GLMDST: Chinese Dialogue State Tracking Framework Driven by LLM

Dialogue State Tracking (DST) is an important component in task-oriented dialogue systems. Accurate dialogue state tracking is essential for generating correct dialogue actions and appropriate natural language responses. However, annotating conversation states is notoriously difficult, leading to a scarcity of DST datasets in languages other than English. With the development of DST, high-quality annotated Chinese conversation datasets have gradually emerged. Despite the emergence of available Chinese conversation datasets, research on the Chinese DST task remains relatively limited.The emergence of Large Language Model (LLM) such as GPT-3 and ChatGPT has sparked considerable interest in evaluating their effectiveness across different tasks. However, the performance of LLM in Chinese DST remains to be explored. In this study, we introduce GLMDST, a Chinese DST framework powered by a smaller open-source bilingual model called chat-glm3-6b. GLMDST demonstrates outstanding performance in experiments on mainstream Chinese DST benchmarks such as CrossWOZ and RiSAWOZ, showcasing excellent adaptability and generalization capabilities.

I. Introduction

With the rapid development of deep learning and natural language processing technologies, and the enormous commercial value embedded in real-world dialogue systems, conversational agents such as Microsoft XiaoIce, Apple Siri, Baidu Xiaodu, among others, have gradually become integrated into daily lives. These systems serve as virtual assistants, offering support for users’ specific goals, such as weather inquiries, navigation, and restaurant reservations. Dialogue State Tracking (DST), which accurately monitors changing user goals and system operations during conversations, is a crucial component of modular task-oriented dialogue systems [1] – [3]. Precise dialogue state tracking is essential for generating correct dialogue actions and appropriate natural language responses, impacting subsequent database searches and dialogue strategy selection operations in task-oriented dialogue systems. Generally, multidomain dialogue states are represented as binary group in the form of (domain-slot, value), such as <restaurant-area, east>. These slot-value pairs are extracted from the context of each turn of dialogue.