13.12.2023

New Paper by our former Scientific Coordinator Dr. Marisa Köllner and former research fellow Dr. Christin Beck - published in the International Workshop on Computational Approaches to Historical Language Change (LChange’23) co-located with EMNLP 2023

GHisBERT – Training BERT from scratch for lexical semantic investigations across historical German language stages

Abstract:

While static embeddings have dominated com- putational approaches to lexical semantic change for quite some time, recent approaches try to leverage the contextualized embeddings generated by the language model BERT for identifying semantic shifts in historical texts. However, despite their usability for detecting changes in the more recent past, it remains un- clear how well language models scale to inves- tigations going back further in time, where the language differs substantially from the training data underlying the models. In this paper, we present GHisBERT, a BERT-based language model trained from scratch on historical data covering all attested stages of German (going back to Old High German, c. 750 CE). Given a lack of ground truth data for investigating lexical semantic change across historical Ger- man language stages, we evaluate our model via a lexical similarity analysis of ten stable concepts. We show that, in comparison with an unmodified and a fine-tuned German BERT- base model, our model performs best in terms of assessing inter-concept similarity as well as intra-concept similarity over time. This in turn argues for the necessity of pre-training historical language models from scratch when working with historical linguistic data.

Here is the link to the full article including the link to the supplementary material: https://aclanthology.org/2023.lchange-1.4/