Kuzushiji is the cursive style of handwriting used across Japan for roughly a thousand years, from the Heian period through the end of the nineteenth century. When Japan modernized its writing system after 1900, kuzushiji fell out of everyday use almost overnight, creating a strange situation: millions of historical documents, from literary masterpieces to family account books, sit in archives that the vast majority of modern Japanese readers cannot decipher. The gap between old and new writing is so stark that reading a page of kuzushiji feels less like reading an older version of the same language and more like encountering a wholly unfamiliar script.
What Kuzushiji Actually Looks Like
Modern Japanese writing uses a standardized set of characters. Each hiragana symbol has one accepted shape, each katakana symbol has one shape, and kanji are rendered with relatively uniform strokes. Kuzushiji throws all of that uniformity out the window. The word itself literally means “crumbling characters,” and the name is apt. Writers used flowing, connected brush strokes that merged individual characters into one another, sometimes reducing a complex kanji to a few elegant curves. Two writers copying the same passage might produce pages that look completely different to an untrained eye.
A key reason for this visual variety is hentaigana, the “variant kana.” Before standardization, a single Japanese syllable could be written using any of several different kanji as its base. The syllable “ka,” for instance, might be derived from one Chinese character in a Kyoto court poem and a completely different Chinese character in an Edo merchant’s ledger. The result was dozens of alternate forms for sounds that today have exactly one hiragana shape each. Researchers working on information-processing standards identified and catalogued these variants, including forms that originated from the same Chinese character but served distinct functions in practice, and prepared a formal proposal to add them to the international character-encoding standard ISO/IEC 10646.
Why Modern Japanese Readers Cannot Read It
The break between kuzushiji and modern Japanese script was not gradual. In the early Meiji era, the government standardized the writing system for use in schools and official documents. One hiragana per syllable was selected, and the rest were dropped. Within a generation, the old cursive forms vanished from daily life. Today, reading kuzushiji is a specialized academic skill, roughly comparable to a modern English speaker trying to read medieval manuscripts written in secretary hand. The letter shapes are technically the same alphabet, but the execution is so different that fluency requires dedicated training.
The problem is compounded by layout conventions. Kuzushiji texts are written vertically, right to left, and characters within a column sometimes run together without clear spacing. Line breaks do not always correspond to word boundaries. Annotations, glosses, and reading marks crowd the margins. For a modern reader accustomed to neatly typeset horizontal text, even locating where one character ends and the next begins can be a challenge.
The Scale of What Remains Unread
Japan’s libraries, temples, and archives hold an enormous volume of pre-modern documents. Estimates commonly cited in digital humanities circles put the number of surviving texts in the millions of volumes. These span every imaginable genre: classical literature, Buddhist sutras, medical treatises, agricultural guides, personal diaries, legal contracts, travel journals, and merchant account books. Only a small fraction have been transcribed into modern Japanese, and an even smaller fraction have been translated into other languages. For historians and scientists, this backlog represents a vast reservoir of untapped primary sources.
The practical consequences are real. Seismologists, for example, rely on historical records to understand earthquake frequency and intensity in regions where instrumental data only goes back about a century. Many of the most detailed eyewitness accounts of pre-modern earthquakes exist only in kuzushiji. Without transcription, that data stays locked in brushstrokes that machines and most humans cannot interpret.
Crowdsourcing Transcription
One response to the transcription bottleneck has been to enlist the public. Minna de Honkoku, a Japanese citizen-science project, began as an effort to transcribe earthquake-related historical materials held in the library of the University of Tokyo’s Earthquake Research Institute. The platform gives users a document image alongside a vertical text editor where they can type out the kuzushiji they see. A ranking system tracks how many characters each participant has transcribed, which helps keep volunteers motivated. As of October 2024, more than 9,700 people had registered for the project, and the total number of characters transcribed had reached about 41 million.
