Human-to-Human Conversation Dataset for Learning Fine-Grained Turn-Taking Action

Kehan Chen, Zezhong Li, Suyang Dai, Wei Zhou, Haiqing Chen

Abstract: Conducting natural turn-taking behavior takes a crucial part in the user experience of modern spoken dialogue systems. One way to build such system is to learn those behaviors from real-world human-to-human dialogues, which have the most diverse and fine-grained turn-taking actions than any manual constructed sessions. In this paper, we propose a Dataset — FTAD which could be used to learn turn-taking policies directly from human. First, we design an annotation mechanism to transform existing human-to-human dialogue session into structural data with most fine-grained turn-taking actions reserved. Then we explored a set of supervised learning tasks on it, showing the challenge and potential of learning complete fine-grained turn-taking policies based on such data.