Komonjo Camera, the AI App for Reading Cursive Japanese Characters

Overview

In June 2023, TOPPAN Corporation released an application for iOS, the Komonjo Camera 古文書カメラ. This was followed by an Android version in October of the same year. This is good news for Japanese historians and literary scholars. 

Komonjo Camera is a smartphone application with AI-OCR technology that assists in the deciphering of kuzushi-ji くずし字 (cursive) characters. For humanities researchers in the field of Japanese studies, the issue of reading kuzushi-ji is an unavoidable one. Most Japanese documents from before World War II are written in complicated cursive font known as kuzushi-ji, but it is said that only about 0.01% of native Japanese speakers have the ability to understand them.[1] It is essential to undergo a reasonable period of specialised training in order to read and understand kuzushi-ji characters. Together with the MIWO みを app released by the Center for Open Data in the Humanities in 2021, the Komonjo Camera could be of great help to researchers and students who need to read and understand kuzushi-ji characters.[2]

Source: YouTube @fuminoha

Development Process

Amongst the historical documents written in kuzushi-ji that remain in Japan are those that contain valuable information on disaster prevention, the creation of tourism resources and the revitalisation of local communities, including records of disasters and local culture, which are directly related to contemporary social issues. However, as mentioned above, the fact that the kuzushi-ji script is extremely difficult to decipher for the majority of modern people is a major barrier to utilising records and documents from that period. In many cases, old documents owned by individuals are even destroyed because their contents are not known, and they are often stored in a state of deterioration and at risk of loss without being deciphered.

Based on an awareness of such issues, in 2015, in collaboration with the National Institute of Japanese Literature, TOPPAN developed a Document Image Retrieval System and a kuzushi-ji OCR system. Since then, they have developed and demonstrated the OCR technology with several research institutions and platforms, including Keio University and the Minna de Honkoku みんなで翻刻 project. The company has succeeded in realising a highly accurate lapsed character AI-OCR by having the AI learn from its accumulated database of character forms.[3] Currently, in addition to Komonjo Camera, they are offering Fuminoha Viewer ふみのはビューアー, which can superimpose original images and deciphered text, released in 2017, and Fuminoha Zemi ふみのはゼミ, an online kuzushi-ji deciphering support system, released in 2021, as a service for deciphering ancient documents and utilising kuzushi-ji materials, Fuminoha ふみのは®.[4]

The Fuminoha AI “learns” kuzushi-ji glyphs based on data provided by the National Institute of Japanese Literature, Minna de Honkoku, Mitsui Bunko, National Institute of Japanese Language and Linguistics, Kyoto City Library of Historical Documents, Tsubouchi Memorial Theatre Museum at the Waseda University, Keio University Library, Yamaguchi Prefectural Archives, the Printing Museum, and others.

Features

The Komonjo Camera app can be downloaded free of charge. The decoding function using AI-OCR is free for up to 30 times a day, and it is expected that in-app purchases will enable the number of times the decoding function can be used to be unlocked in the future[5] 

The application is very easy to use: simply photograph the kuzushi-ji characters with our smartphone and decipher them on the spot, making it easy for even kuzushi-ji beginners to come into contact with ancient documents. Conventional kuzushi-ji decoding tools, such as paper-based dictionaries, allow users to search for kuzushi-ji by collocation[6] or by brush stroke usage,[7] but even the search itself requires a certain degree of skill. In this respect, it is very convenient that this simple system of deciphering a document is now possible. 

It is possible to select either the full-automatic mode for decoding all characters in the image or the range selection mode for decoding only specific characters depending on the purpose.  It is very useful to be able to use the two different methods when you want to quickly reprint the entire document, or when you want to read only the parts that you cannot read.

After deciphering has been completed, if you want to redo a part of the transcription or make corrections manually, you can tap on the target text frame and edit it. This is able to further improve the accuracy of the AI. 

One of the features of Komonjo Camera is that it is equipped with two high-precision AI-OCR engines: the Komonjo AI 古文書AI and the Kotenseki AI 古典籍AI.[8] The former specialises in deciphering handwritten text and the latter in deciphering printed text, and as such the app is capable of deciphering a wider range of cursive characters with high accuracy, which is truly groundbreaking considering that the alternative choice of app, MIWO, tends to have extremely high reading accuracy for kuzushi-ji in printed books called hanpon 版本, but somewhat lower accuracy for handwritten kuzushi-ji in letters, diaries, and memos, etc.[9] This function is truly groundbreaking. 

Komonjo Camera will be useful for all researchers on Japanese studies who work with historiographies written in kuzushi-ji script.

Points to Note

As mentioned above, Komonjo Camera is extremely useful, but there are a few points to bear in mind. 

First, although the accuracy of the AI-OCR in Komonjo Camera is of a high standard, it is not perfect. The AI-OCR is a system that presents characters that are similar in shape from among the character data it has already learnt. Therefore, characters that the AI has not yet learnt, or characters written in unusual ways or in messy handwriting, may also degrade the AI’s recognition ability. This means that there remain challenges in deciphering variant characters, ligatures and special characters. Therefore, it is essential not to be overconfident in the decoding results, but to check them with our own eyes. As of June 2023, AI-OCR specialises in the deciphering of kuzushi-ji characters mainly from the Edo period. This means that the reading accuracy of documents from outside the Edo period, i.e. ancient, mediaeval, and modern times, may be reduced. In particular, researchers whose research focuses on periods other than the early modern period will need to bear this in mind.

