Multi-Attention Network for Aspect Sentiment Analysis

Aspect sentiment analysis is a fine-gained task in sentiment analysis. In this paper, we propose a novel LSTM network model, which combines multi-attention and aspect contexts, i.e. LSTM-MATT-AC. Multi-attention mechanism that integrates the factors of location, content and class could adaptively capture important information in the contexts with the supervision of aspect targets. In other words, the model is more robust against irrelevant information. Simultaneously, aspect context mechanism extends differentiate left and right contexts given aspect targets and strengthens the expressive power of the model for handling more complication by mining deeper semantic information. Experiment results on SemEval2014 Task4 and Twitter datasets show that the accuracy of sentiment classification reaches 80.6%, 75.1% and 71.1% respectively. Compared to previous neural network-based sentiment analysis models, the accuracy has been further improved.