SLS: Stochastic dialogue management via multi-layer classification

Many of the current dialogue systems have employed rule-based or template-based dialogue management strategies to highly control the system response and the dialogue flow. There have also been some studies on stochastic dialogue generation such as POMDP. Both have pros and cons, as it is hard to achieve a perfect dialogue performance relying on one sole solution.

In this project, we will explore automatic dialogue and response generation via the combination of stochastic approaches and heuristic methods. The student will explore a stochastic model to learn the best dialogue hypothesis based on crowd-generated data. We will start with a simple prototype framework such as an XML schema for defining tasks/actions, and then shift it towards automatic generation when we collect sufficient conversational data with the initial system setting. A hierarchical infrastructure will be investigated, where the higher layers are designed for categorizing the context of the conversation and the lower layers are designed for domain-specific word-based semantic understanding.

Requirements: machine learning background and Java expertise. This project could also start as a Super UROP project and turns into an MEng thesis later (see the UROP openings page). If interested, please send a CV to Jingjing Liu (