Poster D43, Friday, August 17, 4:45 – 6:30 pm, Room 2000AB

Single trial analysis of EEG elicited by expected and implausible words in sentences

Seana Coulson1, Megan Bardolph1, Cyma Van Petten2;1UC San Diego, 2Binghamton University

To gauge the sensitivity of the scalp recorded brain response to lexical retrieval versus integration, the present study utilized a regression-based approach to predict the amplitude of single trial EEG to words in sentences during time windows associated with the N400 and LPC. Unlike previous studies focused on highly constraining sentences, here we explore the full range of constraint. To help dissociate retrieval from integration, we utilized plausibility ratings (integration), and two corpus-derived measures of semantic similarity, contextual LSA and best completion LSA (retrieval). LSA (latent semantic analysis) represents the similarity between the sentence-final word and either the words in the sentence frame (contextual LSA), or, the sentence final word and the expected ending (Best Completion LSA). EEG was recorded as healthy adults read sentence frames such as “Gary doesn’t think a husband should cheat on his”, that ended either with the best completion (BC), such as “wife,” or an implausible ending (Implausible), such as “cement”. Participants’ task was to answer comprehension questions following 25% of the sentences. Sentences ranged in constraint from 2-100%, and were divided into categories via a median split. Conventional ERP analyses involved mean amplitude measurements of ERPs 300-500ms post-onset over central-parietal electrodes (N400), and 600-900ms at six frontal and central-parietal electrodes (LPC). Relative to BC, Implausible endings elicited more negative N400 [Ending x Electrode, p<0.001], and more positive LPC [Ending x Anteriority, p<0.01]. Independent of the Ending effect, critical words elicited more positive LPC in low than high constraint sentences. Single trial N400 data in each condition were analyzed via linear mixed effects models with random intercepts for subjects (n=26) and for items (n=280). The initial model included factors of cloze probability, constraint, plausibility, contextual LSA, and (for Implausible endings) Best Completion LSA. Model comparison suggested the BC condition was predicted by cloze probability (F = 5.95), contextual LSA (F = 3.18), and their interaction (F = 7.60). Among high constraint sentences, N400 amplitude was an inverse function of cloze probability, with the most predictable endings eliciting the most positive (least negative) N400; among low constraint sentences (cloze < 0.5), contextual LSA improved the model as words more similar to their sentence contexts elicited less negative N400. Thus corpus-derived measures of a word’s contextual fit can capture variance among low constraint sentences that goes undetected by the cloze task, and indicate N400 amplitude is sensitive to contextual fit. N400 to Implausible completions was related to Best Completion LSA (the similarity of the final word to the best completion ending). As Best Completion LSA increases, the N400 elicited by Implausible endings was less negative. Results are argued to support a context-sensitive retrieval process underlying the N400. For the LPC in the BC condition, the only significant predictor was cloze probability. The more predictable a word was, the more positive the LPC. For the LPC in the Implausible condition, the only significant predictor was plausibility (F=4.51). The more implausible the ending, the larger the LPC. Results are consistent with the suggestion that LPC indexes lexical integration.

Topic Area: Meaning: Combinatorial Semantics