Construction and Evaluation of a Self-Attention Model for Semantic Understanding of Sentence-Final Particles

Abstract

In order to achieve a richer understanding of the meaning of words beyond the connection between words and images (visual information) alone, we wanted to make the model learn the relationship between words and various subjective senses (including vision). For this purpose, we proposed Subjective BERT, a self-attention model that takes various subjective senses in addition to language and images as input, and attempted to understand utterances such as Ringo da yo. `I want to tell you that here is an apple.' and Banana da ne. `I want to make sure that we both know here is a banana.', paying particular attention to the acquisition of function words (sentence-final particles). Simulation experiments revealed that the constraints imposed by the sentence-final particles were acquired and content words were learned based on them.