hide
Free keywords:
-
Abstract:
INTRODUCTION: Language processing requires comprehenders to establish links amongst words. This often results in non-adjacent dependencies (NADs) that can span multiple words. Research on statistical learning (Buiatti et al., 2009; Kabdoben et al., 2015) found that NADs can be inferred from artificial grammars (AG) by children and adults. Here, we ask whether the NAD learning is constrained by the boundaries of multi-word chunks in the stimulus. This tests the prior hypothesis that dependencies do not cross the boundaries of chunks in working memory (Christiansen & Chater, 2005). METHODS: We recorded the EEG while native speakers of German underwent an AG learning experiment. During an initial learning phase, subjects listened to isochronous six-syllable chunks that were presented auditorily in eight 3-min streams. Chunks were interrupted by 80-ms gaps to ensure chunk learning. Importantly, stimuli contained NADs that would span two syllables, either within or across chunks. In the subsequent test phase, participants listened to 144 stimuli that either contained within– or across-chunk NADs or not; subjects also answered a question about whether they heard a certain sequence after each trial. EEG preprocessing was done by applying the Harvard Automated Preprocessing Pipeline (Gabard-Durnam et al., 2018). To assess the learning of chunks and boundaries, we computed Evoked Power (EP) and Inter-trial Phase Coherence (ITPC) for the learning phase. EP and ITPC at the chunk frequency were compared to the neighboring frequency bins via one-way ANOVA. To compare the NAD processing within and across chunks, an ERP analysis was conducted on the data from the test phase. A nonparametric statistical analysis was conducted across all electrodes to compare the ERP to syllables that constitute the second element of an NAD to the ERP to syllables that do not finish an NAD. PREDICTION: For the learning phase, EP and ITPC should show frequency peaks at the chunk frequency. For the testing phase, the difference wave between NADs and non-NADs will be smaller in the across-chunk condition than in the within-chunk condition. RESULTS: Preliminary results from N = 10 are reported here. In the learning phase, EP and ITPC peaks at the chunk frequency were observed, suggesting that participants learn the chunks and boundaries. ERPs in the test phase showed a marginally significant (p = 0.053) early negativity in the within-chunk condition, but not the across-chunk condition. In line with the hypothesis, processing of NADs across chunks appears to be harder than within chunks. The results help better understand the relationship between segmentation of speech into memory chunks and the formation of inter-word dependencies. If stable in the full sample, our findings have implications for the interplay of chunking and dependency formation in human language acquisition and processing.
