[5E1-GS-6d] Language media processing

9:30 AM - 9:45 AM JST(12:30 AM - 12:45 AM UTC)

[5E1-GS-6d-03] Visualization of Collocational Naturalness in Japanese Verb Pairs

〇Riko Fukami¹, Mina Akaishi², Miki Hanazaki² (1. Hosei University Graduate School of Computer and Information Sciences, 2. Hosei University Faculty of Computer and Information Sciences)

Keywords:

Natural Language Processing,Collocation,Visualization,BERT

In natural language, semantically similar words may differ in acceptability depending on their collocations with surrounding words, and understanding such differences is often difficult for second language learners. This study aims to clarify differences in collocational patterns and usage contexts of semantically similar Japanese verb pairs using contextual prediction probabilities from BERT. We target several near-synonymous verb pairs and combine each verb with a shared set of accusative-case (“o”-marked) nouns to create many phrases. For each phrase, we compute the average log probability of the masked verb based on BERT’s predictions and visualize the co-occurrence probabilities of paired verbs with each noun in a two-dimensional space. We also collect human acceptability judgments for the same phrases and examine their relationship with the system’s predictions. The results indicate a significant correlation between predicted probabilities and native speakers’ judgments of naturalness, suggesting that verb–noun combinations favored by BERT tend to be perceived as more natural. These findings imply that visualizing collocational naturalness can support learners in acquiring native-like intuitions about Japanese verb usage.