Aspect-based sentiment analysis with enhanced aspect-sensitive word embeddings

摘要

Aspect term sentiment classification (ATSC) aims at identifying sentiment polarities towards some aspect terms described in a text. One of the challenges in the ATSC is that the same word may express different sentiment polarities for distinct aspects. For instance, if the word “high” is used to describe the quality of a product, this word is most likely used to express a positive opinion towards the product. However, if the aspect term is about the price of the product, the same word “high” is quite likely used to represent a negative sentiment polarity. Such aspect-sensitive word features are also useful for the ATSC when the comparative forms are used to express opinions. What sentiment or opinion is expressed largely depends on who we compare with and how we compare. We describe a weakly supervised method to create an aspect-sensitive lexicon for each aspect, which is a relatively accurate representation of the sentiments that are related to that aspect. We also propose a sentiment analysis model enhanced with the learned aspect-sensitive word embeddings, and extensive experiments show that this model achieved state-of-the-art performances on multiple datasets.