Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Earlier evoked response for lexical surprisal in L1 compared to L2 during naturalistic listening
Poster Session D, Saturday, October 26, 10:30 am - 12:00 pm, Great Hall 3 and 4
Jonathan Brennan1, Tzu-Yun Tung2; 1University of Michigan, 2University of Chicago
INTRODUCTION A growing body of work demonstrates alignment between next-word predictions from Large Language Models (LLMs) like GPT2 and human neural responses associated with predictive language processing in fMRI, ECoG, MEG and EEG. The present project extends this to the comparison of predictive processing in a first vs. a second language. The role of prediction in second language processing has received significant attention. While there is consensus that language proficiency modulates prediction in some way, there are differences in the nature by which predictions are changed (e.g. is the timing of prediction delayed? Are predictive cues weighted differently?), and also ongoing debate about how different kinds of bilingual language experiences (sequential, balanced, etc.) might affect predictive processing. Here we, first, replicate prior work showing effects of next-word predictability (surprisal) on evoked signals within 200 ms in L1 speakers listening to an audiobook. We then evaluate responses of the same speakers when they listen to an audiobook in a second language. METHODS N=29 L1 speakers of Mandarin Chinese (Exp. 1) and N=19 L1 Mandarin, L2 English (Exp. 2; overlapping with the first group) participated in this study. Participants in both experiments had EEG recorded from 31 active electrodes at 500 Hz (0.1-200 band-pass) while listening to chapters from an audiobook story in either L1 (Mandarin, Exp 1) or L2 (English, Exp 2). Data were epoched around word onset (-200-1000 ms), cleaned of eye-artifacts with ICA, and epochs and channels with excessive noise were marked with the autoreject algorithm. All stimuli came from translations of the same audiobook (total length: ~1 h), but participants listened to different, counterbalanced, chapters in the two languages. Surprisal estimates were derived from pre-trained English GPT2 and Chinese GPT2. Epochs corresponding to content words were sorted into equal-sized bins based on surprisal (high, med, low) and averaged to create ERPs. RESULTS Mandarin L1 datasets show a robust evoked response for surprisal such that higher surprisal leads to more negative-going waveform between 200 and 400 ms; this effect is strongest at anterior sites and replicates prior work comparing surprisal to evoked responses during naturalistic listening with EEG. L2 datasets, in English, also show more negative evoked responses for higher surprisal words at anterior sites, but this response was observed at a delay such that the waveform difference was strongest from 350-450 ms. CONCLUSION This work replicates prior research demonstrating effects of surprisal on evoked responses during naturalistic listening in L1 English and extends that work to L1 Mandarin. We also find surprisal effects with a similar topography and polarity in L2 processing (L2 English, L1 Mandarin) but with a latency delay of ~100 ms. This is consistent with the theory that predictive processing may be slower in L2 comprehension and sets the stage for future work leveraging surprisal derived from LLMs to probe the mechanisms of predictive processing across different levels of language proficiency.
Topic Areas: Multilingualism, Computational Approaches