Teaching computers how to really talk to humans
Last year, the buzz was that chatbots would achieve the unthinkable: using artificial intelligence, they would handle customer service in a way that did not drive customers totally nuts. That hype has died down, but the advances in natural language processing that led to it were real, and entrepreneurs continue to test new applications.
Amir Konigsberg, CEO of Tel Aviv-based Twiggle, is one such entrepreneur. His firm uses AI to teach retail search engines to better understand what potential customers want when they type into an online engine.
Why Twiggle matters: Backed by Alibaba, Twiggle says it is partnering with three of the world's top 20 retailers to improve their search capabilities. (Konigsberg says he can't divulge names because the product is still being tested by these companies). It wants the interaction between a customer and a search engine to resemble that between a customer and a sales person, solving difficult natural language processing problems along the way. Axios asked Konigsberg where this science is headed.
Axios: Tell us about your background, and why you think improving search is an interesting problem and a good business proposition
Konigsberg: Search combines several things that I've been pursuing for a long time. I have a PhD in interactive decision theory, which is basically a combination of game theory and psychology, insofar as it relates to how people process and relate to information. I spent some time at Google, back in 2005 when it was setting up its operations outside the U.S., and helped launch its Israel operation. I also spent four years at General Motors working on user experience technology for autonomous vehicles, specifically on search and speech systems, and other kinds of adaptive behaviors inside the vehicle that result from autonomous driving scenarios.
How did the idea for Twiggle come about?
I met my co founder Adi Avidor at Google, where he was responsible for projects like Google's suggestions mechanism and Google spell checker. Our experience led us to realize that search could be so much more and better than what it is today.
We wanted to take the experience, where you have to actually sit down in front of your computer and think what to type so that Google will understand you. We thought, "Let's shift that burden, and make sure that this search engine is smart, and adaptive enough to understand you no matter how you type."
We realized that applying this specifically to e-commerce made sense because retailers will see a direct, measurable benefit when customers can communicate with them better.
What is an example of how search can be improved using AI?
If you try to type something simple into most retail search engines like "black dress shirt," most engines will give you results that include a lot of dresses. That's an example that is very clear to you: I'm asking for a shirt, not a dress. But computers don't understand this.
In what we have created, you can ask, "I want an inflatable pool for my yard, I want it to be 2 meters wide, and I want it to be about 40 cm high, and I want it to be blue, and child safe. You'll get search results that mirror what you ask for. You don't have to waste time iterating. It happens immediately because we are able to understand what you want and the products.
How does it work?
We collect large amounts of data from search engines, retailer websites, and manufacturer websites, and feed that into our deep learning technology to create a methodology of tying search queries to product information.
How will the way humans interact with computers in retail situations, or otherwise, change in the next five to ten years?
The hardest problem we face is teaching computers to understand what customers want. You can talk with any retailer's search engine today and say that you want a pair of jeans. Then it shows you jeans, but you want to respond, "No, I want a slim fit, and in darker blue."
To understand that simple sentence, and provide you with new results, the engine has to understand that what you're saying relates to the previous search. Tying those two things together is a very complicated task today that no one has managed to accomplish.
After we've solved problems like that, it will be easier to then combine the solution with things like voice search, so that can have conversations with computers that feel like real conversations.