Language Technologies Thesis Proposal
- Gates Hillman Centers
- Reddy Conference Room 4405
- TIANCHENG ZHAO
- Ph.D. Student
- Language Technologies Institute
- Carnegie Mellon University
Learning Generative End-to-end Dialog Systems with Knowledge
Dialog systems are intelligent agents that can converse with human in natural language and facilitate human. Traditional dialog systems follow a modular approach and often have trouble expanding to new or more complex domains, which hinder the development of more powerful future dialog systems. This dissertation targets at an ambitious goal: to create domainagnostic learning algorithms and dialog models that can continuously learn nto converse in any new domains and only requires a small amount of new data for the adaption. Achieving this goal first requires powerful statistic models that are expressive enough to model the natural language and decision-making process in dialogs of many domains; and second requires a learning framework enabling models to share knowledge from previous experience so it can learn to converse in new domains with limited data.
End-to-end (E2E) generative dialog models based on encoder-decoder neural networks are strong candidates for the first requirement. The basic idea is to use an encoder network to map a dialog context into a learned distributed representation and then use a decoder network to generate the next system response. These models are not restricted to hand-crafted intermediate states and can in principle generalize to novel system responses that are not observed in the training. However, it is far from trivial to build a full-fledged dialog system using encoder-decoder models. Thus in the first stage of this thesis, we develop a set of novel neural network architectures that offer key properties that are required by dialog modeling. Experiments prove that the resulting system can interact with both users and symbolic
knowledge bases, model complex dialog policy and reason over long discourse history.
We tackle the second requirement by proposing a novel learning with knowledge (LWK) framework that can adapt the proposed system to new domains with minimum data. Two types of knowledge are studied: 1) domain knowledge from human experts 2) models' knowledge from learning related domains. To incorporate these knowledge, a domain description that can compactly encode domain knowledge is first proposed. Then we develop novel domain-aware models and training algorithms to teach the system learn from data in related domains and generalize to unseen ones. Experiments show the proposed framework is able to achieve strong performance in new domains with limited, even zero, in-domain training data In conclusion, this dissertation shows that by combing specialized encoderdecoder models with the proposed LWK framework, E2E generative dialog models can be readily applied in complex dialog applications and can be easily expanded to new domains with extremely limited resources, which we believe is an important step towards future general-purpose conversational agents that are more natural and intelligent.
Maxne Eskenazi (Chair)
Dilek Hakkani-Tur (Google)