초록 열기/닫기 버튼
Recently, neural language models (NLMs) trained on vast amounts of raw text have been shown to acquire abstract linguistic representations. In this study, we tackled the robustness of these abstractions by focusing on discourse information determined by implicit causality verbs. Comparing the observed behaviors from human experiments and L1 NLMs, we evaluated the extent to which discourse structure could be acquired by the L2 NLM. We especially probed whether implicit verb biases could influence the probability of pronouns. We discovered that the L2 Long-Short Term Memory NLM is unable to demonstrate knowledge of implicit causality when resolving reference. In other words, the L2 LM has difficulty encoding semantic information in its abstract representational space. The mismatches in syntactic representations and behavior suggest that the L2 LM passes over the learned abstract categories from the data. Our results further indicate that the observed the L2 LM’s behaviors can contradict learning representations of discourse/semantic information, pointing to limitations of language modeling.
