Topic-Aware Multi-turn Dialogue Modeling

In the retrieval-based multi-turn dialogue modeling, it remains a challenge to select most appropriate response according extracting salient features in context utterances. As conversation goes on, topic shift at discourse-level naturally happens through continuous context. However, all known systems are satisfied with exploiting local words for utterance representation but fail capture such essential global topic-aware clues discourse-level. Instead of taking topic-agnostic n-gram as processing unit matching purpose existing systems, this paper presents novel solution which segments and extracts utterances an unsupervised way, so that resulted model is capable capturing need thus effectively track flow during conversation. Our modeling implemented by newly proposed segmentation algorithm Topic-Aware Dual-attention Matching (TADAM) Network, matches each segment dual cross-attention way. Experimental results on three public datasets show TADAM can outperform state-of-the-art method, especially 3.3% E-commerce dataset has obvious shift.