CS Events

PhD Defense

Coherence as a Key Ingredient to Learn Effective Communication Strategies

 

Download as iCal file

Wednesday, May 17, 2023, 09:00am - 11:00am

 

Abstract: Coherence is a very important aspect of communication and helps keep track of the ongoing discussion. It gives structure to discourse and can be categorized through a set of qualitative coherence relations [57]. Human interlocutors collaborate with each other to engage in coherent communication and utilize a wide range of communication strategies, e.g. clarifications and corrections, to resolve any ambiguities in the context and achieve shared objectives [25]. Conversational models, based on MDP or end-to-end neural networks, assume that coherence can be estimated implicitly from the conversational data. In this work, we empirically show that modelling coherence explicitly helps in learning flexible and context-specific communication strategies. We also show that coherence-based state tracking mechanisms allow us to build simulation environments which can be used to empirically analyze different communication aspects such as clarifications. This allows us to identify the usefulness of techniques like reinforcement learning (RL) and identify cases where such techniques are most useful. Lastly, we present experiments which show how explicit use of coherence signals can help in building effective evaluation models.

 

Speaker: Baber Khalid

Location : CoRE 305

Committee

Professor Matthew Stone (Chair)

Professor Yongfeng Zhang

Professor Karl Stratos

Professor Matthew Purver (Queen Mary, University of London)

 

Event Type: PhD Defense

Abstract: See above

Organization

Rutgers University

School of Arts & Sciences

Department of Computer Science

Contact  Professor Matthew Stone