Detecting Opinions Using Deep Syntactic Analysis

Proceedings of Recent Advances in Natural Language Processing, pages 392–398, Hissar, Bulgaria, 12-14 September 2011. Detecting Opinions Using Deep Syntactic Analysis Caroline Brun Xerox Research Centre Europe Meylan, France Caroline.Brun@xrce.xerox.com Abstract In this paper, we present an opinion detection system built on top of a robust syntactic pars- er. The goal of this system is to extract opi- nions associated with products but also with characteristics of these products, i.e. to per- form feature-based opinion extraction. To car- ry out this task, and following a target corpus study, the robust syntactic parser is enriched by associating polarities to pertinent lexical elements and by developing generic rules to extract relations of opinions together with their polarity, i.e. positive or negative. These rela- tions are used to feed an opinion representa- tion model. A first evaluation shows very en- couraging results, but numerous perspectives and developments remain to be investigated. 1 Introduction Opinion mining (or sentiment analysis) arouses great interest in recent years both in academia and industry. With the emergence of discussion groups, forums, blogs, web sites compiling con- sumer reviews on various subjects, there is a huge mass of documents containing information expressing opinions. This constitutes a very im- portant data source for monitoring various appli- cations (business intelligence, product and ser- vice benchmarking, technology watch). Conse- quently, numerous research works at the cros- sroads of NLP and data mining, are focusing on the problem of opinion detection and mining. In this paper, we present an opinion detection sys- tem developed in the framework of the European Project Scoop1 1 http://www.scoopproject.eu/overview.html . This system uses a robust parser specifically adapted for opinion detection, and we focus here on recent developments made for English. Our goal is to extract opinions related to the main concepts commented in the reviews (e.g. products, movies, books...), but also on the features associated to these products (such as certain characteristics of the products, their price, associated services, etc...). After a brief review of related work, we describe a corpus analysis conducted on a first target cor- pus consisting of reviews about printers, copiers and scanners. The following section describes in details the building of the opinion detection sys- tem, which makes an intensive use of syntactic information. Finally, we present a preliminary evaluation of the performances of this system and conclude on our perspectives. 2 State of the Art Besides works about lexical resources acquisi- tion for opinion mining, discussed in section 4.3.2, two main types of works can be distin- guished: those aiming at classifying texts accord- ing to an overall polarity (positive, negative and sometimes neutral), generally based on super- vised approaches (such as (Pang et al. 2002), or (Charton and Acuna-Agost 2007)), and those aiming at extracting precise information about positive or negative aspects of a given product or topic. The latter consider that the main concept (e.g. a product) is related to several features (e.g. quality, print speed and resolution for a printer), that can be evaluated separately. Our system be- …

Content truncated to 500 words