Transformer models, such as BERT, are often taken off-the-shelf and then fine-tuned on a downstream task. Although this is sufficient for many tasks, low-resource settings require special attention. We demonstrate an approach of performing an extra stage of self-supervised task-adaptive pre-training to a number of Croatian-supporting Transformer models. In particular, we focus on approaches to language, domain, and task adaptation. The task in question is targeted sentiment analysis for Croatian news headlines. We produce new state-of-the-art results (F1 = 0.781), but the highest performing model still struggles with irony and implicature. Overall, we find that task-adaptive pre-training benefits massively multilingual models but not Croatian-dominant models.

Transfer learning is one of the prevailing approaches towards training language-specific BERT models. However, some languages have uncommon features that may prove to be challenging to more domain-general models but not domain-specific models. Comparing the performance of BERTić, a Bosnian-Croatian-Montenegrin-Serbian model, and Multilingual BERT on a Named-Entity Recognition (NER) task and Masked Language Modelling (MLM) task based around a rare phenomenon of indeclinable female foreign names in Serbian reveals how the different training approaches impacts their performance. Multilingual BERT is shown to perform better than BERTić in the NER task, but BERTić greatly exceeds in the MLM task. Thus, there are applications both for domain-general training and domain-specific training depending on the tasks at hand.