Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Small Open Models Achieve Near Parity with Large Models in Low Resource Literary Translation at a Fraction of the Cost

Created by
  • Haebom

Author

Mihai Nadas, Laura Diosan, Andreea Tomescu, Andrei Piscoran

Outline

This paper presents TINYFABULIST TRANSLATION FRAMEWORK (TF2), an integrated framework for literary translation in the low-resource language Romanian. TF2 is an integrated framework for dataset generation, fine-tuning, and evaluation centered on generating and releasing a compressed fine-tuned language model (TF2-12B) and large-scale synthetic parallel datasets (DS-TF2-EN-RO-3M and DS-TF2-EN-RO-15K). Based on an existing large-scale synthetic English fable dataset (DS-TF1-EN-3M), we generate 15,000 high-quality Romanian reference data items and fine-tune the model using directive fine-tuning and adapter compression on a 12 billion-parameter open-weighted model. Evaluation is performed by combining corpus-level BLEU and a five-dimensional LLM-based evaluation metric (accuracy, fluency, coherence, style, and cultural adaptation). Experimental results show that the fine-tuned model achieves fluency and relevance comparable to the best-performing large-scale proprietary models, while remaining open-source, accessible, and cost-effective. The model, dataset, script, and evaluation prompts are all publicly available.

Takeaways, Limitations

•
Takeaways:
◦
Providing an efficient and reproducible pipeline for literary translation in low-resource languages.
◦
Utilizing an open model, we present the potential for widespread adoption of culturally significant literary content translations from low-resource languages.
◦
Enabling research by releasing high-quality, large-scale synthetic datasets.
◦
Development of a lightweight model that rivals the performance of large-scale proprietary models.
◦
Validating the effectiveness of directive fine-tuning and adapter compression techniques.
•
Limitations:
◦
Because it relies on synthetic data, it may not fully reflect the complexity of real-world literary translation.
◦
Since the evaluation scale is LLM-based, limitations of LLM may affect the evaluation results.
◦
Currently limited to English-Romanian translations, generalizability to other language pairs requires further research.
◦
A 12 billion parameter model still requires significant resources, so development of a more lightweight model may be necessary.
👍
Made with Slashpage