‘Algorithms are no match for human intelligence, especially when it comes to natural language’

A conversation with Geoff Stead, chief product officer at Babbel, on innovations in language tech, ‘deep learning’, automated bias in machine translation, and much more…

Name: Geoff Stead

Job title: chief product officer, Babbel

Professional bio: Geoff has led edtech innovation teams in the UK, US and now Berlin, Germany; building a diverse and playful range of learning apps including global quizzes, MOOCs, augmented reality (AR) artworks, audio tours, 3D videos, and museum guides. He’s an expert in how new technologies are changing the way people work and learn, as well as how to ignite the creativity needed to build such products. 

Twittter: @geoffstead

Q. How has the pandemic and the subsequent events of the last 12 months impacted language learning and translation apps?

At Babbel, we’ve seen a huge increase in people signing up and learning with us over the last 12+ months. This was most visible at the start of the pandemic, when we saw a surge in both subscriptions and daily usage. What’s great to see is that the pattern of ongoing learning within the app has continued, with more time being spent using the product even today.

More widely, this is also true across the digital industry and we’ve been able to benefit from this because of the wide range of levels we cover, and the mix of different ways to learn. Some of our less-digital competitors, like traditional language schools and language assessment organisations, have really struggled during the pandemic. This is partially why we launched our new platform, Babbel Live, to offer video-based classes with expert linguists to support learners looking for a more human-faced experience.

Funnily enough, despite the lack of travel opportunities, when we ask our learners why they want to speak another language, travel remains one of the top motivators across all countries we sell in – we all want what we can’t get! I’m predicting a mad frenzy of second language holidays as soon as the borders open up again.

Q. Machine translation has come on leaps and bounds in recent years. Can you walk us through the developments of the last 15 years in a nutshell?

Firstly, machine translation is pretty amazing. Especially when combined with text and voice recognition and text-to-speech. I can scan a menu, or record someone’s voice, and get a plausible enough translation spoken back to me. There have been many major breakthroughs that have got us here, but these stand out.

If you think back to before machine translation tools like Google Translate, it was very obvious when something was auto-translated, or translated by a human. Around 15 years ago there was a big breakthrough when Google used vast quantities of real, human translated documents to build a ‘phrase-based translation’ model (instead of earlier word-based). This was a major breakthrough at the time, which combined the power of the machine (comparing millions of docs), and the human (creating meaningful phrases).

“Then, came the second big development of ‘neural mapping’ to understand context…Suddenly, the same phrase inside two different paragraphs might be translated quite differently. This also makes it possible to translate between two languages even if you don’t have any training data to connect them.”

Then, came the second big development of ‘neural mapping’ to understand context. This was five years ago, when the Translate team made a huge step forward by using machine learning and word mapping to try and understand the full meaning behind a sentence. Suddenly, the same phrase inside two different paragraphs might be translated quite differently. This also makes it possible to translate between two languages even if you don’t have any training data to connect them. A fun example of this in action is to try translating “to kill two birds” into German, which comes out correctly as “zwei Vögel zu töten”.  However, “to kill two birds with one stone” translated into German comes out as “zwei Fliegen mit einer Klappe erschlagen”, or “to kill two flies with one swat” – which is the equivalent phrase, even though the words differ.

Q. How do innovations like 'deep learning' and 'neural machine translation' shape language learning and translation technologies?

Both are solid enablers for modern language learning. They provide the necessary tools that we use at Babbel when analysing the ways in which people interact with our app. In addition, such innovations also help us offer targeted advice to learners on how they can improve their speaking and writing.

“The more exposure people have to other languages, the more they are motivated to learn them, and the introduction of this technology is about deepening human interaction”

They also underpin several great translation apps that help all language adventurers feel more confident when exploring a new culture and language. The more exposure people have to other languages, the more they are motivated to learn them, and the introduction of this technology is about deepening human interaction.

Q. What are the general pros and cons of machine translation?

There are of course many pros; advancements in machine translation have made languages more rapidly accessible to people everywhere, bridging the gap of understanding between humans via technology – all at the click of a button. Digital tools also transform learning and teaching for millions worldwide, whether it’s free access to high quality digital resources, or optimising the time spent with a teacher to focus on active learning, rather than passive knowledge transmission.

“Advancements in machine translation have made languages more rapidly accessible to people everywhere, bridging the gap of understanding between humans via technology – all at the click of a button”

The best type of learning is when you get meaningful feedback on your mistakes and guidance on what you should focus on to correct them. When learning a new language this is quite complex to do in an automated way, given the wide range of potential mistakes a new speaker might make. Artificial intelligence (AI) however, can help with this. There is a field of AI called NLP (Natural Language Processing) that both looks at long strings of text and tries to understand both the meaning and different parts of speech used. We can use this to offer feedback to learners as they progress. A long-form version of this is writeandimprove, offering online english essay feedback, but the same principle also works with shorter texts too.

However, there are also a few drawbacks to machine learning. While instant translation technology can provide an emergency band-aid in times of need, it’s limited in the sense that it represents a short-term fix, rather than being a long-term substitution for properly learning another language. AI is continually developing, but languages are deeply complex and despite the hype around machine learning, algorithms are yet no match for human intelligence, especially when it comes to natural language

“While instant translation technology can provide an emergency band-aid in times of need, it’s limited in the sense that it represents a short-term fix, rather than being a long-term substitution for properly learning another language”

This is partly because machine translation often ignores the subtle nuances between different languages and can’t take into account expression. This is why language students are often not taught how to use translation tools, as they somewhat optimistically think they are always correct (clue: they are often not!).

