Elsevier

Brain Research

Volume 1146, 18 May 2007, Pages 75-84
Brain Research

Research Report
Multiple effects of sentential constraint on word processing

https://doi.org/10.1016/j.brainres.2006.06.101Get rights and content

Abstract

Behavioral and electrophysiological studies have uncovered different patterns of constraint effects on the processing of words in sentences. Whereas response time measures have indicated a reduced scope of facilitation from strongly constraining contexts, event-related brain potential (ERP) measures have instead revealed enhanced facilitation for semantically related endings in such sentences. Given this disparity, and the concomitant possibility of functionally separable stages of context effects, the current study jointly examined expectancy (cloze probability) and constraint effects on the ERP response to words. Expected and unexpected (but plausible) words completed strongly and weakly constraining sentences; unexpected items were matched for contextual fit across the two levels of constraint and were semantically unrelated to the most expected endings. N400 amplitudes were graded by expectancy but unaffected by constraint and seemed to index the benefit of contextual information. However, a later effect, in the form of increased frontal positivity from 500 to 900 ms post-stimulus-onset, indicated a possible cost associated with the processing of unexpected words in strongly constraining contexts.

Introduction

A substantial body of literature using a variety of behavioral and neurophysiological measures attests to the impact of sentence context information on word processing. Congruent context information has been found to speed word naming and lexical decision (word/nonword) judgment times (Fischler and Bloom, 1979, Kleiman, 1980, Schuberth et al., 1981, Stanovich and West, 1979) and to enhance word perception in the Reicher–Wheeler task (Jordan and Thomas, 2002). During natural reading, contextually constrained words are more likely to be skipped and less likely to be regressed to, and, when fixated, are viewed for less time than less constrained words (Ehrlich and Rayner, 1981). A word in a supportive context also elicits a reduced N400 (Kutas and Hillyard, 1980), an event-related brain potential component that has been linked to the access and integration of meaning information (see review by Kutas and Federmeier, 2001).

Although it seems clear that context affects word processing, the mechanisms by which it does so remain controversial. A critical issue concerns when context has its effects, and – on the assumption that the effects of context may be multiple – whether there are different mechanisms involved at different points in time. For example, context effects might occur relatively late, at the point in which the word currently being analyzed is integrated with information stored in working memory, or much earlier in the processing stream (e.g., if the system anticipates likely upcoming words, features, or concepts and makes preparations to process them prior to their occurrence); they might involve enhancement (facilitation) or suppression (inhibition) or both; and they might be differentially impacted by matching and mismatching information. A wealth of empirical data has been brought to bear on such issues, yielding an interesting pattern of convergence and divergence across tasks and measures.

Whether measured in terms of response times, eye-movement patterns, or brain electrophysiology, facilitation has been observed for moderately or highly predictable words in congruent contexts, and this facilitation is graded by the degree of fit between a particular word and a particular sentence. Such fit is often defined empirically using a cloze procedure, in which participants are asked to complete a sentence fragment with the word that first comes to mind; the frequency with which a particular word is produced is its cloze probability for that context (Taylor, 1953). For all measures, benefits of context (i.e., advantages for words in predictive contexts relative to wholly unpredictive – sometimes called neutral – contexts) have generally been shown to be monotonically graded with cloze probability: the greatest benefit (or N400 reduction) is seen to items with high cloze probability, sometimes called “best completions”, but some facilitation is observed even for items of only moderate predictability (e.g., Jordan and Thomas, 2002, Kutas and Hillyard, 1984, Rayner and Well, 1996, Schuberth et al., 1981).

The effects of context on the processing of unexpected words, however, have been less clear, and are of great theoretical import for determining the scope and nature of context's influence. Behavioral investigations, primarily using lexical decision tasks, have found that such effects are influenced by a number of factors, including the semantic similarity between the unexpected word and the best completion for that sentence and the context's constraint, e.g., the degree to which it narrows down the range of possible continuations. Schwanenflugel and her colleagues observed facilitation for the processing of unexpected endings related to an expected completion, but only when these items completed weakly constraining contexts (e.g., “She cleaned the dirt from her sandals,” where “shoes” is the expected but low cloze probability ending) and not when they appeared in strongly constraining contexts (e.g., “On a hot summer's day, many people go to the lake,” where “beach” is expected with high cloze probability) (Schwanenflugel and LaCount, 1988, Schwanenflugel and Shoben, 1985). These results have been interpreted as suggesting that strongly constraining contexts yield a narrower scope of activation than do weakly constraining contexts. In particular, Schwanenflugel and her colleagues (Schwanenflugel and LaCount, 1988, Schwanenflugel and Shoben, 1985) have hypothesized that more constraining contexts establish a larger set of featural restrictions for possible upcoming words. Since, on their proposal, facilitation occurs only when the semantics of an incoming word matches on all of the featural descriptions established by the context, the greater number of featural restrictions exacted by strongly constraining contexts increases the probability of a mismatch and thus makes facilitation less likely.

