The paper presents TURNA, a language model designed for the low-resource language Turkish. The model is capable of both understanding and generating natural language tasks. TURNA is pre-trained using an encoder-decoder architecture based on the unified framework UL2 and a diverse corpus curated specifically for this purpose. The model has been evaluated with three generation tasks and five understanding tasks for Turkish, showing that it outperforms several multilingual models and competes with monolingual Turkish models in understanding tasks.
Publication date: 26 Jan 2024
Project Page: hf.co/boun-tabi-LMG/turna
Paper: https://arxiv.org/pdf/2401.14373
