신경망 언어 모델과 인간 언어 사용자의 주어와 목적어 관계절 처리 비교 연구

Abstract

This paper is to investigate the distinct aspects of processing subject and object relative clauses (SRC and ORC) in neural language models (LMs) and humans using the materials, which are constructed by manipulating two RC types (SRC vs ORC) and two intervening PP types (locative vs temporal). Surprise values are collected from the transformer GPT-2 and BERT language models. Reading time data for humans are taken from Lowder and Gordon’s (2021) eye-tracking experiment. According to Lowder and Gordon (ibid.) that take as a critical region the matrix verb after the RC in the subject position, for humans the locative PP contained in the RC register longer reading times ORCs than SRCs, while the temporal PP does so for SRCs than ORCs. By contrast, for the two neural LMs, surprisals are higher for ORCs than SRCs regardless of whether the PP in question is locative or temporal. There was no statistically significant linear fit between human and the LMs’ responses. The result shows that to the extent that the distinction between locative and temporal PPs in RCs is syntactico-semantic, neither of the two neural language models is able to acquire human-like sensitivity to such a distinction.