A Position-aware Bidirectional Attention Network for Aspect-level Sentiment Analysis

Proceedings of the 27th International Conference on Computational Linguistics, pages 774–784 Santa Fe, New Mexico, USA, August 20-26, 2018. 774 A Position-aware Bidirectional Attention Network for Aspect-level Sentiment Analysis Shuqin Gu1, Lipeng Zhang2, Yuexian Hou1∗and Yin Song1 1School of Computer Science and Technology, Tianjin University, Tianjin, China 2School of Computer Software, Tianjin University, Tianjin, China {shuqingu,lpzhang, yxhou, songyin}@tju.edu.cn Abstract Aspect-level sentiment analysis aims to distinguish the sentiment polarity of each specific as- pect term in a given sentence. Both industry and academia have realized the importance of the relationship between aspect term and sentence, and made attempts to model the relationship by designing a series of attention models. However, most existing methods usually neglect the fact that the position information is also crucial for identifying the sentiment polarity of the aspect term. When an aspect term occurs in a sentence, its neighboring words should be given more attention than other words with long distance. Therefore, we propose a position-aware bidirec- tional attention network (PBAN) based on bidirectional GRU. PBAN not only concentrates on the position information of aspect terms, but also mutually models the relation between aspect term and sentence by employing bidirectional attention mechanism. The experimental results on SemEval 2014 Datasets demonstrate the effectiveness of our proposed PBAN model. 1 Introduction Sentiment analysis, also known as opinion mining (Liu, 2012; Pang et al., 2008), is a vital task in Natural Language Processing (NLP). It divides the text into two or more classes according to the affective states and the subjective information of the text, and has received plenty of attention from both industry and academia. In this paper, we address the aspect-level sentiment analysis, which is a fine-grained task in the field of sentiment analysis. For instance, given the mentioned aspect terms {menu, server, specials}, and the sentence is “The menu looked good, except for offering the Chilean Sea Bass, but the server does not offer up the specials that were written on the board outside.”. For aspect term menu, the sentiment polarity is positive, but for aspect term server, the polarity is negative while for specials, the polarity is neutral. One important challenge in aspect-level sentiment analysis is how to model the semantic relationship between aspect terms and sentences. Traditional approaches have defined rich features about content and syntactic structures so as to capture the sentiment polarity (Jiang et al., 2011). However this kind of feature-based method is labor-intensive and highly depends on the quality of the features. Compared with these methods, neural network architectures are capable of learning features without feature engineering, and have been widely used in a variety of NLP tasks such as machine translation (Cho et al., 2014), question answering (Andreas et al., 2016) and text classification (Lai et al., 2015). Recently, with the development of the neural networks, they are also applied to target-dependent sentiment analysis1, such as Target-Dependent LSTM (TD-LSTM) (Tang et al., 2015) and Target-Connection LSTM (TC-LSTM) (Tang et al., 2015). However, these neural network-based methods can not effectively identify which …

Content truncated to 500 words