Eugen Rochko
Just gonna throw this out here - any Ruby gem for relatively quick & accurate language detection that you know of?
Apr 18, 2017 at 4:41pmWeb
Show threadEugen RochkoApr 18, 2017

I didn't expect anyone to think I am asking for a General AI.

I just want semi-accurate heuristic to spare users having to tag language manually.

Proprietary APIs like Google Translate are out of the question. Already using WhatLanguage, it's not working well. If I can't find anything, welp, there's that. Not a big deal.

Show thread葉月Apr 18, 2017
@Gargron Wait so Mastodon is guessing our languages by random? That creates all sorts of problem for us CJK users. (See Han Unification) We should at least have an option to manually pick the correct language for the text.
Show threadEugen RochkoApr 18, 2017
@hatsuki It's guessing, but there are no filters yet.
Show thread葉月Apr 18, 2017
@Gargron Oops, now I see that none of the text has lang attribute. This is a bug, CJK languages will not display correctly without lang attributes, due to Unicode Han Unification. I know this will clutter UI but I strongly recommend looking into this before Mastodon gets critical mass among Chinese-speaking communities, that will be a very hard problem to solve...
Show threadMatthew SkalaApr 18, 2017

@hatsuki @Gargron seems to me if it contains Unified Han and kana you call it "Japanese"; if it contains hangul with or without Unified Han you call it "Korean"; and if it contains Unified Han but no significant amount of those other things, you call it "Chinese." Done, and close enough. Telling Swedish from Norwegian is much harder.

As for needing a "lang" tag to choose fonts, nearly all the Web has that problem anyway.

Show thread葉月Apr 18, 2017

@mattskala @Gargron It's broken everywhere, so don't fix it. Sounds strange enough, but I am not the one doing the work, so I have nothing to complain.

On detecting languages though, this assumes single language posts. At least it should be clear to the user which lang tag will be added when tooting.

Show threadMatthew SkalaApr 18, 2017
@hatsuki @Gargron Well, Twitter seems to survive without good language detection. It *is* a little weird that "Haitian Creole" seems to be their default for unrecognizable stuff.
Show thread葉月Apr 18, 2017
@mattskala @Gargron Unicode is a sad story.
Show threadJookiaApr 18, 2017
@hatsuki @gargron @mattskala Han Unification + lack of language tags in Unicode is a reason why I'm aim to support ShiftJIS as well
Show threadBrendan MolloyApr 18, 2017
@Gargron if it's not, it's something I've been meaning to work on for some time. Wrote a python port ages ago, should modernise.
Show threadFrancescaApr 18, 2017
@Gargron only thing that comes to mind is Python's Natural Language Toolkit. There's probably something similar for Ruby.
Show threadOrión MontoyaApr 18, 2017
@Gargron https://github.com/jmhodges/rchardet is kind of old but it is not so old that the algorithm should be broken. Uses the same heuristics as web browsers: perhaps worth a shot.
Show threadOrión MontoyaApr 18, 2017
@Gargron omg nevermind, I'm a dingus: encoding != language, nothing to see here
Show threadRick Minerich 🔥Apr 18, 2017
@Gargron there's an opportunity here to get the Mastodon users to train something custom that can be open source.
Show threadNetzSoOcApr 18, 2017
@Gargron you can train a system, the problem is there are too many languages
Show threadNetzSoOcApr 18, 2017
@Gargron if they ate tagging already you can use these data to train that system
Show threadFrancis ChongApr 18, 2017
@Gargron you don't need something like ai. https://github.com/BYVoid/uchardet There should gems too
Show threadCSApr 18, 2017
@Gargron Have you looked at the implementation of WhatLanguage? If this #Python library looks like it is sufficiently different/better, we could port it to #Ruby: https://github.com/Mimino666/langdetect
Show threadCSApr 18, 2017

@Gargron I tried out this library on a couple Japanese toots from the federated timeline and it detected they were Japanese every time.

The algorithm used looks pretty solid as well. Would you use a #Ruby port?

Show threadFrancis ChongApr 18, 2017
@Gargron you do not need something like ai, something like statistic can do it. I forget the one I used, but something similar https://github.com/peterc/whatlanguage
Show threadFrancis ChongApr 18, 2017
@Gargron I think this is it https://github.com/igrigorik/language_detector
Show threadCSApr 18, 2017
@siuying @Gargron This looks quite good. 👍
Show threadNicki H ☑️Apr 18, 2017
@Gargron Thank you for trying!
Show threadpaulcbettsApr 18, 2017
@Gargron cld2 is accurate given enough text, which for a toot will sometimes work and sometimes not
Show threadL3viathanApr 18, 2017

@Gargron #NLProc‌er here, Have you taken a look at https://github.com/diasks2/ruby-nlp?

