Slovak Coreference Resolution with Pretrained Encoders

This paper investigates coreference resolution in Slovak using pretrained encoder-based language models and a hybrid approach combining linguistic rules with neural classification. We introduce SlovakCoref, a modular system evaluated on 370 annotated Slovak paragraphs containing 6,492 mentions. The proposed approach achieves a CoNLL-style F1 score of 0.808 under a custom evaluation protocol, establishing an experimental baseline for further research into Slovak natural language understanding.