The Language Barrier Is About to Fall | Hacker News Reader

The Language Barrier Is About to Fall

wsj.com

35 pointsTrisell57 comments

This article is BS. Machine translation has been improving but it's still nowhere close to being able to provide real-time translation of spoken languages, and I bet anything it won't be there after another decade. I know for a fact that Baidu is throwing lots of money and people at fundamental R&D in translating Asian languages, and it's no piece of cake.

As for "earpieces whispering nearly simultaneous translations", it would be a cool party trick but I doubt they'll work in real life.

These moonshot projects have been around for decades, and they make for some cool entertainment at conferences and demos in very tightly controlled settings. As for the predictions that "the software in the cloud connected to the earpiece in your ear will re-create the voice of the speaker", the author should take up writing bad science fiction.

It depends on what the bar for "good enough" and "fast enough" is. Remember that Google Translate, just one example, was pretty feeble to start but has proven itself to be adequate for most needs now even if a lot of the meaning gets lost in translation.

Could that be paired up with Google Voice? Absolutely.

It's not easy, but it's inevitable.

I wouldn't say there are that many situations where Google Translate is good enough, to be honest. It's not so different from an automatic way of looking words or phrases up in a dictionary.

I can't remember who said it, but someone said technology changes less in 1 year than anyone expects, but more in 10 years.

Translation, however, is a tricky bit. Due to the nature of the structure of language, it would have to be "say a sentence, wait, respond, wait" which would slow down the course of a conversation significantly.

I think you could get to usefulness for simple transactions today if you're google or apple for a romance language in a standard accent. You would probably have to download something beforehand to preload the soundex for that set of accents, but it's within the power of your phone to do so. You'd probably need the top 3000 words + food + touristic places + top 5000 idioms in a set of common accents for the region, but that's not too bad compared to talking about something like politics.

The top phones have the power to do that today.

I'll just leave these here.

July 2014: https://www.youtube.com/watch?v=XuWN5AxlgxU

http://www.skype.com/en/translator-preview/

I'm currently living in a country where I don't speak the language. I can't believe how incredible translation tools are. Even though I don't understand when people speak to me, I can write and read emails, and accomplish things that would be incredibly more difficult without this ability (e.g. talking to my insurance agent). And when I say I can read and write emails, I mean it can do so in about 1.5x longer than a native one. I have developed very basic abilities in the language (in the 2 months I've been here), and so I constantly use translation tools and check in both directions -- translate to the language, modify it to make sense according to my understanding -- translate it back to English -- all within the space of a couple of minutes.

I still miss out on learning more idiomatic expressions, but I imagine those will come in time. In the meantime, I am able to get a surprising amount of business done, in ways that would have been impossible before.

Not to mention that continuously doing all this translation of my written communications is making me learn the spoken language much faster. Although, for that, nothing replaces actually speaking with people, of course! ;) But it's much easier to bootstrap actual conversation when you've been learning individual words by reading and writing.

Overall I would say that automatic translation tools have made it possible for my life to move in directions that where language would have been a serious roadblock in a previous era.

What is the other language, and how similar is it to English (or other languages you know)? (Beofre you answer, my predicted answer is "very similar".)

Similar enough, certainly. More similar to my second language. (I know French, I'm learning Spanish.)

It definitely lowers the barrier, but don't think there's no barrier just because it's a "similar" language. I'm claiming that automatic translation has accelerated my learning, not that learning it would be impossible otherwise.

Nonetheless it is quite difficult to live somewhere where you don't speak the language, no matter what tools you have available. Mostly it doesn't matter, but when you find yourself arguing with administrators... or the police...

(I almost got into a bike accident the other day. I wasn't even all that worried about my health, I was mostly scared of having to talk to people, hospital/police/insurance..! or argue with the guy who hit me...)

I suspect real-time machine translation is within reach if the speakers are trained to use a restricted subset of their native language. Moreover, I suspect that this approach could be both realistic and useful.

This restricted subset situation is de facto already the case; existing tools don't handle regional dialects and colloquialisms. I just tried "嘘じゃねいぞ" (uso ja nei zo!) in Google Translate and it came up with rubbish because it doesn't know that "nei" is a variant of "nai".

