LitLat BERT is a trilingual model, using xlm-roberta-base architecture, trained on Lithuanian, Latvian, and English corpora. Focusing on three languages, the model performs better than multilingual BERT, while still offering an option for cross-lingual knowledge transfer, which a monolingual model wouldn't.

We compare LitLat BERT with multilingual BERT (mBERT), XLM-RoBERTa (XLM-R) and monolingual Latvian BERT (LVBERT) (Znotins and Barzdins, 2020). The report the results as a macro F1 score of 3 named entity classes shared in all three datasets: person, location, organization.