Workshop on Machine Learning in Speech and Language Processing

September 13, 2016
San Francisco, CA, USA

Speaker: Ryan Lowe (McGill University)

Title: Modern challenges in learning end-to-end dialogue systems

Abstract:
Very recent work in building dialogue systems has focused on end-to-end systems that learn directly from conversational data, without relying on pre-specified state or action representations. This has become feasible at a large scale only in the past few years with the rise of deep recurrent neural network (RNN) methods for modelling text. Although these systems hold more promise for scaling to general domains compared to traditional modular systems, there are still significant challenges to be overcome before these systems are able to converse naturally with humans about any subject. In this talk, I will discuss some of the most pressing issues facing end-to-end dialogue systems, including the need for large datasets, a lack of standardized evaluation methods, and the generation of generic responses, as well as some of our work addressing these problems.