Lang-8

» see lang-8.com as it lived, on the wayback machine

~ the eulogy ~

my english was not built by anybody correcting me. it came out of documentation and forum threads, read at volume across twenty years, and that method gets you fluent while leaving a particular residue. you end up with a large vocabulary and a set of small habits nobody ever flagged. Lang-8 existed precisely to flag those, and i never once used it.

The site started in august 2006 as a research project by Yangyang Xi, born in China and raised in Japan, working with a classmate. Lang-8 Inc. was formally set up in Tokyo on 29 june 2007. The idea was almost too simple to fund: you write something in the language you are learning, and people who speak it natively correct what you wrote.

The correction tools were the good part. A native speaker could strike out your phrasing and set their version beside it with the change colored in, so what you got back was the shape of your own mistake rather than a grade. Then you turned around and did the same for somebody writing in whatever language you had grown up in.

By 2017 it held more than 750,000 registered accounts reaching over 190 countries, covering something like 90 languages. It was never large by social network standards and was never trying to be.

~ the dossier ~

~ how it actually worked ~

The currency was reciprocity, which is rarer online than it sounds. Nobody paid for corrections. You earned them by giving them, and the whole economy ran on the fact that every person on the site was an expert in exactly one thing, their own first language, and a beginner in another.

That arrangement also produced the thing Lang-8 is now best known for among people who never used it. Every correction was a pair, the learner's sentence next to a native speaker's repair of it, which happens to be exactly the shape of training data a machine needs in order to learn error correction. The NAIST Lang-8 Learner Corpora came out of that pile and became one of the largest second language learner corpora anyone had assembled.

~ the unraveling ~

A site whose product is human attention has an obvious hole in it. Spam accounts worked out that Lang-8 was a place to put text in front of real people, and the company running it was small enough that the moderation load fell on a handful of staff who were also trying to build something else.

The something else was HiNative, a question and answer service the same company launched for quick language questions. It grew faster. By the end of september 2016 it had around 240,000 users while Lang-8 was aging, and a small company cannot carry two products when only one of them is moving.

Lang-8 closed on 29 february 2024, a leap day, so the anniversary of its death only comes around every four years. The company was direct about the reason. It was too small to keep fighting the bots while carrying a second product.

~ what it left ~

lang-8.com still resolves, and it sends you to a page on hinative.com explaining where everything went. That is a gentler ending than most of the graves here got.

The corrections outlived the site. Google Research published a cleaned version of the corpus called cLang-8, holding over two million corrected english sentences, and grammar correction models have trained on that material ever since. Seventeen years of strangers patiently fixing each other's writing got compressed into a dataset, which now sits underneath tools that correct your sentences with nobody on the other end. I do not think that betrays the original idea. It is not the same idea either.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »