RAG-Powered ARI Restores Historical Documents with Expert Precision
New framework combines LLMs with external knowledge to fix illegible historical texts.
Historical documents are invaluable archives but often become illegible due to physical deterioration. Existing restoration methods using masked language models rely on local context and struggle with named entities (e.g., people, places) that require external historical knowledge. To solve this, researchers developed ARI, a framework that combines large language models with retrieval-augmented generation (RAG). ARI dynamically retrieves relevant historical facts from external sources and feeds them to the LLM, enabling accurate restoration of context-dependent proper nouns. This hybrid approach leverages both the LLM's pre-trained knowledge and explicit external references.
Extensive experiments on Korean historical documents show that ARI substantially outperforms baseline methods across both general characters and named entities. The framework was validated through comprehensive evaluations, including expert assessments from domain historians, who confirmed its practical utility. ARI can accelerate the analysis of degraded records by reducing manual restoration effort. The work has been accepted as a Findings paper at ACL 2026, signaling strong peer recognition. This approach opens the door to applying RAG-based LLMs for cultural heritage preservation and archival science.
- ARI combines an LLM's implicit knowledge with RAG to retrieve external historical context for restoring illegible text.
- Achieves substantial gains over masked language models, particularly for named entities like people and places.
- Accepted to ACL 2026 Findings; expert assessments validate it as a practical tool for historians and archivists.
Why It Matters
Historians can restore centuries-old damaged documents with AI, accelerating analysis of cultural heritage records.