Authors - Bayangali Abdygalym, Madina Sambetbayeva, Aigerim Yerimbetova, Elmira Daiyrbayeva, Duman Telman Abstract - Dense retrievers achieve strong results on lexical synonymy but degrade on hierarchical (BT/NT) and associative (RT) relations in specialized multilingual thesauri. This paper addresses typed thesaurus relation extraction by introducing a sparse–neural pipeline that combines a domain-tokenized BM25 index with morphological prefix tokens for agglutinative Kazakh and inflectional Russian, one-hop graph expansion over partially observed thesaurus structure, a KazBERT relation classifier over six labels (BT, NT, RT, SYN, LE, NONE) trained with hard-negative mining over taxonomic siblings, and an LE-mediated synonym closure that propagates synonymy through translation edges. The pipeline is evaluated on a curated trilingual military thesaurus of 5,262 canonical entries and 64,374 manually validated relation edges, using a 15% term-level holdout (711 query terms). The proposed pipeline reaches BT Hit@5 = 0.796, NT Hit@5 = 0.602, RT Hit@5 = 0.726, and SYN Hit@5 = 1.000. The results indicate that sparse lexical retrieval remains highly competitive for hierarchical thesaurus relation discovery in low-resource multilingual settings.
