Matching Intentions for Discourse Parsing in Multi-party Dialogues.

Discourse parsing in multi-party dialogues is a fundamental task in dialogue systems, which involves identifying relations between elementary discourse units (EDUs), where an EDU is an utterance in a dialogue. Despite the variety of approaches that have been proposed to enhance context information in utterances, detecting related utterances that share only a few common words remains a challenging task. This paper proposes a novel insight that utilizing intention matching can significantly improve the modeling of utterance relations. Specifically, the intention of a speaker’s utterance represents a high-level abstraction of the utterance semantics, including the dialogue context and speaker context. In our model, a graph neural networks is utilized to encode each utterance’s intention, then the graph contrastive learning is used to study the match between utterances and intentions. Furthermore, we design three layers of intention to capture the meaning of utterances at different granularities. We conduct extensive experiments on two standard benchmark datasets, and the results show that our proposed model achieves state-of-the-art performance compared to current approaches, including the GPT-3.5 model.