Adversarial reading networks for machine comprehension

Machine reading has recently shown remarkable progress thanks to differentiable reasoning models. In this context, End-to-End trainable Memory Networks (MemN2N) have demonstrated promising performance on simple natural language based reasoning tasks such as factual reasoning and basic deduction. However, the task of machine comprehension is currently bounded to a supervised setting and available question answering dataset. In this paper we explore the paradigm of adversarial learning and self-play for the task of machine reading comprehension. Inspired by the successful propositions in the domain of game learning, we present a novel approach of training for this task that is based on the definition of a coupled attention-based memory model. On one hand, a reader network is in charge of finding answers regarding a passage of text and a question. On the other hand, a narrator network is in charge of obfuscating spans of text in order to minimize the probability of success of the reader. We experimented the model on several question-answering corpora. The proposed learning paradigm and associated models present encouraging results.