Overhyping Neural Machine Translation: “It’s very easy to beat out of the box Google”

New people come in through our office every day, we’ve kind of gotten used to it. But, when one of the fathers of machine translation walks in through the door, it’s reason enough to jump out of one’s skin.

That’s what happened when Professor Andy Way of the ADAPT Center at Dublin City University joined us in Lisbon for the first episode of Understand with Unbabel.

Professor Andy Way has been working with machine translation since 1988 and he is one of the world’s top researchers in the field. However, as he likes to point out, even though he has been working on machine translation for long enough to be called one of the fathers, people were doing machine translation a long time before he began.

---

The beginning of Machine Translation and the skepticism around it

Machine translation has gained much traction during the last decade but, believe or not, people have been working on it for almost 60 years. Back in 1949, according to Professor Andy:

“There was a famous letter written by Warren Weaver talking about the possibilities of machine translation, that we could look at some of the things that were learned from the war about coding and decoding and encoding secret messages, and try to use the same sort of technology to process human languages”.

That’s when it all started.

However, the initial enthusiasm didn’t last long and was naturally replaced by skepticism, once an infamous report essentially predicted that they’ll be no future for machine translation because of the problem of ambiguity that machines will never be able to disambiguate.

Words like pen, whether it’s a writing implement or whether it’s an enclosure that you keep animals in would be to much of challenge for machine translation to work.

Despite the skepticism, researchers were able to turn the tables since then and machine translation has become one of the biggest trends in tech over the last decades, particularly Neural Machine Translation (NMT). So much so that it seems to be the next big thing that everyone is talking about.

---

Neural Machine Translation: will it remove humans from the equation?

“Neural Machine Translation has come along and really looks like sweeping everything before. It’s rapidly becoming the new state-of-the-art”.

Neural Machine Translation is like having a computer system acting like a brain simply by imitating biological neural networks. This kind of system feeds itself with data in order to progressively learn and consequently improve its translation.

That’s actually the system we use and develop at Unbabel. But, will it replace humans in the translation process?

Despite the big frenzy, according to Professor Andy “this hype needs to be treated with some caution”, because it can’t work on its own to provide an accurate human translation.

Watch the interview below:

Google has even said that they were coming very close to human translation quality in a recent paper from 2016. “As soon as you hear that, translators start to get a bit freaked” said Andy, and “the last thing you want to do is alienate the translators”. Why? Because “the human in the loop will always be the most important part of that translation pipeline”.

Therefore, humans are a crucial part of the equation, and Professor Andy Way tells us why:

“The machine translation quality is often very, very good indeed, sometimes, misleadingly good. So, neural machine translation can produce very fluid outputs which have nothing to do with the actual source language. They’re not good translations, but those can be very hard. For a human to detect when they’re trying to validate the final document before it gets sent off to a client”.

For neural machine translation to work “we need human translations to feed into our systems” and train them. Once the system receives the data it starts to learn patterns and to produce better translations.

“If we think for one moment where we would be without translators, we wouldn’t even be able to start”, says Andy. The secret is pretty much to “let the machine do what it’s really good at, and then, let the human elements kind of do the validation, do the correction, do like updating of the source data.”

But what about Google and its human translation quality? “I always say to my students, Google is not a machine translation company” replies Andy.

“I’m never afraid to go up against Google if I’m talking to potential clients or industry partners because in a well defined scenario where we have customised user data we can always train an engine and feedback insights from the human translators to try to improve the engine and you can’t do that with Google.”

In practice, according to Professor Andy Way it’s actually “very easy to beat out of the box Google”.

---

Watch the whole interview here and subscribe to our newsletter to receive the upcoming episode of Understand with Unbabel, a series where we dive deep into the issues, topics and challenges faced as we accelerate towards a world without language barriers.

Source: https://unbabel.com/blog/overhyping-neural-machine-translation-easy-beat-box-google/