Geographically‑Informed Multilingual Neural Machine Translation

This work introduces the approach of integrating geographic coordinates into a multilingual neural machine translation architecture, alongside special tokens (linguistic tags). The approach enables modeling of language continua and hypothetical language varieties through geospatial interpolation across supported languages. We fine-tuned a Transformer model on a custom dataset of 31 languages annotated with geographic vectors and three types of tags (family, group, script), enabling the model to condition translations on spatial and linguistic features. Our experiments demonstrate that geographic embeddings encourage more coherent language clustering in the model’s latent space, facilitating smoother interpolation between mother than two related languages (e.g., across the Germanic or Slavic continua). Additionally, the model exhibits capabilities, such as performing partial transliteration between scripts. However, given the amount of data and training used, the model's capabilities are insufficient for generating non-existent hypothetical language varieties under unusual conditions (such as Balkan Germanic).