an alternative predictive inference can be enhanced with more supporting information that comes in during reading comprehension. According to findings in previous studies, predictive inferences are activated in a general way and are modified or become quickly lost if not supported by the following information (Keefe and McDaniel, 1993). Readers might initially draw general and flexible predictive inferences with more than one possibility. They try to make them more specific predictive inferences as the reading process goes on (McKoon and Ratcliff, 1986; Lassonde and O’Brien, 2009). Moreover, when the amount of information supporting the alternative predictive inferences increases, the activation of the once low-competitive predictive inferences will be enhanced. This assumption could be supported by Harmon (2005), in which the researcher adopted a probe-naming task to detect the activation of a primary predictive inference. The results indicated that the primary predictive inference could be activated if there was low distracting information before the inference-evoking sentence. It was concluded that with less distracting information, the probability of a predictive inference would be increased. Conversely, the more distracting information there was, the less probable the predictive inferences might happen. Successful integration detected at a later stage suggests that a little bit more supporting information could specify and further activate the alternative predictive inference to make it more relevant to the current situation model, thus easing its integration. Thirdly, the delayed integration or revision could also be viewed as a spill-over effect, in which readers do not begin resolving the inconsistency until after moving past the contradictory sentence. The delayed integration or revision might be related to a spill-over effect in contradiction solution. Previous studies constructed post-contradictory sentences to check the spill-over effects of predictive inference activation (e.g., Klin et al., 1999a, 1999b; Weingartner et al., 2003). Consistent with the findings of previous studies, the revision process is more complete after the contradiction appears and some integration processes have been shown to be delayed until the end of the sentence (Rayner et al., 1989; Guzmán and Klin, 2000). In other words, the delayed integration or revision aligns with the findings of previous studies that have found observed activation of predictive inferences with a longer delay. For instance, in Experiment 1 of Weingartner et al. (2003), the Inter-Stimulus Interval between the critical context and the naming probe was increased from the 500 ms in Klin et al. (1999a) to 1,500 ms. With this longer delay, evidence of predictive inferences was provided in the distractor version. Lastly, the revised information may still exists in the long-term memory representation of readers and exerts influences on subsequent reading processes. According to the KReC framework, this deactivated information can still be reactivated, even though it has lost activation and becomes less accessible during the comprehension of follow-up information. Prior studies have found that readers have difficulties removing the incorrect information from their working memory (e.g., Wilkes and Leatherbarrow, 1988; Johnson and Seifert, 1994; Nahatame, 2014). The outdated information might be reactivated whenever supported by incoming information (e.g., Campion, 2004; Guéraud et al., 2005). Similarly, Weingartner et al. (2003) discovered that the targeted inference was not deleted in the presence of alternative consequences.

In addition, the conflicts between global and local constraints could help further explain the results in our current study. For materials used in Xu et al. (2023), there were strong contextual constraints for eliciting the primary predictive inferences. However, its contextual support was not sufficient for the activation and encoding of the alternative predictive inferences, because predictive inferences generated by the global context (the introduction) conflicted with the local context when readers read containing contradictory information against the global context. According to the fruitful results concerning the interplay between global and local context in research areas such as the processing of counterfactual conditionals (e.g., Ferguson et al., 2008; Nieuwland and Martin, 2012), fictional contexts (e.g., Nieuwland and Van Berkum, 2006; Filik and Leuthold, 2008) and concessive sentences (e.g., Xiang and Kuperberg, 2015; Rich and Harris, 2022), the global context serves to constrain the effect of the local context on predictability, even at the earliest stages of lexical processing. Although the traditional view held that obtaining evidence for the activation of predictive inferences only occurred when the strong supportive context immediately preceding the point at which activation was measured (e.g., McKoon and Ratcliff, 1986; Keefe and McDaniel, 1993), recent studies have proved that activation does not depend on the immediate local context and information from earlier portions of a text can play an important role in the overall pattern of activation (e.g., Cook et al., 2001; Guéraud et al., 2008). According to the incremental processing point of view, contextual support can come from anywhere in the text and global context can mediate expectations generated by local constraints (Lassonde and O’Brien, 2009; Rich and Harris, 2022). Facilitation for predictable words has been found in the presence of highly constraining local contexts (words) directly preceding the critical word (e.g., Fitzsimmons and Drieghe, 2013; Chow and Chen, 2020). Therefore, results of the current study evidence that when the global context is in conflicts with the local context, the global context constrains the effects of the local contexts on predictive inference making and revision.