AI Building AI: How Our Bots Are Now Training Themselves 100x Faster

There is one universal truth when it comes to bot building: training is a pain.

SAP Conversational AI
Chatbots Magazine

--

There is one universal truth when it comes to bot building: training is a pain. We decided to tackle this issue by making our bots train themselves! When you enter one sentence, our AI suggests 5 new ones. All hail collaborative deep learning, right?

Earlier today, I came across this awesome infographic detailing the future of AI: what will happen, what might happen, and what will definitely remain science fiction for a very long time.

Despite all the hype, HAL is not yet around the corner. What is already possible today is AIs and humans working together on tasks simple enough to be automated and data-intensive applications. In fact, pattern recognition is currently one of the hottest fields in AI. It works well because it frees up time experts (doctors, traders…) spend on repetitive tasks and makes them more efficient. While keeping the precision of human supervision, these AIs operate at a machine-scale. Moreover, because the AI is used as a screening method, false positives are avoided.

As an AI startup involved in the creation of human-machine interfaces, it’s thrilling to think that we might be making history any day, and I feel that today, we’re getting closer.

It has to be said: training AI is the biggest pain

Powerful AIs all have one thing in common: an extensive training based on qualified datasets. AIs are like babies; they need to learn by example.

A very clear demonstration of that is Tay, Microsoft’s Twitter bot who learned from what users told him. It took him only 18 hours to be disabled because he started mirroring user behavior and posting very offensive tweets.

We’re well aware of the training challenge in today’s AI industry. What’s the difference between small AI firms and Google? Data.

To train your systems efficiently, you need a sizable dataset. How you get it is up to you! In bot building, the most common way is to type sentences by hand. But having a big dataset isn’t the only thing that matters: it needs to be qualified. Having random incorrect sentences won’t make your AI smarter, it might even make it dumber. That’s why you should spend time making sure you have a clean set of sentences to train your conversational systems.

Collaboration is a game-changer

We’ve established training is a major part of bot building and probably the most time-consuming. In real life, when things get hard, what do you do? You get help, you get collaborators.

Many companies have done it over the years: AirBNB, Uber, BlablaCar have based their business on people working together.

Why can’t AI do the same? Why can’t you get thousands of people creating your training dataset? That’s right, you can, and that’s just what we’ve done.

Right from the start, Recast.AI was designed as a collaborative bot platform. All our users participate in the global training of our models when building their bots. Our own NLP technology, powered by deep learning, is growing stronger and smarter everyday! And because we have two years of extensive shared training behind us, we are now able to drastically increase bot time-to-market by 100 for all our developers and clients with Expression Suggestions.

Introducing Expression Suggestions: your semi-automatic bot training

Let’s get back to basics: when building a bot, you first create intents. Intents can be seen as boxes in your bot’s brain, defining the different topics it can understand (greetings, weather, food, drinks, etc). You need to enter a certain amount of sentences in each intent to train your bot. It’s a tedious process; you have to think of sentences meaning the same thing that are all different. You have to picture all the different ways a question can be asked. You sometimes have to use slang, sometimes formal speech. And you have to do that in different languages!

Step back, have a coffee: now, Recast.AI can do that for you.

When entering one sentence, our AI suggests 5 others that might match. And the more you use this feature, the sharper it’ll be. This is a truly innovative way to think about bot building; what better than AI to train AI?

Want to try it out? Go on Recast.AI, create a bot, an intent, and see the magic.

What are you going to do with all this free time? Add new features to your bots? Another language? Make a whole new one? I’m sure you’ll figure that out on your own.

If you’d like to know more, please reach out to our support team! And now that you can build bots in the blink of an eye, please send us your creations! We’re always proud to promote our community’s work.

Originally published at Recast.AI Blog.

--

--

Bot building software for the enterprise. Formerly known as Recast.AI, startup acquired by @SAP in Jan 2018 to transform customer experience with #bots and #AI