Elsevier

Journal of Memory and Language

Volume 83, August 2015, Pages 118-139
Journal of Memory and Language

The effect of contextual constraint on parafoveal processing in reading

https://doi.org/10.1016/j.jml.2015.04.005Get rights and content

Highlights

  • The boundary paradigm was used to assess preview benefit with stimuli from Schotter (2013).

  • Previews/targets were embedded into constrained sentences (Schotter, 2013 used neutral).

  • Semantic associates showed early preview benefit and late cost in constrained sentences.

  • Within-subject experiment replicated effects for both neutral and constrained sentences.

  • Results compatible with E-Z Reader assumption of early oculomotor “hedged bet.”

Abstract

Semantic preview benefit in reading is an elusive and controversial effect because empirical studies do not always (but sometimes) find evidence for it. Its presence seems to depend on (at least) the language being read, visual properties of the text (e.g., initial letter capitalization), the type of relationship between preview and target, and as shown here, semantic constraint generated by the prior sentence context. Schotter (2013) reported semantic preview benefit for synonyms, but not semantic associates when the preview/target was embedded in a neutral sentence context. In Experiment 1, we embedded those same previews/targets into constrained sentence contexts and in Experiment 2 we replicated the effects reported by Schotter (2013; in neutral sentence contexts) and Experiment 1 (in constrained contexts) in a within-subjects design. In both experiments, we found an early (i.e., first-pass) apparent preview benefit for semantically associated previews in constrained contexts that went away in late measures (e.g., total time). These data suggest that sentence constraint (at least as manipulated in the current study) does not operate by making a single word form expected, but rather generates expectations about what kinds of words are likely to appear. Furthermore, these data are compatible with the assumption of the E-Z Reader model that early oculomotor decisions reflect “hedged bets” that a word will be identifiable and, when wrong, lead the system to identify the wrong word, triggering regressions.

Introduction

Recently, researchers have debated whether, and to what extent, readers obtain semantic preview benefit from the upcoming word during reading (Hohenstein and Kliegl, 2014, Hohenstein et al., 2010, Rayner, 2009, Rayner and Schotter, 2014, Rayner et al., 2014, Schotter, 2013, Schotter et al., 2012, Yang et al., 2010). Semantic preview benefit refers to the phenomenon in the boundary paradigm (Rayner, 1975) in which reading times on a fixated target word are faster when a preview word previously in its location (i.e., before it was fixated) is semantically related to the target, compared to unrelated. Semantic preview benefit is controversial because its presence varies depending on which language is tested; originally, semantic preview benefit was not observed in English (Rayner et al., 1986, Rayner et al., 2014) but was observed in German (Hohenstein and Kliegl, 2014, Hohenstein et al., 2010) and Chinese (Yan et al., 2012, Yan et al., 2009, Yang, 2013, Yang et al., 2010). The inconsistency with which semantic preview benefit is observed raises the issue of, not whether it is real, but rather what conditions are necessary and sufficient for it to be observed, which will potentially lead us to a better understanding of the reading process as a whole.

Recently, Schotter (2013) did find evidence for semantic preview benefit in English, but only when the preview and target were synonyms (e.g., startbegin), not if they only shared an associative semantic relationship (e.g., readybegin). Furthermore, Rayner and Schotter (2014) tested whether an orthographic property that differs between English and German (i.e., capitalization of the first letter of nouns) might partially explain these cross-language differences. They found that, in English, initial letter capitalization increased the magnitude of preview benefit generally, and led to semantic preview benefit (for similar effects in German, see Hohenstein & Kliegl, 2014).

