Language Log

A #StableDiffusion model trained on images of Japanese Kanji characters came up with “Fake Kanji” for novel concepts like Skyscraper, Pikachu, Elon Musk, Deep Learning, YouTube, Gundam, Singularity, etc.

They kind of make sense. Not bad! https://t.co/ibegk4XszN pic.twitter.com/qkNcA9AAWb

— hardmaru (@hardmaru) January 6, 2023

[This is a guest post by Bill Benzon]

Everybody's talking about it.

"Meta has developed an AI translator for a primarily-spoken language

It only translates between Hokkien and English for now, but offers potential for thousands of languages without official written systems."

By Amanda Yeo, Mashable (October 20, 2022)

If true, this technology could be an enormous boon for illiterates everywhere.  It also has important theoretical and linguistic implications.

New article by Nyri Bakkalian in Unseen Japan (9/17/22):

"New App Promises Greater Convenience in Reading Old Japanese Cursive:

Kuzushiji, the 'crushed letters' found in historical Japanese documents, have long been the bane of scholars. A new app may change all that."

The author bemoans:

During my graduate education in Japanese history, interpreting handwritten primary source material from the 19th century and earlier was one of my greatest challenges. Typeset historic documents exist, especially in my period of focus during the Bakumatsu-Meiji transition. But the further back in time one’s research focus is situated, the rarer these documents become. There is a plethora of handwritten documents, written in historic cursive, but learning how to read them is a significant investment of time and resources beyond the means of most people who might otherwise have the inclination to learn.

"Google Translate learns 24 new languages"
Isaac Caswell, Google blog (5/11/22)

==========

[This is a guest post by Jenny Chu]

It came up today when I was reading this somewhat viral post on Facebook. 

Now it's getting interesting:

"China’s internet police losing man-versus-machine duel on social media"

Stephen Chen, SCMP (11/14/21)

    Hordes of bot accounts using clever dodging tactics are causing burnout among human censors, police investigative paper finds
    Authorities may respond by raising a counter-army of automated accounts or even an AI-driven public opinion leader

There may well be a dogma out there stating that all languages are equally complex, but I don't believe it, especially not if it has to be "drummed" into our minds.  I have learned many languages.  Some of them are exceedingly hard (because of their complexity) and some of them are relatively easy (because they are comparatively simple).  I have often said that Mandarin is the easiest language I ever learned to speak, but the hardest to read and write in characters (though very easy in Romanization).  And remember these posts:

"Difficult languages and easy languages" (3/4/17)

"Difficult languages and easy languages, part 2" (5/28/19)

With this post, we are already acquainted with Inspur's Yuan 1.0, "one of the most advanced deep learning language models that can generate coherent Chinese texts."  Now, with the present article, we will delve more deeply into the potentials and pitfalls of Inspur's deep learning language model:

"Inspur unveils GPT-3 equivalent for Chinese language", by Wei Sheng, TechNode (1026/21)

The model is trained with 245.7 billion parameters—the number of weights in an artificial neural network, according to the company. This is more than the Elon Musk-backed GPT-3 language model for English, which has 175 billion parameters. Inspur said the Yuan model was trained with 5 terabytes of datasets.

…

New article in EnterpriseAI (October 21, 2021):

"Language Model Training Gets Another Player: Inspur AI Research Unveils Yuan 1.0",  by Todd R. Weiss

From Pranav Mulgund:

Given this:

"Measure words for robots" (9/4/21)

and this:

"Arigatō" (9/3/21),

I could not help but think of this:

Christian Horn was reading an article in Japanese Endgadget (8/11/21) about the introduction of a new kind of robot called a "Cyberdog".

Says Christian:

You don't need to know Japanese to understand the fascinating part:  in Japanese, when counting things, the type of "thing" you are counting is relevant.  So you count "flat things" differently than "long shaped" things.  Or machines, fish, or animals.

The article states that Cyberdog is aimed at developers, and is limited to "1000台(匹?)", showing hesitation over which measure word to use, dai 台 (counter for machines, including vehicles) or hiki 匹 (counter for small animals​; counter for rolls of cloth; counter for horses​).  If you use dai 台 as a measure word for counting Cyberdogs, it would indicate that you think of them as machines.  If you use hiki 匹 for counting them, it would indicate that you regard Cyberdogs as animals.

Article by Holly Else in Nature (8/5/21):

"‘Tortured phrases’ give away fabricated research papers

Analysis reveals that strange turns of phrase may indicate foul play in science"

Here are the beginning and a few other selected portions of the article:

In April 2021, a series of strange phrases in journal articles piqued the interest of a group of computer scientists. The researchers could not understand why researchers would use the terms ‘counterfeit consciousness’, ‘profound neural organization’ and ‘colossal information’ in place of the more widely recognized terms ‘artificial intelligence’, ‘deep neural network’ and ‘big data’.

Further investigation revealed that these strange terms — which they dub “tortured phrases” — are probably the result of automated translation or software that attempts to disguise plagiarism. And they seem to be rife in computer-science papers.

Research-integrity sleuths say that Cabanac* and his colleagues have uncovered a new type of fabricated research paper, and that their work, posted in a preprint on arXiv on 12 July¹, might expose only the tip of the iceberg when it comes to the literature affected.

[*VHM:  Guillaume Cabanac, a computer scientist at the University of Toulouse, France]