A conflict opinion recognition method based on graph neural network in Aspect-based Sentiment Analysis
Pan Li, Wenbing Chang, Shenghan Zhou, Yiyong Xiao, Chaofan Wei, Runze Zhao · 2022 5th International Conference on Data Science and Information Technology (DSIT) · 2022
Aspect-based sentiment analysis is a research direction of fine-grained sentiment analysis, and is usually used in comment texts. We found that most studies ignored the conflict sentiment during analysis. However, comments with conflicting emotions are usually longer, contain more information, and can reflect the changes of users' opinions. If we can effectively recognize the conflict sentiment in online review texts, we can better help merchants find the shortcomings of products and make improvements. Therefore, based on the research of others, we propose a new D-MA-EGCN model to promote the accuracy of conflict sentiment recognition. The model uses pre-trained BERT model to encode the sentences, and uses edge-convolutional neural network to extract the relationship between aspect word and sentiment word, so as to avoid the misclassification problem caused by the long distance between aspect words and emotion words or multiple sentiments at the same time. The experiment on SemEval dataset shows that our model can dramatically improve the recognition accuracy of conflicting sentiments.