A Lightweight Unified Generative Framework with Prompting for Chinese and English Spoken Language Understanding

Spoken Language Understanding (SLU) encompasses two subtasks: Intent Detection (ID) and Slot Filling (SF), aiming to achieve semantic analysis of user dialogue. Currently, Chinese and English hold crucial positions in the research and application of SLU systems. However, due to significant differences in linguistic structure and grammar between Chinese and English, existing SLU methods often struggle to efficiently handle SLU tasks in both languages simultaneously. To address this issue, we propose a Lightweight Unified Generative Framework with prompting for Chinese and English SLU called LightM-UG. This framework transforms the ID and SF tasks into text generation tasks. It includes a shared semantic guidance mechanism to efficiently acquire semantic features between intents and slots. Additionally, we design a lightweight strategy to finely adjust the model to meet the specific vocabulary requirements of Chinese and English SLU tasks, effectively reducing the model's parameters and achieving lightweight model construction. Our experiments and ablation studies on three Chinese SLU datasets (CAIS, SMP2019, SMP2020) and two English multi-intent SLU datasets (MixSNIPS, MixATIS) demonstrate that our method improves overall accuracy by 0.15-4.2 % in full-data scenarios and by 1.2-2.5 % in few-shot settings, showcasing outstanding few-shot learning capabilities.