In this project, we use deep learning to detect deception through online conversations in the interaction-based party game Mafia. We present various neural network frameworks (LSTM, bidirectional LSTM, CNN) to generate predictions of deceptive text. The results of this paper suggest the capability of neural network architectures to effectively detect deception in the Mafia setting, and offer qualitative analysis in agreement with current deception research findings.
We have trained tensorflow frameworks based on unidirectional LSTM, bidirectional LSTM, as well as CNN, that successfully perform deception detection on online Mafia game interactions using the Mafiascum dataset . All three models performed above baseline accuracy (81.6%), with 88.8% (unidirectional LSTM), 94.7% (bidirectional LSTM), and 94.9% (CNN). This suggests that deeplearning techniques may be able to effectively identify linguistic cues that signal deception.
This is a relatively recent paper (March 18, 2019). I wonder how this compares against good players here.