Omisión de efectos dinámicos de vorticidad

More recently, the degree of coupling of regressive eye movements and linguistic guid- ance was further investigated by von der Malsburg and Vasishth in two articles (von der Malsburg & Vasishth, 2011; 2013).

Based on the results of the three experiments discussed above (Frazier & Rayner, 1982; Meseguer et al., 2002; Mitchell et al., 2008), von der Malsburg and Vasishth pro- posed that the focus on single fixations might not be very informative with regard to the pattern of regressions that result when reanalysis begins. Thus, they developed a method that allowed them to analyze the scanpath of regressive eye movements during sentence

8 _{Mitchell and colleagues leave it open whether at least some regressions are guided by a}

35 reading and to quantify the similarity of these patterns. In order to investigate the rea- nalysis processes in garden path sentences, they used this new approach to re-analyze the data from Meseguer et al.’s study.

In their scanpath analysis, von der Malsburg and Vasishth were able to identify three different reading patterns that occurred after making a regression out of the dis- ambiguating (entraran / entraron) or post-disambiguating region. The first pattern (A) contained a long regressive saccade to the beginning of the sentence and subsequent re- reading of the sentence. The second pattern (B) was characterized by a regression to the beginning of the sentence without rereading9_{, and the third pattern (C) consisted of a}

short regression that targeted the disambiguating material. The analysis revealed that only pattern A (rereading) occurred significantly more often in the garden path condi- tions whereas patterns B and C were equally distributed over the two sentence types. Thus, von der Malsburg and Vasishth concluded that rereading is the preferred strategy when reanalyzing the sentence. In addition, the results suggest individual differences be- tween readers so that for each reading pattern some readers were found who had a pref- erence for this strategy.

The detected rereading pattern (A) stands in clear contrast to the predictions of the selective reanalysis hypothesis and questions the findings of Meseguer et al. who provided evidence for increased re-inspections of the matrix verb dijo (region 2) and the adverb cuando (region 6).Von der Malsburg and Visishth interpreted this contradiction in the way that the reported regressions of Meseguer et al. targeting regions 2 or 6 were just diverged in the course of the following fixations and did not build own patterns. In addition, the lack of findings of increased regression rates into the sentence initial region by Meseguer et al. is explained by short-falling regressions on their way to the sentence beginning.

Thus, von der Malsburg and Vasishth summarize that their results provide little evidence for the selective reanalysis hypothesis and point instead to a strategy compa- rable to the forward reanalysis hypothesis also outlined by Frazier and Rayner (1982) and where the reader starts reading the whole sentence again from the beginning in or-

9 _{This pattern probably was a result of a missing fixation cross at the beginning of the line}

so that the readers anticipated the presentation of the new sentence. In the follow-up study, a fixation cross was used and subsequently the reading pattern disappeared (see von der Malsburg & Vasishth, 2011; 2013).

36 der to reanalyze it. On the other hand, the results do also contradict the time out hypoth- esis that assumes a random walk through the sentence because there was evidence for a more targeted regression process. Hence, von der Malsburg and Vasishth conclude that the results rather suggest that “saccade programming is indeed influenced by low-level properties of the text while a loosely coupled linguistic system is guiding the overall shape of the trajectory.” (von der Malsburg & Vasishth, 2011, p. 124).

However, because the post-disambiguation region in the Meseguer et al. study was the sentence-final region as well, the results could have been interfered by regres- sions that were triggered by processes beyond the initial parsing of the sentence. Thus, von der Malsburg and Vasishth (2013) conducted a follow-up study using Meseguer et al.’s material but made three substantial adjustments: a) They extended the sentence fi- nal region to distinguish sentence wrap-up effects from parsing effects, b) they included a third, non-ambiguous condition in order to examine the interaction of regressions and parsing strategies, and c) they controlled for differences in working memory by perform- ing an operation span test.

As in the former experiment, von der Malsburg and Vasishth identified three reading strategies that appeared after a regression out of the disambiguation or post- disambiguation region was made, from which two were identical to the ones observed earlier: pattern A (rereading) and pattern C (checking the disambiguating material). In addition, a third strategy was found that contained a short regression from the disam- biguation region back to the pre-disambiguation region. Again, only the rereading pat- tern occurred more often in the high-attachment condition and this effect was more pro- nounced for readers with a high working memory capacity. Von der Malsburg and Va- sishth interpreted their findings as additional evidence that rereading is the dominant strategy when reanalyzing a sentence. However, we think that the results lead to three more important conclusions which we will discuss below.

First, the authors argue that predictions about landing sites of regressions are directly linked to the underlying parsing model. Frazier and Rayner, for example, based their hypotheses on the assumptions of the garden path model and expected compre- hension difficulties on the disambiguation region because the parser incorporates the phrase “a mile” as an object into the main-clause verb “jogs” according to the late closure principle. Whereas there is indeed some evidence that readers parse this sentence in the way the garden path model proposes, the parsing strategies of the Spanish sentences used by Meseguer et al. (and von der Malsburg and Vasishth), for example, are less clear.

