Reproduce “Leveraging External Knowledge for Historical Document Restoration via Retrieval-Augmented Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gabeen Kim, Kyeongpil Kang
An approach from Kangwon National University introduces ARI, a retrieval-augmented large language model framework, for restoring damaged historical Hanja documents by explicitly integrating external knowledge. ARI-32B demonstrated 39.31% Top-1 accuracy for named entities and 80.42% for random characters on synthetic data, and was preferred by experts as a restoration tool on real-world documents.
Key insights
- Explicitly providing external knowledge via RAG and fine-tuning significantly enhances historical document restoration, particularly for knowledge-intensive named entities.
- Lexical similarity-based retrieval (BM25) proved more effective for Hanja character restoration than embedding-based methods, with optimal performance around 20 retrieved documents.
- Prioritizing named entity masking during fine-tuning balances performance improvements on knowledge-intensive segments without degrading general character restoration capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.