Read the full paper "Cross-Lingual Definition Generation from an mT5" included as a PDF in this repo.
Recent progress in definition modeling calls for an expansion into new territory. In this paper, we explore the little-studied prospect of cross-lingual definition generation from a resource-scarce perspective. We show that a small pretrained multilingual Text-to-Text Transfer Transformer (mT5) model can be transformed into a language-agnostic zero-shot definition generator, producing rudimentary English definitions for terms in multiple foreign languages. Throughout the project, several task-specific modifications to the model are devised and tested. We further recommend research paths that may progress the field of cross-lingual definition generation.
Data should be accessed via the Hugging Face datasets library unless the desired dataset is already saved to disk.
Change the working directory to XLdefgen/model, and then run run_model.py as indicated in the model README.