A Dual Knowledge Aggregation Network for Cross-Domain Sentiment Analysis

Abstract

Cross-domain sentiment analysis (CDSA) is an essential subtask of sentiment analysis. It aims to utilize rich source domain data to conquer the data-hungry problem on target domain. Most existing approaches depending on deep learning mainly concentrate on common features or pivots. However, few of them consider the effect of external Knowledge Graph (KG). In this paper, we propose a Dual Knowledge Aggregation Network for Cross-Domain Sentiment Analysis (DKAN), which leverages prior knowledge from two external KGs. Specifically, DKAN comprises two main parts. One is extracting sentence representation features. The other aims to introduce external knowledge better. Also, we use SenticNet to avoid noise from KG by selecting top-n words and inserting special tokens in sentences. We also conduct empirical analyses on the effectiveness of our model on the Amazon reviews dataset. DKAN achieves promising performance compared with other methods.