Because Experiment 1 showed that a manipulation of comprehension question difficulty had a strong impact on the regression behavior at the end of a sentence, we asked whether a different experimental setting would also affect reading behavior. We there- fore compared the reading behavior of subjects lying in a fMRI scanner with those sitting more or less comfortably on a chair and reading sentences on a computer screen. As for
146 the question difficulty manipulation of Experiment 1, we did not find any impact of ex- periment type on first pass reading behavior. In addition and in clear contrast to Exper- iment 1, however, experiment type did not affect the probability of regressions, either. Only the anomaly effect was reduced for the combined fMRI / eye-tracking experiment in the object region (although it was still present), which indicates that people were spending less time rereading the object region. We will discuss this particular pattern below in the context of the single fixation durations.
Surprisingly, our word-based analysis revealed that experiment type neverthe- less had a general impact on reading behavior although it was not visible in the region- based analysis. Specifically, single fixation durations were always longer for the com- bined fMRI / eye-tracking experiment than for the single eye-tracking experiment. This of course raises the question as to why increased single fixation durations did not lead to increased first pass reading times. Obviously, the subjects made more but shorter fix- ations in Experiment 2a, which led in total to approximately the same amount of time spent in a region as in Experiment 2b.
One explanation for this could be that the eye-tracker in the scanner was not as accurate as the one outside the scanner (i.e., combining several shorter fixations into one longer fixation), although both were from the same company (SR Research) and had the same high resolution of 1000 Hz. But in contrast to the tower-mount system used in the single eye-tracking experiment, the eye-tracker in the scanner had to correct for (small) head movements as well. Most importantly, however, the distance from the eyes to the monitor was greater in the scanner, perhaps leading to less precise measurements.
However, it is also possible that the subjects in the scanner in fact applied a dif- ferent basic reading behavior with less single fixations, probably due to the more unfa- miliar reading environment. In any case, since this reading behavior did not affect first pass reading times nor sentence interpretation, we argue in terms of our Information Gathering Framework that experiment type did not adjust the forward threshold. We rather think that the results emphasize the fact that fixations during sentence reading can be best viewed at least as word-driven, but maybe better as even region-driven: Thus, the shape of the eye movement pattern during sentence and text reading seems to be best explained by word-based, or even region-based accounts rather than single fixa- tion accounts.
147 In sum, the results show that experiment type did not affect sentence reading behavior on first pass, which implies that it adjusted neither the forward nor the back- ward threshold in terms of our Information Gathering Framework, in contrast to the question difficulty manipulation in Experiment 1. This is a very important finding be- cause it emphasizes the reliability of the eye-tracking data acquired in the scanner and has methodological implications for future research. Thus, the results indicate that eye- movement data from combined fMRI / eye-tracking experiments can in fact be used to explore the neural correlates of natural reading. In addition, it shows for a further time that the basic mechanisms of sentence interpretation during reading are largely unaf- fected by experimental settings like task or environment.
Another focus in our analysis of the eye-tracking data was set upon the examina- tion of fixation durations before regressions. Our results clearly replicate the findings from Experiment 1 as well as from the literature, and further support the assumptions of the Information Gathering Framework. Again, we were able to show that fixations pre- ceding regressive eye movements were of shorter duration than those preceding pro- gressive eye movements. In particular, whereas this was true for all regions, we once again found evidence that fixations occurring before regressions at the end of a sentence (which we view as a response to missing expected evidence) were as long as their pro- gressive counterparts. This is a finding which to our knowledge has never been previ- ously reported in the literature and which can be well explained by the Information Gath- ering Framework.
