FuncSA: Function Words-Guided Sentiment-Aware Attention for Chinese Sentiment Analysis

Jiajia Wang, Hongying Zan, Yingjie Han, Juan Cao

Abstract: Sentiment analysis is an important natural language processing application that empowers many other technologies, including product review analysis and recommendation systems. Knowledge has been proven crucial for providing supervision information and improving performance. However, Chinese function words’ knowledge, especially for degree adverbs, negative adverbs, and conjunctions, which may play an essential role in describing the sentiment polarity, is not well investigated in current Chinese sentiment analysis approaches. In this paper, we propose a Function words-guided Sentiment-aware Attention model (FuncSA) for Chinese sentiment analysis to leverage function words’ knowledge. Specifically, we integrate discrete sentiment lexical information using degree adverbs, negative adverbs, and conjunctions in the Chinese Function Word Usage Knowledge Base(CFKB), and improve self-attention to integrate function words’ knowledge into the model. We implement our approach on several open datasets and show that function words are essential in guiding sentiment identification.