“In the end, it’s about human communication. We want to speak to each other, not to machines”

In the end, it’s about human communication. We want to speak to each other, not to machines. Even if you speak to someone who speaks great English but it isn’t their mother tongue, curiosity might lead you to want to learn a bit of their language.

Q. Let's home in on automated bias as a major challenge in language learning tech. In what ways does this impact translation platforms and software?

AI bias is a really big issue across the machine learning industry, with countless examples across face recognition, crime prediction, financial services, and yes, even auto-translation. At the root of this is the fact that computers learn from the data we feed them, and our data is biased. Sometimes because we ourselves are biased.

“At the root of this is the fact that computers learn from the data we feed them, and our data is biased. Sometimes because we ourselves are biased”

Imagine a 1950’s book on home economics. It would probably be filled with painful gender stereotypes that would make us cringe today. Unfortunately, computers aren’t smart enough to cringe. They just absorb that book, as well as every other one they can get their digital ‘hands’ on, and use them to learn that ‘woman’ and ‘kitchen’ have a closer association than ‘man’ and ‘kitchen’. Research has shown that the semantic association between women and cooking for example, is so strong that, for some algorithms that are trained on labelling images, if you give it a picture of a man cooking, it’ll say it’s a woman.

“Research has shown that the semantic association between women and cooking for example, is so strong that, for some algorithms that are trained on labelling images, if you give it a picture of a man cooking, it’ll say it’s a woman”

In the same way that AI can learn, for example, that ‘dog’ and ‘cat’ and ‘animal’ are all related, or that ‘apple’ and ‘banana’ are related, it also considers ‘business’, ‘office’ and ‘salary’ to be systematically closer to words associated with men, such as ‘uncle’ and ‘father’. “Man is to woman as king is to…” will give you “queen”. But when you say “Man is to woman as pilot is to…”, it gives you “flight attendant.”

So clearly, the industry has a long way to go…We don’t want to be contributing to that bias-loop, but rather be helping our learners experience the real language, as used by local speakers today.

language learning
Geoff Stead, CPO, Babbel

Q. How do we get past automated bias in translation technology?

It will take a long time and a lot of hard work. What we as consumers can do is be aware and keep putting pressure on the big machine learning companies to take the ethics of machine learning seriously (most are, albeit reluctantly). But in a way, that is no different from us humans. We too have many biases encoded in the language we speak. As citizens of the world, we are also continually evolving our own use of language as we learn more about topics such as gender equality and racial bias.

“What we as consumers can do is be aware and keep putting pressure on the big machine learning companies to take the ethics of machine learning seriously (most are, albeit reluctantly)”

Q. Do you have any personal favourite translation tech 'fails'?

In the past, there were many sites (like Bad Translator) or hashtags (like #GoogleTranslateFails) celebrating these, but as the translations have improved, the opportunity for translator-jokes seems to have dried up.

This doesn’t mean it’s perfect though. In fact, Google’s own team will tell you “State-of-the-art systems lag significantly behind human performance in all but the most specific translation tasks”. Google Translate struggles with the formal and informal forms required by some European languages: for example, it battles with the various past tenses of the romance languages. It also does a poor job with languages where there is less written material to learn from.

Q. With exponential improvements in such innovations (deep learning, neural machine translation, etc.), how will translation apps and software continue to change over time? What will translation tech look like in another 15 years, for example?

Translation quality will continue to improve. As part of that, computers will also get better at creating legitimate seeming pieces of long form writing. The ways we interact with translations will become rapidly more seamless.

Already I can scan something with my phone and hear it translated via a bluetooth earpiece. We can even pair up with our phones and speak to each other in different languages, both with live translating earpieces. This is pretty much the Babel fish of Hitchhikers Guide. These things will continue to get smaller, better and more ubiquitous.

Q. Where will such improvements leave more 'traditional' forms of language learning?

Viewing this as auto-translation vs. language learning is somewhat of a false dichotomy. Auto-translation is a great new tool that can help people learn languages more easily, but not a substitute for human conversation. 

“Viewing this as auto-translation vs. language learning is somewhat of a false dichotomy. Auto-translation is a great new tool that can help people learn languages more easily, but not a substitute for human conversation”

If you take a purely functional view of language and assume the only reason I would learn German is to exchange basic information, then yes, Google Translate could, in fact, replace my need to learn the language. But reality is wildly different to that: language = culture = conversations. Once upon a time, I myself learnt German to talk to a cool woman and her friends that I met on the beach. I wanted to understand their jokes and better understand all those quirky differences between us. We are now married! If I’d depended on a translation app, I’m sure we wouldn’t still be together today.

“Machine translation will never fully master the subtleties and nuances of human communication”

Machine translation will never fully master the subtleties and nuances of human communication. Humans have a seemingly infinite ability to create new words, or new meanings for old words, or multiple meanings for the same words. Even inside one household, my teenage kids sometimes seem to speak in a different language. This playfulness is how we find our own people and is some of the magic you get from learning a new language and getting started with having a real conversation in that language.  

Q. I totally acknowledge what you say about the 'false dichotomy', but many critics still fear for machine translation 'taking over' human language learning – what would you say to those people?

“Bring on the robots!” They help to get the party started.

Most of the language learners I’ve met tell me that their greatest fear is getting a real conversation going, then making an embarrassing mistake. The cool thing with machine translation is that you have backup. You can check the thing you wanted to say before saying it, or check your understanding of what the person you were chatting to said before answering. It’s like pitching up at a party, feeling a bit shy, and having someone help to make the basic introductions. But once the ‘wheels are greased’, so to speak, the basic human need to communicate and interact will take over from where the robots kicked things off.

Leave a Reply

Upcoming webinar

The digital student journey

How to include everyone from recruitment to graduation and beyond

Wednesday, May 19

11AM (BST)