One of the reasons semantic preview benefit is controversial is because many researchers presume that firm evidence for its presence would either be incompatible with the E-Z Reader model (a serial attention shift model of oculomotor control in reading) or would only be possible under very specific and rare circumstances (Hohenstein & Kliegl, 2014). However, Schotter, Reichle, and Rayner (2014) used E-Z Reader to simulate the data from Schotter (2013) and demonstrated that semantic preview benefit is not incompatible with its architecture. The components of the model can be grouped into five broad categories: (1) pre-attentive visual processing, which occurs for multiple words so long as they are within the limits of the perceptual span; (2) two serial stages of lexical processing (L1 is an early, cursory stage and L2 is a late, full identification stage); (3) attention allocation; (4) post-lexical integration of word meanings; and (5) two stages of saccade programming (M1, which is labile (cancellable) followed by M2, which is non-labile). The completion of the first lexical processing stage (L1) initiates both the second stage (L2) and the beginning of saccade programming (M1). Thus, because saccade programming is triggered by the completion of only cursory lexical processing, these saccade decisions are considered “dumb” in that they are not initiated based on complete lexical identification of the word. The model accounts for preview benefit by means of the relative timing of the completion of lexical processing on the current word and the saccade away from it. For example, saccade programming requires about 125 ms from the start of planning until execution; if the completion of lexical identification (i.e., L2, which is initiated at the same time as saccade programming) occurs before this time, attention shifts to the upcoming word and begins cursory lexical processing (i.e., L1) on the upcoming word before the eyes move to it.

Schotter et al. took a new perspective on modeling the boundary paradigm, with the E-Z Reader model and thus a new consideration of how parafoveal processing influences the reading process. Specifically, they used the model to estimate how far into lexical processing of the parafoveal word the model had progressed. In the simulations of Schotter’s (2013) data, the model estimated that preprocessing of the preview had reached the L2 stage of the model (i.e., semantic processing) a modest but non-trivial 8% of the time. Importantly, although not discussed by Schotter et al. (2014), a corollary of this finding is that in these cases the model had also reached the stage of saccade programming away from the parafoveal word because the end of the L1 stage initiates both the start of L2 and the start of M1 (i.e., saccade programming). We will return to this idea in the General Discussion.

Given that the other languages that demonstrate semantic preview benefit did not exclusively use synonym previews, the type of semantic relationship cannot be the only explanation of cross-language differences. Schotter, 2013, Laubrock and Hohenstein, 2012 suggested that these cross-language differences in the presence of semantic preview benefit might be explained by differences in orthography; languages with an orthography that is shallow (e.g., German) or non-alphabetic (e.g., Chinese) might be more likely to show the effect because semantics may be accessed sooner during preview, due to less time spent decoding phonology (compared to in English). As a consequence, the earlier access to semantic information from the preview in German and Chinese might allow more time for spreading activation among semantic representations in the linguistic system, allowing for something akin to semantic priming (Schotter, 2013). A direct test of this hypothesis is difficult because it is not possible to rigorously control all the differences across these languages while manipulating the theoretically relevant variables. Instead, in the present study we turn to a different contributing factor to the reading process (and semantic pre-activation) that has received little attention in relation to semantic preview benefit thus far, but may be an important consideration: contextual constraint or expectations of upcoming words (see below).

While it is well-demonstrated that the meaning generated from the prior sentence context1 exerts an influence on language processing, the exact nature of this effect is poorly understood. Part of the lack of clarity surrounding the effect of sentence context is that, across studies, researchers have manipulated it in different ways, measured it with different methods, and used different theoretical constructs to discuss it.

DeLong, Troyer, and Kutas (2014) provide an overview of the distinction between different theoretical constructs, highlighting the following terms, which we will group further as they pertain to reading. 1. Prediction suggests an active, conscious, effortful process in which a single item is predicted and there are benefits for success (i.e., if the predicted word is encountered) and costs for failure (i.e., if any other word is encountered). This construct is more likely to operate in more strategic tasks that require an overt judgment, for which the response time is slower than the typical eye movement in reading (e.g., longer than 250 ms). 2. Pre-activation/Anticipation suggests a less specific process, for which ‘landscapes’ of pre-activated items (i.e., multiple possible words) are dynamically shifted by all available information sources. This construct is a more apt description of what may be happening during reading in that parafoveal (and subsequently foveal) visual information may be what reshapes the “landscapes” of pre-activated semantic concepts that were generated by the context. 3. Lastly, DeLong et al. (2014) note that predictability “can be divorced from prediction, in that an item can be predictable even if it is not predicted” (p. 633), which leads us to the issue of how one can measure expectations generated by the sentence context and the degree to which such measures relate to processes involved in reading.