The kinds of restrictions I'm thinking of are deeper: avoidance of the full variety in sentence structure, and excessive complexity, like compounded of relative clauses. Not to mention language-dependent tricks like double meanings depending on puns, and generally all figures of speech that do not translate.

A restricted subset of your native language is something which you already understand, and learning to speak it is easier than a whole different language.

People doing business abroad would just study that for a few weeks as part of the workflow.

The dialect could become richer and thus less restricted as time goes on, making it easier to use. It could be customizable as well, to a particular speaker's idiosyncrasies. If a given speaker often uses some figure of speech out of habit, it could just be defined for that speaker, with a hand-crafted translation to various languages.

> A restricted subset of your native language is something which you already understand, and learning to speak it is easier than a whole different language.

It's interesting to note that before machine translation, there have been attempts to try to standardize restricted subsets of English as a way of making it easier for non-native English speakers (for example "Basic English" and "Basic Global English").

However, I think both in the case of communication with non-native speakers and in the case of machine translation, it's actually not that easy for native speakers to anticipate what phrasings will be problematic.

These existing efforts have had the problem that reducing vocabulary can actually make the resulting language more reliant on things such as verb colocations that are actually harder for non-native speakers. Machine translation has an advantage that it can presumably understand a rich vocabulary, so it might be possible to get somewhat better results by using as precise words as possible while keeping the sentence structure simple. Still, especially with modern statistical machine translation, and perhaps even more so in the future with recurrent neural network-based translation, it may not be that easy to predict potential pitfalls in an attempt to work around them. It actually might have been easier to train people to get better results (at least for specific language pairs) with older, rule-based translation systems, since they were more predictable.

In either case, there presumably needs to be a certain level of accuracy before people can start attempting to tweak it by their choice of wording. I don't think we are there yet, especially for English <-> Japanese. I just tried entering "I'm not going to tell him that." in Google translate, and it gave me "私は彼のことを言うつもりはありません。" ("I'm not going to mention him.") This is a fairly simple sentence, with no colloquialism, and yet Google Translate fails utterly. This is much more concerning than something that can be easily worked around, such as not knowing colloquialisms such as ”じゃねー".

Obviously the program must have figured out that there are two clauses: "I'm not going to X", where X is "tell him that". How the heck does it go from there to "kare no koto wo iu". If we translate just "tell him that" alone we get ことを彼に伝えます which is okayish. If it would just stick some version of that into the "watashi ha ... tsumori wa arimasen" pattern, it would be progress.

Google Translate can't handle even perfectly idiomatic Japanese. It's so bad I had to write my own translation aid software (shameless plug: http://ichi.moe).

For example, Google can't handle an extremely common Japanese word なんて. It translates it as "Nantes" (the city in France). Even though the city name would certainly be written in katakana and not hiragana.

I've been learning Japanese as of late, and automated translation tools are near useless for anything but sentences that are just a few words stringed together (and even then, nonsense is often returned as soon as a tiny bit of slang or metaphorical language is used). I also have friends who speak languages that are not part of the top 10 most spoken in the world (eg Finnish, Hungarian, etc) and trying to translate their Facebook statuses returns complete garbage.

Between languages that are fairly close to one another (such as French and Spanish), sure - you can pipe text from one language to another in Google Translate and the results will be mostly understandable (although still very awkward). But anything else, and automated translation still has ways to go. Turns of phrases, idioms, general syntax etc. are fairly similar between languages that are closely related, so automated translation has an easy job there (with the caveat that it starts falling apart as soon as slang or more colloquial expressions are used). But step any further than that, and it's a lost cause.

The article also assumes that there's a 1:1 mapping between everything you can express in every language, and that it's just a matter of software finding the right mapping for whatever you're saying into the destination language. But any translator will tell you that this couldn't be further from the truth - there are many things I can say in English to my Californian friends that don't really work in a French conversation with my Parisian friends (let alone in Mandarin or Japanese). At best you'll be met with blank stares, and at worst you'll be breaking etiquette in major ways. Part of learning a new language is learning the space of what can be expressed and how in that language.

