For the first two tasks, we fine-tuned two RoBERTa and XLM-RoBERTa models for (predominantly) English and multilingual datasets, respectively. Gururangan et al. (2020) show that continuing pre-training language models on task-relevant unlabeled data might contribute to improve the performance of final fine-tuned task-specific models-in particular, in low-resource situations. Considering the fact that the affiliation strings' grammar has its own structure, which is different from the one that would be expected to be found in free natural language, we explore whether our affiliation span identification and NER models would benefit from being fine-tuned from models that have been further pre-trained on raw affiliation strings for the masked token prediction task.

We adatap models to 10 million random raw affiliation strings from OpenAlex, reporting perplexity on 50k randomly held-out affiliation strings. In what follows, we refer to our adapted models as AffilRoBERTa (adapted RoBERTa model) and AffilXLM (adapted XLM-RoBERTa).

Specific details of the adaptive pre-training procedure can be found in Duran-Silva et al. (2024).

We report masked language modeling loss as perplexity measure (PPL) on 50k randomly sampled held-out raw affiliation strings.

AffilGood-AffilRoBERTa achieves competitive performance to 2 tasks in processing affiliation strings, compared to base models