Digitizing Nepal's Written Heritage: A Comprehensive HTR Pipeline for Old Nepali Manuscripts

2025-12-18

AI Summary

This study addresses the challenge of digitizing endangered historical manuscripts in resource-scarce Old Nepali. We propose the first end-to-end handwritten text recognition (HTR) pipeline for this language, designed to support large-scale archival preservation. Methodologically, we employ a CNN-Transformer/LSTM encoder-decoder architecture, complemented by three novel components: (i) a script-adapted data augmentation strategy tailored to low-resource historical scripts; (ii) a lexicon-guided decoding mechanism to improve lexical accuracy; and (iii) token-level confusion analysis for fine-grained error diagnostics. Our contributions are threefold: (1) the first complete line-level HTR pipeline for Old Nepali historical documents; (2) open-sourcing of all training code, model configurations, and evaluation scripts; and (3) achieving a 4.9% character error rate (CER) on a confidential test set—substantially outperforming existing baselines. The framework establishes a reusable technical paradigm for HTR in low-resource historical language settings.

Technology Category

• Natural Language Processing: Learning & Optimization for NLP
• Machine Learning: Large Multimodal Models (LMMs)
• Planning, Routing, and Scheduling: Activity and Plan Recognition

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

This paper presents the first end-to-end pipeline for Handwritten Text Recognition (HTR) for Old Nepali, a historically significant but low-resource language. We adopt a line-level transcription approach and systematically explore encoder-decoder architectures and data-centric techniques to improve recognition accuracy. Our best model achieves a Character Error Rate (CER) of 4.9%. In addition, we implement and evaluate decoding strategies and analyze token-level confusions to better understand model behaviour and error patterns. While the dataset we used for evaluation is confidential, we release our training code, model configurations, and evaluation scripts to support further research in HTR for low-resource historical scripts.

Problem

• Develops first HTR pipeline for Old Nepali manuscripts
• Improves recognition accuracy using encoder-decoder architectures
• Analyzes model behavior and errors for low-resource scripts

Innovation

• End-to-end HTR pipeline for low-resource Old Nepali manuscripts
• Encoder-decoder architectures with data-centric techniques for accuracy
• Line-level transcription and decoding strategies for error analysis