Abstract
With the rapid digitization of Classical Latin texts and the recent breakthroughs in Large Language Models (LLMs), this paper contributes to ancient language research by participating in the EvaLatin 2026 shared task. We present Team uOttawa's system description and results for the Named Entity Recognition (NER) challenge. The task comprises two subtasks: coarse-grained NER with 11 classes and fine-grained NER with 28 classes, each evaluated under strict and fuzzy matching regimes. By leveraging prompt engineering on commercial LLMs such as Gemini 2.5 Pro and Claude Sonnet 4.5, we demonstrate that the underrepresented ancient Latin language can benefit significantly from cross-lingual transfer learning, capitalizing on advances made by the broader LLM development community. Our methods achieve outstanding results, securing first place in both NER subtasks and attaining the highest scores across all evaluation metrics and regimes among all submissions.
Introduction
The growing availability of digitized Classical Latin resources, combined with the unprecedented capabilities of Large Language Models (LLMs), opens new avenues for computational analysis of ancient languages. However, such languages remain underrepresented in many NLP resources and models. This work addresses this gap by participating in EvaLatin 2026, a shared task focused on Natural Language Processing for ancient Latin. Specifically, we tackle the Named Entity Recognition (NER) task, which involves identifying and classifying entities such as persons, locations, and organizations within Latin texts.
The task is structured into two subtasks:
- Coarse-grained NER: 11 entity classes (e.g., person, place, group)
- Fine-grained NER: 28 entity classes (e.g., specific subtypes like "person:mythological" or "place:geographical")
Each subtask is evaluated under two regimes: strict (exact boundary matching) and fuzzy (partial credit for overlapping spans).
Methodology
We approach the NER task using transfer learning via large language models, specifically through prompt engineering. Our system relies on commercial LLMs—Gemini 2.5 Pro and Claude Sonnet 4.5—which have demonstrated strong performance across various NLP tasks, including low-resource and historical languages. By crafting task-specific prompts, we adapt these general-purpose models to perform NER on Classical Latin without requiring fine-tuning on large annotated datasets.
Our prompt design includes:
- Clear instructions with examples for each subtask
- Format specifications to ensure consistent output structure
- Contextual cues to handle the linguistic nuances of Classical Latin (e.g., inflections, flexible word order)
We also employ ensemble techniques, combining predictions from both models to improve robustness and accuracy.
Results
Our system achieved top performance in both subtasks of the EvaLatin 2026 NER shared task. Specifically, we recorded the highest F1 scores under both strict and fuzzy evaluation regimes. The success underscores the effectiveness of LLM prompting as a transfer learning strategy for low-resource ancient languages.
Key results:
- Coarse-grained NER: First place across all metrics
- Fine-grained NER: First place across all metrics
- Overall: Best scores among all submissions for both subtasks and both evaluation regimes
Discussion
The strong results demonstrate that commercial LLMs, when properly prompted, can serve as powerful tools for ancient language processing. This is particularly valuable for Latin, which lacks the abundant labeled data available for modern languages. Cross-lingual transfer learning enables the models to leverage their extensive training on modern languages and apply that knowledge to the target domain.
Furthermore, the use of prompt engineering avoids the need for extensive computational resources, making it an accessible approach for researchers in digital humanities and computational classics.
Conclusion
This paper presents a successful application of LLM prompting for Named Entity Recognition in Classical Latin, achieving state-of-the-art results in the EvaLatin 2026 shared task. Our findings suggest that transfer learning from modern-language LLMs is a viable and effective strategy for underresourced ancient languages, paving the way for further NLP research in this domain.
Acknowledgments
We thank the organizers of EvaLatin 2026 and the LT4HALA workshop for providing the evaluation framework and datasets. We also acknowledge the support of the University of Ottawa.
References
- arXiv:2608.04015 [cs.CL], submitted on 26 May 2026.
- EvaLatin (LT4HALA@LREC), ELRA, May 2026, Palma De Majorque, Spain.
via ArXiv CL+LG