Send message to Authors

Paper short abstract

We present a preliminary analysis toward an RAG-based search system for Japanese classical texts using OCR-generated texts. Rather than focusing only on OCR accuracy, we examine how text normalization influences search performance. We report our initial findings, and discuss how they will guide future system development.

Paper long abstract

Japanese classical texts are an essential resource for literary and historical studies. Researchers search for examples of specific expressions across large collections to investigate language use and historical change. In recent years, the National Institute of Japanese Literature (NIJL) has promoted the large-scale digitization of Japanese classical texts, and more than 500,000 works have been digitized to date. Although these digitization efforts have made huge collections of classical texts available, it remains very difficult to search them because OCR-generated texts contain recognition errors and historical orthographic variations.

This study presents a preliminary analysis for the development of an example-based search system using RAG for Japanese classical texts. Rather than evaluating OCR correction by transcription accuracy, we investigate how different text normalization and OCR error correction methods influence search performance. To this end, we compare multiple document collections, including original OCR texts, rule-based normalized texts, LLM-corrected texts, and manually corrected texts, using a dense retrieval model based on ColBERT.

We report preliminary results on how different text normalization methods affect search performance for Japanese classical texts. The analysis provides an initial perspective on how text normalization and OCR correction relate to search. Based on these preliminary findings, we discuss the design of a future RAG-based search system for Japanese classical texts.