On this view, then, strongly constraining contexts yield a robust but focused facilitation, which does not extend even to those unexpected words that share semantic features in common with the best completion (although at least one study (Schwantes, 1985), using a naming task, did find facilitation for unexpected words in strongly constrained sentences). Facilitation from weakly constraining contexts, in contrast, does seem to extend to other, related words/concepts. In Schwanenflugel and LaCount's (1988) study, there was no benefit for unexpected and unrelated words in any context, suggesting a limit to the scope of facilitation even for weakly constrained sentences. Indeed, for unexpected words that are anomalous in their contexts, inhibitory effects have sometimes been observed—most often in lexical decision tasks (Fischler and Bloom, 1979, Fischler and Bloom, 1980) but also in naming tasks under some conditions of stimulus degradation (Stanovich and West, 1979, Stanovich and West, 1983). It has been argued that these inhibitory effects of sentential context arise in a different (relatively late) processing stage.

However, N400 responses – which, as already mentioned, are graded with cloze probability in a manner that parallels behavioral data – have shown a pattern of sensitivity to contextual constraint and semantic relatedness that is intriguingly different from the behavioral effects. Initial reports found little influence of contextual constraint: Kutas and Hillyard (1984) examined ERP responses to low cloze probability words in strongly, moderately, and weakly constraining sentence contexts (as defined by the cloze probability of the best completion for that sentence), and found no difference in N400 amplitude. Thus, whereas in behavioral studies the amount of facilitation observed for an unexpected word also seemed to depend on the predictability of other words implied by the context (but never actually presented), N400 amplitudes seemed to pattern with cloze probability, independent of contextual constraint.

However, these initial studies did not systematically control for the degree of semantic relationship between the unexpected word and the best completion, a factor that had clear effects on the behavioral patterns. Federmeier and Kutas (1999b) examined the effects of relatedness and constraint by recording ERPs to unexpected sentence final words that had greater or lesser degrees of semantic overlap with the most expected completion. In an attempt to control for the expectancy of the two unexpected ending types, both were designed to be implausible completions for the sentence pairs. In sentence pairs such as, “They wanted to make the hotel look more like a tropical resort. So along the driveway they planted rows of …” N400 responses were facilitated for expected (“palms”) relative to unexpected completions, but among unexpected completions were smaller to those from the same semantic category (“pines”) than those from a different category (“tulips”). Further, this pattern interacted with constraint: greater facilitation was found for these “within category violations” in strongly than in more weakly constraining sentences, an effect that went in the opposite direction from the rated plausibility of these words in their contexts. Federmeier and Kutas (1999b) interpreted their findings as providing support for the hypothesis that listeners use context information to actively prepare for – i.e., to predict – semantic features of upcoming items. Facilitation is then a function of both the strength of the prediction (which varies with constraint) and the amount of semantic overlap between the predicted word and the one actually presented (greater for within-category than for between-category violations).

Note that the effects of constraint observed on N400 amplitude measures by Federmeier and Kutas (1999b) are opposite the pattern found for lexical decision times by Schwanenflugel and LaCount (1988). Rather than showing a narrowed scope of facilitation for strongly constraining sentences, the electrophysiological data suggest that, at least at some processing stages, strongly constraining sentences more strongly facilitate unexpected but semantically related words — including those that are actually implausible in their contexts. Such a pattern makes sense on the assumption that strongly constraining contexts can engender both greater benefits and greater costs for word processing. As emphasized by Federmeier and Kutas (1999b), increased constraint could mean increased pre-activation of some types of information, and thus increased benefit for words whose features overlap with those elicited by the context. However, as emphasized by Schwanenflugel and LaCount (1988), increased constraint could also increase the possibility of a mismatch and/or entail costs associated with altering the contextually-induced pattern of activation when an unpredictable item is encountered. In other words, the narrowed scope evident for strongly constraining contexts in behavioral measures may arise from processing that occurs downstream of the N400. Critically, the difference in the pattern seen across the two measures suggests that there may be multiple effects of context that differentially modulate processing at different times.