The most common way to determine the effect of the sentence context on word recognition in reading is to create sentences in which a particular word is highly expected in a particular location (e.g., “The children went outside to…”). To confirm the manipulation, researchers use a modified cloze task (see Taylor, 1953 for the original cloze task) in which the prior sentence context is provided (as above) and subjects supply the word that they think comes next. Typically, the way these data are analyzed is by coding the response as correct (e.g., 1) if the subject supplied the word that the experimenter intended (e.g., “play” in the example, above) and incorrect (e.g., 0) if the subject supplies any other word; the predictability of a word is determined as the proportion of “1” responses for that word. There are some modifications to this coding scheme (e.g., accepting the intended word’s plural/singular counterpart or other small morphological variations), but for the most part this procedure is used to determine how expected a particular word form is, given the context. In order to create a comparison of predictability, reading times on these highly expected words are contrasted with reading times on a word that was never or rarely supplied by the subjects in the norming procedure, but is nonetheless a plausible word that makes sense in the exact same sentence (e.g., “swim” in the example above; generally, this is confirmed with another norming task in which subjects rate how acceptable/sensible the sentence is).

Using this manipulation, several reading studies have found that both the likelihood of fixating a word and the durations of fixations are modulated by cloze predictability (Balota et al., 1985, Drieghe et al., 2005, Ehrlich and Rayner, 1981, Kliegl et al., 2004, Rayner et al., 2011, Rayner and Well, 1996, Zola, 1984). Initially, researchers hypothesized that the effect of predictability was the manifestation of the reading system “guessing” the upcoming word (e.g., McClelland and O’Regan, 1981a, McClelland and O’Regan, 1981b; cf. Rayner & Slowiaczek, 1981), similar to the “prediction” account described by DeLong et al. (2014). On this view, the system uses parafoveal preview to confirm this guess (i.e., that there is a linear relationship between predictability and reading time; see Smith & Levy, 2013 for a discussion). This is the underlying assumption in both the E-Z Reader (Reichle, Pollatsek, Fisher, & Rayner, 1998) and SWIFT (Engbert, Nuthmann, Richter, & Kliegl, 2005) models of oculomotor control in reading.

Recently, however, there have been several findings that question the interpretation that individual words are guessed or predicted. First, the relationship between cloze predictability and reading time is not linear, but rather varies across the predictability spectrum: in reading studies, words that have a high or moderate cloze predictability are fixated for less time than words that have a low cloze predictability, but there does not seem to be much distinction between words with high or moderate cloze predictability (Rayner & Well, 1996; see also, Ehrlich and Rayner, 1981, Hyönä, 1993). This inconsistency could be due to limitations of the cloze task in measuring small differences in absolute predictability, or due to a logarithmic relationship between predictability and reading time. Indeed, a logarithmic relationship was reported by Smith and Levy (2013), who concluded that this relationship suggests “the comprehension system must be able to simultaneously pre-activate large portions of its lexicon in a quantitatively graded fashion” (p. 311). Perhaps then, despite the fact that cloze predictability demonstrates a very strong quantitative relationship with reading times, measuring the effect of context on word processing in this way may not fully capture how readers use context to make reading more efficient. Therefore, we must consider other ways in which the effect of context can be measured.

