Classificatio Sine Iactu – That Is, Zero-Shot NERC in Latin

Abstract

This paper presents a zero-shot approach to Named-Entity Recognition and Classification (NERC) in Latin, applied to the EvaLatin shared task. Given the novelty and granularity of the annotation guidelines, which preclude the use of existing annotated resources, we employ the zero-shot model GLiNER2, a general information extraction system capable of CPU-efficient inference, within a cross-lingual pipeline. Latin texts are first translated into English via the Google Translate API, processed by the model, and the resulting annotations are aligned back to the original Latin using word-alignment techniques. Rule-based post-processing addresses labelling inconsistencies and low-confidence predictions. We evaluate two model variants, a large monolingual and a multilingual model, under both strict and fuzzy evaluation. The large model delivers the best results for the coarse-grained task (F1: 0.590 fuzzy), while the multilingual model outperforms it on the fine-grained task (F1: 0.432 fuzzy). Results indicate that multilingual embeddings confer an advantage for fine-grained semantic distinctions, that English embeddings introduce systematic bias in cross-lingual transfer, and that zero-shot NER represents a viable, reproducible baseline for low-resource historical languages. Fine-tuning on guideline-compliant annotated data remains a priority for future work.

Resources

Details