[4H2-E-5] Human interface, education aid: human evaluation

[4H2-E-5-01] Recognition of Kuzushi-ji with Deep Learning Method

A Case Study of Kiritsubo Chapter in the Tale of Genji

〇Xiaoran Hu¹, Mariko Inamoto², Akihiko Konagaya¹ (1. Tokyo Institute of Technology, 2. Keisen University)

キーワード：Deep Learning, Kuzushi-ji, the Tale of Genji

Reading ancient documents is one of fundamental works on the study of national literature. However, due to the use of kuzushi-ji (classical cursive handwriting characters) in the ancient documents, it requires a lot of knowledge and labor to read. This paper uses an End-to-End method with attention mechanism to recognize the continuous kuzushi-ji in phrases. Compared with the traditional recognition model with Connectionist Temporal Classification (CTC), our approach can get higher accuracy of recognition. The method can recognize phrases written by kana (47 different characters) with the accuracy 78.92% and recognize phrases containing both kanji (63 different characters) and kana with accuracy 59.80%.