Syntactic Information and Multiple Semantic Segments for Aspect-Based Sentiment Classification

Abstract

Aspect-based sentiment classification (ASC) is a task to determine the sentiment polarities of specific aspects in a review. Syntactic information like dependency relation has been proven effective when extracting the sentiment features. On the other hand, multiple semantic segments in a review may influence the sentiment polarity. Thus, we propose a neural network based on dependency relation and structured attention (DRSAN) to fuse both dependency relation features and multiple semantic segments with different attention mechanisms. To verify the performance of DRSAN, we build a Chinese Mobile Phone Review (CMPR) dataset. To our knowledge, we are the first to explicitly integrate dependency relation and structured attention for the ASC task. The experimental results on SemEval 2014 Task 4, Twitter, CMPR, and several other cross-lingual and cross-domain datasets show that the proposed model outperforms all other benchmark models.