Fine-Grained Sentiment Analysis Tasks Guided by Domain Knowledge

Aspect-Based Sentiment Analysis (ABSA) analyzes the emotion distribution of text from aspect feature granularity. In the previous ABSA task, only the emotion polarity corresponding to the aspect term was judged, but the reason for this emotion polarity was not further explored. To solve these problems, this paper proposes SOKB(Sentiment opinions knowledge BERT), which is a BERT based method. We extracted the aspect sentiment triplet and further constructed it into a sentiment knowledge graph. By combining the knowledge graph with BERT, we obtained the SOKB model with prior information, which enhanced the semantic representation of the model and made it possess certain common sense and reasoning ability. We have introduced a Self-attention layer in the downstream task module that outputs comments about aspect category entities. Experiments show that the proposed SOKB method is effective in the aspect level sentiment analysis task. SOKB model can be well applied to the search system, which helps to improve the accuracy of search results and provide users with better search services.