Aspect-based sentiment analysis via fusing multiple sources of textual knowledge

A unified model which exploits and incorporates multiple sources of knowledge to improve its ability on ABSA is proposed and results validate the effectiveness of the method and prove that multiple source of knowledge can collaboratively enhance the model.

Abstract

The aim of aspect-based sentiment analysis (ABSA) is to predict sentiment polarity of text toward a specific aspect. Although existing neural network models show promising performances on ABSA, their capabilities can be unsatisfactory in cases where the amount of training data is limited. In this paper, we propose a unified model which exploits and incorporates multiple sources of knowledge to improve its ability on ABSA. Structure knowledge is extracted via clause recognition and fused in the model through the generation of multiple context representations to force the model to capture aspect-specific context information. Sentiment knowledge is exploited by means of training a general classification model with the sentiment labels of documents and fused through pretraining specific layers to extract contextual features and predict sentiment polarities more accurately. In addition, information of conjunctions is fused in our model to capture the relations between clauses and provide additional sentiment features. Experimental results on five publicly available ABSA datasets validate the effectiveness of our method and prove that multiple sources of knowledge can collaboratively enhance our model.