Improvement of Japanese Text Emotion Analysis by Active Learning Using Transformers Language Model
Tatsuya Ikeagami, Xin Kang, Fuji Ren · 2022
With the increasing popularity of Twitter in recent years, many people are using Twitter to communicate on a daily basis. Recent research suggests that a large part of the texts posted by Twitter users contain various emotions, which can further be used for the Web marketing and information recommendation systems. However, because the language usage habits differ among Twitter users, the research of emotion analysis on Twitter texts has been considered to be a difficult problem. In addition, the number of Twitter texts is so large that it is impossible to analyze all of them. In this paper we propose to create a high-quality Japanese Tweet Emotion corpus by using active learning and employ a Transformers language model to improve the emotion analysis. Experiment results suggest that a fine-tuned Transformer language model tuns better than the previous model in both Twitter emotion analysis and Twitter emotion corpus creation.