As an alternative to cloze predictability, which focuses primarily on differing degrees of expectations for different words given the same context, some researchers have employed manipulations of the context itself. One of the limitations of using the cloze task to measure these manipulations is that the primary way it is coded and used to generate stimuli for experiments obscures information about the distribution of responses, which may be of theoretical importance when considering the reading process. For example, a moderately constraining sentence generally leads to cloze probabilities of around .3–.6, meaning that a particular word form is supplied by one third to two thirds of the subjects tested. But given that very same outcome, there could be many different underlying distributions. Fig. 1 demonstrates the cleanest mathematical relationship one could imagine (i.e., constraint works by decreasing the standard deviation of the distribution of responses, honing on words surrounding a particular meaning or event). It is most straightforward to compare a very high constraint sentence in which one word form is highly favored (e.g., a cloze probability of .8 or higher) and almost no other word forms are provided (e.g., cloze scores of <.1) with a low constraint sentence in which there is almost no consistency in the responses, with all word forms having low cloze scores. The interpretation of such a comparison would lend support to the “prediction” hypothesis, described above, because the constraint serves to change the expectations for a single word. However, it is not clear that extremely high cloze sentences are that common, given that they are relatively uninformative (i.e., do not add any new information since the comprehender can infer the completion). One could argue that very high constraint sentences (e.g., “The opposite of black is…”) constitute a qualitatively different type of sentence, rather than an extreme on the constraint continuum (however, adjudicating between these possibilities is beyond the scope of this paper). Thus, it may be more informative to consider sentences with a moderate degree of constraint. That is, such sentences raise the expectations for a particular word form with some consistency, but do not make that word obligatory.

More important for the present study, however, is the question of what the nature of the other responses in moderate cloze sentences looks like. For example, if provided with the sentence, “At the dog show, Spot had the most points and won the coveted…” the experimenter’s intended target word could be “prize,” and in our norming task (see below) this word form was provided 40% of the time—aligning with the conclusion that the sentence provides a moderate degree of cloze. However, inspection of the other responses suggests that crucial information is missing from this coding scheme: “award,” “medal,” and “ribbon” were provided 20%, 10% and 10% of the time, respectively. If one assumes that theses are acceptable synonyms of the target, then we might conclude that the constraint of the sentence is actually quite high (i.e., including all of these responses because the meaning was predicted leads to a cloze score of 80%). Additionally, one could argue that the remaining responses “bone” and “blue” satisfy this coding principle, as well, assuming that a bone would be a nice prize for a dog and that blue was provided because the subject imagined a “blue ribbon” as the prize. Thus, this sentence actually generates a high degree of expectation surrounding the writer’s intended idea (i.e., a reward for a good job), but that this idea can manifest itself as a variety different word forms.

An interesting potential consequence of sentence constraint acting in this way is that we would expect, in contrast to the “prediction” hypothesis, that increasing constraint to a moderate level (i.e., not to the point where one word is obligatory) would make it easier to read, not only the highest cloze word, but also words that are related to the meaning that the reader expects. Indeed, this is what Kutas and Hillyard (1984) found when they measured event related brain potentials (ERPs) to the final words in sentences that were constrained to varying degrees. In this study, Kutas and Hillyard manipulated both the sentence constraint and the target word orthogonally, allowing them to investigate the effect of both variables. However, they note that, because these variables are intrinsically tied, it is impossible to fully dissociate them; cloze probabilities of the best completion in the high constraint sentences were much higher (.92) than in the moderate (.63) and low constraint sentences (.29). When they compared the ERPs across conditions, they found that the most expected word in the most constrained context did not elicit a significant N400 (considered an index of the difficulty of semantic processing; DeLong et al., 2014), but low-cloze words in all sentence constraint conditions elicited N400s of approximately similar amplitudes. In contrast, when comparing ERPs to the best completions across degrees of sentence constraint they found graded effects related to the amplitude of the N400 component, which lends support for the idea of landscapes of pre-activation.

Additionally, Kutas and Hillyard (1984) found the largest N400 to words that were unrelated to the best completion (e.g., “Don’t touch the wet dog” when the word with the highest cloze was “paint”) and a relatively reduced N400 for words that were semantically related to the best completion (e.g., “coffee” in the sentence “He liked lemon and sugar in his ____” where the best completion was “tea”), even though both words had extremely low cloze probabilities and were perfectly acceptable completions. From these data it is unclear whether the reduction in the N400 to these semantically related words stems from distributed pre-activation (i.e., that “coffee” was, to some degree, pre-activated) or to semantic spreading activation from the most highly provided response. However, one could argue that these two interpretations are part and parcel of the same process of distributed expectations.

