From Traditional Taggers to LLMs: A Comparative Study of POS Tagging for Medieval Romance Languages

Researchers conducted a systematic evaluation of large language models for part-of-speech tagging in Medieval Romance languages, comparing them against traditional taggers. The study demonstrates that LLM-based approaches with fine-tuning and cross-lingual transfer learning significantly outperform conventional methods, offering practical applications for digital humanities research on historical texts.

This research addresses a specialized but meaningful challenge in natural language processing: applying modern AI techniques to historically fragmented texts where orthographic inconsistency and sparse training data create substantial obstacles. The study evaluates how contemporary LLMs handle Medieval Occitan, Catalan, and French—languages with limited digital resources and high morphological complexity—against older statistical and rule-based approaches that dominated NLP for decades.

The findings emerge within a broader trend of LLMs demonstrating unexpected versatility across linguistic domains far removed from contemporary internet text. Traditional POS taggers, built on fixed rule sets or trained on modern corpora, struggle with medieval texts because language evolves significantly over centuries. LLMs, trained on diverse data and capable of few-shot adaptation, bridge this gap more effectively than previous generations of NLP tools.

For the digital humanities and historical linguistics communities, these results validate deploying modern neural architectures to unlock insights from manuscript collections and historical corpora. The emphasis on cross-lingual transfer learning proves particularly valuable because medieval languages share etymological roots and grammatical structures, allowing knowledge from resource-rich varieties to improve tagging accuracy for under-resourced ones. The research demonstrates that proximity between languages matters more than sheer data volume—a finding that suggests careful transfer strategy design yields better outcomes than brute-force multilingual training.

Looking forward, this work establishes a replicable framework for applying LLMs to other low-resource historical languages and scripts. The public release of code and models enables broader adoption in academia, potentially accelerating digitization projects and computational analysis of medieval texts. This methodology could extend to other linguistic tasks—parsing, named entity recognition, semantic annotation—multiplying its impact across digital humanities research.

• →LLM-based POS tagging substantially outperforms traditional rule-based and statistical taggers on medieval Romance language texts.
• →Cross-lingual transfer learning effectively improves tagging accuracy for under-resourced medieval language varieties.
• →Linguistic proximity between source and target languages proves more important than dataset size in designing transfer learning strategies.
• →Fine-tuning LLMs on domain-specific historical texts yields larger accuracy improvements than zero-shot or few-shot prompting alone.
• →Released code and models enable reproducible research and practical deployment in digital humanities projects studying historical texts.