Presentations

Abstract

Subject-verb agreement is one of the most common
grammatical encoding operations in language
production. In many languages, morphological
inflection on verbs code for the number of the head
noun of a subject phrase (e.g., The key to the cabinets
is rusty). Despite the relative ease with which subjectverb
agreement is accomplished, people sometimes
make agreement errors (e.g., The key to the cabinets
are rusty). Such errors offer a window into the early
stages of production planning. Agreement errors are
influenced by both syntactic and semantic factors, and
are more likely to occur when a sentence contains either
conceptual or syntactic number mismatches. Little
is known about the timecourse of these influences,
however, and some controversy exists as to whether
they are independent. The current study was designed
to address these two issues using EEG. Semantic and
syntactic factors influencing number mismatch were
factorially-manipulated in a forced-choice sentence
completion paradigm. To avoid EEG artifact associated
with speaking, participants (N=20) were presented with
a noun-phrase, and pressed a button to indicate which
version of the verb ‘to be’ (is/are) should continue
the sentence. Semantic number was manipulated
using preambles that were semantically-integrated or
unintegrated. Semantic integration refers to the semantic
relationship between nouns in a noun-phrase, with
integrated items promoting conceptual-singularity.
The syntactic manipulation was the number (singular/
plural) of the local noun preceding the decision. This
led to preambles such as “The pizza with the yummy
topping(s)... “ (integated) vs. “The pizza with the tasty
bevarage(s)...” (unintegrated). Behavioral results showed
effects of both Local Noun Number and Semantic
Integration, with more errors and longer reaction times
occurring in the mismatching conditions (i.e., plural
local nouns; unintegrated subject phrases). Classic ERP
analyses locked to the local noun (0-700 ms) and to the
time preceding the response (-600 to 0 ms) showed no
systematic differences between conditions. Despite this
result, we assessed whether difference might emerge
using multivariate pattern analysis (MVPA). Using the
same epochs as above, support-vector machines with a
radial basis function were trained on the single-trial level
to classify the difference between Local Noun Number
and Semantic Integration conditions across time and
channels. Results revealed that both conditions could
be reliably classified at the single subject level, and
that classification accuracy was strongest in the epoch
preceding the response. Classification accuracy was
at chance when a classifier trained to dissociate Local
Noun Number was used to predict Semantic Integration
(and vice versa), providing some evidence of the
independence of the two effects. Significant inter-subject
variability was present in the channels and time-points
that were critical for classification, but earlier timepoints
were more often important for classifying Local Noun
Number than Semantic Integration. One result of this
variability is classification performed across subjects was
at chance, which may explain the failure to find standard
ERP effects. This study thus provides an important first
test of semantic and syntactic influences on subject-verb
agreement with EEG, and demonstrates that where
classic ERP analyses fail, MVPA can reliably distinguish
differences at the neurophysiological level.

Abstract

Prominent theories of predictive language processing assume that language production processes are used to anticipate upcoming linguistic input during comprehension (Dell & Chang, 2014; Pickering & Garrod, 2013). Here, we explore the converse case: Does a task set including production in addition to comprehension encourage prediction, compared to a task only including comprehension? To test this hypothesis, we conducted a cross-modal naming experiment (Experiment 1) including an object naming task and a self-paced reading experiment (Experiment 2) that did not include overt production. We used the same predictable (N = 40) and non-predictable (N = 40) sentences in both experiments. The sentences consisted of a fixed agent, a transitive verb and a predictable or non-predictable target word (The man drinks a beer vs. The man buys a beer). Most of the empirical work on prediction used sentences in which the target words were highly predictable (often with a mean cloze probability > .8) and thus it is little surprising that participants engaged in predictive language processing very easily. In the current sentences, the mean cloze probability in the predictable sentences was .39 (ranging from .06 to .8; zero in the non-predictable sentences). If comprehenders are more likely to engage in predictive processing when the task set involves production, we should observe more pronounced effects of prediction in Experiment 1 as compared to Experiment 2. If production does not enhance prediction, we should observe similar effects of prediction in both experiments. In Experiment 1, participants (N = 54) listened to recordings of the sentences which ended right before the spoken target word. Coinciding with the end of the playback, a picture of the target word was shown which the participants were asked to name as fast as possible. Analyses of their naming latencies revealed a statistically significant naming advantage of 106 ms on predictable over non-predictable trials. Moreover, we found that the objects’ naming advantage was predicted by the target words’ cloze probability in the sentences (r = .411, p = .016). In Experiment 2, the same sentences were used in a self-paced reading experiment. To allow for testing of potential spill-over effects, we added a neutral prepositional phrase (buys a beer from the bar keeper/drinks a beer from the shop) to each sentence. Participants (N = 54) read the sentences word-by-word, advancing by pushing the space bar. On 30% of the trials, comprehension questions were used to keep up participants' focus on comprehending the sentences. Analyses of participants’ target and post-target reading times revealed numerical advantages of 6 ms and 20 ms, respectively, in the predictable as compared to the non-predictable condition. However, in both cases, this difference was not statistically reliable (t = .757, t = 1.43) and the significant positive correlation between an item’s naming advantage and its cloze probability as seen in Experiment 1 was absent (r = .037, p = .822). Importantly, the analysis of participants' responses to the comprehension questions, showed that they understood the sentences (mean accuracy = 93%). To conclude, although both experiments used the same sentences, we observed effects of prediction only when the task included production. In Experiment 2, no evidence for anticipation was found although participants clearly understood the sentences and the method has previously been shown to be sensitive to measure prediction effects (Van Berkum et al., 2005). Our results fit with a recent study by Gollan et al. (2011) who found only a small processing advantage of predictive over non-predictive sentences in reading (using highly predictable sentences with a cloze probability > . 87) but a strong prediction effect when participants read the same sentences and carried out an additional object naming task (see also Griffin & Bock, 1998). Taken together, the studies suggest that the comprehenders' task set exerts a powerful influence on the likelihood and magnitude of predictive language processing. When the task set involves language production, as is often the case in natural conversation, comprehenders might engage in prediction to a stronger degree than in pure comprehension tasks. Being able to predict words another person is about to say might optimize the comprehension process and enable smooth turn-taking.