Research on the similarity calculation of short text in the terminology domain based on siamese BERT model

This study focuses on the task of term-domain short-text similarity computation. It addresses two main challenges: dataset scarcity and insufficient deep semantic extraction. To solve these issues, we first construct the multi-domain Terminology Definition Similarity (TDS) dataset using an automated pipeline. This pipeline combines data generation (based on the GPT-4 model) with a manual verification process involving expert quality control. The design ensures the production of high-quality data. We then present an innovative model named SDQKC (Sbert + dynamic QK + contrast). The model optimizes the Siamese BERT network through a dynamic QK co-attention mechanism and enhances its deep-level semantic understanding by incorporating contrastive learning. Experimental results show that the SDQKC model achieves Pearson correlation coefficients of 0.69384 on the BQ Corpus (Bank Question Corpus) and 0.69511 on the TDS dataset. These results significantly outperform other baseline models, demonstrating the effectiveness and superiority of the proposed methodology.