AS-GSI: Aspect-Level Sentiment Analysis Integrating Global Semantic Information

The purpose of aspect-level sentiment analysis is to predict the sentiment polarity of a given aspect in context. At present, most of the existing models use the attention mechanism to capture sentiment information with a given aspect. That did not fully consider the global semantic information, resulting in low accuracy of sentiment polarity recognition. To address this problem, this paper proposes an aspect-level sentiment analysis model that integrates global semantic information(AS-GSI). Firstly, AS-GSI captures the semantic information of context by using BiLSTM. The self-attention mechanism and aspect-based attention mechanism are introduced into AS-GSI, which to obtain global semantic information and sentiment information. Then, a simple and effective fusion mechanism is designed to fuse the two pieces of information fully. Finally, the fused information is input into the softmax-based aspect level sentiment analyzer to calculate the sentiment score. The validity of AS-GSI is verified on the Restaurant14, Restaurant15, and Restaurant16 of SemEval. Experimental results show that the accuracy of AS-GSI can reach 78.27%, 78.14%, and 80.33% on three data sets.