Fine-tuning multilingual sentence transformers for low-resource skill-concept matching across Kazakh, Russian, and English

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Introduction Matching short, specialized skill expressions across English, Russian, and Kazakh is challenging because general-domain multilingual encoders underperform on terse, code-mixed, domain-specific phrases, particularly in the low-resource Kazakh setting. Methods We fine-tuned a multilingual Sentence Transformer using a staged Multiple Negatives Ranking objective on a trilingual paraphrase corpus, including Russian augmentation pairs. We evaluated skill matching and semantic similarity across languages and assessed the resulting embeddings through downstream skill-taxonomy clustering. Results Fine-tuning preserved English performance while improving Russian and Kazakh similarity quality. Russian cosine Pearson correlation increased from 0.8125 to 0.8221, while Kazakh cosine Pearson increased from 0.5989 to 0.6050 and Kazakh dot-product Pearson from 0.4487 to 0.4912. Agglomerative clustering improved mean silhouette from 0.27126 to 0.2851 and reduced erroneous clusters from 19.74% to 13.97%. Discussion The results provide evidence consistent with cross-lingual transfer as an important mechanism of improvement for Kazakh. They also motivate language-specific threshold calibration and demonstrate that intrinsic similarity improvements translate into a cleaner downstream skill taxonomy.