Hacker News Reader

Comment by est31 | Hacker News Reader

Before you say that Google has world-class ML engineers and Mozilla doesn't even have a chance of competing, consider that the quality of the startup deepl's translator is comparable and possibly even better to the quality of giants like Google or Microsoft. So this seems like one of those problems where small companies can still be competing with GAFAM-scale ones.

View on HN

So much of machine translation is the quality of your parallel corpus and your preprocessing. I've got a theory that Google actually has too much data, and maybe too much imperfect data. Also, Deepl came out of Linguee (https://en.wikipedia.org/wiki/Linguee), who appear to have a massive number of parallel texts...I really believe whoever who has the (well-curated) texts has the (NMT) world.

As an aside, I don't know who needs to know this, but we are really lucky right now in that we now have tools that enable pretty much anyone to train a language model and translate with it, not just giant or small companies. It's not necessarily going to be any good, but the tools are all right there for us plebes and it's pretty fun.

I pulled together a tutorial just this weekend walking through doing this on Google Colab (https://sevenminuteserver.com/post/2019-10-17-machine-transl...) and EC2 Spot instances (https://sevenminuteserver.com/post/2019-10-17-machine-transl...) if anyone wants to play around with it.

> and possibly even better

It is better.

Google is now just a sluggish tech bully and only shows world-beating performance through acquisitions (such as, sadly, their acquisition of DeepMind).

The Google of 1999 - that a lot of us thrilled to - is long dead.

I am in the translation business, and although we do not use MT so far for the verticals we focus on, perhaps what shocks me most is Deepl Alexa 90-day trend [0]. It is so smooth it seems faked.

[0] https://www.alexa.com/siteinfo/deepl.com

Never heard of that website so I give it a try.

It says "[t]ranslate from any language", but doesn't seem to support Chinese or Japanese at all.

To me, translating other Western languages to English is never a problem. I can use any half-decent online translation services and the result is totally understandable.

> I can use any half-decent online translation services and the result is totally understandable

But in the case of deepl, the result is not only understandable but also often (but not always) nearly perfect.

When you're just trying to read stuff to understand it, "understandable" is good enough. I'll be waiting for a looong time for CN/JP/KR translation to reach anywhere close to Not Utterly Incoherent.

Me, too, me, too. Knowing a little bit of Chinese, I can understand that this is not an easy job for an AI. Too much context dependend stuff. Or maybe, there are simply not enough bilingual documents for the training. That's the great advantage of the European union :)

"Translate from any language" in the context of deepl.com website means "detect language".

AFAIK, only the devil can translate from any language (not even Google Translate can translate from any language, in spite of attributed evilness to Google).

All the Google ML experts in the world don't matter if Google translate isn't staffed(no idea if it is or it isn't). It's quite possible that deepl has more engineers and ML experts working on the problem than Google does. Google's deep pockets and easily transferrable roster does have the ability to let it play catch-up very quick though, if anyone with influence thinks deepl has gotten too far ahead.

> It's quite possible that deepl has more engineers and ML experts working on the problem than Google does.

According to this article, they had 22 employees at launch and back then it was already lauded as quite good: https://www.gruenderszene.de/allgemein/deepl-maschineller-ue...

It's even possible that the Google ML experts left Google and went to work somewhere else.

Reply on news.ycombinator.com