Structure-Preserving Document Translation via Multi-Stage LLM Pipeline: A Case Study in Marathi

A new translation framework designed for Indian government documents preserves layout and formatting when converting Marathi-language PDFs into English, addressing a gap left by conventional text-only machine translation systems. The framework, detailed in a paper submitted on 27 Jun 2026, integrates layout-aware optical character recognition, coordinate-based text extraction, large language model based translation, and structured document reconstruction through HTML representations [1]. Government documents in India are predominantly issued in regional languages such as Marathi, creating substantial accessibility barriers for non-native readers, interstate administrative bodies, and policy analysts [3]. Existing translation systems largely neglect document structure, formatting integrity, and domain-specific terminology, limiting their usefulness for official documentation [1]. The system uses Chandra OCR, a model available on Hugging Face, to extract both textual content and layout information [3]. Unlike traditional OCR systems that return plain text, this model produces structured HTML containing positional metadata and document hierarchy [3]. By enforcing spatial alignment constraints and preserving hierarchical document elements, the framework ensures structural consistency between the source and translated documents [4]. Experimental evaluation on real-world Marathi government PDFs demonstrated improved structural preservation, translation coherence, and terminological consistency compared to conventional text-only translation pipelines [1]. The results indicate that preserving spatial metadata significantly improves document usability [4]. The paper treats document translation as a combined structural and linguistic transformation problem, preserving both meaning and formatting [3]. India's broader artificial intelligence landscape provides context for such work. The AI market in India is projected to reach $8 billion by 2025, growing at a 40% compound annual growth rate from 2020 to 2025 [2]. Government initiatives such as NITI Aayog's 2018 National Strategy for Artificial Intelligence have supported AI development across sectors including healthcare, finance, and education [2]. The challenge of translating Indian-language legal and government documents has drawn attention from multiple research groups. A separate 2025 study explored handwritten legal document translation for Marathi, evaluating both traditional OCR-machine translation pipelines and vision large language models for digitizing records such as FIRs, charge sheets, and witness statements in India's district and high courts [5]. The proposed framework contributes toward scalable multilingual accessibility solutions for e-governance and administrative document processing [1]. The authors present the system as a way to address the limitations of conventional text-only translation systems through a multi-stage pipeline that maintains layout fidelity during end-to-end document transformation [4].