For these reasons, I don't really buy the whole "earpiece that automagically lets you speak with anyone in the world" anytime in the next few decades. Which is fine, because learning languages is fun :)

On the other hand, I'm currently in the Netherlands for a conference, and everyone speaks flawless English. I've had a similar experience in Germany, Sweden, and a bunch of other countries (but not in my country of origin, France, where being terrible at English seems to be a point of pride). So maybe taking a page from their book when it comes to education is a better idea than waiting for technology to let us be lazy.

Exactly. I've been taking Japanese recently, too, and the grammatical structure of the language alone is enough to show that the "real-time whispering earpiece" as an impossible concept: in English, the verb usually comes near the beginning of a sentence. In Japanese, it's near (or at) the end. Translating one language to the other therefore requires a mental "tape delay", while you map the one grammar onto the other.

Even when you hear bilingual, fluent translators converting Japanese to English, there's often a weird disassociation between the words the speaker is using and the words coming out of the translator. The translator is buffering the words and translating in logical units. Even if you grant the benefit of the doubt and call this "real time" translation, I know of know translation system that is capable of any sort of similar kind of judgment with an unbounded input stream. Deciding where to begin/end translation requires knowing what the original sentence intended to express. And computers still suck at that.

German, one of English's closest languages, also does the "verb at end" for many constructs (e.g. past tense, future tense, "must"/"should"/"can")

I think people are overestimating the word "language barrier". People often forget that the language barrier is not a one-sided problem. Communication is a collaborative effort (unless we're just spying them). I bet that something like keyword spotting for commonly used 3000 words is already pretty feasible, and that would greatly reduce the language barrier. Perfect translation is neither possible nor necessary to induce mutual effort.

(By the way, I've seen so many US tourists in Europe who don't even try to speak a single word in their language. They don't seem to know that a simple "thank you" in their native tongue would greatly reduce the barrier and change their sentiment towards the US. In this sense, I believe that technology is more or less neutral to mutual understanding, because not knowing a language (or the idea that not knowing a language is acceptable) can have adverse effects.)

Edit: a few grammatical errors. Damn my Japaneseness!

This. Two years ago I visited Paris for the first time. I knew four phrases: hello, please, thank you, and "my name is.." and I expected ridicule and people being rude to me left and right.

It turned out that attempting to be polite in French bought me enough good will and sympathy that I was able to communicate without getting rude looks or annoyed responses. I may have a different experience in the French countryside - less tourists, less patience? - but I was pleasantly surprised the whole time.

> (the idea that not knowing a language is acceptable)

It's not? I'm in big trouble. There are more than 6,000 languages I do not know.

The point is probably "it's not acceptable to assume 'everyone knows English, but most are just too lazy to use it'" rather than "of course you're supposed to be fluent in any language" ;)

Yet another journalist who doesn't understand how languages actually work.

Other languages aren't just "English with different words and syntax". A single word can have dozens of different meanings, the set of different meanings of a word needn't be the same across languages. And even if you know the context the meaning doesn't have to be unambiguous. A perfect translation would need to recognize plausible ambiguities and retain them across languages.

Language is not code. Language is an expression of thoughts in a form that builds upon vaguely defined tokens. It's amazing that human language communication works this well at all given how ambiguous and faulty it actually is in practice.

Even English isn't English as innuendo and alternate meanings cloud understanding of it constantly. Even the simple phrase "I'd hit that" means any number of things depending on the speaker and the context.

This doesn't even touch on how sometimes British English needs to be "translated" to American English, like in the case of Harry Potter being adapted with many words changed.

Likewise it's almost impossible to express the nuance of something like "that's so fucking fucked" in a language lacking a word of similar versatility.

I'd argue there's no such thing as a "perfect translation". That's why people who translate novels have a considerable amount of work to do to pick, from all possible translations, the one that best represents the tone and intent of the original author.

I still think it's possible to come up with a passable translation automatically in real-time. It's inevitable.

Sure, but translations are lossy. No 1-to-1 translation can ever completely emulate perfect fluency.

