Text Sentiment Analysis with Event Information
Xue Jun Zhang, Hui Zhao, Jingjing Chang, Di Ying · Research Square · 2023
Abstract In recent years, the research of text sentiment analysis has recognized the importance of associating events with sentiment polarity, that is, the sentiment triggered by events can be detected through direct or indirect modeling of events. However, there are two problems with the current study: (1) There is no publicly available dataset; (2) Existing models fail to better model event information. To address these issues, we created an event-based sentiment analysis dataset (ESCO), where each sentence is labeled with event triplet representation and sentiment polarity labels. Based on this dataset, we propose a sentiment analysis model Event-GCN-Attention (EGA) based on event-enhanced semantics. We use Bi-directional Long-Short Term Memory (Bi-LSTM) and Graph Convolutional Network (GCN) to model event information. At the same time, Bi-LSTM is used to model the context information of sentences, and multi-polarity orthogonal attention is used to capture the difference attention weight under different polarities, which can be used as a significant discriminant feature. Finally, we assign different weights according to the importance of event information and sentence context information and combine them to get the final sentence representation. These representations can be used to guide downstream sentiment analysis tasks. Studies show that our proposed model outperforms existing baseline models in our constructed dataset. The dataset (ESCO) is publicly available on https://github.com/cjj-sunshine/event-based-implicit-sentiment-analysis.