4. DECEPTION
[This project is a part of my tenure at University of South Carolina as an affiliate Researcher]
A project funded by the US Airforce office of scientific Research
Problem Understanding
Deception in a simpler term is a lie. To study deception is complex topic since capturing it involves text, speech, facial expressions, body movements and much more.
Deeper Understanding(Literature Review)
There are three well-known methods for identifying untruthfulness
Omission, involves leaving out pertinent information whereas the other method, referred to as commission, involves generating false information (Hample, 1982 )
The vast majority of studies on the behavioral correlates of lying have used paradigms requiring participants to generate lies of commission (example, Vrij, Edwards, Roberts, & Bull, 2000)
Using a novel laboratory paradigm, Ratliff, Berman, and Barry (2006, unpublished manuscript) found that, when presented with a social interaction in which deception was required, participants were much more likely to commit lies of omission than lies of commission.
Proposed Solution
Applying ML models on the deceptive dataset to predict the different form of deception
Implementation
Understanding forms of deception using the definitions from psychology, focussing on a deeper understanding of omission
Data Collection: Data is collected from a well-known Twitter handle of a media house in India called "Times of India" and a mixture already existing fake news dataset
Data Annotation: 40+ annotators worked on classifying different forms of deception and lies as proposed in the flowchart
An analysis of links between 5W(what, who, when, where, why) and form of deception was observed
Building a scalable multitask learning model to detect different form of deception
My Contribution
Studied lies of omission from a 5W-based SOTA Semantic Role Labeler
Identified different forms of omission namely speculation, bias, distortion, and opinion on tweets/Fake news dataset
While identifying different forms of deception, I observed a pattern with the missing Ws and how certain words seemed a little biased making me believe that it's non-deceptive.
Experimenting with a multi-task learning model to predict different forms of deception and lies as proposed in the flowchart above
Result
Preprint available on Arxiv and continuing experimenting with multimodal data to go deeper in the next iteration [Paper Link]