Target Information Fusion Based on Memory Network for Aspect-Level Sentiment Classification

Abstract

Aspect-level sentiment classification is a fine-grained task that provides more complete and deeper analysis results. Attention networks are widely used for aspect-level sentiment classification task. However, when multiple target words in a sentence contain opposite sentiments or the expressions of different targets are similar, the network tends to perform poorly. Our studies find that the method of averaging a sentence or target words weakens the capacity of key words. A target information fusion memory network is proposed to solve this problem in this paper. Firstly, the feature of sentences is extracted through a Bi-LSTM network. Then, the feature of target is extracted incorporated into the sentence feature extracted. Then, the memory information of the specific target is formed by the position coding. Finally, the recurrent attention network is utilized to extract the sentiment expression from the memory. Compare with IAN, the method proposed achieves 1.5% and 1.9% accuracy improvement on SemEvil2014 restaurant dataset and self-defined Chinese mobile phone dataset, respectively. A further extend experiment proves that the proposed method can effectively improve the performance in the case of complex sentences.