Will Machines Ever Master Translation?
Language translation is proving to be one of the hardest tasks to automate—and one of the most important
Steven Cherry: Hi, this is Steven Cherry for IEEE Spectrum’s Techwise Conversations.
January is a time for prediction, so here’s one: Google will have driverless production cars on the road before Google Translate gets used at the UN. Why? Because self-driving cars is a hard problem, but translation is a really hard problem. Let me give you an example.
Most of us are familiar with the King James Bible’s rendering of the 23rd Psalm, the one that goes, “Yea, though I walk through the valley of the shadow of death, I will fear no evil: For thou art with me; thy rod and thy staff, they comfort me.” Listen to two other translations:
First, Robert Young’s 1862 version, usually called Young’s Literal Translation:
Also—when I walk in a valley of death-shade, I fear no evil, for Thou [art] with me, Thy rod and Thy staff—they comfort me.
A bit stilted, you say? How about something more modern, like the Contemporary English Version:
I may walk through valleys as dark as death, but I won’t be afraid. You are with me, and your shepherd’s rod makes me feel safe.
Neither of them is wrong, yet they sound wrong. As I said, translation is really hard. How does Google Translate do? It’s so bad, it’s too painful to read. Instead, here’s another bit of the 23rd Psalm, the part that in the King James version goes, “Surely goodness and mercy shall follow me all the days of my life: and I shall dwell in the house of the LORD for ever.” Here’s Google’s version, translated from the Hebrew:
And yet, goodness and mercy will pursue—all—days of my life, and I returned home—Lord took days.
Like a talking dog, the surprise isn’t that it does it well, it’s that it does it at all. Language translation is really hard. But unfortunately, in an ever-shrinking and ever-more-connected world, it’s more important than ever, as a new book, Found in Translation: How Language Shapes Our Lives and Transforms the World, makes clear. Along the way, the book is chock full of funny anecdotes and tragic tales. It was published by Penguin Group in October, and it’s written by Jost Zetzsche and Nataly Kelly, who is my guest today.
She’s the chief research officer at Common Sense Advisory, a market research firm at the intersection of language services and technology. She also blogs about translation for The Huffington Post. She joins us by phone.
Nataly, welcome to the podcast.
Nataly Kelly: Thank you for having me, Steven.
Steven Cherry: I want to get to Google Translate, but first there’s a very quiet little story that you tell in the book about the 2010 Haiti earthquake. I think it’s a good reminder of how important translation is and how technology can make it better.
Nataly Kelly: Yes, well, that’s one of my favorite stories in the book. It’s about how translation actually helps save lives in the aftermath of the earthquake in Haiti. And so essentially there was a service that was built to enable volunteers to text-message their translations via mobile phone so that the relief workers and disaster responders could actually understand what people were saying on the ground. So this was a project that was by Rob Munro from Stanford, and he had been developing methods to process large volumes of SMS text messages in less-common languages, and it just so happened developed that they were able to use this for helping the rescuers save victims of the earthquake in Haiti.
Steven Cherry: And the problem here was that the Haitians speak a language, a Creole, that almost none of the rescuers spoke.
Nataly Kelly: Exactly. Most of those outsiders that were coming from other countries were coming from the United States or Europe and did not speak Haitian Creole, which is the language that is spoken there, which is heavily influenced by French but still is a separate language.
Steven Cherry: I think if people have a picture in their minds of translation, I think for many of us it comes from the movie Judgment at Nuremberg. Movies usually skip over the whole question of foreign languages because it really slows a movie down, but Judgment at Nuremberg took the time to show questions asked in English, translated into German for the prisoners, German response, and then the translation back to English. It’s kind of ironic that this is our picture of translation, because the trial itself in Germany used some very innovative technology. Maybe you could tell us about it?
Nataly Kelly: Sure. Well, this was one of the first times that simultaneous interpretation was used for a major world event like this. And it was necessary because of the number of languages involved, as well as the need to have communications with so many different people. So this was quite challenging. There were very few interpreters that actually had those skills, to be able to interpret, simultaneously listening and speaking at the same time. Prior to that, most interpretation was done consecutively—you know, someone would listen and then wait and then interpret afterward. So this was done primarily to speed things up and to enable other people to interpret simultaneously into other languages. So this was a pretty important milestone for the history of interpreting.
Steven Cherry: Just to give people a sense of the level of detail in your book, there’s an ironic, almost cruel, twist to the actual trial at Nuremberg—not the movie, the trial—concerning one of the translators. What was that?
Nataly Kelly: Yes, well, this is another of my favorite stories in the book about a gentleman named Peter Less. Peter Less was one of the interpreters at Nuremberg, and his entire family was actually killed in Auschwitz, and he was interpreting for the masterminds who were responsible for the deaths of his family members. So talk about a situation in which a human has to become a machine, to disconnect and detach their emotions. That really was a situation that is very difficult for most of us professional interpreters to understand, how he could maintain neutrality and impartiality while interpreting. But he did, and he was well commended for his job as an interpreter there.
Steven Cherry: That’s almost a perfect segue to the future of machine translation. I think it will be news to some of our listeners, but not entirely surprising for them to hear that Google has turned translation into a big data problem. So tell us how they do it.
Nataly Kelly: Yes, well, so basically there are two types of machine translation or automatic translation, no humans involved in the actual translation process. And the first kind is kind of the older kind, which is rules-based, and what that means is that people were creating, coding different rules to make sure that if you put in one type of input, you get a certain type of output. So, for example, I’ll give you a simple example that everyone can understand. Most English speakers know a little bit of Spanish: The word casa, which is “house” in Spanish, and blanca, which is “white” in Spanish, so casa blanca means “white house.”
