Across Generations: A Comparative Analysis of NER for Latin Inscriptions from Classical Machine Learning to LLMs

Abstract

Latin epigraphic texts are a challenging type of historical data for natural language processing (NLP). They are often fragmentary, contain inconsistent spelling, and follow complex Roman naming conventions. This paper investigates Named Entity Recognition (NER) for this domain by comparing several approaches, including feature-based Support Vector Machines, neural models such as BiLSTM and TreeLSTM, pre-trained language models like LatinBERT, fine-tuned Transformer models based on BERT, and large language models used with prompting and supervised fine-tuning. We introduce a manually annotated dataset of 1,000 inscriptions from the Epigraphik-Datenbank Clauss-Slaby, labelled with a fine-grained BIO scheme that captures the internal structure of Roman personal names. Results show that the fine-tuned BERT model achieves the highest performance, with a weighted F1 score of 91.1% and a macro F1 of 68.7%, and clearly outperforms other methods. Additional linguistic features, such as part-of-speech tags and dependency information, yield only limited improvements, likely due to the irregular nature of inscriptional texts. This work provides a new benchmark for NER on Latin inscriptions and offers practical insights into applying modern NLP techniques to historical, non-standardised language.

Resources

Details