You can get pretty close though ;) https://www.youtube.com/watch?v=akbflkF_1zY

The problem is that we still haven't been able to write a system to accurately recognize the words that are being spoken. Differences in diction, dialects, and accents make it difficult to accomplish. Just look at what Microsoft, Google, and Apple are doing - as good as they are, they're nowhere close to where they need to be.

Even once we solve the voice capture issue, I suspect that we'd still run into many areas where the computer would have problems determining the proper context necessary to accurately translate the text. Whatever system is performing the translation would have to maintain a history of the conversation in order to have any hope of understanding context, and at a gathering where you could have multiple conversations occurring at the same time, the process gets even harder.

I think the only part of the author's belief that we're close to achieving is being able to have the computer use the original speaker's voice. Siri and Cortana could already sound more human, but people are currently more comfortable having them sound robotic. I'm sure that will change in time, however.

"...in ten years time." To quote XKCD's translation (sic!), "we haven't finished inventing it yet, but when we do, it'll be awesome." https://xkcd.com/678/

This is the same tired old story that keeps on making rounds for the sixth decade now - "machine translation kind of sucks now, but it will be solved within the next decade!!!!!!!!1!" This is still very much a current paper - note the publication date: http://www.mt-archive.info/Bar-Hillel-1951.pdf

To be fair, translation today is significantly better than sixty years ago. For one, the output is often usable as-is, and with some tweaking is even better.

I've known several people that have learned English by leaning heavily on automated translation until they got a firmer handle on it. If you tried that in the 1970s you'd still be trying to figure out what the word "the" means.

Sure, it's better - in the "pre-MT" and "post-MT" systems (although it's called "crowdsourcing" these days). For standalone MT, the output is just-about-recognizable; nowhere near the babel-fish-like gadget WSJ is presenting as "just about to happen."

真有意思。I see a lot of dismissals of the projections in this article. Thirty years ago, just as I was working for several years as a Chinese-English interpreter, I would have agreed with all those dismissals. I no longer agree. As some of the subcomments here have pointed out, all that automated translation and interpretation have to do is become less expensive than either a) learning the relevant language yourself or b) hiring good-quality human translators and interpreters. Automated translation and interpretation is already well along that path.

A crucial skill for a professional interpreter or a casual language-learner is "lexicalization," figuring out what the correct match is between what is possibly a phrase in one language but only a single morpheme that is part of a larger word in another language. But the huge advantage automated translation solutions have is shared learning and long-lasting memory. If the lexicalization is built into a database, for example a database kept by Apple, Google, IBM, or some other multinational provider of translation services, that learned information can be deployed in products over and over and over again. In the end, that's all that it takes for automated translation and interpretation to do better at equal cost (or as well at less cost) compared to human language-learning or human language services. I got out of interpreting for a living years ago, because I finally came around to the understanding of how little I could compete with the worldwide efforts made with new technology to tackle language problems. I still think learning natural modern human languages is a very intellectually enriching activity, and all my children do that, but we expect to live in a world a decade from now with very good real-time interpreting and translation systems for a wide variety of language pairings.

The idea that one can translate words without understanding what is being said, seems rather similar to an idea that one can infer the meaning of words from their constituent letters. Meaning not only works simultaneously on so many technically inaccessible levels (context, expectations, allusions, previous experiences, etc.), anything mildly interesting (not to mention proper literature), is based by definition on a deviation from established meanings.

So, we need to translate text to an artificial language, which is well understood by both humans and AI, and then render same meaning in native language of listener.

Slovakian or Ukrainian language should be good base for that artificial language, because they are easy to parse and are well developed, with large enough dictionary to cover most cases. Moreover, Ukrainians are very cheap right now, because of war with Russia.

Compare this line from the WSJ article:

> A decade from now, I would predict, everyone reading this article will be able to converse in dozens of foreign languages, eliminating the very concept of a language barrier.

with these excepts from The Guardian at http://www.theguardian.com/technology/2016/feb/10/texas-regi... (referenced two days ago on HN at https://news.ycombinator.com/item?id=11077168 ):

> Y'all have a Texas accent? Siri (and the world) might be slowly killing it

> Voice recognition tools such as Apple’s Siri still struggle to understand regional quirks and accents, and users are adapting the way they speak to compensate.

