Developing specialised language models for low-resource domains typically involves a trade-off between two specialisation strategies: adapting a general-purpose model through continued pretraining or retraining a model from scratch. While adapting preserves the model’s linguistic knowledge, retraining benefits from the flexibility of an in-domain tokeniser – a potentially significant advantage when handling rare languages. This study investigates the impact of tokenisation, specialisation strategy, and pretraining data availability using classical scholarship – a multilingual, code-switching and highly domain-specific field – as a case study. Through extensive experiments, we assess whether domain-specific tokenisation improves model performance, whether character-based models provide a viable alternative to subword-based models, and which specialisation strategy is optimal given the constraints of limited pretraining data. Contrary to prior findings, our results show that in-domain tokenisation does not necessarily enhance performance. Most notably, adaptation consistently outperforms retraining, even with limited data, confirming its efficiency as the preferred strategy for resource-constrained domains. These insights provide valuable guidelines for developing specialised models in fields with limited textual resources.

The quality of automatic transcription of heritage documents, whether from printed, manuscripts or audio sources, has a decisive impact on the ability to search and process historical texts. Although significant progress has been made in text recognition (OCR, HTR, ASR), textual materials derived from library and archive collections remain largely erroneous and noisy. Effective post-transcription correction methods are therefore necessary and have been intensively researched for many years. As large language models (LLMs) have recently shown exceptional performances in a variety of text-related tasks, we investigate their ability to amend poor historical transcriptions. We evaluate fourteen foundation language models against various post-correction benchmarks comprising different languages, time periods and document types, as well as different transcription quality and origins. We compare the performance of different model sizes and different prompts of increasing complexity in zero and few-shot settings. Our evaluation shows that LLMs are anything but efficient at this task. Quantitative and qualitative analyses of results allow us to share valuable insights for future work on post-correcting historical texts with LLMs.