紀要論文

解釈可能な文字の視覚情報と意味情報を考慮した日本語の文書解析

JAPANESE TEXT ANALYSIS CONSIDERING VISUAL AND SEMANTIC INFORMATION OF INTERPRETABLE CHARACTERS

jpn

24368083

https://doi.org/10.15002/00025367

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

63

1

7

2022-03-24

Version of Record

While Asian languages use characters with visual information such as radicals and shapes of kanji characters, japanese also uses characters without visual information such as hiragana and katakana. To take these features of japanese into account, a text analysis model that considers semantic information as well as visual information of characters has been proposed, and has achieved high performance in text analysis. However, while the visual information of the characters can be interpreted as radicals and shapes, the semantic information cannot be interpreted. In this paper, we propose text analysis model that considers highly interpretable visual and semantic information. The visual and semantic information of the character is separated and stored in each dimension of the low-dimensional representation of the character to enhance the interpretability. In addition, we focus on the fact that the acquisition of the meaning of characters in the semantic information depends on the training data. By combining the visual and semantic information, we expect to improve the robustness of the acquisition of the meaning of characters and to achieve high performance in text analysis. In sentiment analysis, we were able to interpret the semantic information of characters such as negative, positive, and user information. The combination of visual and semantic information improved the sentiment analysis task by about 1% and the recognize textual entailment task by about 0.3% compared to using only one of the two.

Article