> ... “I’ve had a bunch of people from Australia and India say they only really get along with Siri if they fake an American accent,” said Lars Hinrichs, a sociolinguist at the University of Texas at Austin.

To be fair, having not tried Siri for a year or two, it has gotten much better at dealing with the Glasgow accent ; )

I recall around 2000 having to fake a US accent in order to get an airline's voice response system to put me through to a human.

You probably have to fake an accent to communicate with some American call-centres as well.

It must work offline. One of the reasons is that tourists often don't buy a data plan (imagine a tour of Europe with a different country every couple of days) and wifi is not where you might need machine translation most, on the road.

Many European mobile contracts allow a bucket of roaming data (400MB per month for me). Not sure what's the status for prepaid SIM cards that tourists from outside of Europe would likely have.

Also, the roaming costs in EU are regulated, with likely price decreases also in the future: https://en.wikipedia.org/wiki/European_Union_roaming_regulat...

A number of prepaid providers are offering data roaming plans now. See this list: http://prepaid-data-sim-card.wikia.com/wiki/European_Union

If I access this article via a google search, it is still paywalled. I see experts-exchange seems to be doing this again also. I thought this was against google's terms & conditions, no?

Google has terms and conditions to be listed in search?

Ya, the page when you click through has to be the same as that indexed by google, I think it might be referred to as cloaking?

>You could host a dinner party with eight people at the table speaking eight different languages, and the voice in your ear will always be whispering the one language you want to hear.

And at this hypothetical dinner party, what would the earpieces whisper in English (or most other languages) when the Japanese guests said いただきます?

Some words just don't have a translation in other languages.

https://www.google.com/search?q=いただきます&tbm=isch

«Bon appétit» would be an adequate translation for most English speakers (even though it's French). A more "pure" English translation would probably be a grace of some kind, which serves more or less the same purpose (although I can also see some atheists bristling at a grace where they wouldn't bristle at いただきます or bon appétit).

Not really. My Japanese roommate would say いただきます when by himself at his desk before eating a candy bar. It would be surprising if even one English speaker in the world regularly says "Bon appétit" in that situation.

The question isn't, "How do you say いただきます in English." The better question is, "Do English speakers have a set phrase they say right before eating." And the answer is no they don't.

In context, the question is in fact what should come out of the English-speaking listener's earpiece to explain the "itadakimasu" utterance from the Japanese speaker, whether or not the English speaker has a custom of uttering anything before any meal or snack.

Suppose English speakers did always say "Bon appétit", even when eating alone and just a snack. It's still not the meaning of "itadakimasu". Wishing for gusto in eating is not the same thing as humbly receiving the meal/snack.

A pure translation would be some sort of grace that is either ambiguous or directed at the host who provides the meal, rather than a god. See my other comment about itadaku.

> Some words just don't have a translation in other languages.

Generally speaking I'm inclined to agree with you but you chose a bad example. The earpieces could translate it as "time to eat!" or "thank you for the food.". It's not a literal translation of course but a lot of the work involved with translations is about finding the right substitutes for these kind of phrases. In japanese there's a lot of mundane phrases that could better illustrate your point, for example how would an earpiece translate 好きです。 as a stand-alone phrase? It could do so literally but it would probably sound very unnatural. To do a natural translation the device would need to be understanding the conversation as a whole and that would be way more difficult to implement since it will involve some actual intelligence :)

懐かしい is another that comes to mind. It's like nostalgia but it's way more positive and can be used as an interjection. Or 我慢する. In fact, I'd say there's a huge array of words that really just don't translate well on their own and yet are often spoken as stand alone sentences.

Meaning just isn't chopped up along the same word boundaries in different languages.

I live abroad and speak far from fluently in a foreign language. I'm personally invested in machine translation improving, and I really hope it does.

> Today’s translation tools were developed by computing more than a billion translations a day for over 200 million people. With the exponential growth in data, that number of translations will soon be made in an afternoon, then in an hour. The machines will grow exponentially more accurate and be able to parse the smallest detail. Whenever the machine translations get it wrong, users can flag the error—and that data, too, will be incorporated into future attempts.