Given the apparent disparity between behavioral and electrophysiological effects of sentential constraint, and the concomitant possibility there are functionally separable stages of context effects on word processing, it makes sense to attempt to further disentangle the effects of cloze probability and constraint on the ERP response to words. In particular, while behavioral studies have examined the influence of contextual information on the response to unexpected and semantically unrelated words, this has not yet been done with ERPs in a systematic and controlled fashion. Such manipulations may yield important data, since mismatch effects and/or costs associated with switching between patterns of activation might be more prevalent for unexpected endings that are more semantically distant from the best completions. One recent study that manipulated lexical- and message-level sources of constraint (Hoeks et al., 2004), observed larger N400 responses to “poor fit” items in sentences with strong constraint (“The javelin was by the athletes summarized”; English translation of a Dutch stimulus sentence) than in sentences with weak constraint (“The javelin has the athletes summarized”). The cloze probability of both of these words was at floor and their rated plausibility did not differ, so the difference in the responses to them would seem to be due to the presence, in the strong constraint condition, of a preferred competitor. However, this study differed from previous studies in that constraint was manipulated syntactically; indeed, a strikingly divergent finding of this study was the lack of an effect of either cloze probability or constraint on the N400 response to “good fit” items (i.e., there was no difference in N400 amplitude between “The javelin was by the athletes thrown”, which had a cloze probability of 77% and “The javelin has the athletes thrown”, which had a cloze probability of 1%). Thus, a study that manipulated constraint semantically would provide a more straightforward comparison to the extant behavioral and ERP data.

In the current study, therefore, we examine ERPs to expected endings of strongly (e.g., “The children went outside to play”) and weakly (e.g., “Joy was too frightened to move”) constraining sentences, and also to plausible, but semantically distinct, unexpected endings in those same contexts (e.g., “look” completing either of the two prior example sentences). For expected endings, cloze probability and constraint cannot be dissociated, since, by definition, only strongly constraining sentences lead to high cloze probability completions (cloze probabilities were 91% and 35%, respectively, for the examples above). However, both types of sentences can accommodate low cloze probability items, and these items can be matched for their rated cloze probability (cloze probability of look was 3% in both of the example sentences) and controlled for their semantic relationship to the expected ending (e.g., look is not associated with or a close semantic relation of either play or move). Any difference in the processing of these unpredictable items can then be attributed to the constraint of the sentence context itself. Of special interest in this study will be whether constraint effects can be seen, when other variables are controlled for, in N400 amplitude patterns or in any other aspects of the ERP response to these semantically-unrelated, unexpected words.

Section snippets

Behavior

Participants correctly recognized an average of 38 of the 160 experimental words in the recognition test (24%) and false alarmed to an average of 4 of the 80 unseen words in the test (5%). They were thus able to discriminate between words they had and had not seen as sentence endings, indicating they were paying attention to the experimental stimuli. Recognition accuracy across the experimental conditions is shown in Table 1. These data were subjected to a repeated measures analysis of variance

Discussion

The goal of this study was to disentangle and examine the influences of expectancy/plausibility (as indexed by cloze probability) and sentential constraint on the unfolding of the brain's response to words embedded in sentence contexts. Prior work indicated intriguing – and still poorly understood – differences between the pattern of constraint effects obtained with behavioral measures and those seen on electrophysiological measures that manifest prior to the behavioral response. Both have

Participants

Thirty-two right-handed native speakers of English at the University of California, San Diego participated in the study in exchange for course credit or cash. Sixteen of the participants were women and sixteen were men. The mean age was 20 years, with a range of 18 to 28 years. All participants were right-handed as assessed by the Edinburgh inventory (Oldfield, 1971); 7 reported having left-handed or ambidextrous family members. All participants reported normal vision and none had a history of

Acknowledgments

We thank David Groppe for determining lexical association in the stimulus set, Marcus Lauer for assistance with cloze norming, and Matt Rambert for help with stimulus development. Support for this research came from grants AG26308 to K. D. F. and from HD22614 and AG08313 to M. K.

References (33)

  • R.C. Oldfield

    The assessment and analysis of handedness: the Edinburgh inventory

    Neuropsychologia

    (1971)
  • P.J. Schwanenflugel et al.

    The influence of sentence constraint on the scope of facilitation for upcoming words

    J. Mem. Lang.

    (1985)
  • S. Coulson et al.

    Right hemisphere activation of joke-related information: an event-related brain potential study

    J. Cogn. Neurosci.

    (2005)
  • A.M. Dale

    Source Localization and Spatial Discriminant Analysis of Event-Related Potentials: Linear Approaches

    (1994)
  • K.D. Federmeier

    The impact of semantic memory organization and sentence context information on spoken language processing by younger and older adults: an ERP study

    Psychophysiology

    (2002)
  • I. Fischler et al.

    Rapid processing of the meaning of sentences

    Mem. Cogn.

    (1980)
  • Cited by (362)

    • The use of eye movement corpora in vocabulary research

      2024, Research Methods in Applied Linguistics
    View all citing articles on Scopus
    View full text