Second, the use of Komonjo Camera requires a communications environment and cannot be used offline. As is well known, historical documents written in cursive script are not only stored places which have the internet connections. As the name jikata monjo 地方文書 [local documents] suggests, it is not unusual for large numbers of historical archives to be found at Shinto shrines, Buddhist temples, and private homes, etc, which are located deep in forests, on mountain tops, or other rural areas without a good reception environment. Old documents in poor condition may deteriorate or be damaged in transit, so the only way to decipher them is to do so on site or to photograph them and digitize them. Komonjo Camera, however, is not suitable for deciphering digital images. It is a shame that they cannot be used in places where there is no internet environment.

Even where internet access is available, use is limited to just 30 times a day. While the company would like to minimise the delivery of materials from the point of view of reducing the risk of deterioration, only 30 times a day is far too inadequate for deciphering large volumes of material such as diaries. The company has not yet announced the details of the paid service for unlocking the maximum number of using times, including the start date and price, so further information will be closely monitored.

Source: Komonjo Camera

One of the biggest problems is that the Komonjo Camera app cannot be downloaded from outside Japan. As is well known, materials written in kuzushi-ji script can be found not only in Japan but also in many other parts of the world. There are many museums and libraries outside Japan that hold manuscripts and printed books in kuzushi-ji characters, as well as art works containing kuzushi-ji script information (ukiyo-e 浮世絵 prints, hanging scrolls, outer boxes of ceramics, etc.), and kuzushi-ji script reading skills are essential for curators and librarians in such institutions. Learning kuzushi-ji is also an important issue for Japanologists living outside Japan, and international kuzushi-ji workshops for young researchers and students are being organised around the world.[10] Restricting downloads to Japan only does not take these demands into account at all. Fortunately, in response to complaints from users outside Japan, Fuminoha‘s Official X (formerly Twitter) account has publicly stated that it will consider expanding the use of the site overseas in the future.[11] It is strongly hoped that downloading the app will be possible from anywhere in the world in the near future for the further development of Japanese.

Another is the issue of personal information and privacy. The Terms of Use of the Komonjo Camera stipulate that the TOPPAN may process, retain and re-use the image data taken by users for AI learning and service improvement, etc.[12] They are not a national, local or other public authority, but only a private, for-profit company. This is a topic that often causes controversy in the shitei kanrisha seido 指定管理者制度 [designated manager system] for public libraries and museums in Japan,[13] have concerns about one for-profit company handling sensitive personal information been adequately addressed? In general, many historians are concerned that significant new discoveries from undisclosed archives will be seen by third parties before the results of their research are published. Alternatively, old documents written in kuzushi-ji often contain detailed local histories and personal information about the ancestors and relatives of people still living, and sometimes sensitive human rights information that could lead to prejudices or discrimination sometimes found. Operators are required to handle information more carefully, keeping these concerns in mind, and explanations for users.

These points need to be noted, but on the whole, the Komonjo Camera is greatly useful and an excellent tool for helping us to read and understand kuzushi-ji characters.

References

[1] MIWO: App for AI Kuzushiji Recognition. Available at: http://codh.rois.ac.jp/miwo/
[2] About MIWO, see James Harry Morris, ‘Practicing Reading Cursive Japanese with Miwo.’ The Digital Orientalist (2022). Available at: https://digitalorientalist.com/2022/06/03/practicing-reading-cursive-japanese-with-miwo/
[3] Komonjo Camera: About Us. Available at: https://camera.fuminoha.jp/aboutus.html
[4] Komonjo kaidoku to kuzushiji shiryō no kaidoku sābisu fuminoha. Available at: https://www.toppan.com/ja/joho/fuminoha/
[5] Komonjo Camera: FAQ. Available at: https://camera.fuminoha.jp/faq.html
[6] Shigeo Negishi, Takayuki Satō, Hiroyuki Yasuike, ed., Shinpen komonjo kaidoku jiten. Tōkyō: Kashiwa Shobō, 1993.
[7] Kōta Kodama, ed., Kuzushiji yōrei jiten. Tōkyō: Tōkyōdō Shuppan, 1993.
[8] Komonjo Camera: Komonjo Camera Features. Available at: https://camera.fuminoha.jp/index.html#feature
[9] Center for Open Data in the Humanities: Miwo towa? Available at: http://codh.rois.ac.jp/miwo/about/
[10] Yukimitsu Hasegawa, ‘Komonjo no kuzushiji ga shizuka na būmu, kenburijji daigaku nado meimonkō mo chūmoku!.’ Diamond Online (2011). Available at: https://diamond.jp/articles/-/257907
[11] X @fuminoha_jp. Available at :
https://twitter.com/fuminoha_jp/status/1717361773765066904?t=qCxaboGKj0CDXJw4H1OeOQ&s=19
[12] Komonjo Camera: Terms of Use. Available at: https://camera.fuminoha.jp/terms-of-use.html
[13] Hirohito Tsuji, ’Consideration of Museum’s Library and Reading Room.’ Bulletin of Museology, Kokugakuin University 43 (2019): 127-138.

Leave a comment