Capsule Network with Interactive Attention for Aspect-Level Sentiment Classification

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5489–5498, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5489 Capsule Network with Interactive Attention for Aspect-Level Sentiment Classification 12Chunning Du, 12Haifeng Sun, 12Jingyu Wang∗, 12Qi Qi, 12Jianxin Liao 12Tong Xu, 12Ming Liu 1State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing 100876, China 2EBUPT Information Technology Co., Ltd., Beijing 100191, China {duchunning,sunhaifeng 1,wangjingyu,qiqi}@ebupt.com Abstract Aspect-level sentiment classification is a cru- cial task for sentiment analysis, which aims to identify the sentiment polarities of specific targets in their context. The main challenge comes from multi-aspect sentences, which ex- press multiple sentiment polarities towards different targets, resulting in overlapped fea- ture representation. However, most existing neural models tend to utilize static pooling op- eration or attention mechanism to identify sen- timental words, which therefore insufficient for dealing with overlapped features. To solve this problem, we propose to utilize capsule network to construct vector-based feature rep- resentation and cluster features by an EM rout- ing algorithm. Furthermore, interactive atten- tion mechanism is introduced in the capsule routing procedure to model the semantic re- lationship between aspect terms and context. The iterative routing also enables encoding sentence from a global perspective. Experi- mental results on three datasets show that our proposed model achieves state-of-the-art per- formance. 1 Introduction Aspect-level sentiment classification is a fine- grained task in the field of sentiment analysis (Pang and Lee, 2008; Liu, 2012), which aims to infer the sentiment polarity (e.g., positive, neutral, negative) of a sentence with respect to the aspect. It demands to differentiate sentiments towards dif- ferent targets when there are multiple targets in one sentence. For example, given the mentioned aspect term {food, price, drinks}, and the sen- tence is “The food was definitely good, but when all was said and done, I just could not justify it for the price including 2 drinks, $100/person.” For aspect term food, the sentimental polarity is posi- ∗Corresponding author. tive, but for aspect term price, the polarity is neg- ative while for aspect term drinks, the polarity is neutral. Recently, with the development of deep learning techniques, various neural networks are designed for this task and obtain promising results (Wang et al., 2016a; Ma et al., 2017; Fan et al., 2018). The main challenge in aspect-level sentiment classification is that one sentence expresses mul- tiple sentiment polarities, resulting in overlapped feature representation. Take the same example above, the sentence simultaneously reviews on ‘food’, ‘price’, and ‘drinks’, and expresses three different sentiment polarities. The highly over- lapped features will confuse the classifier seri- ously. However, most existing methods only keep the most active feature by max-pooling operation or utilize attention mechanism to find the senti- mental words, which fails to distinguish the over- lapped features. Therefore, we propose a novel capsule net- work and iterative EM routing method with in- teractive attention (IACapsNet) to solve this …

Content truncated to 500 words