Learn More from Context: Joint Modeling of Local and Global Attention for Aspect Sentiment Classification

摘要

Aspect sentiment classification identifies the sentiment polarity of the target that appears in a sentence. The key point of aspect sentiment classification is to capture valuable information from sentence. Existing methods have acknowledged the importance of the relationship between the target and the sentence. However, these approaches only focus on the local information of the target, such as the positional relationship and the semantic similarity between the words in a sentence and the target. Moreover, the global information of the interaction of words in sentence and their influence on the final prediction of sentiment polarity are ignored in related works. To tackle this issue, the present paper proposes Joint Modeling of Local and Global Attention (LGAJM), with the following two aspects: (1) the study develops a position-based attention network concentrating on the local information of semantic similarity and position information of the target. (2) In order to fetch global information, such as context information and interaction between words in sentences, the self-attention network is introduced. Besides, a BiGRU-based gating mechanism is proposed to weight the outputs of these two attention networks. The model is evaluated on two datasets: laptop and restaurant from SemEval 2014. Experimental results demonstrate the high effectiveness of the proposed method in aspect sentiment classification.