If this is the reasoning that we will be there in 10 years, it is ridiculously optimistic. I'm pretty sure that there is a exponential fall off from common phrases to a long tail of more nuanced, not as often uttered language. A million people punching in "where is the bathroom?" is not going to make the machine know how to impart the acerbic wit of a phrase from a Palahniuk book to a Japanese business man. Good luck finding someone skilled enough in both languages and cultures to be able to correct the error.

Real skills in multiple languages requires an intellect to parse the situations. While I personally hope that will one day be achievable with AI, I have to think that we are still a long way off.

The benefits to this kind of thing are being able to be very specific and go deeper into the subject matter with one's second or third languages. Plus I might argue that smaller languages won't need to die (if everyone can get along and do business in their own tongue). The drawback is that perhaps people who learn languages in the near future will seem like those who study Latin ever since it became a "dead" language.

Note the qualification of "small-er" languages: you are implying "languages which are still large enough that a translational layer is economically feasible. Small languages (let me pull a number out of thin air and assume <20 million speakers) would still be doomed. Guess what: that means English, French, Spanish, Italian, Russian and Polish for Europe, everyone else kindly GTFO, we don't have resources for implementing your languages.

Sorry, I don't quite follow your intended meaning, but I will try to expand on my initial comment below.

If it no longer matters what language someone speaks, a language with fewer speakers than others wouldn't need to become extinct, as will happen without tech-aided interpretation. Of course, the viability of a language isn't solely connected to the ability for a native speaker to do commerce in it, but I would imagine it's nonetheless an important factor.

From Wikipedia's article 'Lists of endangered langauges':

"While there are somewhere around six or seven thousand languages on Earth today, about half of them have fewer than about 3,000 speakers. Experts predict that even in a conservative scenario, about half of today's languages will become extinct within the next fifty to one hundred years."

I would hope that even languages with 3,000 speakers could be 'saved', in the scenario I propose (assuming a Wikipedic-style worldwide effort - in conjunction with the tech sector, governments and linguists - to get as many languages as possible implemented).

Well, that's the "let's assume unlimited resources (inc. time), suddenly nothing is a major problem" scenario.

The actual scenario tends to be "but you have a choice of all these 20 languages; are you just lazy to learn a proper language that's on the list?" My native language has about 10 million users, so I get this a lot (and consequently, most of my digital UX is in English - significantly less hassle for me).

Also German, Ukrainian, Romanian, and Dutch, and, if you count speakers who are geographically outside of Europe together with those inside it, Turkish.

I speak Arabic, French, English and Spanish; the last 3 are "easier" to translate interchangeably. Good luck breaking the Arabic barrier, each Arab nation has its own dialect and no one hardly speaks "traditional" Arabic anymore. The language itself is very rich and dense, even machine text translation has a long way to go, let alone real-time voice translation. This article is fiction at best.

I thought this was about how the entire world is getting to a point where they are able to speak and understand basic English.

There are certainly countries and population groups that do not speak English well, but it has changed significantly the last 10 years.

I wonder if we will get there before 100% accurate automatic translation.

The entire world*, sure.

(the part of it inside my social bubble; smartphone and internet connection strongly advised, other terms and conditions apply, void where prohibited.)

This "real time whispering earpiece" sounds an awful lot like a bluetooth headset...

For an example of the state of automatic translation c.2010 I give you this:

This is question, engish is faulty therefore the right excused is requested. Thank google to translate to help. SORRY!!!!!

At often, the goat-time install a error is vomit. To how many times like the wind, a pole, and the dragon? Install 2,3 repeat, spank, vomit blows

14:14:01.869 - INFO [edu.internet2.middleware.shibboleth.common.config.profile.JSPErrorHandlerBeanDefinitionParser:45] - Parsing configuration for JSP error handler.

Not precise the vomit but with aspect similar, is vomited concealed in fold of goat-time lumber? goat-time see like the wind, pole, and dragon? This insult to father's stones? JSP error handler with wind, pole, dragon with intercourse to goat-time? Or chance lack of skill with a goat-time?

Please apologize for your stupidity. There are a many thank you

(from https://lists.internet2.edu/sympa/arc/shibboleth-users/2010-... )

Reply on news.ycombinator.com