Abstract
The application of entity linking to the scientific literature in specialised domains is a challenging area, particularly in multilingual and low-resource settings where terminology exhibits severe knowledge base sparsity. The present study investigates scientific entity linking in Kazakh and Russian across four academic domains. The following subjects are covered: Information Technology, Linguistics, Medicine and Psychology. This paper introduces SCIMDIX-EL, an enriched benchmark that aligns domain-specific terms with Wikipedia and Wikimedia Commons identifiers, explicitly handling unlinkable (NIL) entities. Using this benchmark, we evaluate three methodological paradigms: zero-shot autoregressive generation with mGENRE, a hybrid web-retrieval pipeline pairing the Yandex Search API with large language models (Qwen2.5-7B and Sherkala-Chat-8B), and lightweight dictionary baselines, combined with learned abstention thresholds. In this corpus, common-prefix matching achieves a Macro-F1 of 67.72 in Kazakh, statistically indistinguishable from thresholded mGENRE at 67.81, while in Russian, mGENRE exceeds lexical matching by 14.83 percentage points. These results are consistent with a larger lexical contribution in Kazakh and a larger role for semantic disambiguation in Russian, but the language portions also differ in NIL prevalence, dictionary coverage, and mention composition. Error analysis reveals that generative and web-retrieval systems fail in non-overlapping instances, indicating a measurable headroom of 12.64 to 14.04 percentage points for future candidate reranking approaches.