코퍼스와 딥러닝 언어 모델을 활용한 문장 처리의 예측성과 행동 반응 시간과의 관계 연구

Abstract

This study examined whether the predictability is associated with the behavioral reaction times in sentence processing. The information complexity measures have been proposed to quantify the predictability for word-by-word human sentence processing. The most traditional information complexity measure is known as surprisal, which calculates relative unexpectedness at each word in a sentence (Hale 2001, Levy 2005, 2008). The most traditional information complexity measure is known as surprisal, which calculates relative unexpectedness at each word in a sentence (Hale 2001, Levy 2005, 2008), and some studies suggested that surprisal and reading times are positively correlated (Monsalve, Frank and Vigliocco 2012, Smith and Levy 2013). In order to calculate surprisal, the previous studies used one of two ways: Corpus based language models and deep learning based language models. This study, however, used both of them to analyze human reading times, comparing surprisal calculated from corpus-based language models with that calculated from deep-learning-based language models. Many studies partially investigated either of them. In this study, human reading times were analyzed by comparing surprisal calculated from corpus-based language models with that calculated from deep-learning-based language models. The results showed that surprisal calculated from corpus-based language models is more suitable to explain the behavioral reaction time data. Although the deep learning technology performs very well in the field of natural language processing, it does not seem to be human-like processing. Nonetheless, this study can contribute to the development of deep learning technology as well as computational psycholinguistic research in that it tried to compare the outcomes of corpus and deep learning technology with human behavioral responses.