Rethinking Task-Oriented Dialogue Systems: From Complex Modularity to Zero-Shot Autonomous Agent

Abstract

Task-oriented dialogue (TOD) systems are pre-001 dominantly designed to be composed of several 002 functional modules (e.g. dialogue state tracker, 003 dialogue policy, natural language generation) 004 whether they are pipeline or end-to-end archi-005 tectures. However, this modular design not 006 only heavily relies on massive fully-annotated 007 data, but also suffers from many intrinsic draw-008 backs, such as serious error accumulation, poor 009 generalization ability, high customization cost, 010 and low fault tolerance rate. In this paper, we 011 rethink the architecture of the task-oriented di-012 alogue systems and propose a novel fully zero-013 shot autonomous TOD agent, named AutoTOD, 014 where all the delicate modules in traditional 015 TOD systems are deprecated and all it needs 016 is a general-purpose instruction-following lan-017 guage model (e.g. GPT-4). AutoTOD only 018 leverages a simple instruction schema consist-019 ing of the description of tasks and external 020 APIs, and can autonomously decide to what to 021 do at each dialogue turn, including asking for 022 information, calling APIs, summarizing API re-023 sults, and correcting previous mistakes. More-024 over, we propose a simulation-based evalua-025 tion framework to better validate the abilities 026 of TOD models in real-life scenarios. Exten-027 sive experiments conducted on the MultiWOZ 028 and SGD datasets show the superior task com-029 pletion ability and flexible language skills of 030 AutoTOD. 1 031