Enhancing the Quality of Clustering Job Skills Through the Fine-Tuning of the Sentence Transformer Model

This article examines the application of sentence embedding models for clustering skills collected from the online recruitment site hh.kz. Skill clustering can be used to create a skill community within the job vacancy knowledge graph. The main objective of the research is to improve the quality of skill clustering by fine-tuning sentence transformer models on domain-specific data. The experiments conducted showed significant improvements in the model's performance metrics in Russian, including Cosine-Similarity, Manhattan-Distance, Euclidean-Distance, and Dot-Product-Similarity. Applying agglomerative clustering to skill embeddings before and after model fine-tuning demonstrated an enhancement in cluster quality, as evidenced by an increased average silhouette coefficient and a decreased percentage of erroneous clusters. The article also discusses the limitations of the conducted research and suggests directions for future development.