Semantic Features-Based Discourse Analysis Using Deceptive and Real Text Reviews

Abstract

Social media usage for news, feedback on services, and even shopping is increasing. Hotel services, food cleanliness and staff behavior are also discussed online. Hotels are reviewed by the public via comments on their websites and social media accounts. This assists potential customers before they book the services of a hotel, but it also creates an opportunity for abuse. Scammers leave deceptive reviews regarding services they never received, or inject fake promotions or fake feedback to lower the ranking of competitors. These malicious attacks will only increase in the future and will become a serious problem not only for merchants but also for hotel customers. To rectify the problem, many artificial intelligence–based studies have performed discourse analysis on reviews to validate their genuineness. However, it is still a challenge to find a precise, robust, and deployable automated solution to perform discourse analysis. A credibility check via discourse analysis would help create a safer social media environment. The proposed study is conducted to perform discourse analysis on fake and real reviews automatically. It uses a dataset of real hotel reviews, containing both positive and negative reviews. Under investigation is the hypothesis that strong, fact-based, realistic words are used in truthful reviews, whereas deceptive reviews lack coherent, structural context. Therefore, frequency weight–based and semantically aware features were used in the proposed study, and a comparative analysis was performed. The semantically aware features have shown strength against the current study hypothesis. Further, holdout and k-fold methods were applied for validation of the proposed methods. The final results indicate that semantically aware features inspire more confidence to detect deception in text.