Data Augmentation for Fake Reviews Detection

Proceedings of Recent Advances in Natural Language Processing, pages 673–680 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_073 673 Data Augmentation for Fake Reviews Detection Ming Liu and Massimo Poesio Queen Mary University of London Mile End Road London E1 4NS {acw661,m.poesio}@qmul.ac.uk Abstract In this research, we studied the relationship between data augmentation and model accuracy for the task of fake review detection. We used data generation methods to augment two different fake review datasets and compared the performance of models trained with the original data and with the augmented data. Our results show that the accuracy of our fake review detection model can be improved by 0.31 percentage points on DeRev Test and by 7.65 percentage points on Amazon Test by using the augmented datasets. 1 Introduction Online communication has increased the speed and quantity of information sharing between people. While this change has brought a number of benefits, it also increased the opportunities for unscrupulous individuals to deceive (Newman et al., 2003; Hancock et al., 2007; Vrji, 2008). Research shows that on average people tell 1 or 2 lies a day; now, lying has migrated from face-to-face communication to online (Hancock et al., 2004; Mihalcea and Strapparava, 2009). Fake reviews are a particularly problematic type of deceptive online communication, given our reliance on online reviews to guide our purchases (Ott et al., 2011; Fornaciari and Poesio, 2014; Fornaciari et al., 2020). About 1% to 6% positive hotel reviews are estimated to be fake (Ott et al., 2012). Automatic deception detection methods rely on stylometric methods extracting from text hundreds of linguistic features (Newman et al., 2003; Han- cock et al., 2007; Mihalcea and Strapparava, 2009; Fornaciari and Poesio, 2014). More recently, Deep learning has been used (Girgis et al., 2018; Kaliyar et al., 2021; Fornaciari et al., 2021; Salminen et al., 2022). This research also led to the creation of several datasets for training such models (Ott et al., 2011; Fornaciari and Poesio, 2014; Amazon, 2018; Fornaciari et al., 2020). However, these datasets have a number of limitations. They tend to be small, and very domain dependent (a dataset of TripAdvisor reviews is not suitable for training models to detect fake reviews on Amazon and vice- versa). Even more crucially, few of them consist of genuine fake reviews; most were artificially created using crowdsourcing. But crowdsourced fake reviews are known to be different from genuine fake reviews (Fornaciari et al., 2020). In this paper, we focus on the issue of creating suitable datasets for fake review detection research. Data augmentation techniques using text gener- ation would appear to be a potential solution to the problem of generating datasets for fake review detection when we only have a small amount of fake or genuine reviews. And given that modern text generation methods appear to be able to create artificial texts extremely similar to model texts used to prompt them, these methods might be more likely than crowdsourcing of creating artificial fake reviews similar to real fake reviews. In proposals such as (Shehnepoor …

Content truncated to 500 words