Function Analysis of Attention Heads in Japanese Pretrained BERT Using Linguistically Different Inputs

Abstract

The self-attention mechanism (attention mechanism) that Transformers possess internally is being used as an effective method in various fields beyond natural language processing, yet there are many unclear points regarding the interpretation of each attention module. This research proposes a method to analyze the internal behavior of Transformer models by mapping the attention mechanism on a head-by-head basis to linguistic functions, specifically targeting Japanese. Concretely, this involves performing transformations such as swapping tokens that correspond to specific parts of speech, or fixing the vocabulary while only changing the syntactic order, and observing the differences in attention head reactions before and after the transformations. By inputting pairs of sentences that change specific parts of speech or dependency relations and acquiring the differences in attention norm in BERT, it was possible to identify attention heads that are characteristic of specific parts of speech or dependency relations by visualizing these differences