Taking the latest and greatest advances in natural language processing and deep learning, CPSSD students trained a series of models to classify human deception in datasets of user submitted reviews. The output of the research was threefold: a scientific paper describing our research; an open source API providing access to query our trained models; and Lucify, a web interface allowing end users to view an interactive and comprehensive report of review deception and legitimacy.
In our publication we perform an analytic comparison of a number of techniques used to detect fake and deceptive online reviews. We apply a number of machine learning approaches found to be effective, and introduce our own approach by fine-tuning state of the art contextualised embeddings. The results we obtain show the potential of contextualised embeddings for fake review detection, and lay the groundwork for future research in the area.
This is the first research project by CPSSD students and was published in high regard by the Association for Computational Linguistics conference. The ACL conference is “the big one” in compuational linguistics, and presents the greatest NLP and machine learning discoveries of the year. Our submission maxed out a full 5⁄5 from our ACL reviewer’s ‘Recommendation’ scores, and our primary ACL reviewer commented very positively:
This will be a short review because fundamentally I find this to be a fantastic contribution to the student research workshop and I have almost no critiques to provide. The paper establishes the problem well, digs into prior work, is methodologically sound and thorough, adds interesting insights to the conversation on this type of problem, is clear and well-written throughout
As Bachelor level students, we were the most junior authors present at the conference. Even more flattering, it was suggested that we publish our work alongside the main speakers:
The authors should potentially consider withdrawing and working in the reviewer feedback and then submitting it to a main track of a conference. This is good work, well done.
We owe much to our proficient and thorough reviewer and supervisor Jennifer Foster, and to our supervisor Andrew McCarren.
Niall Walsh, Stefan Kennedy, Kirill Sloka