Discovering Protagonist of Sentiment with Aspect Reconstructed Capsule Network

Most existing aspect-term level sentiment analysis (ATSA) approaches combined neural networks with attention mechanisms built upon given aspect to generate refined sentence representation for better predictions. In these methods, terms are always provided in both training and testing process which may degrade aspect-level into sentence-level prediction. However, the annotated term might be unavailable real-world scenarios challenge applicability of methods. this paper, we aim improve ATSA by discovering potential predicted polarity when a test unknown. We access goal proposing capsule network based model named CAPSAR. CAPSAR, categories denoted capsules information is injected through sentiment-aspect reconstruction procedure during training. As result, coherent patterns between aspects sentimental expressions encapsulated capsules. Experiments on three widely used benchmarks demonstrate have exploring from only feeding model. Meanwhile, proposed CAPSAR can clearly outperform SOTA methods standard tasks.