Uni-ETOD: User-Need-Driven Chain of Thought Framework for Fully End-to-end Task-oriented Dialogue System
Abstract: Fully End-to-End Task-Oriented Dialogue Systems (Fully ETOD) retrieve knowledge from a knowledge base in a differentiable manner and generate responses using a language model generator without the need for modular training. However, Fully ETOD faces some challenges. During the retrieval process, the retriever retrieves the knowledge base in a black-box manner, making it difficult for the generator to differentiate the large amount of knowledge obtained by the retriever. This leads to a degradation in the quality of the responses and the trustworthiness of the system. Moreover, as the size of the knowledge base grows, it may exacerbate the risk of this problem. To address this challenge, we first design a dataset for Fully ETOD based on large-scale knowledge bases called FakeRest to solve the scarcity of annotated dialogue data based on large-scale knowledge bases. We also propose a User-need-driven Chain of Thought Framework (Uni-ETOD) for Fully ETOD, which aims to guide LLMs to gradually understand users' thought processes and improve the quality of responses in Fully ETOD. We use ChatGPT, Gemini, Llama3, Mistral, and ChatGLM as the backbone models of the system. On FakeRest, we comprehensively evaluate the capability of each step of Uni-ETOD. The results show that Uni-ETOD will help LLMs better distinguish the retrieved knowledge and enhance the credibility and interpretability of the whole system.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: retrieval, knowledge augmented
Languages Studied: English
Submission Number: 1191
Loading