Large-scale Opinion Relation Extraction with Distantly Supervised Neural Network

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1033–1043, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics Large-scale Opinion Relation Extraction with Distantly Supervised Neural Network Changzhi Sun2, Yuanbin Wu1, 2, Man Lan1, 2, Shiliang Sun1, 2, and Qi Zhang3 1Shanghai Key Laboratory of Multidimensional Information Processing 2Department of Computer Science and Technology, East China Normal University 3School of Computer Science, Fudan University {changzhisun}@stu.ecnu.edu.cn {ybwu,mlan,slsun}@cs.ecnu.edu.cn {qz}@fudan.edu.cn Abstract We investigate the task of open domain opinion relation extraction. Given a large number of unlabelled texts, we propose an efficient distantly supervised framework based on pattern matching and neural net- work classifiers. The patterns are de- signed to automatically generate training data, and the deep learning model is de- signed to capture various lexical and syn- tactic features. The result algorithm is fast and scalable on large-scale corpus. We test the system on the Amazon online review dataset, and show that the proposed model is able to achieve promising performances without any human annotations. 1 Introduction Opinion mining systems aim to detect and ex- tract opinion-related information from texts. With the help of natural language processing algo- rithms and large-scale user generated contents, re- searchers could take a closer look at how peo- ple express their opinions on various objects and topics. Such observations are important both for applications (e.g., recommendation and retrieval) and linguistic studies. In this paper, we address the task of opin- ion relation extraction. The task tries to identify opinion expressions (words indicating sentiments, emotions and comments), opinion targets (objects of opinions) and their relations (what opinion on which target). The following are two examples. 1. The unit is [well designed] and [per- fect reception]. 2. The Passion of The Christ will [touch your heart]. Opinion relations in the two sentences are (“well designed”, “unit”), (“perfect reception”, “unit”), and (“touch your heart”, “The Passion of The Christ”). Extracting opinion bearing relations is usually the first step towards fine-gained analysis of opinion in texts, and plays an important role in other sentiment related applications (e.g., senti- ment summarization). The goal of this paper is to extract opinion relations from open domain large- scale opinion bearing texts. Previous works on fine-grained opinion infor- mation extraction have achieved notable success on many aspects: various domains were exam- ined (Pontiki et al., 2015), different types of re- lations were studied (Ganapathibhotla and Liu, 2008; Narayanan et al., 2009; Wu et al., 2011), and both supervised and unsupervised (pattern- based) algorithms were applied. But we have ob- served some difficulties when trying to use exist- ing methods. The pattern based methods (both lexical patterns and syntactic patterns) are simple, fast, and scalable on large-scale datasets. How- ever, the robustness of patterns is usually question- able in practice. For example, syntactic patterns are sensitive to errors in parse trees, which are common in user generated contents. Lexical pat- terns either have limited coverage (e.g., fixed set of patterns (Riloff and …

Content truncated to 500 words