Analyzing Deceptive Opinion Spam Patterns: the Topic Modeling Approach

Deceptive Opinion Spam commonly takes the form of fake reviews (negative or positive) posted by a malicious web user to hurt or inflate a company's image. As these reviews have been deliberately written to deceive the reader, human reviewers are faring little better than a chance in detecting these deceptive statements. Thus, there is a dire need to address this issue as extracting text patterns from the fraudulent texts with meaningful substructures still remains a challenge. In our research, to obtain a deeper understanding of how lies are expressed in texts, we consider the task as a topic modeling problem, in which we constructed a model to learn the patterns that constitute a fake review, and then explore the outputs of this model to identify those patterns. Topic models may be useful in this task due to their ability to group multiple documents into smaller sets of key topics. As the linguistic cues of the lies are still unknown, a key advantage of this approach is that the algorithm encourages the mixtures composed of only few topics, which makes the representation more interpretable and provides additional opportunities to reveal the patterns and structures within the systems of documents. Our methodology proved to be useful for this study, revealing the lexical cues generally applied by human reviewers to generate deceptive language.