"Collapsed" calligraphy
« previous post | next post »
Responding to this recent post about machine analysis of grammar, "Literary Sinitic / Classical Chinese dependency parsing" (11/27/19), Nicholas Morrow Williams writes:
That reminds me tangentially of something I just heard about, an effort to transcribe Japanese "kuzushiji" (cursive-like) script using AI. This article, which contains some striking illustrations, is about a huge international competition to devise a better method, won apparently by a Chinese team.
Kuzushiji ("character written in a cursive style") takes the following forms in Japanese: じ くずし字, 崩し字, くずしじ. The kanji for kuzushi is 崩 (kuzure) and means "collapse"). At first I thought that was a bit odd for a name that indicates a cursive style of writing. After thinking about the chaotic, tumbling appearance of these characters, which may be seen here, I have to admit that "collapsing" is a pretty good description for them.
Going back to the beginning, what this post is really about is that now machines are being trained to read this script which is beastly difficult for humans to decipher without extensive specialized training.
"Kuzushiji Recognition: Opening the door to a thousand years of Japanese culture", kaggle
Kuzushiji, a cursive writing style, was used in Japan for over a thousand years, beginning in the 8th century. Over 3 million books, on a diverse array of topics such as literature, science, mathematics and cooking are preserved today. However, the standardization of Japanese textbooks known as the “Elementary School Order” in 1900, removed Kuzushiji from regular school curriculum, as modern japanese print became popular. As a result, most Japanese natives today cannot read books written or printed just 120 years ago.
Since Chinese characters entered Japan in the 8th century, the Japanese language has been written using Kanji (Chinese characters in the Japanese language) in official records. However, from the late 9th century, the Japanese began to add their own character sets: Hiragana and Katakana, which derive from different ways of simplifying Kanji. Individual Hiragana and Katakana characters don't contain independent semantic meaning, but instead carry phonetic information (like letters in the English alphabet).
In Kuzushiji documents, Kanji, Hiragana and Katakana are all used. However, the number of character types in each document varies by genre. For example, story books are mostly written in Hiragana while formal records are written mainly in Kanji.
The website then goes on to describe some of the challenges of training machines to read kuzushiji, including the large number and unbalanced nature of unique characters in the dataset, the proliferation of variants of individual characters, similarity of characters, connectedness and overlap between characters, multiplicity of layouts, and so forth.
The hosts of the website state that they need help from machine learning experts to transcribe kuzushiji into contemporary Japanese characters. The larger the dataset that accumulates, the more readily will the investigators be able to develop better algorithms for kuzushiji recognition. The hosts have already made enormous progress toward achieving their goal of creating a workable machine for recognizing kuzushiji. Once they meet their target, the machine will become a fluent reader of kuzushiji, a feat that only 0.01% of modern Japanese natives are capable of. Naturally, this is only a goal toward which the investigators are striving. It is unlikely that they will ever attain anything close to one hundred per cent recognition of kuzushiji texts. But even ninety or 80 per cent will lighten the burden of researchers who wish to convert kuzushiji texts into modern Japanese characters. This will assist in unlocking the treasures hidden in the vast number of kuzushiji texts that have been amassed during the past millennium.
A final note
It's interesting that kuzushi 崩し, くずし is also a term for unbalancing an opponent in Japanese martial arts. The noun comes from the transitive verb kuzusu (崩す), meaning to level, pull down, destroy, or demolish. (Wikipedia)
Whether it's toppling kanji or tumbling bodies, falling can be both graceful and fluid.
Selected readings
"Cursive" (3/30/14)
"Cursive and Characters: Dying Arts" (4/29/11)
[Thanks to Linda Chance]
Rodger C said,
December 3, 2019 @ 10:43 am
Not related, I gather, to kudzu, the vine that pulls down everything.
James Unger said,
December 3, 2019 @ 11:59 am
Quite unrelated, Rodger. The verb /kuzusu/ goes back to /kudusu/ whereas /kuzu/ 'kudzu vine' had original /z/. The "d" in the English borrowing is a relic of an outmoded 19th-century romanization.
