2016 was the year the chatbots took over Silicon Valley. The only trouble was: they didn’t really know how to chat.
As Facebook and so many other Silicon Valley players trumpet the benefits of software that can carry on a conversation—apps that book your plane flights or manage your bank account through SMS-like dialogue—the technology still lags behind. In recent years, using what are called deep neural networks, companies like Facebook, Google, and Microsoft have fashioned services that can reliably identify faces and objects in photos, recognize voice commands on smartphones, and translate from one language to another. But building bots that can truly carry on conversations is still proving elusive. It’s an undertaking that requires a far more varied array of AI techniques; researchers are still trying to figure out how the different approaches all fit together or whether they’ll really work at all.
With these challenges in mind, a team of Facebook researchers has built a new framework for making chatbots chattier—a “training ground” for AI to master a wide range of conversional techniques—not just one or two. “You need to see what a machine learning method can do—which things can it solve, which break it—so we can understand what to fix,” says Jason Weston, a Facebook researcher who specializes in conversational systems. “Just training on one task alone? We don’t think you’re going to get to an intelligent machine that way.”
Facebook’s chatbot training ground is called ParlAI—a play on words well suited to the company’s central artificial intelligence lab, which is littered with French-speaking researchers. And in keeping with its approach to so many of its new technologies, Facebook is sharing this creation with the world at large as an open source tool. The company is offering the software along with a varied collection of public datasets that researchers can use to train their “agents.” The system also ties into Amazon’s Mechanical Turk service, the online retailer’s platform for crowdsourced labor, so that researchers can test their bots in conversation with live humans. In turn, these tests will generate still more data, creating a virtuous circle of chatbot development.
Facebook’s latest move is part of a widespread effort to accelerate the evolution of conversational AI. All the big internet players—from Google to Amazon to Microsoft to IBM—are pushing in this direction, each hoping to fundamentally change the way humans interact with machines. In January, Microsoft acquired a Canadian startup, Maluuba, that specializes in conversational AI research. Amazon is working to build its own datasets for training bots to converse—key to the success of its Alexa platform. For nearly two years, meanwhile, Facebook has been gathering a particularly complex set of data using an experimental digital assistant called Facebook M.
To reach the goal of machines that can truly hold a conversation, each company is taking a slightly different tack. While Facebook is focusing on neural networks that can learn from existing conversations and other datasets, Maluuba specializes in a technique called reinforcement learning, where bots learn by extreme trial and error. But don’t think of these as competing methods. In the end, success will come from a combination of techniques. “We don’t use systems that try to solve everything with one machine learning approach,” says Yunyao Li, who oversees a natural language research lab inside IBM. “Instead, we use the right machine learning method at the right moment.”
This hybrid approach is the thinking that drove the creation of ParlAI. The training ground can help advance neural network research, reinforcement learning, and whatever else may prove useful. It’s designed to drive the development of new technologies not just from one company, but from a world of AI researchers. The ultimate goal is to combine all sorts of methods into a chatbot that can actually chat. “This is not something that any entity—whether it’s Facebook or any other—does all by itself,” says Yann LeCun, the Paris-born researcher who oversees Facebook’s AI Lab. In making ParlAI available as open source, Facebook is signaling once again its belief that there’s more to gain by helping everyone advance toward the same goal than trying to get there alone.