PhraseRNN: Phrase Recursive Neural Network for Aspect-based Sentiment Analysis

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2509–2514, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. PhraseRNN: Phrase Recursive Neural Network for Aspect-based Sentiment Analysis Thien Hai Nguyen Kiyoaki Shirai School of Information Science Japan Advanced Institute of Science and Technology 1-1 Asahidai, Nomi, Ishikawa 923-1292, Japan {nhthien, kshirai}@jaist.ac.jp Abstract This paper presents a new method to iden- tify sentiment of an aspect of an entity. It is an extension of RNN (Recursive Neu- ral Network) that takes both dependency and constituent trees of a sentence into ac- count. Results of an experiment show that our method significantly outperforms pre- vious methods. 1 Introduction Aspect-based sentiment analysis (ABSA) has been found to play a significant role in many applica- tions such as opinion mining on product or restau- rant reviews. It is a task to determine an attitude, opinion and emotions of people toward aspects in a sentence. For example, given a sentence “Except the design, the phone is bad for me”, the system should classify positive and negative as the senti- ments for the aspects ‘design’ and ‘phone’, respec- tively. The simple approach is to calculate a sentiment score of a given aspect as the weighted sum of opinion scores, which are defined by a sentiment lexicon, of all words in the sentence (Liu and Zhang, 2012; Pang and Lee, 2008). This method is further improved by identifying the aspect-opinion relations using tree kernel method (Nguyen and Shirai, 2015a). Other researches have attempted to use unsuper- vised topic modeling methods. To identify the sen- timent category of the aspect, topic models which can simultaneously exploit aspect and sentiment have been proposed, such as TSLDA (Nguyen and Shirai, 2015b), ASUM (Jo and Oh, 2011), JST (Lin and He, 2009) and FACTS model (Lakkaraju et al., 2011). Recursive Neural Network (RNN) is a kind of deep neural network. Using distributed represen- tations of words (aka word embedding) (Bengio et al., 2003; Hinton, 1986), RNN merges word rep- resentations to represent phrases or sentences. It is one of the best methods to predict sentiment la- bels for the phrases (Socher et al., 2011; Socher et al., 2012; Socher et al., 2013). AdaRNN (Adap- tive Recursive Neural Network) is an extension of RNN for Twitter sentiment classification (Dong et al., 2014a; Dong et al., 2014b). This paper proposes a new method PhraseRNN for ABSA. It is an extended model of RNN and AdaRNN, which are briefly introduced in Section 2. The basic idea is to make the representation of the target aspect richer by using syntactic infor- mation from both the dependency and constituent trees of the sentence. 2 Recursive Neural Networks for ABSA In RNN and AdaRNN, given a sentence contain- ing a target aspect, “binary dependency tree” is built from a dependency tree of the sentence. Intu- itively, it represents syntactic relations associated with the aspect. Each word (leaf) or phrase (inter- nal node) in the binary dependency tree is repre- sented as …

Content truncated to 500 words