Post Editing English to Hindi Translation with Fine Tuned Qwen Models
Year: 2025
Type: conference-paper
Abstract: English-to-Hindi machine translation (MT) poses significant challenges due to structural, morphological, and se-mantic complexities between the two languages. This paper introduces a novel automatic post-editing (APE) framework leveraging the Qwen-3 large language model to enhance translation quality. The proposed methodology integrates three key components: (i) data augmentation through contextual synonym replacement to enrich linguistic diversity, (ii) prompt engineering tailored to encapsulate complex linguistic and semantic post-editing instructions, and (iii) parameter-efficient fine-tuning using QLoRA to adapt Qwen-3 for the APE task without excessive computational overhead. The system was evaluated on the multi-domain EnIndic dataset, consisting of MT outputs and human post-edited references. The experimental results show a significant improvement in the performance with BLEU 61.48, ChrF++ 93.87, TER 23.09, WER 1.3857, which is higher as compared to baseline results of BLEU 2.16, ChrF++ 2.49, TER 116.14, and WER 0.7311. The results highpoint the helpfulness of our outline in generating linguistically accurate and semantically reliable translations. However, there are silent limitations to generalizing the process to low-resource language pairs and defining its applicability to other datasets and domains. These questions will proposal future re-search opportunities to the NLP community, as they will provide a casual to progress more powerful multilingual post-editing systems.
Authors: Priyanka Suram, Pramoda Patro
Institution: SR University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7167019286