Leveraging Different Context for Response Generation through Topic-guided Multi-head Attention

Multi-turn dialogue system plays an important role in intelligent interaction. In particular, the subtask response generation in a multi- turn conversation system is a challenging task, which aims to generate more diverse and contextually relevant responses. Most of the methods focus on the sequential connection between sentence levels by using hierarchical framework and attention mechanism, but lack reflection from the overall semantic level such as topical information. Previous work would lead to a lack of full understanding of the dialogue history. In this paper, we propose a context-augmented model, named TGMA-RG, which leverages the conversational context to promote interactivity and persistence of multi-turn dialogues through topic-guided multi-head attention mechanism. Especially, we extract the topics from conversational context and design a hierarchical encoder-decoder models with a multi-head attention mechanism. Among them, we utilize topics vectors as queries of attention mechanism to obtain the corresponding weights between each utterance and each topic. Our experimental results on two publicly available datasets show that TGMA-RG improves the performance than other baselines in terms of BLEU-1, BLEU-2, Distinct-1, Distinct-2 and PPL.