From Traditional Taggers to LLMs: A Comparative Study of POS Tagging for Medieval Romance Languages
A new research paper published on arXiv presents a systematic empirical evaluation of Large Language Models (LLMs) for Part-of-Speech (POS) tagging in Medieval Romance languages. The study addresses the challenges posed by orthographic variation, morphological complexity, and scarce annotated resources in Medieval Occitan, Catalan, and French. Authors Matthias Schöffel and Esteban Garces Arias compare traditional rule-based and statistical taggers against modern open-source LLMs using various settings, including zero-shot and few-shot prompting, monolingual fine-tuning, and cross-lingual transfer learning. Results indicate that LLM-based approaches consistently outperform traditional methods, with fine-tuning and multilingual training providing the most significant improvements. Notably, cross-lingual transfer learning proves highly beneficial for under-resourced varieties, while targeted bilingual training can surpass broader multilingual configurations for specific languages. The findings emphasize the importance of linguistic proximity and dataset characteristics in designing transfer strategies for historical Natural Language Processing. This work offers practical guidance for deploying LLM-based pipelines in digital humanities research and releases all code, models, and datasets to ensure reproducibility.
Wire timeline
From Traditional Taggers to LLMs: A Comparative Study of POS Tagging for Medieval Romance Languages
A new research paper published on arXiv presents a systematic empirical evaluation of Large Language Models (LLMs) for Part-of-Speech (POS) tagging in Medieval Romance languages. The study addresses the challenges posed by orthographic variation, morphological complexity, and scarce annotated resources in Medieval Occitan, Catalan, and French. Authors Matthias Schöffel and Esteban Garces Arias compare traditional rule-based and statistical taggers against modern open-source LLMs using various settings, including zero-shot and few-shot prompting, monolingual fine-tuning, and cross-lingual transfer learning. Results indicate that LLM-based approaches consistently outperform traditional methods, with fine-tuning and multilingual training providing the most significant improvements. Notably, cross-lingual transfer learning proves highly beneficial for under-resourced varieties, while targeted bilingual training can surpass broader multilingual configurations for specific languages. The findings emphasize the importance of linguistic proximity and dataset characteristics in designing transfer strategies for historical Natural Language Processing. This work offers practical guidance for deploying LLM-based pipelines in digital humanities research and releases all code, models, and datasets to ensure reproducibility.
cs.AI updates on arXiv.org