A semantic linking model for Turkic languages: a fairness-constrained cross-lingual approach for low-resource settings

Abstract

We propose a semantic linking model for the Turkic language family, most members of which are considered low-resource from the computational linguistics viewpoint. The theoretical foundation of the model is Liddy’s layered scheme of natural language processing, compressed here into four levels: graphemic-morphological, syntactic, semantic, and cross-lingual. Rather than treating expert-built knowledge resources as hard constraints, we introduce them as a noisy supervision signal weighted by a per-language reliability coefficient ????????; this construction subsumes both classical retrofitting (???????? → 1) and purely distributional models (???????? → 0) as special cases of a single scheme. The optimization problem is augmented with a fairness constraint bounding the gap between the best and worst per-language quality, which prevents the model from drifting toward resource-rich languages at the expense of low-resource ones. A two-level algorithm is derived that combines Adam updates for embeddings, a closed-form orthogonal Procrustes step for cross-lingual mappings, and softmax reparametrisation for the language-weight simplex with a projected subgradient step for the dual variable; a standard ????(1/√????) bound is stated for the outer loop. The evaluation protocol includes five baselines that isolate the contribution of each model component, intrinsic metrics (Spearman correlation on SimRelUz, Precision@1 and Precision@5 for bilingual lexicon induction) and extrinsic metrics (chrF++ for downstream machine translation), with 95% bootstrap confidence intervals and Holm–Bonferroni correction for multiple comparisons.