紀要論文

文字の構成要素を考慮した文書分類の再考

RETHINKING DOCUMENT CLASSIFICATION CONSIDERING CHARACTER COMPONENTS

jpn

24368083

https://doi.org/10.15002/00030748

法政大学大学院理工学研究科

法政大学大学院紀要. 理工学研究科編

65

1

5

2024-03-24

Version of Record

In Japanese and Chinese natural language processing, character-based natural language processing that takes into account the radicals of Chinese characters contributes to the improvement of document analysis performance. In order to take character shapes into account, many end-to-end models of character encoders and document classifiers based on convolutional neural networks (CNNs) have been proposed in the past. In this study, we propose Vision Transformer - Character-level Transformer (ViT-CLT), which consists of a Vision Transformer (ViT) as a character encoder and a character-level Transformer (CLT) as a document classifier, in order to realize high-performance document classification that takes into account relationships among character sub-components. In evaluation experiments, we confirmed that ViT-CLT improved prediction performance by 20.0% compared to previous character encoder and document classifier models using CNNs in a categorization task using Japanese news articles. Furthermore, the visualization results of attention in ViT-CLT’s character encoder showed that ViT-CLT was able to consider the components of Chinese characters better than the previous model.

Article