FSAM: A Flexible Short Similarity Analysis Model for Chinese Medical Short Texts

Automated semantic similarity analysis of medical short texts is crucial for applications such as clinical decision support and intelligent medical consultation, yet it remains challenging due to the prevalence of specialized terminology, complex contextual semantics and the trade-of between accuracy and inference speed in existing models. To address this, we propose FSAM, a flexible similarity analysis model for Chinese medical short texts. The core innovation of FSAM lies in its CTF module, which integrates part-of-speech masking, a medical knowledge-enhanced similarity matrix, and an interactive attention mechanism to achieve multi-granularity semantic representation and alignment. When evaluated on two medical-domain benchmarks CHIP-STS and CQSJD, FSAM achieves state-of-the-art accuracy of 86.45% and 92.46%, respectively, while maintaining significantly lower computational overhead than competitive interactive models. The model’s strong generalization capability is further demonstrated on three public datasets from general and insurance domains, where it consistently delivers competitive performance. Ablation studies con- firm the substantial contribution of each component within the CTF module. These results collectively indicate that FSAM efectively balances accuracy and inference speed in medical short text analysis and generalizes well to out-of- domain datasets, highlighting its promising potential for practical deployment in healthcare NLP systems.