What Google is doing is something completely different: It’s called “statistical machine translation.” So what Google and Microsoft and others are doing is relying on large amounts of data in order to train their machine-translation engine. So it’s looking at the fact that if it sees casa blanca appearing, you know, 500 000 times and it’s translated as “white house” most of the time, then it’s going to show you that result. So it’s basically looking at the number of occurrences along with other data and contextual information to give you what it’s predicting to be the correct translation. There are times when it might appear in an article, or has other context information, that it might translated as Casablanca the movie title, you know, and maybe it will say, “If there isn’t a space between casa and blanca, then we’ll leave it in Spanish.” So with statistical machine translation, you’re basically relying on massive amounts of data, whereas with rules-based, you’re getting down to the nitty gritty and trying to capture the critical rules of the language that’s in question and how they work.
Steven Cherry: And extra rules for all of the exceptions…
Nataly Kelly: Absolutely. And that’s where it gets really tricky, because language is full of exceptions [laughs].
Steven Cherry: So just to be clear: The data that Google and Microsoft and others are relying on tend to be things like U.N. documents that are translated into multiple languages, or maybe Web pages that have been translated by humans from one language to the next. And they take all of these documents, which in the case of big languages—that is to say, popular languages—might be lots of data, and in the case of other languages, might not be that much data. And I guess the big successes come when either there’s a lot of data that is direct translations between the two languages.
Nataly Kelly: Not only the data, but the quality of the data. If it’s relying on a bunch of translations that were poorly done, then it could be giving poor results, because statistically there are more instances of that poor translation than there are of a good translation. Of course, the measure of what is a good translation and what isn’t is just as subjective as which of those passages—the translated passages you read— is the best [laughs]. You know, one person might think that the King James version is the best because it’s more popular and more well known. Others might prefer other versions or other translations. So that’s where it gets very complicated. But generally you might see over time that with a system like Google Translate, you get a poor translation, but then you come back a month later, a year later, and you get a better translation. What’s interesting about this is as more data comes into the system, it’s getting better and better, so it is changing, and you’re absolutely right: Google Translate can’t even launch a language unless there is a sufficient amount of data to train the engine and to ensure a certain level of quality.
Steven Cherry: Yeah, and you note in the book that Google usually waits for their translation engine to get to a certain level of quality before releasing it at all between two languages. So they don’t always wait though, right? In the Haitian earthquake, and you also mention in the book the 2009 Iranian elections, they didn’t hold back; they released the engine a little earlier than they might have. Isn’t that kind of risky? Elsewhere in the book you describe some of the dangers that bad translations can do: People giving the wrong kind of medication, serial rapists going unpunished, even fanning the flames of religious warfare in one case.
Nataly Kelly: Yes, and in fact those were actually due to human translation errors, all of those examples. But, yes, a single mistranslation can be devastating, and this is why a lot of people are right to caution the use of machine translation for different applications.
So machine translation can be extremely useful in legal settings, for example, where there’s a discovery phase, and throughout that process, hundreds or thousands of cases have to be mined to see if there are instances of a certain charge or a certain situation. And so it can be very useful to kind of scan and get the gist of information, and it can actually be useful.
I don’t want to make machine translation all be equated with publicly available free translation like Google Translate. Rules-based translation and even some statistical and hybrid machine translation tools have been used very successfully in certain contexts. But definitely when you have a situation, like a patient comes into an emergency room, you don’t want to be using machine translation for that. And we have heard of cases where doctors are trying to use Google Translate on their phones to communicate with a patient about his symptoms, and that is certainly not a recommended use at this point in the development of the technology.
Steven Cherry: In a recent blog post, you call attention to the fact that Google just hired Ray Kurzweil, who, it turns out, this is kind of a strange development. Kurzweil is about as big a cheerleader for technology as there is, but he’s something of a skeptic when it comes to machine translation. Why is that?
Nataly Kelly: Well, he actually believes that machine translation will get to human translation quality, but he believes that, at least when I interviewed him for the book and for another Huffington Post article that I wrote, he doesn’t believe that that means that human translators will go away. He believes that their profession will evolve, and in fact it is already evolving, you know, more and more human translators are gaining experience with machine translation.
We did a recent survey at Common Sense Advisory, and it showed that 20 percent of freelance translators had already used machine translation on a project. But in terms of Ray Kurzweil, he does believe that machine translation has great promise, and in fact he, when I interviewed him, he mentioned Franz Och, who is the mastermind behind Google Translate. He mentioned that he knew of him from his time when he was at the university and he was developing his research in that area, and he thought that statistical machine translation had great promise. So it’s interesting now to me that the two are coming together.
Steven Cherry: It is indeed. Well, Nataly, it’s a huge area. As you note in the book, it’s a $33 billion industry right now, and as commerce and culture continue to go global, it’s only going to grow, so thanks for writing such an enlightening and entertaining book about it, and thanks for joining us today.
Nataly Kelly: It’s my pleasure. Thanks, Steven.
Steven Cherry: We’ve been speaking with Nataly Kelly, coauthor of the new book Found in Translation: How Language Shapes Our Lives and Transforms the World, published by the Penguin Group’s Perigee Press [Perigee Trade].
For IEEE Spectrum’s “Techwise Conversations,” I’m Steven Cherry.
NOTE: Transcripts are created for the convenience of our readers and listeners and may not perfectly match their associated interviews and narratives. The authoritative record of IEEE Spectrum’s audio programming is the audio version.