LHPF: An LLM-Based Hierarchical Pipeline Framework for Spoken Language Understanding
Xudong Zhu, Yuanping Chen, Xiaohui Rong · 2024
In this study, we propose a hierarchical task learning framework based on Large Language Models (LLMs) to improve the performance in spoken language understanding (SLU). Our approach introduces an Alternative Entity Extract Model designed to assist LLMs in accurately extracting entities and locations from natural language utterances. Additionally, we develop an LLM-Based Hierarchical Pipeline Framework that integrates a small-scale model and an LLM in a pipeline, where the small model generates intermediate results, and the LLM refines the final output through prompts and intermediate results. By fine-tuning the LLM efficiently, we adapt it to downstream tasks, ensuring better generalization across both single-domain and cross-domain scenarios. Extensive experiments and quantitative analysis demonstrate that our proposed method not only excels in single-domain SLU tasks but also achieves robust performance in cross-domain settings. We achieve unification of the models on both single-domain and cross-domain problems