In contrast to Experiment 1, we also found (despite the significant effect of ex- periment type, see above) that anomaly led in general to increased fixation durations for anomalous sentences in Region 3 (verb), 4 (object–determiner) and 5 (object–noun), although the effect was not very strong compared to the effects of saccade and experi- ment type. This finding is nonetheless not unexpected, given the fact that reading times for these regions were also increased. However, one could argue that in terms of the In- formation Gathering Framework, fixations in anomalous sentences should be shorter than in non-anomalous sentences because they reflect integration difficulties. Note how- ever, that the Information Gathering Framework makes the prediction that only fixations before regressions should be shorter compared to progressions when these regressions are caused by integration difficulties. And this is the exact pattern we found in our data, because the effect of sentence type (anomalous vs. non-anomalous) did not override the effect of saccade type. In addition, the forward threshold (which triggers a regression in
148 the case of integration difficulties) is not defined as a time threshold but as a threshold of confidence level. Thus, for anomalous sentences it often (but not necessarily) happens that the computation of the confidence level already reveals some problems in the lower linguistic levels, which leads to increased fixation durations. Since we did not find any reliable effects of sentence type on fixation durations in Experiment 1, this could suggest that the processing difficulties on the single fixation level were simply not strong enough to increase single fixation durations in general although these effects were clearly visible in first pass time.
In addition to the general impact of anomaly, we found that the anomaly effect was reduced for the combined fMRI / eye-tracking experiment, indicated by a significant interaction with experiment type in region 3 (verb), 5 (object–noun), 7 (pre-final) and 8 (final). In the object region, this interaction was also found in go-past time in the region- based analysis. Together with the lack of an interaction in first pass regression probabil- ity in all regions, this suggests that these reduced fixation durations are probably due to a shorter rereading of these regions instead of a general reduced probability to reread (i.e., by making less regressions). Thus, subjects in the combined fMRI / eye-tracking ex- periment appeared to spend less time in rereading sentence material when faced with anomalous sentences compared to subjects in the single eye-tracking experiment. It could be that this was just because the subjects in the scanner wanted to finish the ex- periment earlier. Note, however, that the effect was still small and failed to reach signif- icance in the region-based analysis for region 4 (spill-over) and 5 (final). Thus, we should consider these results with caution. Importantly, experiment type did not influence the detection of an anomaly because the anomaly effect was still present in all regions. Con- sequently, they do not question the general reliability of the eye-tracking data, because language processing was not affected (note also – although we did not examine this in more detail here – there were no differences with regard to comprehension accuracy between the two experiment types). Nor do the results provide any difficulties for the Information Gathering Framework because the model is not specified with regard to sec- ond pass reading as yet.
In our experiments we also employed German counterparts of the English SRA sentences used in Experiment 1 to compare the eye movement pattern and language processing mechanisms between these two languages. Our data replicates the finding from Experiment 1: Even a high association between verb and object did not eliminate the detection of the anomaly. Thus, the anomaly effect was clearly visible in the object
149 region, although a sentence structure with a reversal of thematic roles (likely to induce a plausible interpretation of the anomalous sentences if the reader adopted a superficial reading strategy) was chosen. These results again suggest that sentence interpretation is always complex and not underspecified on the syntactic level, which we interpret as a hint against the assumptions of the good enough approach that proposes that readers adopt a sentence interpretation that is just “good enough” for the task at hand. But note that in contrast to Experiment 1, we did not explicitly manipulate the task (unless we view the experiment type as some kind of task). Thus, the predictions of the good enough approach for Experiment 2 are not as strong as for Experiment 1.
Besides these replications, our experiments also revealed some important differ- ences between the processing of the English and German SRA sentences. In particular, whereas the reading times on the verb region were increased for the English SRA sen- tences, we were not able to find any such effect for their German counterparts in first pass time. An initial interpretation might be that in the German sentences the anomaly effect was delayed and not detected in first pass time on the verb. However, this inter- pretation can hardly be correct as our analysis also showed an increased probability for first pass regressions out of the verb region. It is therefore possible to conclude that the detection of the anomaly is not delayed but rather occurs earlier in German than in Eng- lish. Thus, the early detection of the anomaly may have caused an early regressive eye movement without increasing fixation duration. This dissociation between increased fix- ation duration and regression behavior is one of the core principles of the Information Gathering Framework and fits well to the findings of reduced fixation durations before regressions which has been discussed above.