1Annual Review of Earth and Planetary Sciences. Minna de Honkoku: Citizen-Participation Transcription Project for Japanese Historical DocumentsThose 41 million characters are not just useful for earthquake research. The paired data, kuzushiji images matched to their modern text equivalents, has begun feeding into machine-learning systems as training data for artificial intelligence. In that sense, every volunteer who transcribes a page is also teaching a computer to read. The project demonstrates how citizen science can simultaneously advance historical scholarship and accelerate computational approaches to the same problem.
2Annual Review of Earth and Planetary Sciences. Minna de Honkoku: Citizen-Participation Transcription Project for Japanese Historical DocumentsTeaching Machines to Read Cursive
The machine-learning side of kuzushiji recognition has accelerated rapidly since around 2018. A landmark step was the creation of benchmark datasets specifically designed for this task. Researchers introduced Kuzushiji-MNIST, a dataset modeled on the well-known MNIST handwritten-digit benchmark but featuring kuzushiji characters instead. Two larger and more challenging companion datasets, Kuzushiji-49 and Kuzushiji-Kanji, followed, giving the research community standardized problems to test their algorithms against.
3arXiv. Deep Learning for Classical Japanese LiteratureThese datasets sparked a wave of activity. Competitions on kuzushiji recognition attracted researchers from around the world, and a variety of recognition methods were proposed in a short span of time.
4arXiv. Survey on Deep Learning-based Kuzushiji RecognitionSome approaches treat the problem as character-level classification, where the system identifies individual characters one at a time. Others tackle entire pages, attempting to segment columns, locate individual characters, and classify them in a single pipeline. More recent work has explored focusing on character parts rather than whole characters, using neural network architectures to extract features that capture the distinctive sub-strokes of kuzushiji forms.
5Pattern Recognition. Japanese historical character recognition by focusing on character partsThe results on clean, well-preserved documents are genuinely impressive, good enough to be useful as a first-pass reading aid for scholars. But the real world of historical archives is not clean.
When the Documents Fight Back
Historical manuscripts age. Paper yellows, foxing spots appear, ink bleeds through from the reverse side, and edges crumble. Many Japanese documents also bear vermillion seals stamped by owners, libraries, or government offices over the centuries. These seals overlap with the text, sometimes obscuring entire characters under red ink. For a human expert, context and experience fill in the gaps. For an OCR system trained on pristine images, a seal smeared across a column of text can be catastrophic.
Researchers have pointed out that although existing OCR methods perform well on clean pre-modern Japanese documents, they often fail when confronted with document degradation and seals that significantly affect recognition accuracy. Until recently, no publicly available dataset specifically addressed these challenges. New benchmark datasets featuring degraded documents with seals have begun to fill that gap, giving the field realistic test cases to push accuracy forward on the kinds of messy, imperfect pages that actually sit in archives.
6arXiv. DKDS: A Benchmark Dataset of Degraded Kuzushiji Documents with Seals for Detection and BinarizationDocument condition is only part of the difficulty. The sheer variability of kuzushiji across time periods, regions, and individual writers means that a model trained on Edo-period printed books may stumble on a Kamakura-period scroll, and vice versa. Genre matters too: a poetry anthology uses different conventions than a financial ledger. Building systems robust enough to handle this diversity remains an active area of research.
Hentaigana and the Encoding Problem
For digital preservation to work, computers need to be able to store and display kuzushiji characters, not just as images but as searchable text. This is harder than it sounds. The Unicode standard, which underpins virtually all modern text processing, originally included only the standardized set of modern kana. The variant forms, hentaigana, simply did not exist in the digital character universe. If you transcribed a historical document that used a specific hentaigana form, you had no way to represent that exact form as text. You could only approximate it with the modern equivalent, losing information about which variant the original writer chose.
The effort to fix this involved carefully defining a set of hentaigana suitable for the exchange of academic information. Researchers collected characters from letterpress printing and digital fonts, cataloguing variants and distinguishing between forms that, while derived from the same Chinese character, served recognizably different functions in historical usage. A formal proposal was prepared for submission to the ISO/IEC 10646 standard, the international framework that feeds into Unicode.
