Stance Classification With Large Language Models in Estonian Written Media
Large language models, like BERT, have proved to be useful for under-resourced smaller languages by requiring less training data. We trained such large language models to identify different stances towards immigration in Ekspress Grupp and Uued Uudised Estonian news media between 2015–2022. We manually annotated immigration related sentences and used them to fine-tune large language models. The study provides a new dataset; predictions based on the models trained on that dataset, that show the changes in stances towards immigration across time in different Estonian news media; and the limits of this method.