THOTH seminar at Inria Montbonnot
14th February 2017, 11:00AM
Julien Perez: "Machine Reading, Gated Memory Networks and Non-Markovian Decisions"
Abstract: During the last couple of years, the field of natural language understanding has started to benefit from recent advances in deep learning and attention-based models. In this context, the field of machine reading has emerged from the observation that natural language remains the largest source of information and knowledge available. Machine reading escapes from the necessity of explicit intermediate encoding of knowledge. It consists in developing and evaluating optimizable learning models that are able to answer questions about a given text using (text,question,answer) triples. In the first part of this talk, we provide a survey of models and corpora recently developed in the domain. In the second part, we show how this approach enables innovative directions for learning dialog systems, from the state tracking task to end-to-end learning. We also introduce a novel model called the Gated Memory Network that achieves state of the art performance on several of these corpora. Finally, we apply this model as a general framework for non-Markovian decision making.