Sentiment and Emotion Based Representations for Fake Reviews Detection

Proceedings of Recent Advances in Natural Language Processing, pages 750–757, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_087 750 Sentiment and Emotion Based Text Representation for Fake Reviews Detection Alimuddin Melleng Queen’s University Belfast amelleng01@qub.ac.uk Anna-Jurek Loughrey Queen’s University Belfast a.jurek@qub.ac.uk Deepak P Queen’s University Belfast deepaksp@acm.org Abstract Fake reviews are increasingly prevalent across the Internet. They can be uneth- ical and harmful. They can affect busi- nesses and mislead customers. As opin- ions on the Web are increasingly relied on, the detection of fake reviews has become more critical. In this study we explore the effectiveness of sentiment and emo- tions based representations for the task of building machine learning models for fake reviews detection. The experiment performed with three real-world datasets demonstrate that improved data represen- tation can be achieved by combining sen- timent and emotion extraction methods, as well as by performing sentiment and emo- tion analysis on a part-by-part basis by segmenting the reviews. 1 Introduction The Internet has evolved into a content creation platform where people express their opinions and experiences. Online reviews written by users have significant impact on customers and companies. Potential customers often consult reviews before making a purchase. Reviews help potential cus- tomers to gain insights from other people’s ex- periences, particularly in making choices on pur- chasing products or services. At the same time, companies need reviews on their products or ser- vices in order to get feedback and maintain good reputation. However, not all reviews available in the Internet are genuine. Profusion of reviews of questionable quality increase concerns about their trustworthiness. Moreover, users with mal- intent often post fake reviews (FR) to mislead cus- tomers by promoting or demoting products or tar- get stores. Authors of FR can sway customer choices towards companies with which they are associated, or against competitors making fake re- views a lucrative business. There has been an in- crease in FR profusion lately. According to the report of the Harvard Business School (Luca and Zervas, 2016) the percentage of fake reviews on YELP1 increased from 5 % in 2006 to 20% in 2013. This makes FR detection an important chal- lenge to be addressed. FR were firstly categorized by Jindal et al. (2008) into three groups: (1) Untruthful opin- ions: mislead readers by giving positive reviews to promote or demote target object, (2) Reviews on brands only: the reviewer focus on the brands, producers or sellers of a product or service without commenting on the product or service, (3) Non- reviews: the reviews are irrelevant to the product and do not contain opinions but advertisements or questions. The first category is the most challeng- ing type to detect, and that is the focus of our pa- per. Given the large numbers of reviews posted daily, automatic methods would be preferred over manual ones as illustrated in (Ott et al., 2011). Re- cent years have witnessed an increased impetus on machine learning methods for data-driven FR de- tection (Mukherjee et al., 2013; Ott et al., 2011; …

Content truncated to 500 words