Topic enhanced sentiment co-attention BERT

Shiyu Wang, Gang Zhou, Jicang Lu, Jing Chen, Yi Xia

Abstract: In the sentiment analysis tasks, the idea of introducing external information to improve the prediction performance is sprouting up. The Topic Sentiment Joint Model shows that as the carrier of fine-grained sentiment, topics have an auxiliary effect on sentiment analysis. However, the current research on fine-grained view focuses on recognizing sentiment, but pays little attention to their impact on the overall feeling. To use topic information to assist sentiment analysis, this paper proposes a topic-enhanced sentiment analysis model. Through the multi-task learning framework, the topic information is learned and used to guide the model for sentiment classification. The multi-stage learning strategy guarantees the accuracy of prior knowledge. With the help of the alternating co-attention mechanism, the essential topics and emotional expression are concerned. In this paper, the labels of Chinese bank forum data and restaurant review data are automatically transformed to obtain supervised topic information. The experimental results show that the model has achieved state-of-the-art over the baseline in two datasets. The Kappa coefficient of sentiment prediction has increased by more than 1%. The model also has apparent advantages in interpretability and noise resistance.

External IDs:dblp:journals/jiis/WangZLCX23