News for hackers. Stuff that matters.

Baidu Research Announces Breakthrough in Simultaneous Translation

https://simultrans-demo.github.io/

We are excited to announce STACL, the first simultaneous translation system w/ integrated anticipation & controllable latency. This is a major breakthrough in #NLP due to word order differences between languages and simultaneity requirement. Demo & paper: https://bit.ly/2yTCvjE
( 17 comments | 130 points | on HN )
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by sarabande on Tuesday October 23, 2018 @04:56PM (#18287405)
Really cool demo! Just a note, the last example is corrupted at least for me in Firefox on mobile (but not Firefox on a computer).

Curious how this would work translating between a language like German where verbs sometimes are placed at the very end. Or even if clauses make things complicated. You wouldn't know the meaning for a long time -- or would you guess and amend?

To make this concrete (DE <-> ZH) with EN provided for clarity, apologies for errors:

[EN] I didn't go out last evening, because my mother had warned me about feral cats on the street.

[FROM ZH] 昨晚我不去出因为我妈妈警告我街上有野猫。

[TO DE] Gestern Abend bin ich nicht rausgegangen, weil meine Mutter mich vor Wildkatzen auf der Straße gewarnt hatte.

In that German sentence, the key verb in the second clause "to warn [warnen]" happens nearly at the end, but in the beginning in Chinese [警告].

So if I'm halfway through the second clause of my Chinese sentence, ...因为我妈妈警告我, the simultaneous translation in German is "weil meine Mutter [.....]" and then all at once "mich vor Wildkatzen auf der Straße gewarnt hatte" when the Chinese sentence finally ends.

I suppose you could split the clause again to give some "early" meaning in the sentence in real-time: "weil meine Mutter mich davor gewarnt hatte, dass [...]" and then the bit about the wild cats when it shows up in the Chinese sentence, but I'm sure you can construct more pathological examples so my question still stands.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by zmk_ on Wednesday October 24, 2018 @11:59AM (#18293563)
Reminds me of the joke about interpreters at the European Parliament. It goes something like this: there are always 3 waves of laughter when an MEP says something funny. First are the natives, then most of others who rely on translation, last are the Germans who wait for the verb.
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by captainmuon on Wednesday October 24, 2018 @01:19PM (#18294407)
You can turn it around, it's the English who have to wait for the object.

I think in such a translation system (to text) you could just insert a placeholder and replace it later. Or (in speech) you'd just delay the sentence until it is complete. If you reach a timeout, you just render it ungramatically in the wrong word order. Especially when speaking freely and making up the sentence on the spot, you could say something like "because my mother... about wild cats on the street... [uhhm, you know] ... she warned me".

Chinese also sometimes has "strange" word order for western ears, when you just state the "topic" upfront instead of building a relative clause like you would in English or German. "[Talking about] wild cats on the street, my mother warns me".

A live translation system will always have a tradeoff between accurracy+elegance and latency, just like real human translators.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by comex on Wednesday October 24, 2018 @03:54PM (#18295791)
The example in the "Demo IV" video seems to be like that. I don't know Chinese, but this is directly from the video:

–

Word-by-Word Translation: river zemin correct united states president of speak express regret

Simultaneous Translation (wait 3): jiang zemin expressed his welcome to the us president 's remarks .

Simultaneous Translation (wait 5): jiang zemin expressed his regret over the us president 's remarks .

Baseline Translation (greedy): jiang zemin expressed regret over the us president 's remarks .

–

The "wait 3" model and "wait 5" model have a fixed 3 and 5 words, respectively, of "buffer" before they have to start translating. In this case, the Chinese word corresponding to "regret" comes between 3 and 5 words after its corresponding position in the English sentence. "Wait 5" is able to see it, but "wait 3" simply has to guess what goes in that position – and what it guesses is almost the exact opposite of the actual meaning! And it doesn't seem to be able to revise its translation after the fact.

To be useful in practice, you would definitely need to be able to do that. It seems like it would also help to dynamically vary the buffer size based on the amount of ambiguity – but according to the paper on arxiv, previous systems already did that, while theirs intentionally throws it away in favor of a fixed buffer, apparently because it's simpler to train and because they prioritize low latency.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by paraditedc on Wednesday October 24, 2018 @11:16AM (#18293142)
This is not as hard as you think, and your example in Chinese is not complete.

