A unified position-aware convolutional neural network for aspect based sentiment analysis

Abstract

Aspect based sentiment analysis (ABSA) can be divided into aspect category sentiment analysis (ACSA) and aspect term sentiment analysis (ATSA). To detect the sentiment polarity in a context given an aspect, both ACSA and ATSA highly rely on aspect-centroid sentiment information. However, aspect terms in ATSA appear in the context, but aspect categories in ACSA may not. This make it difficult to use important aspect position information and further design effective aspect injection strategy when modeling ACSA and ATSA in a unified framework. To address this problem, we propose a novel Unified Position-aware Convolutional Neural Network (UP-CNN). There are two major modifications in UP-CNN. Firstly, to handle the absence of aspect position in ACSA, we propose an aspect detection network with prior knowledge. Thus ABSA can be solved in a unified view with the important aspect position. Secondly, to fit CNN in ABSA, an aspect mask is proposed to construct aspect-aware context representation. Experimental results on seven datasets demonstrate that our model performs effectively and efficiently on both ACSA and ATSA tasks.