Deciphering Japanese Labels on Old Natural History Specimens

For natural history specimens, metadata such as collection locality, date, and collector’s name are essential information ensuring the specimens’ scientific value. However, specimens collected decades ago may have missing or illegible metadata. The Takashi Kurobe’s Rock and Mineral Collection*1 consists of 883 rock, mineral, and fossil specimens stored in the Soil Science Laboratory in the Faculty of Agriculture at the Tokyo University of Agriculture and Technology. Of these, 611 specimens were collected through purchase, while the remaining 272 were collected by faculty members, students, and university associates. The specimens collected by faculty members are mainly accompanied by hand-written labels, and many of them contain partial information or are difficult to read because of the cursive scripts. The Japanese language is represented by 48 Hiragana and Katakana characters each, and more than 10,000 Kanji characters. In addition, the labels of old natural history specimens were written in English and Romanized Japanese using the alphabet. Such a variety of characters makes it difficult to decipher handwritten labels on the specimens even for native Japanese speakers. This study reports on how label information in the Takashi Kurobe’s Rock and Mineral Collection was deciphered to recover the scientific value of its specimens. Following this study, the Takashi Kurobe's Rock and Mineral Collection*1 has been made available as a digital archive with the photographs and metadata of each specimen on the web page of Nature and Science Museum, Tokyo University of Agriculture and Technology. CASE 1 In Takashi Kurobe’s Rock and Mineral Collection, there was a rock specimen with a single label with five characters in cursive scripts (Fig. 1). It was clear from their number and shapes that these five characters likely represent the location where the specimen was collected, but it was illegible due to the cursive scripts. The Center for Open Data in the Humanities is developing an application “miwo” that uses AI to recognize cursive scripts in documents. However, the accuracy of “miwo” is not yet high enough. Due to the above situation, deciphering the labels of this specimen was a dead end. Therefore, we decided to look not only at the label but also at the specimen itself. This specimen was a dark green volcanic rock with uneven pores throughout. On the inner surface of some of the bubbles, fine yellowish-gray crystals were found. These features are common in