Abstract: Targeted sentiment analysis(TSA)is a crucial task for fine-grained public opinion mining,which focuses on predicting the sentiment polarity towards a specific target in a given sentence. Most of existing works ignore the syntactic structure of the context sentence,and may pay attention to irrelevant context words when making sentiment judgments. To tackle the problem,a novel syntax aware model is proposed for TSA,which integrates the pre-trained bidirectional encoder representation from transformers models and a graph convolutional network over the dependency tree of the sentence to capture the context information and syntactic structure information of the sentence respectively. The proposed model uses the multi-head attention mechanism to aggregate the information to obtain the final target sentiment representation. The proposed model is also combined with the existing domain adaptive method to introduce domain knowledge and syntactic knowledge,which further improves the performance. The experimental results on several widely-used benchmark datasets demonstrate the effectiveness of the proposed model.

Key words: targeted sentiment analysis, graph convolutional network, bidirectional encoder representation from transformers, dependency tree