Most NLP stuff isn't in Ruby (Python, Java, etc. are more prevalent).

Show threadFabrice DesréApr 18, 2017
@Gargron Look at N-gram based language detection libraries. There are many available, eg. https://github.com/optimaize/language-detector
Show threadarrufatApr 18, 2017
@Gargron might this be of your interest? https://github.com/greyblake/whatlang-rs
Show threadkasran, fourier transfemmeApr 18, 2017
@Gargron it seems like something a neural network might be good for! you'd probably need to train it for a long time though
Show threadSpacedragon Apr 19, 2017
@Gargron I'm actually planning on running a project (self-funded) to attempt to detect dominant languages on public Mastodon instances, as part of Radar.
Show threadjd Ⓐ★😼🚀🌍🇪🇺🇭🇺Apr 19, 2017

@wogan @Gargron I have noticed that a few people claiming they are learning esperanto – but I do not see anyone writing here in it.

Social media actually does reflect the real world haha

Show threadOlivier AuberApr 20, 2017
@Gargron "A semi-accurate heuristic to spare users having to tag language manually"... Have you heard about Pierre Lévy's #IEML? Here's an aborted project envisioned at the time of Identica (2009)
http://perspective-numerique.net/wakka.php?wiki=PropositionCodageIEML (Sorry French)
Show threadParker HigginsApr 18, 2017
@Gargron you could just do a regex search for "ananas"?
Show threadEugen RochkoApr 18, 2017
@xor it should support Japanese and the various variants of Chinese.
Show threadmattn✅Apr 18, 2017
@Gargron @xor whatlanguage?
Show threadEugen RochkoApr 18, 2017
@mattn @xor That's what v1.2 is using, and when it encounters Japanese text, it pretty much reports a random language each time.
Show threadmattn✅Apr 18, 2017
@Gargron @xor because some unicode code point is shared between japanese/chinese.
Show threadEugen RochkoApr 18, 2017
@mattn @xor uh no, it doesn't support either of those, so it just reports it as french, german, italian, polish, etc randomly
Show threadmattn✅Apr 18, 2017
@Gargron @xor like this? https://github.com/peterc/whatlanguage/issues/1
Show threadmattn✅Apr 18, 2017
@Gargron @xor Ah, it doesn't support japanese yet. :-(
Show threadMunin, Keeper of LoreApr 18, 2017
@Gargron ...I think that, perhaps, that might not be a good thing to roll into the core functionality? Perhaps instead build an API for modules other people can build for functionality like that if they so choose?
Show threadashleyApr 18, 2017
@munin @Gargron and perhaps a client setting for which language the user is using?
Show threadSpencer Alves [backup]Apr 18, 2017
@Gargron You can't seriously expect a machine to do that accurately, can you?
Show threadEugen RochkoApr 18, 2017
@impiaaa accuracy is not required, only a close guess.
Show threadEugen RochkoApr 18, 2017
@impiaaa alternative is marking the language as whatever the user chose their UI to be in, or worse even, adding yet another control where you pick the language of the toot (bleh, I don't want even more controls!)
Show threadpolarity 🇩🇪Apr 18, 2017
@Gargron @impiaaa magic tag #ger or 🇩🇪
Show thread葉月Apr 18, 2017
@Gargron What are you trying to achieve? Tag the toots with the correct lang attribute?
Show threadAndrew NesbittApr 18, 2017
@Gargron this one is pretty nice: https://github.com/peterc/whatlanguage
Show threadCSApr 18, 2017
@andrewnez @Gargron as mentioned already, that doesn't work on Japanese: https://mastodon.social/@Gargron/3134956
Show threadjomoApr 18, 2017
@andrewnez @Gargron
> It works well on texts of over 10 words in length (e.g. blog posts or comments) and very poorly on short or Twitter-esque text, so be aware.
Show threadoptikfluffel✨Apr 18, 2017
@Gargron there should be something useful in here https://github.com/arbox/nlp-with-ruby/blob/master/README.md
Show threadvalApr 18, 2017
@Gargron There is probably some binding to the wordnet database; it's quite popular in the field
Show threadBrendan MolloyApr 18, 2017
@Gargron https://github.com/hashwin/scylla ? any textcat derivative is a good start