Joint Opinion Relation Detection Using One-Class Deep Neural Network

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 677–687, Dublin, Ireland, August 23-29 2014. Joint Opinion Relation Detection Using One-Class Deep Neural Network Liheng Xu, Kang Liu and Jun Zhao National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences, Beijing, 100190, China {lhxu, kliu, jzhao}@nlpr.ia.ac.cn Abstract Detecting opinion relation is a crucial step for fine-gained opinion summarization. A valid opin- ion relation has three requirements: a correct opinion word, a correct opinion target and the linking relation between them. Previous works prone to only verifying two of these requirements for opinion extraction, while leave the other requirement unverified. This could inevitably intro- duce noise terms. To tackle this problem, this paper proposes a joint approach, where all three requirements are simultaneously verified by a deep neural network in a classification scenario. Some seeds are provided as positive labeled data for the classifier. However, negative labeled data are hard to acquire for this task. We consequently introduce one-class classification problem and develop a One-Class Deep Neural Network. Experimental results show that the proposed joint approach significantly outperforms state-of-the-art weakly supervised methods. 1 Introduction Opinion summarization aims to extract and summarize customers’ opinions from reviews on products or services (Hu and Liu, 2004; Cardie et al., 2004). With the rapid expansion of e-commerce, the number of online reviews is growing at a high speed, which makes it impractical for customers to read throughout large amounts of reviews to choose better products. Therefore, it is imperative to automatically gener- ate opinion summarization to help customers make more informed purchase decisions, where detecting opinion relation is a crucial step for opinion summarization. Before going further, we first introduce some notions. An opinion relation, is a triple o = (s, t, r), where three factors are involved: s is an opinion word which refers to those words indicating sentiment polarities; t is an opinion target, which can be any entity or aspect of an entity about which an opinion has been expressed; r refers to the linking relation between s and t. As in Example 1, s={clear}, t={sceen}, and there is a linking relation between the two words because clear is used to modify screen. Example 1. This mp3 has a clear screen. For a valid opinion relation, there are three requirements corresponding to the three factors: (i) the opinion word indicates sentiment polarity; (ii) the opinion target is related to current domain; (iii) the opinion word modifies the opinion target. Previous weakly supervised methods often expand a seed set and identify opinion relation either by co-occurrence statistics (Hu and Liu, 2004; Hai et al., 2012) or syntactic dependencies (Popescu and Etzioni, 2005; Qiu et al., 2009) following the assumption below. Assumption 1. Terms that are likely to have linking relation with the seed terms are believed to be opinion words or opinion targets. For example, if one has an opinion word seed clear (which satisfies requirement i), and one finds that it modifies the word …

Content truncated to 500 words