The example you gave is idiomatic Chinese:

> 昨晚我不去出 , 因为 我妈妈警告我街上有野猫 。

However, this is also idiomatic Chinese:

> 我妈妈警告我街上有野猫 , 所以 昨晚我不去出。

The order of two parts can be interchangeable depending on what you want to emphasize.

So I think any AI that is complex enough can handle this difference in ordering quite easily.

Also, if you see the gif before Demo I, it implies that the word "meet" is inferred before it actually appeared, which is exactly what you described.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by paulgrant999 on Tuesday October 23, 2018 @08:55PM (#18288816)
translation keeps correcting itself.

:)

same thing you do when you listen and try to guess where the conversation is headed. ;)

also with the same flaws when you guess wrong.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by 2pointsomone on Tuesday October 23, 2018 @08:01PM (#18288557)
Is the code for your paper available in a Github repo?
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by dis-sys on Wednesday October 24, 2018 @11:19AM (#18293161)
just wondering when there will be a paper submission to a top ranking conference in the area actually AI translated from the native language of its authors. or maybe it has already been done before? :)
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by ramblerouser on Monday October 29, 2018 @11:18PM (#18333784)
So this article is the only thing this poster has ever submitted, which is a bit suspicious in itself, but additionally the account being non-civil in attacking my criticism has basically only commented on this one article. Pretty strange I'd say.
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by SamBam on Wednesday October 24, 2018 @02:04PM (#18294840)
There's a quote from a book that I can't quite recall, but related to German professors droning on with ever-more deeply nested sentences, and finally finishing with all the verbs at the end, leaving the poor students utterly lost as to which verb went with which sub-clause.

Anyway, great story, I'm sure, but it seem apropos.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by daminglu on Tuesday October 23, 2018 @07:34PM (#18288432)
Awesome demo!
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by ramblerouser on Wednesday October 24, 2018 @12:01PM (#18293593)
In Demo IV it's clear that this is intended to be used without correcting previous output. So with a look-ahead of 3 words it outputs the exact opposite of the intended meaning. For this to be practical, the look ahead probably has to be the whole sentence, defeating the purpose.

The propaganda push in Western media to build up and legitimize Chinese academia and technology ventures is staggering. Check out that map at the bottom. This article is more popular in EU and US than China itself. 99.9% of the chinese economy is built around stolen or borrowed technology. So what would be a minor development by a western researcher is a "breakthrough" for Baidu.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by corey_moncure on Wednesday October 24, 2018 @02:40PM (#18295157)
This is one reason that interpreting between SVO (e.g. English) and SOV (e.g. Japanese) ordered languages can be so fatiguing. You have to constantly buffer the sentence in your working memory while you wait for the completion before you can output a conventional sounding interpretation. It's even worse with drop-heavy languages, where key words are frequently omitted from the sentence and left to contextual assumption.
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by tareqak on Wednesday October 24, 2018 @03:29PM (#18295588)
I don't see where your accusation about this being a Western media issue comes from in this particular case: the link is the project's GitHub repository. Moreover, it makes sense that USA and EU together have more media companies than China does by itself. Finally, this project is about better language translation technology, which is something that can help bridge cultural divides instead of leaving cultures in isolation.

I don't know enough about languages to be able to tell if some advance is a major breakthrough or not, but people on Hacker News share and applaud minor improvements in many personal or professional projects, so I don't see why a Chinese project should be held to a different standard on this forum in that regard.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by withhighprod on Wednesday October 24, 2018 @01:48PM (#18294697)
Keep thinking in that way, it’s great for your mental health
Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by ramblerouser on Wednesday October 24, 2018 @02:01PM (#18294818)
I know you mean well, but please take a look at the Hackernews guidelines linked at the bottom of the page, specifically the following two passages.

>Don't be snarky. Comments should get more civil and substantive, not less, as a topic gets more divisive.

>Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something.

These are great guidelines, I really appreciate the thoughtfulness of whoever wrote them.

Re:Baidu Research Announces Breakthrough in Simultaneous Tra... by withhighprod on Wednesday October 24, 2018 @05:11PM (#18296433)
Sorry since it’s hate speech in the first place, I couldn’t help