open access

https://hdl.handle.net/2115/95944

Utilizing BERT with Auxiliary Sentences Generation to Improve Accuracy of Japanese Aspect-based Sentiment Analysis Task

英語

Aspect Sentiment Classification

2024-02-06

Wydawnictwo Naukowe UAM

Human Language Technologies as a Challenge for Computer Science and Linguistics

2023

348-352

Aspect-based Sentiment Analysis (ABSA) is an sentiment analysis task to determine the polarity of opinions regarding a particular aspect of a text. ABSA is a fine-grained sentiment analysis that can handle more detailed and useful information, hence it continues to attract attention in the NLP field. Although many studies have focused on English as the target language, there are still few studies focused on Japanese. Therefore, in this study, we aim to improve the accuracy of the Aspect Extraction (AE) subtask and the Aspect Sentiment Classification (ASC) subtask of the Japanese ABSA task. We improved the method proposed for the English ABSA task, which is to automatically generate auxiliary sentences and then fine-tune the BERT with sentence pairs that combine target sentences and the generated auxiliary sentences, and applied it to the Japanese ABSA task. As a result of the performance evaluation experiment based on a Japanese ABSA dataset called chABSA, compared to the results of the previous study, the F1-score improved by 7.54 points in the AE subtask and the accuracy improved by 3.1 points in the ASC subtask, confirming the effectiveness of the proposed method.

会議発表論文

Version of Record

Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International