情報処理学会 第86回全国大会 会期:2024年3月15日~17日

1W-06
Evaluating the Effect of Sub-Word Training on Syntactic Language Models -A Linguistic Perspective-
○岡村康平(筑波大)
This study evaluates the effect of incorporating sub-word units (morphemes) in training large language models (LLMs), highlighting its role in reducing vocabulary size and managing unknown and compound words more effectively. Beyond these established benefits, the research also emphasises the importance of considering the hierarchical morphological structures of languages and their generative capacities. It addresses how simplistic assumptions underpinning syntactic competence, which focus solely on surface linear strings, struggle to resolve linguistic problems such as ambiguities (e.g., un-lockable versus unlock-able) and long-dependencies (e.g., en-joy-able versus *joy-able). The research posits that an improved understanding of morphological structures, grounded in the Hierarchical Morphological Processing Hypothesis, can enhance the performance and accuracy of LLMs.