Leveraging External Knowledge for Historical Document Restoration via Retrieval-Augmented Large Language Models

The paper presents ARI, a retrieval-augmented framework for restoring damaged historical documents. It combines the implicit knowledge of pretrained large language models with explicitly retrieved historical context, targeting proper nouns that masked-language-model approaches cannot reliably infer from local context alone. Experiments on Korean historical documents reportedly show substantial improvements over baselines for both general character restoration and named-entity restoration. Expert evaluations also indicate that ARI could be useful to domain specialists analyzing historical records. The abstract does not provide the numerical results, dataset scale, or detailed retrieval and evaluation settings.