A Neural Model for Aspect-Level Sentiment Classification of Product Reviews Assissted by Question-Answering

Aspect-level sentiment analysis targets at judging the associated polarity of the opinion holder’s view towards given aspect category, and it usually contains two sub-tasks, namely aspect term extraction and aspect-related sentiment classification. However, the aspect term extraction, an auxiliary sub-task, receives too much emphasis, and it is labor-consuming. With the development of deep learning theory and language models, quite a few problems in natural language processing (NLP) domain have been solved. To address the problems mentioned above, we propose to remove the aspect term extraction component and regard aspect categories as guidelines to apply the question answering theory. Simultaneously, we utilize BERT as the base support unit to sufficiently encode the context information. For the purpose of evaluating our method, we also construct our Chinese review corpus for aspect-level sentiment analysis assignment. Experimental results show that our proposed method achieves an appreciable performance on our dataset.