Victor Mair said,
December 3, 2019 @ 1:13 pm
From David Lurie:
I've been hearing about these projects for years and agree that they are promising, although I'm less sanguine about the value of 80% accuracy, especially since the more-difficult-to-decipher graphs are often logographic, and thus of much greater semantic significance (and much lower redundancy) than the phonographs that surround them…
This usage of the character 崩 is a wonderful example of the uneven mapping of meaning between Chinese graphs/morphemes and the Japanese words they are linked to; since the transitive-intransitive pair of verbs kuzusu-kuzuru (modern kuzureru) mean "break down," but also by extension "disorder," "dishevel," "relax" (etc.), this is the first time I've realized that from a Chinese perspective it is odd for kuzushiji to be written 崩し字… Quickly, here are a few relevant modern kuzusu examples, from some of the Kenkyūsha English-Japanese dictionaries:
膝を崩す sit at ease; make oneself comfortable [at home].
気をつけの姿勢を崩す relax one's ramrod posture
彼は崩した着こなしがうまい. He's good at dressing casually.
VHM: I should have said that even 50% would be an amazing achievement.
cameron said,
December 3, 2019 @ 9:44 pm
How well does current OCR technology handle printed Japanese from recent decades?
Michael Watts said,
December 4, 2019 @ 4:46 am
Does this differentiate this script from any other Japanese script?
Antonio L. Banderas said,
December 4, 2019 @ 8:09 am
How do Kuzushiji in particular compare to handwritten Japanese texts in general?
Robot Therapist said,
December 4, 2019 @ 10:24 am
Is this the kind of script that would be found on a kakemono (for example)?
Johannes Choo said,
December 7, 2019 @ 1:40 pm
It would be easy for an AI that accomplishes transcription to annotate these transcriptions with a degree of uncertainty: YouTube's auto-generated subtitles represent this with opaqueness. Another thing that is easily done as well is to correspond the transcription to an image of the original text
With these, an interactive side-by-side view of transcription beside a photo would seem very feasible, and useful even if it achieves low accuracy, say, 50%.
If the algorithm were learning on-line it could even allow scholars to suggest and learn from them even as it predicts the bulk of the text.
Josh R said,
December 9, 2019 @ 2:13 am
Michael Watts said,
Does this differentiate this script from any other Japanese script?
Antonio L. Banderas said,
How do Kuzushiji in particular compare to handwritten Japanese texts in general?
Robot Therapist said,
Is this the kind of script that would be found on a kakemono (for example)?
I am not sure how much of this applies to Chinese writing, but here's a quick primer on Japanese handwritten and calligraphic scripts.
Historically (and in modern calligraphic writing), there were three kinds of brush-written scripts: kaisho (block writing), gyousho (running writing), and sousho ("grass" writing). Kaisho was writing with each stroke clearly delineated. But the strokes could be run together and/or abbreviated in common and regular ways. Sousho is writing with the entire character in this run-together, abbreviated manner. Gyousho was writing along the pretty wide spectrum between kaisho and sousho.
Kuzushiji, then, though it can broadly refer to any kind of writing in which the block form of the character is "broken down", is used specifically as an umbrella term for gyousho and sousho, or even slightly "broken down" kaisho in brush writing.
Pre-war, IIRC, Japanese students would learn the common and regular patterns of gyousho and sousho, which is necessary to know how to read them. Post-war, however, compulsory education only deals with block writing, and in as much as students learn brush writing, it is largely limited to kaisho.
Modern handwritten Japanese is most similar to gyousho, but the various ways of running together and abbreviating the strokes is more ad hoc and personal. Also, it's generally done with a pen rather than a brush. Handwriting that is flowing and calligraphic is often called "tappitsu 達筆, "accomplished brush", which can be an ironic way of saying, "This writing is so accomplished I can't even read it."
Related, but somewhat apart of this, is artistic calligraphy, which is what you typically see on scrolls. This is essentially gyousho and sousho (sometimes kaisho), but with added artistic intent, from both the calligrapher's own personal style as well as the content of the words. This added artistic intent can render the characters even more illegible to laypeople.