However, if this assumption holds true, we have to ask the question as to why there is an earlier detection of the anomaly in German than in English. We consider a plausible explanation may be that the two languages differ with regard to their underly- ing form-to-function mappings (MacWhinney et al., 1984). Thus, in English, the actor of a sentence can primarily be identified on the basis of word order information, whereas in German case marking and animacy are more reliable to identify the actor. If we apply these considerations to the SRA sentences used in our studies, we should expect different processing patterns, especially on the verb.
Specifically, in both languages the inanimate initial argument (the flower / der Fehler), which is unambiguously marked as the actor (either by word order as in English
150 or by case marking as in German), causes an animacy violation on the verb position be- cause the verb requires an animate actor. In German, however, actor identification relies primarily on case marking and animacy. Thus, these information cues are evaluated quite early in the argument interpretation and consequently lead to an early error signal. In English by contrast, the animacy information is not used to develop a first interpreta- tion of the argument because sentence interpretation primarily relies on word order in- formation and this information does not lead to any difficulties. In only a later step, the sentence interpretation is matched with world knowledge and an error signal is pro- duced, resulting in longer reading times or regressions. These results are therefore in line with crosslinguistically motivated language processing models such as the extended Argument Dependency Model (eADM; Bornkessel & Schlesewsky, 2006; Bornkessel- Schlesewsky & Schlesewsky, 2009) which claim a different time course in argument in- terpretation for English and German (see also Bornkessel-Schlesewsky & Schlesewsky, 2008, for a discussion of crosslinguistic differences in ERP effects elicited by SRA sentences).
More evidence for different processing patterns between English and German is gleaned from go-past time on the object position. Whereas in both languages we found an almost identical reading pattern in first pass time with only slightly reduced reading times for high compared to low associated anomalous sentences, this pattern diverged in go-past time. In English, the reduced first pass reading times actually led to increased go-past times which suggest that the anomaly effect was detected later for high associ- ated sentences compared to low associated sentences. In German by contrast, a high as- sociation generally led to reduced reading times and the size of the anomaly effect was not shaped. This indicates that in German the anomaly effect was detected regardless of lexical association but based on animacy evaluations. In English by contrast, a high asso- ciation led to a delayed effect of anomaly due to the important role of word order.
Finally, our eye-tracking analysis was able to replicate findings from Experiment 1 with regard to the regression target pattern. Again, we found evidence that the major- ity of regressive eye movements were short, although the results were not as clear as in Experiment 1. Nevertheless, two different target patterns of regressions in the sentence were again apparent. One group of regressions targeted the sentence beginning whereas the second group targeted a position much closer to the current fixation position. This suggests that regressions in general may aim to either reread the sentence from the be- ginning (c.f. the rereading pattern from von der Malsburg & Vasishth, 2011, 2013; and
151 the forward reanalysis strategy proposed by Frazier & Rayner, 1982) or to reread a por- tion that is within 15-20 characters to the left of the current fixation. Although still more research is needed to examine the landing site distributions of regressions in more de- tail, these results are generally in line with the Information Gathering Framework that differentiates between a strategy based target selection and a language based target se- lection whereas the latter can only be applied for words within the perceptual span (which comprises about 15 characters).
In addition, the inspection of the landing site pattern again revealed evidence for a linguistically controlled target selection. Thus, we found increased re-inspections of the subject and verb region in anomalous sentences. Given that these regions cause the integration difficulties because they constitute the animacy violation, this is a very rea- sonable regression pattern. Note, however, that the subject region was also the initial region of the sentence. Thus, it is not possible to distinguish regressions that aim to re- read the sentence from the beginning, from those regressions targeting the location where the potential error source has been expected. Also, the fine-grained analysis of the proportions suggests that the largest difference of targeting probability was found on the subject determiner. Thus, in anomalous sentences the determiner became the regres- sion target more frequently than in non-anomalous sentences. This fits well to the role of the determiner in German because the determiner contains the case marking. In the SRA sentences used in our study, for example, an accusative determiner “den” would have solved the animacy violation. However, whereas this could be interpreted as more evidence for a linguistically constrained target selection, we again have to view these results with caution because the determiner was also the first word in the sentence. Thus, we have to leave it for future research to examine this pattern in more detail.