Cross-domain sentiment classification based on key pivot and non-pivot extraction

Abstract Cross-domain sentiment classification (CDSC) aims to predict the sentiment polarity of reviews in an unsupervised target domain by utilizing a classifier learned from a supervised source domain. Most existing methods mainly construct a knowledge transfer model to enhance the category consistency between the source and target domains. These methods not only have little interpretation of transferable information, but are also weak at capturing the sentiment characteristics of a specific domain. For cross-domain reviews, pivots, that is, the domain-shared sentiment words and non-pivots, that is, the domain-specific sentiment words, are critical sentiment clues in the CDSC task. In this study, we comprehensively use domain-shared and domain-specific sentiment information to construct a knowledge transfer model with interpretability for the CDSC task. Specifically, we construct a novel hierarchical attention network called KPE-net, which automatically extracts pivots between two domains. When pivots are used as the bridge, a joint attention learning network called NKPE-net is built to capture non-pivots of different domains. Finally, combining the sentiment factors generated by key pivots and non-pivots, a sentiment-sensitive network model (SSNM) is proposed to realize the transfer of attention to emotions across domains. Experiments on the Amazon review dataset demonstrated the superiority of the proposed model.