Most of us talk to our computers on a semi-regular basis, but that doesn’t mean the conversation is any good. We ask Siri what the weather is like, or tell Alexa to put some music on, but we don’t expect sparkling repartee — voice interfaces right now are as sterile as the visual interface they’re supposed to replace. Facebook, though, is determined to change this: today it unveiled a new research tool that the company hopes will spur progress in the march to create truly conversational AI.
The tool is called ParlAI (pronounced like Captain Jack Sparrow asking to parley) and is described by the social media network as a “one-stop shop for dialog research.” It gives AI programmers a simple framework for training and testing chatbots, complete with access to datasets of sample dialogue, and a “seamless” pipeline to Amazon’s Mechanical Turk service. This latter is a crucial feature, as it means programmers can easily hire humans to interact with, test, and correct their chatbots.
Abigail See, a computer science PhD at Stanford University welcomed the news, saying frameworks like this were “very valuable” to scientists. “There’s a huge volume of AI research being produced right now, with new techniques, datasets and results announced every month,” said See in an email to The Verge. “Platforms [like ParlAI] offer a unified framework for researchers to easily develop, compare and replicate their experiments.”
In a group interview, Antoine Bordes from Facebook’s AI research lab FAIR said that ParlAI was designed to create a missing link in the world of chatbots. “Right now there are two types of dialogue systems,” explains Bordes. The first, he says, are those that “actually serve some purpose” and execute an action for the user (e.g., Siri and Alexa); while the second serves no purpose, but is actually entertaining to talk to (like Microsoft’s Tay — although, yes, that one didn’t turn out great).
“What we’re after with ParlAI, is more about having a machine where you can have multi-turn dialogue; where you can build up a dialogue and exchange ideas,” says Bordes. “ParlAI is trying to develop the capacity for chatbots to enter long-term conversation.” This, he says, will require memory on the bot’s part, as well as a good deal of external knowledge (provided via access to datasets like Wikipedia), and perhaps even an idea of how the user is feeling. “In that respect, the field is very preliminary and there is still a lot of work to do,” says Bordes.
It’s important to note that ParlAI isn’t a tool for just anyone. Unlike, say, Microsoft’s chatbot frameworks, this is a piece of kit that’s aimed at the cutting-edge AI research community, rather than developers trying to create a simple chatbot for their website. It’s not so much about building actual bots, but finding the best ways to train them in the first place. There’s no doubt, though, that this work will eventually filter through to Facebook’s own products (like its part-human-powered virtual assistant M) and to its chatbot platform for Messenger.