MFDG: A Multi-Factor Dialogue Graph Model for Dialogue Intent Classification

Jinhui Pang, Huinan Xu, Shuangyong Song, Bo Zou, Xiaodong He

Abstract: Interest in speaker intent classification has been increasing in multi-turn dialogues, as the intention of a speaker is one of the components for dialogue understanding. While most existing methods perform speaker intent classification at utterance-level, the dialogue-level comprehension is ignored. To obtain a full understanding of dialogues, we propose a Multi-Factor Dialogue Graph Model (MFDG) for Dialogue Core Intent (DCI) classification. The model gains an understanding of the entire dialogue by explicitly modeling multi factors that are essential for speaker-specific and contextual information extraction across the dialogue. The main module of MFDG is a heterogeneous graph encoder, where speakers, local discourses, and utterances are modelled in a graph interaction manner. Based on the framework of MFDG, we propose two variants, MFDG-EN and MFDG-EE, to fuse domain knowledge into the dialogue graph. We apply MFDG and its two variants to a real-world online customer service dialogue system on the e-commerce website, JD, in which the MFDG can help achieving an automatic intent-oriented classification of finished service dialogues, and the MFDG-EE can further promote dialogue comprehension with a well-designed knowledge graph. Experiments on this in-house JD dataset and a public DailyDialog dataset demonstrate that MFDG performs reasonably well in multi-turn dialogue classification.