Target-based Sentiment Analysis in Finance with Domain Knowledge

In order to apply the target-based sentiment analysis to finance domain, this paper constructs a corpus with eight types of entities and four types of emotions. Compared to previous works, the corpus proposed in this paper is more fine-grained. Furthermore, we propose a complete and universal framework for target-based sentiment analysis, which contains two subtasks, i.e. named entity recognition and entity-level sentiment classification. We incorporate the pre-trained language model with financial domain knowledge and achieve significant performance improvement.