Distance Metric Learning for Aspect Phrase Grouping

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2492–2502, Osaka, Japan, December 11-17 2016. Distance Metric Learning for Aspect Phrase Grouping Shufeng Xiong1,3, Yue Zhang2, Donghong Ji1∗, Yinxia Lou1 1Computer School, Wuhan University, Wuhan, China 2Singapore University of Technology and Design, Singapore 3Pingdingshan University, Pingdingshan, China {xsf, dhji}@whu.edu.cn yue zhang@sutd.edu.sg Abstract Aspect phrase grouping is an important task in aspect-level sentiment analysis. It is a challenging problem due to polysemy and context dependency. We propose an Attention-based Deep Dis- tance Metric Learning (ADDML) method, by considering aspect phrase representation as well as context representation. First, leveraging the characteristics of the review text, we automatically generate aspect phrase sample pairs for distant supervision. Second, we feed word embeddings of aspect phrases and their contexts into an attention-based neural network to learn feature repre- sentation of contexts. Both aspect phrase embedding and context embedding are used to learn a deep feature subspace for measure the distances between aspect phrases for K-means clustering. Experiments on four review datasets show that the proposed method outperforms state-of-the-art strong baseline methods. 1 Introduction For aspect-level sentiment analysis (Hu and Liu, 2004; Pang and Lee, 2008), aspect identification from the corpus is a necessary step. Here aspect is the name of a feature of the product, while an aspect phrase is a word or phrase that actually appears in a sentence to indicate the aspect. Different aspect phrases can be used to describe the same aspect. For example, “picture quality” could be referred to “photo”, “image” and “picture”. All aspect phrases in the same group indicate the same aspect. In this paper, we assume that all aspect phrases have been identified by using existing methods (Jin et al., 2009; Kobayashi et al., 2007; Kim and Hovy, 2006), and focus on grouping domain synonymous aspect phrases. Most existing work employed unsupervised methods, exploiting lexical similarity from semantic dic- tionary as well as context environments (Zhao et al., 2014; Zhai et al., 2011a; Guo et al., 2009). The context for an aspect phrase is formed by aggregating related sentences that mention the same aspect phrase. Thereafter, aspect phrase and context environment are represented using bag-of-word (BoW) models separately, and integrated into a unified learning framework. One limitation of the existing methods is that they do not model the interaction between aspect phrases and their contexts explicitly. For example, in the review “the picture is clear, bright and sharp and the sound is good”, the words “clear”, “bright” and “sharp” are related to the aspect phrase “picture”, while the word “good” is related to the aspect phrase “sound”. By the traditional model, these words are not differentiated when they are taken for the context, thereby causing noise in the grouping of the two aspect phrases. To address this issue, we propose a novel neural network structure that automatically learns the relative importance of each context word with respect to a target/aspect phrase, by leveraging an attention model (Luong et al., 2015; Rush et al., …

Content truncated to 500 words