Detecting Deception in Non-Cooperative Dialogue: A Smarter Adversary Cannot be Fooled That Easily

Recent work has learned non-cooperative dialogue behaviour within a stochastic trading game, including dialogue moves such as bluffing and lying. Here, we introduce an adversary which can detect deception based on logical contradictions between dialogue moves. Being caught in deception, the adversary will penalise this behaviour by either refusing to trade or declaring victory. We compare our results to a learning agent trained with a gullible adversary and show that a more realistic adversary decreases the chances of winning by over 20%, if the penalty for cheating is to lose the game. In future work we will re-train the learning agent within this more challenging environment.