An Interpretable Neuro-Symbolic Reasoning Framework for Task-Oriented Dialogue Generation

We study the interpretability issue of task-oriented dialogue systems in this\npaper. Previously, most neural-based task-oriented dialogue systems employ an\nimplicit reasoning strategy that makes the model predictions uninterpretable to\nhumans. To obtain a transparent reasoning process, we introduce neuro-symbolic\nto perform explicit reasoning that justifies model decisions by reasoning\nchains. Since deriving reasoning chains requires multi-hop reasoning for\ntask-oriented dialogues, existing neuro-symbolic approaches would induce error\npropagation due to the one-phase design. To overcome this, we propose a\ntwo-phase approach that consists of a hypothesis generator and a reasoner. We\nfirst obtain multiple hypotheses, i.e., potential operations to perform the\ndesired task, through the hypothesis generator. Each hypothesis is then\nverified by the reasoner, and the valid one is selected to conduct the final\nprediction. The whole system is trained by exploiting raw textual dialogues\nwithout using any reasoning chain annotations. Experimental studies on two\npublic benchmark datasets demonstrate that the proposed approach not only\nachieves better results, but also introduces an interpretable decision process.\n