The first on-line experiment assessed effects of purely linguistic context on processing effort of the critical target nouns, in terms of reading times.
An “implausible” condition as well as a spill-over region following the verb’s argument in all conditions were added only for this experiment (e.g. the man soon spills the book at the restaurant). The implausible condition served as a reference measure for effects of
1Note that the relevant nouns were pre-selected from the DeReKo corpus (in order to control for frequency) and not collected from the cloze task answers. Hence the relatively high SD for these particular nouns in the cloze task.
implausibility on reading times, and as sanity check for the design. This condition was expected to clearly elicit longer reading times for the – in the context of the verb spill – highly unpredicted and hence highly surprising noun (the book).
The added adverbial phrases served as post target spill-over regions for the time-dependent measure, which requires longer time windows (as compared to ERPs or the ICA).
The verb-noun manipulation resulted in a 2x3 design in which constraining (spill) and unconstraining (order) verbs were paired with objects that were more plausible in the con-straining verb context (water, i.e., most predictable and least surprising as compared to the other objects used), less plausible (ice cream, i.e., mid predictable and surprising) and implau-sible (book, i.e., unpredictable and most surprising). As reflected by the plausibility rating ran prior to the experiment, all three objects were equally plausible in the unconstraining verb context while the target noun’s plausibility differed only in the constraining verb context.
36 experimental and 36 filler items were distributed across six lists, using the Latin square design in such a way that each participant would see each item in only one condition. 24 native speakers of German (students of Saarland University) gave informed consent before participating in this study for monetary reimbursement. Their age ranged from 18 to 32 years (M = 22.71).
Sentences were presented as a whole, in the centre of the screen (Times New Roman, 20 pt), with a drift correct fixation point, shown at the top left corner in order to avoid initial fixations at the sentence. Participants were instructed to read for comprehension, at their own pace.
Predictions In the recent psycholinguistic literature, it is well established that in human sentence processing, the probability of a word to appear in its previous (linguistic) context largely affects the time it takes to read this word. This correlation has further often been quantified using information theoretic surprisal. Van Berkum et al. (2005), for example, measured significantly increased reading times after expectation-inconsistent adjectives in self-paced reading. Smith and Levy (2013) further even suggest that the quantitative form of the relationship between a word’s reading time and its predictability, that is, the strength of a comprehender’s prediction for a word, is indeed logarithmic. Only recently, Goodkind and Bicknell (2018) demonstrated that the degree of predictive power of word surprisal, as derived from different language models for reading times can even be a reliable, linear function of the respective language model’s quality.
Based on this corpus of studies, longer reading times reflecting increased processing effort were expected on or after the most surprising and least predictable implausible target nouns, only when appearing in a context allowing for predictions, namely when following
the constraining verbs. If, however, the verbal constraint alone was not enough to elicit (lexical) predictions about the target nouns, no differences between the object conditions in the constraining verb context were expected. According to many studies suggesting a strong context dependence of predictions, no differences in processing effort were expected on the verb, although the verbal constraint could cause participants to make more detailed assumptions about the target noun (e.g., something "spillable"). However, compared to studies using bigger linguistic context to strengthen expectations about target words prior to them, the stimuli used in this experiment are not embedded in a wider context. The only information driving possible expectations or even specific predictions about the target noun hence come from the verb’s constraint.
Analysis If not stated differently statistical analyses of the data collected in this and all following experiments were conducted using the R statistical programming environment (RCoreTeam, 2013) and the lme4 package (Bates et al., 2015). The dependent measures is all experiments were always analysed within two non-overlapping time windows on the critical words. Since the verb’s information was hypothesized to possibly drive a reduction of mainly visual objects not matching its constraints in the later VWP experiments (Experiments3 to 7), the time window critical for analysis was on the verb. In order to not bias results towards the visual context having an influence, a comparison for processing effort on the verb without visual context was needed. The verb window was hence also analysed in the first two experiments not featuring visual context, although the purely linguistic context was not expected to provide strong enough constraints for participants to decide against certain nouns.
In the second critical time window, namely the target noun itself, differences in processing effort and surprisal were expected in all experiments as soon as expectations about the noun can be informed by either linguistic or visual information, or even by the combination of both.
InExperiment 1, reading times were hence measured and analysed within the two non-overlapping time windows on the verb and on the target noun. Since reading time differences are often measured with a slight delay, the spill-over regions following the respective critical words were additionally analysed. Time measures were log-transformed due to the natural skewness of reaction time data, and entered as dependent variables into linear mixed-effects models. The contrast-coded Object and Verb conditions as well as the scaled length of the target word (measured in characters) were entered as fixed factors. Following Barr et al.
(2013), the models were run with the maximal converging random effects structure, including intercepts and slopes for Subject and Item.
Figure 3.2Total dwell time results in all levels of the conditions in Experiment 1. Error bars reflect 95% confidence intervals (CI).