영어 보충어 강요 문장 처리 연구: 전산 심리언어학 관점에서

Abstract

This paper investigated complement coercion in English from computational psycholinguistic perspectives. In previous psycholinguistic behavioral studies, when humans processed sentences such as The author began the book, it took longer in reading time to process entity noun phrases (e.g., the book) after the next coercing verbs (e.g., begin) that require event arguments (e.g., writing the book) than the ones after non-coercing verbs (e.g., write) that do not. We leveraged one of the recent neural language models, GPT-2, to examine how it as a computational manifestation of the human mind engages in processing complement coercion in English. We found that GPT-2 yielded the higher surprisal value for coerced complements than for non-coerced complements, which is in keeping with the results from previous psycholinguistic behavioral studies. This finding shows that there is a close link between reading time in humans and the informational-theoretic measure of surprisal in GPT-2 during processing sentences containing complement coercion in English, and that in light of their correlation, GPT-2 seems to have learned relevant linguistic information bearing on complement coercion in English.