Dialogue State Tracking Using Large Language Models

Published:

*Data Science Capstone Research at NYU Shanghai by Bale Chen, Peiyang Wu, and Xiaocheng Yang, supervised by Prof. Wilson Tam. *

This project explores the task of Dialogue State Tracking (DST) in task-oriented dialogue systems. DST involves understanding and tracking user intentions in multi-turn dialogues, typically represented as slot-value pairs. The study reviews various DST methods, highlighting limitations in classification-based and span extraction-based approaches. The paper adopts a generative approach, leveraging large language models (LLMs) to improve performance. Experimenting with different formulations, model scaling, and data scaling, the study demonstrates the feasibility of using LLMs for DST, achieving a new state-of-the-art on MultiWOZ 2.2 and 2.4 benchmarks. The slot-level question-answering formulation is shown to enhance LLM performance, reducing hallucination. Additionally, the paper provides insights into the influence of model scaling and data scaling on DST performance, contributing to a comprehensive understanding of the subject.

[Code] [Report]