GraphMemDialog: Optimizing End-to-End Task-Oriented Dialog Systems Using Graph Memory Networks

Effectively integrating knowledge into end-to-end task-oriented dialog systems remains a challenge. It typically requires incorporation of an external base (KB) and capture the intrinsic semantics history. Recent research shows promising results by using Sequence-to-Sequence models, Memory Networks, even Graph Convolutional Networks. However, current state-of-the-art models are less effective at history KB in following ways: 1. The representation is not fully context-aware. dynamic interaction between seldom explored. 2. Both sequential structural information can contribute to capturing semantics, but they studied concurrently. In this paper, we propose novel Network (GMN) based Seq2Seq model, GraphMemDialog, effectively learn inherent hidden history, model KBs. We adopt modified graph attention network rich whereas context-aware entities learnt our GMN. To exploit interaction, design learnable memory controller coupled with entity memories recurrently incorporate context through multi-hop reasoning mechanism. Experiments on three public datasets show that GraphMemDialog achieves performance outperforms strong baselines large margin, especially datatests more complicated information.