Talking Robots: How Does Google Make Their Artificial Intelligence Sound Human?Last updated September 5, 2018
How do I love thee? Let me calculate the ways…
Technology speaks to us every day: to give recommendations on where to go, directions on how to get there, or weather updates for when we arrive. But we’re no longer just asking simple questions and receiving simple responses — we’re starting to have much more complex conversations with the artificially intelligent virtual assistants that run our lives.
As our apps get chattier, one of the thorniest problems for engineers is in how to facilitate more natural-sounding conversations between humans and machines. It’s one thing being able to tell you the weather, and quite another to sound like a friend. Currently virtual assistants have a limited ability to contextualise information, as well as clunky and often unnatural sounding speech patterns:
You: what’s the weather like today?
Virtual assistant: the weather today is sunny, with highs of 23 degrees.
The conversation above is nothing like the more natural discussion you’d have with a friend:
You: nice weather today?
Friend: it’s lovely! Hottest day of the year, I reckon.
If technology — and particularly virtual assistants — are going to fit seamlessly into your life, they need to become more like you. And that means speaking in the same way you do. This becomes even more pressing in the world of sextech, where interactions are more personal and intimate: robots that sound too robotic won’t just be annoying — they’ll be a huge turn-off. Imagine asking your virtual assistant to dim the lights and put on some sexy music, only for it to insist on Google searching ‘sexy’ or replying with some ‘dim light’ facts.
Talking robots don’t sound like people
Chatbots are computer programs written with the Turing Test in mind — you type questions and statements into a box and your chatbot replies in as ‘human’ a way as possible. The idea is that if they are good enough, they’ll convince the human who is speaking to them that they’re a human too.
But of course if you’ve ever conversed with one you’ll see just how difficult it is to emulate human conversation. The example below was taken from AI chatbot Mitsuku — she’s a pretty good bot, but she still fails in what humans would consider to be quite simple tasks.
She’s very forgetful, for one. If you follow a line of questioning, it’s hard for her to contextualise words like ‘it’ or ‘he’ — eventually forgetting your original topic and wandering off onto something new. Mitsuku mentioned James Brown in one of her answers, so I asked her:
Me: Do you like his music?
Mitsuku: Yes. It’s really good.
Me: What’s your favourite of his songs?
Mitsuku: My favourite song is Surfin’ Bird, by the Trashmen
Alongside her poor memory, Mitsuku also, understandably, struggles with colloquial terms and unusual phraseology. Some of her responses — clearly pre-programmed — are incredibly good, but the lack of variety means she’d make a fairly boring friend. I tested her skills as a potential amorous assistant, quizzing her on a few love-related questions:
Me: What is love?
Mitsuku: Love is giving someone the power to destroy you but trusting that they don’t.
Me: Have you ever been in love?
Mitsuku: I love mousebreaker [her creator]. But I don’t really feel love like you humans.
Me: Help me think of something romantic.
Mitsuku: I am always thinking.
The fact that I could have asked her anything and her answers usually made sense speaks to the fact that she’s a pretty good bot. But she’s still following strict rules in response to specific prompts — and that’s clear from reading her answers. This is a problem for all virtual assistants — although Siri might respond quickly when you ask it for ‘pizza restaurants near me’, there are hundreds of other ways to ask it for that information, and it’s impossible to pre-program them all. Most people who have virtual assistants know this, of course, and after a while will remember the best ways to speak to their devices to make sure they get the information they need. But engineers are working on ways to make talkative technology even better.
How do you make a talking robot sound more human?
Enter Google’s artificial intelligence team, which is trying to solve the problem of the ‘robotic’ robot. They are developing their artificial intelligence with the help of thousands of romance novels. They feed the novels into their neural network, which learns as it digests the stories and gradually becomes more conversational.
Speaking to Buzzfeed news, Andrew Dai explained: “In the Google app, the responses are very factual. Hopefully with this work, and future work, it can be more conversational, or can have a more varied tone, or style, or register.”
Romance novels work best for this, apparently, because the plot is formulaic but the styles and language can vary drastically — meaning it’s easy for Google’s bot to learn lots of different, conversational ways of saying the same thing. Does this mean that Siri of the future will be keen to chat us all up? Well, Dai explains that their AI is now ‘quite sexy’ as well as ‘imaginative’, but the aim of it isn’t to create a virtual sex chat operator — just to make conversation with virtual assistants more natural.
However, we can see the benefit of this style of learning in the sex tech space. Although sex robots are a long way off, virtual assistants are already used in most areas of our lives — it’s not hard to imagine them helping out in the bedroom too: ‘Hey Siri, dim the lights.’ or ‘Cortana, recommend me some romantic music.’
The Google team’s research isn’t there to help ‘teach’ their artificial intelligence network to be romantic — the romance novels just happen to have one of the most useful formulas for helping a neural network pick up linguistics. But by giving the talking robot a more ‘natural’ language, Google is helping to ensure that all conversations with your phone will sound more natural in future — whether you’re asking for directions, the weather, or a little help setting the mood.