Developing Better Models for Dialogue Threads and Responses

Thumbnail Image
Li, Tianda
Dialogue System , Response Selection , Multiple Participants , Dialogue Disentanglement , Deep Contextualized Utterance Representations , Natural Language Processing
In this thesis, we focus on developing machine learning algorithms to model human dialogues and conversations by investigating three basic tasks. We first study the development of retrieval-based chatbots, in which the models aim to recommend the most appropriate responses for users based on the dialogue context. We propose a pretrained-based SA-BERT method to incorporate the identities of speakers which outperforms all existing models on four benchmark datasets. In the second task, we aim to help users to better access the dialogue history by proposing novel models for dialogue disentanglement. Specifically, multiple threads of discussions in the same dialogue are disentangled and organized to single topic threads. This research can help users to more easily follow up on the topics discussed in complex dialogues. We propose a hierarchical model by considering both single utterance and context semantics, which achieves new state-of-the-art performance on two datasets. In the third task, we are concerned with the outcome of dialogues --- determining if users' questions or problems have been addressed or answered in a dialogue, and if so, where they are addressed in the dialogue. Different from the previous study in which the contribution of each utterance towards the outcome is a black-box, we perform a pilot study by adding interpretability through determining if each utterance is contributing to addressing the user's questions. Our attention-based model is shown to achieve the best-reported performance on the evaluation dataset.
External DOI