UWaterloo at SemEval-2016 Task 5: Minimally Supervised Approaches to Aspect-Based Sentiment Analysis

Proceedings of SemEval-2016, pages 372–377, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UWaterloo at SemEval-2016 Task 5: Minimally Supervised Approaches to Aspect-Based Sentiment Analysis Olga Vechtomova and Anni He University of Waterloo, Canada {ovechtomova, anni.he}@uwaterloo.ca Abstract This paper describes our system for Aspect- Based Sentiment Analysis (ABSA), task 5 of SemEval 2016. To conduct sentence level ABSA, we employed minimally supervised approaches for each type of extracted infor- mation. The system uses Word2Vec to derive word semantic similarities, and relies on ex- ternal review corpora as training data. The re- sults of the 2016 evaluation are discussed and suggestions for improvements are given. 1 Introduction In this paper, we describe our system for Aspect- Based Sentiment Analysis (ABSA), Task 5 of Se- mEval 2016. The task involves classifying consumer reviews into existing aspect category labels (slot 1), identifying the opinion target expression (OTE) cor- responding to the aspect category label (slot 2), then assigning sentiment polarity expressed about the as- pect category (slot 3). The aspect category (slot 1) consists of two parts: Entity (e.g. food, ambience, restaurant) and Attribute (e.g. price, quality). Our system tackles the problem in three stages ad- dressing each slot individually for the restaurant do- main and laptop domain (slot 3 only). The strengths of the system include no lexicon resources and min- imal amount of labeled corpora. For the aspect category identification and OTE extraction, the ap- proach is semi-supervised utilizing Word2Vec mod- els (Mikolov et al., 2013) to derive word semantic similarities. The lists of words ranked by their se- mantic similarities to each entity and each attribute are generated for the aspect category identification. For the polarity assignment, we build from the ap- proach of (Vechtomova et al., 2014) using corpora of consumer-rated electronics and restaurant reviews. The method does not require any sentiment labels at the word or sentence level or sentiment lexicons. The paper describes each stage of the system in detail in section 2. In section 3, results of the system are presented and conclusions are drawn in section 4. 2 System Description 2.1 Corpora At different stages of our system we used various resources we automatically generated from two cor- pora, described below. In both of these corpora we used only the original text of the review and the over- all review ratings assigned by the consumer. For the Restaurant domain, we used a corpus of 157,865 restaurant reviews from one of the major business review websites (Vechtomova, 2014). The collection contains reviews for 32,782 restaurants in the United States. The average number of words per review is 64.7. We will refer to this corpus as the Restaurant corpus throughout the paper. For the Laptop domain, we used a subset of the Amazon corpus (Jindal and Liu, 2008), containing 138,504 reviews of products in the category Con- sumer Electronics. We will refer to this corpus as the CE corpus throughout the paper. 2.2 Aspect Category Detection For this slot, the goal is to identify …

Content truncated to 500 words