The Ability of an English Language Model to Understand of World Knowledge with Backshift

초록

This study evaluated whether an English language models using artificial intelligence represent world knowledge, focusing on the backshift phenomenon. Backshift refers to the past tense being used in indirect speech, with direct speech being in the present tense. This study argued that language models capture grammatical phenomena that interact with world knowledge and can represent knowledge beyond mere grammatical knowledge. This study used BERT and mBERT models to measure the surprisals of verbs in indirect speech. Surprisal is a measurement that increase in proportion to the difficulty of language processing. Experimental results demonstrated that BERT and mBERT models were sensitive to the backshift phenomenon, revealing that the artificial intelligence language model captured backshift in indirect speech. As backshift requires an understanding of truth, knowledge, and commonsense, these results indicated that the language model understood world knowledge.

키워드