Towards Accurate and Smart Task-Oriented Dialogue Systems

Dialogue systems such as Apple Siri, Microsoft Cortana are designed to help people in many aspects. In particular, task-oriented dialogue systems (TOD) assist humans in finishing various tasks such as setting alarms and making recommendations. Keeping track of user intention and providing information accurately and smartly with minimum conversational turns is a big challenge. In this first work, I propose a discriminative nearest neighbor intent detection model to accurately identify user intentions and recognize unsupported or unrelated out-of-scope (OOS) user queries with limited training examples. In the second work, I further investigate whether pre-trained Transformers are robust in few-shot intent detection w.r.t. general and relevant OOS examples on our newly constructed and released datasets. In the third work, I propose a few-shot intent detection model through contrastive pre-training and fine-tuning to accurately identify both general and fine-grained user intents. I propose a dual strategy for slot-value predictions on dialog state tracking across multiple domains in the fourth work. Furthermore, since pipeline-based systems require lots of annotations and are hard to scale, the errors are also easy to propagate to different modules. We thus want to build accurate and smart end-to-end TOD based on seq2seq language models to leverage sophisticated NLU and NLG. In the fifth work, I present an end-to-end TOD framework via simple 'database' to handle all the components and add flexibility to update intents and slots dynamically.