Enhancing intent recognition in conversational human–computer interaction using LoRA-adapted Qwen2.5-7B

Abstract

Intent recognition is a fundamental task in conversational human‑computer interaction (HCI) systems, enabling machines to understand user goals from natural language utterances. Most existing approaches rely on small pretrained encoders that struggle with the long‑tail distribution of intent classes. This paper investigates the use of Qwen2.5‑7B, a 7.6‑billion‑parameter large language model, to perform intent recognition on two widely used HCI benchmarks: the Airline Travel Information System (ATIS) and the Snips Natural Language Understanding (SNIPS) datasets. We employ low‑rank adaptation (LoRA) to inject lightweight trainable matrices into the Transformer layers and noisy embedding fine‑tuning (NEFTune) to improve generalisation. Experiments show that the LoRA‑adapted Qwen2.5‑7B model achieves state‑of‑the‑art accuracy on both benchmarks, and that adding noise during fine‑tuning further boosts accuracy and macro‑F1 scores.