Adversarial Networks for Machine Reading - Naver Labs Europe


Machine reading has recently progressed remarkably with a help of 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 training of machine comprehension models commonly requires an annotated question-answer dataset for supervised learning. In this paper we explore adversarial learning and self-play for developing machine reading comprehension. Inspired by the success in the domain of game learning, we present a novel approach to train machine comprehension models based on a coupled attention-based memory model. In our approach, a reader network is in charge of finding answers to the questions regarding a passage of text, while a narrator network tries to obfuscate spans of text in order to minimize the probability of success of the reader. We tested the model on several question-answering corpora. The proposed learning paradigm and associated models show promising results.