Further evidence that multiple meanings are being pre-activated comes from a study investigating (self-paced) reading times; Roland, Yun, Koenig, and Mauner (2012) created moderately constraining contexts (e.g., “The aboriginal man jabbed the angry lion with a/an ____”) that yielded a distribution of responses in the modified cloze task, generally surrounding a particular semantic feature (e.g., pointed objects). They compared reading times on words in these sentences with words in sentences that were less semantically constraining (e.g., replacing the word “jabbed” with “attacked” creates less of a preference for a pointed instrument completion). Roland et al. (2012) found that the degree of semantic similarity (measured via latent semantic analyses—LSA) between the most expected word (Mcloze(p) = .24) and the word provided (in the reading task) significantly predicted processing time; reading times were shorter when the provided word was semantically similar to the expected word than when it was semantically dissimilar, even though both words were in the low-cloze probability range (Mcloze(p) = .02).

There is one further issue when considering the use of cloze probability to measure the effect of the sentence context, which relates to the distinction between predictability and prediction. Mainly, when subjects provide responses in the cloze task, they (almost always) provide a single word that satisfies the semantic and syntactic constraints imposed on that sentence position (cf. Roland et al., 2012). That is, the sentence “The children went outside to…” must be completed with a verb to be grammatical. However, in the scenario that the subject imagines in order to complete this task, she may imagine not only the action she will provide, but also auxiliary event-related information. For example, if the subject completing the cloze task were in a northern location in winter, she might imaging that “play” in this scenario entails building a snowman, having a snowball fight, sledding, etc. Therefore, even though almost all the subjects would respond with “play” (perhaps because of its high lexical frequency, or because the idea is fairly cliché) does not mean that they are not also imagining these other entities that are not grammatically licensed, which may be part of elaborative inference or other discourse comprehension processes.

There is empirical data to support this suggestion. Metusalem et al. (2012) conducted an ERP experiment in which subjects read short, two-sentence scenarios that set up a particular event (e.g., “A huge blizzard ripped through town last night. My kids ended up getting the day off from school.”) They then read a third sentence with one of three possible target words (e.g., They spent the whole day outside building a big ____ in the front yard.”) and compared the ERP waveforms across those words. The three words were (1) a highly expected word (snowman; Mcloze(p) = .81), (2) a contextually anomalous word that was related to the event (jacket; Mcloze(p) = .00), or (3) a contextually anomalous word that was unrelated to the event (towel; Mcloze(p) = .00). Metusalem et al. (2012) found a reduced N400 for the event-related word compared to the event-unrelated word, even though neither word was ever produced in the cloze norming task. Thus, these data support the idea that sentence (or in this case, discourse) contexts can generate expectations for event-related words beyond just the next, grammatically- and semantically-licensed word.

Taken together, these findings (Kutas and Hillyard, 1984, Metusalem et al., 2012, Roland et al., 2012) suggest that the sentence context pre-activates (and consequently facilitates processing of) a non-singular set of ideas and words. Given that both contextual constraint and parafoveal preview exert robust effects on reading behavior, the question remains how their influences interact to affect online processing. Crucially, the use of the boundary paradigm (Rayner, 1975) allows us to dissociate the information the reader has about the preview word (in parafoveal vision) from the target word (once it is fixated). An advantage of this experimental design is that it allows us to provide the reading system with previews of words that may not make sense grammatically in the sentence, which change to sensible target words once the display change occurs. Therefore, any potential disruption to comprehension that is created by the hedged bet based on a preview of a semantically associated but grammatically inappropriate word can be repaired if the reader makes a regression to view the perfectly sensible target word. It is likely, then, that the identity of the preview would have more of an influence on early reading measures (e.g., single fixation duration and gaze duration) than on later reading measures (e.g., total reading time) when the system has had more time to encounter the target.

