Finding and Classifying Names – Challenges and Opportunities for Fine-tuned Transformers

Transformer models have become ubiquitous in Computational Linguistics due to their high performance and flexibility. However, they have mainly been applied to  modern-day texts or newspapers from the past few centuries. In addition, they have not yet been applied to onomastic tasks other than Named Entity Recognition (NER). Thus, in order to investigate the usefullness with regard to a more historical time frame on the one hand and a novel onomastic task on the other, this paper shows how the German transformer model gbert performs when fine-tuned on two different classification tasks. The first one is NER for Middle High German where an F1 score of .96 is reached with 2400 samples. The second one is a novel one concerning the classification of surnames found in Germany by their historically underlying naming motive where an F1 score of .72 is reached with 8000 samples. The latter one is also accompanied by an ablation study illustrating that an increase in training data does not always yield better performances.

Keywords: transformer models, fine-tuning, NER, text classification, onomastics