37 Von der Malsburg and Vasishth discussed the predictions of five parsing models (garden- path model, construal model, good-enough parsing, parallel model, unrestricted race model) with regard to the sentences they used in their experiment and showed that all of them should result in different eye movement patterns. In particular, the models differ with regard to expected parsing difficulties which are assumed to trigger a regressive eye movement. Since regressions did occur in both garden-path and non garden-path sentences, this raises the question if rereading is restricted to reanalysis or if reanalysis is required in either sentence type. This, of course, should influence the expected regres- sion pattern. Thus, in order to understand regressions in the context of sentence parsing it seems necessary to control more carefully with which interpretation the parser comes up with. We think this is a very important issue.

Second, although von der Malsburg and & Vasishth were also focusing on com- prehension difficulties that showed up in first pass reading times and tried to identify the underlying parsing principles (they interpreted their findings as evidence for the good enough approach), they treated regressions out of the (post-)disambiguation re- gion as reanalysis processes. Especially, these reanalysis processes were assumed to be carried out in order to find the correct interpretation of the sentence. But the very low accuracy in the comprehension questions for high-attached sentences (58%) shows that a large number of participants failed to parse these sentences correctly although some of them regressed to reread earlier parts. Since von der Malsburg and Vasishth did not examine the relationship of regression patterns and accuracy, three important questions remain open:

a) Were the readers able at all to diagnose the error in their interpretation cor- rectly, independently of their reanalysis behavior? This ability is a prerequi- site for the selective reanalysis hypothesis but due to the very poor accuracy results it seems likely that while many readers came up with comprehension difficulties in first pass reading, they were not able to localize the problem. b) Was there a preference to use a certain regression strategy when readers

knew the source of the error compared with when they just had no idea? Again, the very poor accuracy results could indicate that rereading was the preferred strategy because many readers were completely overburdened and were not able to develop hypotheses on what might have went wrong.

38 c) Was there a reading strategy that was more efficient and led to better com-

prehension (and accuracy results)? Maybe rereading did occur most fre- quently but rarely helped to find the correct interpretation whereas another pattern was more efficient but due to the small number of high-proficient readers it did not show up often.

Finally, the results of the first analysis (von der Malsburg & Vasishth, 2011) showed that readers may have a preference for one reading strategy although the reason for this pref- erence was not specified. In addition, the results of experiment 2 revealed that the reader’s working memory capacity might influence the parsing decisions (i.e., that low- capacity readers tended to leave the attachment-decision open till the disambiguating information was apparent10) and led to different reading strategies, too. These findings

suggest that individual differences between readers could influence parsing decisions and reading strategies. Thus, we should take into account that it might be hard to find one guiding principle for all readers that allows to predict the scanpath signature of re- gressive eye movements.

10 _{This also fits in nicely with findings from the ERP literature (see, e.g., Gunter, Wagner,}

39

3. A new approach: The Information Gathering Framework

After we have reviewed the basic findings on regressive eye movements in sentence reading, we can sum up by saying that the picture they reveal is not very consistent: Alt- hough there is little doubt that regressions are linguistically driven, the strength of this linguistic control remains unclear. Whereas approaches like the selective reanalysis hy- pothesis assume a very strong linguistic guidance on triggering and landing sites of re- gressions, other experimental findings have shown that this hypothesis cannot fully ac- count for the variability in target positions of regressions. Also, it seems implausible that regressive eye movements are just a response to difficulties in integrating linguistic ma- terial since regressions do not only occur in the context of processing problems. Thus, none of the approaches discussed so far may capture the full pattern of regressive eye movements during reading and an elaborated model of regressive saccades is still a ma- jor desideratum, even after many decades of reading research.

The goal of the current thesis is therefore to formulate a new framework that may provide a general tool for our understanding of regressive eye movements, without limiting it to a small range of linguistic phenomena. As a starting point, we will use the model of falling confidence proposed by Bicknell and Levy (2010). But instead of focus- ing on theoretical considerations about reading strategies, the current aim is to develop a realistic model of human reading behavior, which means that the model should be able to cover findings from the existing literature as well as make further testable predictions about reading behavior. This, however, requires some substantial modifications in the architecture of the FC model, so that we will call the new account the Information Gath- ering Framework.

Note that in contrast to the model of falling confidence, the Information Gather- ing Framework is not incorporated into a computational model as yet that allows for simulating reading behavior. Instead, the Information Gathering Framework takes into account more cognitive and linguistic properties of eye movement control than the for- mer model does. But the current considerations should be used by future research to combine these two approaches and to develop a computational version of the Infor- mation Gathering Framework as well.

40

3.1 The architecture of the Information Gathering Framework

In the following, we will first briefly outline the basic ideas of the model, before laying out the new assumptions of the Information Gathering Framework and clarifying the modifications from the FC model in more detail. Afterwards, we will apply this frame- work to several empirical findings reported in the literature before we formulate a cou- ple of predictions based on the model.