A Multitask XLM-R-Based Hybrid Approach for Named Entity Recognition, Coreference Resolution, and Genealogical Relation Extraction from Uzbek Historical Documents

Automatically extracting genealogical information from historical documents in the Uzbek language requires identifying personal names, linking various expressions referring to the same individual, and determining kinship relations between persons. However, this process is complicated by factors such as variations in the spelling of personal names, the presence of archaic vocabulary, and the distribution of information regarding a single individual across different parts of the text. This study proposes a multi-task hybrid approach based on the XLM-RoBERTa-large model to solve these problems. In the proposed system, named entities are identified using BiLSTM–CRF, expressions referring to the same individual are linked via a coreference model, and genealogical relations are determined using a combination of a neural model and mechanisms based on kinship rules. The approach was evaluated on the ArchiveGene-Real dataset of 171 authentic Uzbek historical documents; every configuration was retrained with five random seeds, and differences were tested with a paired document-level bootstrap and Holm correction. The full system reached F1 = 0.624 ± 0.021 for named entity recognition, CoNLL F1 = 0.570 ± 0.010 for coreference, and F1 = 0.441 ± 0.026 for genealogical relation extraction. With gold person mentions, coreference rose to 0.802 ± 0.009 and relation F1 to 0.509 ± 0.029; with gold mentions and gold coreference, the kinship grammar reached 0.627, which shows that most of the remaining error arises upstream of the relation stage. Neither the neural relation head (0.242) nor lexical rules alone (0.150) came close to the full system, and disabling coreference, entity refinement or cluster splitting lowered relation F1 significantly. All retrained components (domain adaptation, BiLSTM, CRF, cluster attention, and the auxiliary mention and relation losses) improved NER significantly. A single-task NER model scored higher than the multi-task model (0.669 vs. 0.624; p = 0.125), but replacing the NER stage with the strongest standalone encoder left relation F1 unchanged (0.440). On 13 external documents, NER and coreference transferred well, whereas relation F1 fell to 0.186; this gap was traced to the coverage of the kinship grammar rather than to the neural components. The results show that combining neural document-level modelling with explicit kinship knowledge is an effective and transparent way to extract genealogical information from historical documents. The main limitations are the small size of the corpus and of the external set, the use of manually corrected transcriptions (the OCR error rate was not measured), and the restriction of mentions to named references.