Multi-Task Multi-Head Attention Memory Network for Fine-Grained Sentiment Analysis

Sentiment analysis is widely applied in personalized recommendation, business reputation monitoring, and consumer-driven product design and quality improvement. Fine-grained sentiment analysis, aimed at directly predicting sentiment polarity for multiple pre-defined fine-grained categories in an end-to-end way without having to identify aspect words, is more flexible and effective for real world applications. Constructing high performance fine-grained sentiment analysis models requires the effective use of both shared document level features and category-specific features, which most existing multi-task models fail to accomplish. In this paper, we propose an effective multi-task neural network for fine-grained sentiment analysis, Multi-Task Multi-Head Attention Memory Network (MMAM). To make full use of the shared document level features and category-specific features, our framework adopts a multi-head document attention mechanism as the memory to encode shared document features, and a multi-task attention mechanism to extract category-specific features. Experiments on two Chinese language fine-grained sentiment analysis datasets in the Restaurant-domain and Automotive-domain demonstrate that our model consistently outperforms other compared fine-grained sentiment analysis models. We believe extracting and fully utilizing document level features to establish category-specific features is an effective approach to fine-grained sentiment analysis.