Modeling Emotional Dialogs with Sequence to Sequence Networks
Event details
Date | 27.08.2018 |
Hour | 09:00 › 11:00 |
Speaker | Yubo Xie |
Location | |
Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Martin Rajman
Thesis advisor: Prof. Pearl Pu
Co-examiner: Prof. Wulfram Gerstner
Abstract
In the recent years, the successful application of sequence to sequence neural networks to statistical machine translation, including the usage of attention mechanism for the aim of translation alignment, has inspired a huge amount of work in dialog generation. Most of the studies are focused on how to improve the content quality of the dialogs generated, for example, the diverse beam search algorithm to increase the diversity in the responses, the persona-based model to handle the issue of speaker consistency, and the hierarchical recurrent encoder-decoder network to model the context of multi-turn dialogs. However, not much attention has been put on the affective/emotional aspects of the dialogs. In this proposal, we introduce some existing literature on emotion handling in dialog systems. Since the current work mainly deals with emotional responses in single-turn dialogs, we then discuss how we could potentially extend to multi-turn settings, ideally with the capability of tracking the change of emotions when the dialog progresses.
Background papers
Sequence to Sequence Learning with Neural Networks, by Ilya Sutskever, et al.
Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models, by Iulian Vlad Serban, et al.
Affective Neural Response Generation, by Nabiha Asghar, et al.
Exam president: Prof. Martin Rajman
Thesis advisor: Prof. Pearl Pu
Co-examiner: Prof. Wulfram Gerstner
Abstract
In the recent years, the successful application of sequence to sequence neural networks to statistical machine translation, including the usage of attention mechanism for the aim of translation alignment, has inspired a huge amount of work in dialog generation. Most of the studies are focused on how to improve the content quality of the dialogs generated, for example, the diverse beam search algorithm to increase the diversity in the responses, the persona-based model to handle the issue of speaker consistency, and the hierarchical recurrent encoder-decoder network to model the context of multi-turn dialogs. However, not much attention has been put on the affective/emotional aspects of the dialogs. In this proposal, we introduce some existing literature on emotion handling in dialog systems. Since the current work mainly deals with emotional responses in single-turn dialogs, we then discuss how we could potentially extend to multi-turn settings, ideally with the capability of tracking the change of emotions when the dialog progresses.
Background papers
Sequence to Sequence Learning with Neural Networks, by Ilya Sutskever, et al.
Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models, by Iulian Vlad Serban, et al.
Affective Neural Response Generation, by Nabiha Asghar, et al.
Practical information
- General public
- Free
Contact
- EDIC - [email protected]