Feature Distillation Network for Aspect-Based Sentiment Analysis

Abstract A proliferation of user-generated content on the web is fueling research into sentiment analysis for improved extraction of human emotional information. Aspect-based sentiment analysis (ABSA) is currently at the focus, which seeks to predict the sentiment of certain aspects in text. The primary challenge is to recognize the relevant contexts for different aspects. Most prior approaches combining recurrent neural networks and attention mechanisms inevitably introduce extraneous noise and diminish prediction accuracy. Furthermore, the sentiment of some context words varies with different aspects and cannot be inferred from themselves alone, which is another challenge that prevents attention mechanisms from performing properly. In this study, we propose a feature distillation network (FDN) for reducing noise and distilling aspect-relevant sentiment features. A novel double-gate mechanism is designed to implement the interactions between aspects and their corresponding contexts at a fine granularity. We introduce a contextual nonlinear projection layer before the double-gate mechanism to generate aspect-specific word representations, which enables the double-gate mechanism to accurately distinguish between sentiment features of the same context word that corresponds to the different aspects. Experiments show that the FDN achieves state-of-the-art performance and improves accuracy from 1.0 percent to 2.0 percent on all benchmarks for ABSA task.