HiTKG: Towards Goal-Oriented Conversations via Multi-Hierarchy Learning

Human conversations are guided by short-term and long-term goals. We study how to plan goal sequences as coherently humans do naturally direct them an assigned in open-domain conversations. Goal a series of knowledge graph (KG) entity-relation connections generated KG walkers that traverse through the KG. The existing recurrent attention based either insufficiently utilize conversation states or lack global guidance. In our work, hierarchical model learns planning learning framework. present HiTKG, transformer-based walker leverages multiscale inputs make precise flexible predictions on paths. Furthermore, we propose two-hierarchy framework employs two stages learn both turn-level (short-term) global-level (long-term) Specifically, at first stage, HiTKG is trained supervised fashion sequences; second tries approach via reinforcement learning. addition, MetaPath backbone method for path representation exploit entity relation information concurrently. further Multi-source Decoding Inputs Output-level Length Head improve decoding controllability. Our experiments show achieves significant improvement performance compared with state-of-the-art baselines. Additionally, automatic human evaluation prove effectiveness planning.