7Journal of Information Processing and Management. The past, present, and future of hentaigana: Standardization for information processingThat proposal succeeded. In 2017, Unicode 10.0 added 285 hentaigana characters, meaning that for the first time, scholars could encode historical kana variants as proper text rather than image workarounds. The practical impact is significant: transcriptions can now preserve the orthographic choices of the original writer, enabling computational analysis of how writing styles varied across regions and centuries. It also means that search engines and databases can index these characters, making vast archives of transcribed kuzushiji text more findable.
Learning Kuzushiji Today
If you are interested in reading kuzushiji yourself, several paths exist. Japanese universities offer courses in classical paleography, and some community education centers in Japan run introductory workshops. Outside Japan, a handful of graduate programs in East Asian studies include kuzushiji training, though it tends to be reserved for advanced students working with primary sources.
Online resources have expanded access considerably. The Kuzushiji learning app developed by the Center for Open Data in the Humanities (CODH) in Japan gamifies character recognition, letting users practice identifying characters at their own pace. Several Japanese libraries have released high-resolution scans of historical documents alongside modern transcriptions, creating de facto self-study materials. The National Institute of Japanese Literature’s database of classical texts is one major portal, offering side-by-side images and transcriptions for thousands of works.
The learning curve is steep. Even with good resources, gaining functional fluency takes years of practice, partly because the range of character variants is so large and partly because reading comprehension requires not just recognizing shapes but understanding classical Japanese grammar and vocabulary, which differ substantially from the modern language. Most learners begin with printed woodblock texts, where the characters are relatively standardized, before graduating to handwritten manuscripts where individual style adds another layer of difficulty.
Where AI and Human Readers Meet
The most productive workflow right now is collaborative. AI tools generate a first-pass transcription of a kuzushiji page, and a human expert reviews and corrects it. This approach dramatically speeds up the process compared to purely manual transcription while maintaining accuracy that purely automated systems cannot yet guarantee on their own, especially for degraded or idiosyncratic manuscripts.
The citizen-science model adds a third layer. Platforms like Minna de Honkoku produce transcriptions that serve double duty: they are useful to researchers studying the content of the documents, and they simultaneously generate labeled training data that improves the AI tools. As the AI improves, it can pre-populate transcription fields for volunteers, reducing the skill barrier for participation. Volunteers then correct the machine’s guesses, which in turn generates better training data. The feedback loop is slow but directionally powerful.
For fields like seismology, the payoff is tangible. Japan sits on one of the most seismically active zones on Earth, and its written record stretches back over a millennium. Each transcribed earthquake account adds a data point to the long-term seismic record, helping scientists calibrate recurrence models for events that happen on timescales longer than instrumental monitoring can capture. The text generated by Minna de Honkoku is already being used for seismological research and, as a side benefit, helps raise public awareness of historical disaster patterns.
8Annual Review of Earth and Planetary Sciences. Minna de Honkoku: Citizen-Participation Transcription Project for Japanese Historical DocumentsKuzushiji Beyond Japan
Japan is not the only country grappling with a gap between historical handwriting and modern literacy. Similar challenges exist wherever a society’s writing conventions changed dramatically: Arabic manuscripts in varying calligraphic styles, medieval Latin documents in insular or Beneventan scripts, and Chinese texts written in grass script all present analogous difficulties. What makes the kuzushiji case distinctive is the combination of scale, the sheer volume of surviving documents, and the relatively sudden historical break that made them inaccessible.
The tools and methods being developed for kuzushiji have potential applications beyond Japanese documents. The deep-learning architectures designed for cursive character recognition, the crowdsourcing platforms, and the strategies for handling degraded source material are in principle transferable to other historical scripts. Researchers in handwritten text recognition across multiple languages have looked to the kuzushiji benchmarks as models for how to organize datasets and structure competitions that push the state of the art forward. In that sense, the effort to unlock Japan’s pre-modern archives is also quietly advancing the broader field of historical document analysis worldwide.