Adversarial Language Games for Advanced Natural Language Intelligence

We study the problem of adversarial language games, in which multiple agents with conflicting goals compete each other via natural interactions. While games are ubiquitous human activities, little attention has been devoted to this field processing. In work, we propose a challenging game called Adversarial Taboo as an example, attacker and defender around target word. The is tasked inducing utter word invisible defender, while detecting before being induced by attacker. Taboo, successful need hide or infer intention, induce defend during conversations. This requires several advanced abilities, such pragmatic reasoning goal-oriented interactions open domain, will facilitate many downstream NLP tasks. To instantiate game, create environment competition platform. Comprehensive experiments on baseline attack defense strategies show promising interesting results, based discuss some directions for future research.