Dialogue and Interactive Systems
Dialogue systems aim at interacting with human users with natural languages for certain purpose, such as the accomplish specific tasks or just to have a good conversation. Current research efforts mostly focus on utterance understanding and generation, approaching them as a static natural language understanding and generation problems. In contrast, we pay attention to the essence of interaction property of dialogue systems, working towards building a system that can really interact with real users. The key techniques to be developed are dialogue state tracking and dialogue policy making.
Our current research lies on two fronts. In the first, we study dialogue state tracking which is the key for robust dialogues in multi-turn interactions. We insert an intermediate component (that can either be CNNs, RNNs or MLPs) between the encoder and decoder in a typical sequence-to-sequence (seq2seq) structure. It employs two copyNets , where one copies the necessary information from the encoder to the dialogue state tracker and the other copies the dialogue state tracker to the decoder. This enables the dialogue state tracker to be seamlessly fitted into the seq2seq model, and able to be jointly optimized with the tasks of utterance understanding and generation. This results in our Sequicity  and SEDST  models.
The second front aims to combine the dialogue systems with existing applications (such as the recommendation or QA systems etc.) by utilizing the interaction mechanism to achieve better performance. A successful case is our EAR model  for conversational recommendation. It integrates a recommendation component and a conversational component to support each other. The recommendation component estimates the user preferences and find the uncertain preferences for the conversational component to consult with the user; while the conversational components get the user feedback through interacting with the user and feed back to the recommendation components for better user preferences estimation. As such, EAR can explicitly capture the dynamic user preferences and able to deliver better recommendations. Note that, this research only focuses on dialogue (i.e., conversation) state and policy, setting the EAR and users (or user simulators) to interact through template instead of natural languages.
In the future, we would continue to investigate the above two research directions. Specifically, for the first direction, we will train a task-oriented dialogue system that can robustly interact with the users based on the Sequicity and SEDST structures. To do that, we need to build a user simulator to interact with the dialogue system where reinforcement learnings are employed for optimization. For the second direction, we plan to introduce conversational mechanisms to more tasks such as the recommender systems. We will equip the conversational QA system with the ability to consult with the users by asking clarification questions, thus enabling the conversational QA system be able to answer the question more precisely.
- Wenqiang Lei, Xisen Jin, Zhaochun Ren, Xiangnan He, Min-Yen Kan and Dawei Yin: Sequicity: Simplifying Task-oriented Dialogue Systems with Single Sequence-to-Sequence Architectures. ACL 2018.
- Xisen Jin, Wenqiang Lei, Zhaochun Ren, Hongshen Chen, Shangsong Liang, Yihong Zhao and Dawei Yin: Explicit State Tracking with Semi-Supervision for Neural Dialogue Generation. CIKM 2018: 1403-1412.
- Jiatao Gu, Zhengdong Lu, Hang Li and Victor O. K. Li: Incorporating Copying Mechanism in Sequence-to-Sequence Learning. ACL (1) 2016
- Wenqiang Lei, Xiangnan He, Yisong Miao, Qingyun Wu, Richang Hong, Min-Yen Kan and Tat Seng Chua (2020). Estimation–Action–Reflection: Towards Deep Interaction Between Conversational and Recommender Systems. In WSDM 2020