Hierarchical Human-Like Strategy for Aspect-Level Sentiment Classification with Sentiment Linguistic Knowledge and Reinforcement Learning.
Min Yang,Qingnan Jiang,Ying Shen,Qingyao Wu,Zhou Zhao,Wei Zhou
DOI: https://doi.org/10.1016/j.neunet.2019.05.021
IF: 7.8
2019-01-01
Neural Networks
Abstract:Aspect-level sentiment analysis is a crucial problem in fine-grained sentiment analysis, which aims to automatically predict the sentiment polarity of the specific aspect in its context. Although remarkable progress has been made by deep learning based methods, aspect-level sentiment classification in real-world remains a challenging task. The human reading cognition is rarely explored in sentiment classification, which however is able to improve the effectiveness of the sentiment classification by considering the process of reading comprehension and logical thinking. Motivated by the process of the human reading cognition that follows a hierarchical routine, we propose a novel Hierarchical Human-like strategy for Aspect-level Sentiment classification (HHAS). The model contains three major components, a sentiment-aware mutual attention module, an aspect-specific knowledge distillation module, and a reinforcement learning based re-reading module, which are consistent with the stages of the human reading cognitive process (i.e., pre-reading, active reading, and post-reading). To measure the effectiveness of HHAS, extensive experiments are conducted on three widely used datasets. Experimental results demonstrate that HHAS achieves impressive results and yields state-of-the-art results on the three datasets.