This paper presents elaborate testing of various LLMs on their generalization capacities. We finetune six encoder models that have been pretrained with very different data (varying in size, language, and period) on a challenging event detection task in Early Modern Dutch archival texts. Each model is finetuned with 5 seeds on 15 different data splits, resulting in 450 finetuned models. We also pre-train a domain-specific Language Model on the target domain and fine-tune and evaluate it in the same way to provide an upper bound. Our experimental setup allows us to look at underresearched aspects of generalizability, namely i) shifts at multiple places in a modeling pipeline, ii) temporal and crosslingual shifts and iii) generalization over different initializations. The results show that none of the models reaches domain-specific model performance, demonstrating their incapacity to generalize. mBERT reaches highest F1 performance, and is relatively stable over different seeds and datasplits, contrary to XLM-R. We find that contemporary Dutch models do not generalize well to Early Modern Dutch as they underperform compared to crosslingual as well as historical models. We conclude that encoder LLMs lack temporal generalization capacities and that bigger models are not better, since even a model pre-trained with five hundred GPUs on 2.5 terabytes of training data (XLM-R) underperforms considerably compared to our domain-specific model, pre-trained on one GPU and 6 GB of data. All our code, data, and the domain-specific model are openly available.

This paper discusses the re-usibility of existing approaches, tools and automatic techniques for the annotation and detection of events in a challenging variant of centuries old Dutch written in the archives of the Dutch East India Company. We describe our annotation process and provide a thorough analysis of different versions of manually annotated data and the first automatic results from two fine-tuned Language Models. Through the analysis of this complete process, the paper studies two things: to what extent we can use NLP theories and tasks formulated for modern English to formulate an annotation task for Early Modern Dutch and to what extent we can use NLP models and tools built for modern Dutch (and other languages) on Early Modern Dutch. We believe these analyses give us insight into how to deal with the large variation language showcases in describing events, and how this variation may differ accross domains. We release the annotation guidelines, annotated data, and code.

Neural Network (NN) architectures are used more and more to model large amounts of data, such as text data available online. Transformer-based NN architectures have shown to be very useful for language modelling. Although many researchers study how such Language Models (LMs) work, not much attention has been paid to the privacy risks of training LMs on large amounts of data and publishing them online. This paper presents a new method for anonymizing a language model by presenting the way in which MedRoBERTa.nl, a Dutch language model for hospital notes, was anonymized. The two-step method involves i) automatic anonymization of the training data and ii) semi-automatic anonymization of the LM’s vocabulary. Adopting the fill-mask task where the model predicts what tokens are most probable in a certain context, it was tested how often the model will predict a name in a context where a name should be. It was shown that it predicts a name-like token 0.2% of the time. Any name-like token that was predicted was never the name originally present in the training data. By explaining how a LM trained on highly private real-world medical data can be published, we hope that more language resources will be published openly and responsibly so the scientific community can profit from them.

Electronic Health Records contain a lot of information in natural language that is not expressed in the structured clinical data. Especially in the case of new diseases such as COVID-19, this information is crucial to get a better understanding of patient recovery patterns and factors that may play a role in it. However, the language in these records is very different from standard language and generic natural language processing tools cannot easily be applied out-of-the-box. In this paper, we present a fine-tuned Dutch language model specifically developed for the language in these health records that can determine the functional level of patients according to a standard coding framework from the World Health Organization. We provide evidence that our classification performs at a sufficient level to generate patient recovery patterns that can be used in the future to analyse factors that contribute to the rehabilitation of COVID-19 patients and to predict individual patient recovery of functioning.