Japanese Braille Translation Using Deep Learning - Conversion from Phonetic Characters (Kana) to Homonymic Characters (Kanji).

Abstract

A blind student writes and submits reports in Braille word processor, which is difficult for teachers to read. This study's purpose is to make a translator from Braille into mixed Kana-Kanji sentences for such teachers. Because Kanji has homonyms, it is not always possible to get correct results when converting. To overcome this difficulty, we used deep learning for translation. We built a training dataset composed from 15,000 pairs of Braille codes and mixed Kana-Kanji sentences, and a validation dataset. In training, we got an accuracy of 0.906 and a good Bleu score of 0.600. In validation, we found 5 mistaken words in selecting homonymous Kanji by examining translation mistakes from 100 pairs of the verification sentences. The choice of homonymous Kanji depends on the context. For decreasing such type of errors, it is necessary to introduce of translation of paragraphs by increasing the scale of the network model in deep learning, and to expand the network structure.