A Prompt-Guided and Quality-Aware Robust Text-Audio Intent Recognition Framework for Elderly Care

In natural language understanding, intent recognition plays a central role in human-computer interaction. However, in elderly-care scenarios, acoustic signals are often affected by atypical speech patterns, slower speaking rates, and environmental noise, making audio information less reliable and reducing the effectiveness of conventional text-audio fusion methods. To address this problem, we propose a prompt-guided and quality-aware text-audio intent recognition framework. Specifically, a χ²-based intent prototype soft prompt is introduced to enhance the semantic representation of text. Then, a residual-free text-guided cross-attention module is designed to refine degraded acoustic features using textual semantics as reliable guidance. In addition, a dynamic fusion gate is developed to adjust the contributions of text and audio based on modality reliability and intent-related information. Experiments on the MIntRec dataset with simulated acoustic degradation show that the proposed model achieves 60.90% accuracy, 60.80% weighted F1, and 57.90% macro-F1, outperforming several competitive baselines. These results indicate that the proposed framework can improve the robustness of intent recognition under challenging acoustic conditions in elderly-oriented interaction scenarios.