The present study uses the boundary paradigm and the preview manipulation employed by Schotter (2013; identical, synonym, semantically associated, and unrelated) and introduces a manipulation of contextual constraint. That is, Schotter exclusively used neutral sentences with very low cloze probabilities for any of the preview or target words in order to investigate the influence of parafoveal preview alone. But a crucial part of much of the theorizing about the reading process (described above) suggests that contextual constraint may change how preview benefit effects are manifested (i.e., by changing expectations about what words should or could appear). Therefore, it is important to empirically test whether and how expectations about and parafoveal preview of words change the way they are read.

Section snippets

Subjects

Forty undergraduates at the University of California San Diego participated in the experiment for course credit. All subjects were native English speakers with normal or corrected-to-normal vision and were naïve to the purpose of the experiment.

Apparatus

Eye movements were recorded with an SR Research Ltd. Eyelink 1000 eye tracker (with a sampling rate of 1000 Hz) in a tower setup that restrains head movements with forehead and chin rests. Viewing was binocular, but only the movements of the right eye

Experiment 2

In order to directly test how contextual constraint influences preview benefit effects in reading, we conducted a second experiment in which both the constraint and preview manipulations were implemented in a within-subjects design.

General discussion

In the present study, we recorded reading times on preview/target words from Schotter (2013) that were embedded in sentences that were moderately constrained so that, rather than a particular word form, the idea shared by the target/synonym was highly predicted (75% of the time). We found that, in contrast to the data reported by Schotter (2013), semantically associated previews yielded faster reading times compared to an unrelated preview during early (first-pass) reading time measures and

Conclusion

The results of the experiments reported here contribute evidence that the presence of an observable semantic preview benefit depends on certain conditions being met but, importantly, we are now beginning to be able to predict when it will and will not be observed. Mainly, semantic preview benefit is unlikely to be observed when reading in a deep orthography, such as English, and there are no cues from the text to support it. This is important to contextualize cross-language differences with

Acknowledgments

This research was supported by the Atkinson Family Endowment Fund and Grant HD065829 from the National Institutes of Health. Portions of these data were presented at the Chinese International Conference on Eye Movements in Beijing, China in May, 2014. We thank Jeff Elman, Denis Drieghe, and two anonymous reviewers for helpful feedback on previous drafts. Note: Keith Rayner passed away in January 2015 after a long and valiant fight against cancer. The first draft of this paper was submitted

References (43)

  • B. Angele et al.

    Processing the in the parafovea: Are articles skipped automatically?

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2013)
  • Bates, D., Maechler, M., Bolker, B. M., & Walker, S. (2014). Lme4: Linear mixed-effect models using S4 classes (R...
  • K.A. DeLong et al.

    Pre-processing in sentence comprehension: Sensitivity to likely upcoming meaning and structure

    Language and Linguistics Compass

    (2014)
  • D. Drieghe

    Parafoveal-on-foveal effects on eye movements during reading

  • D. Drieghe et al.

    Eye movements and word skipping during reading revisited

    Journal of Experimental Psychology: Human Perception and Performance

    (2005)
  • J.L. Elman

    On the meaning of words and dinosaur bones: Lexical knowledge without a lexicon

    Cognitive Science

    (2009)
  • R. Engbert et al.

    SWIFT: A dynamical model of saccade generation during reading

    Psychological Review

    (2005)
  • S. Hohenstein et al.

    Semantic preview benefit during reading

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2014)
  • S. Hohenstein et al.

    Semantic preview benefit in eye movements during reading: A parafoveal fast-priming study

    Journal of Experimental Psychology: Learning, Memory, and Cognition

    (2010)
  • J. Hyönä

    Effects of thematic and lexical priming on readers’ eye movements

    Scandinavian Journal of Psychology

    (1993)
  • R. Kliegl et al.

    Length, frequency, and predictability effects of words on eye movements in reading

    European Journal of Cognitive Psychology

    (2004)
  • Cited by (0)

    Deceased.

    View full text