Industrial Automatic Control Systems and Controllers Annotation << Back Method of Detecting Design Patterns Using a Language Model on a Transformer Architecture and Contrastive Learning Jameleh Asaad Asaad, Аvksentieva E.Yu. This study proposes a method for detecting software design patterns using transformer-based language models, contrastive learning, and threshold-based classification. The dataset is derived from DPDAtt corpus, containing 1,645 labeled Java fi les covering 13 GoF design patterns. CodeT5 generates contextualized embeddings that capture syntactic and semantic properties of code. Contrastive learning trains these embeddings, bringing similar patterns closer while separating different ones, with early stopping to prevent overfitting. Four methodological variations were explored. Code 1 used a simple projection network with on-the-fl y pair generation. Code 2 introduced precomputed pairs and threshold-based classification. Code 3 grouped patterns into Behavioral, Creational, Structural, and Unknown categories. Code 4 combined balanced pair generation, early stopping, thresholding, and categorical grouping, achieving the best results: 87.1 % accuracy and a macro F1-score of 81.7 %. The approach reliably classifies major design patterns while handling unknown or ambiguous patterns effectively. Transformer embeddings provide rich contextual representations, contrastive learning ensures meaningful separation, and thresholding identifies rare patterns. Overall, this methodology enhances both accuracy and robustness in software design pattern detection. Future work may improve handling of rare patterns using semi-supervised learning or adaptive thresholds. Keywords: design pattern detection, language models, transformer architecture, contrastive learning, automated code analysis, machine learning. DOI: 10.25791/asu.10.2025.1616 Pp. 20-23.