Dual-Word Embedding Model Considering Syntactic Information for Cross-Domain Sentiment Classification

Abstract

The purpose of cross-domain sentiment classification (CDSC) is to fully utilize the rich labeled data in the source domain to help the target domain perform sentiment classification even when labeled data are insufficient. Most of the existing methods focus on obtaining domain transferable semantic information but ignore syntactic information. The performance of BERT may decrease because of domain transfer, and traditional word embeddings, such as word2vec, cannot obtain contextualized word vectors. Therefore, achieving the best results in CDSC is difficult when only BERT or word2vec is used. In this paper, we propose a Dual-word Embedding Model Considering Syntactic Information for Cross-domain Sentiment Classification. Specifically, we obtain dual-word embeddings using BERT and word2vec. After performing BERT embedding, we pay closer attention to semantic information, mainly using self-attention and TextCNN. After word2vec word embedding is obtained, the graph attention network is used to extract the syntactic information of the document, and the attention mechanism is used to focus on the important aspects. Experiments on two real-world datasets show that our model outperforms other strong baselines.