Leveraging External Knowledge for Historical Document Restoration via Retrieval-Augmented Large Language Models

2025-01-03 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

The article introduces ARI, a novel Retrieval-Augmented Large Language Model (RAG-LLM) framework designed for historical document restoration, specifically addressing the challenge of restoring named entities that require external knowledge. ARI combines the implicit knowledge of pre-trained LLMs with explicitly retrieved external context. Extensive experiments on Korean historical documents, including the Annals of the Joseon Dynasty (AJD) and the Journal of the Royal Secretariat (JRS), demonstrate that ARI significantly outperforms baseline methods in restoring both general characters and named entities. ARI-32B, fine-tuned on Qwen3 32B, achieved superior performance, particularly for named entities like Person (PER) and Location (LOC), and was preferred by human experts. The framework utilizes chronological metadata, few-shot prompting, and BM25-based retrieval with an 80% deduplication threshold, incurring a computational cost of approximately 1,500 H200 GPU hours.

Key takeaway

For AI Scientists developing historical document restoration systems, you should integrate retrieval-augmented generation (RAG) with fine-tuned LLMs. This approach significantly improves named entity restoration, which traditional masked language models struggle with. Specifically, prioritize BM25 for retrieval and apply an 80% deduplication threshold to external documents. This strategy, exemplified by ARI-32B, enhances accuracy and provides a practical tool for domain experts, reducing manual effort in analyzing ancient archives.

Key insights

Retrieval-augmented LLMs effectively restore damaged historical documents, especially named entities, by integrating external knowledge.

Principles

• External knowledge is crucial for named entity restoration.
• RAG and fine-tuning combine to enhance LLM restoration.
• Lexical similarity (BM25) is key for Hanja character restoration.

Method

The ARI framework fine-tunes a Qwen3 LLM with RAG, integrating chronological metadata, few-shot examples, and BM25-retrieved, deduplicated external documents to restore masked Hanja characters.

In practice

• Use BM25 for retrieving Hanja-specific external context.
• Apply an 80% similarity threshold for RAG document deduplication.
• Prioritize named entity masking in 25% of training data.

Topics

• Historical Document Restoration
• Retrieval-Augmented Generation
• Large Language Models
• Hanja Characters
• Named Entity Recognition
• Joseon Dynasty Archives

Code references

• EvelynKimm/ARI
• rapidfuzz/RapidFuzz

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Büyük Dil Modellerinin Hafızasını Güçlendiren Yaklaşım:Retrieval-Augmented Generation (RAG) — RAG enhances LLMs by providing external, real-time context, reducing hallucinations and improving factual accuracy.
• Decompose, Retrieve, Cite: A RAG Pipeline for Structured Report Generation from Technical Documentation — A RAG system for technical docs uses prompt decomposition and specialized parsing to generate structured, cited reports with high accuracy.
• Leveraging External Knowledge for Historical Document Restoration via Retrieval-Augmented Large Language Models — ARI uses RAG with LLMs to restore historical documents, excelling at named entities by combining implicit and explicit knowledge.
• Retrieval-Augmented Generation as a Bridge Between Classical NLP Pipelines and Modern LLM Systems — Retrieval-augmented generation (RAG) addresses the limitation of static knowledge in large language models (LLMs) by connecting them to external, up-to-date knowledge sources.
• Howard University-AI4PC at SemEval-2026 Task 8: Query Reformulation and Dense-Lexical Retrieval Fusion for Multi-Turn Retrieval-Augmented Generation — A training-free hybrid RAG system combines multi-signal query reformulation and fused retrieval for competitive multi-turn performance.
• Multi-Field Hybrid Retrieval-Augmented Generation for Maritime Accident Root Cause Analysis — Structuring RAG architectures around domain-specific document fields significantly enhances downstream generation quality and factual consistency.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.