Facebook has rolled out ParlAI (pärˌlē), which is expected to revolutionize conversational AI systems across different platforms.
Yann LeCun, Facebook Artificial Intelligence Researchers (FAIR) head, explained, “Ultimately one of the objectives of this is to have your own digital friend, your virtual assistant that is basically customized for you and under your control.”
For ParlAI, the FAIR team worked closely with the people who developed Facebook M, the social media company’s smart assistant for its messaging service. The aim is to make chatbots more responsive, articulate, and eventually, more efficient.
Researchers will also be able to customize ParlAI to be used in different technologies as the source code will soon be released by Facebook.
The system itself boasts of 20 built-in languages. Once perfected, there won’t be any questions out there that won’t be answered by the chatbots. Initially, examples of Q&A from Microsoft, Stanford, and Facebook are incorporated in the data sets.
They could also extrapolate the meaning in the question, taking into account the nuances of each language. The trick is to create an algorithm that is capable of machine-learning the complexities of the language, and to adapt accordingly.
While this new development is not exactly a breakthrough in natural language, ParlAI is nevertheless an important step toward better communication between chatbots and humans.
Jason Weston, researcher at FAIR, said that Facebook’s ParlAI is not exactly new technology, as researchers before have already made advances in question-answering systems. However, any progress they made were ignored because first, they were too narrow, and second, they were micro-focused on a single task.
It’s also a case of “once burned, twice shy” as industries have been promised before by researchers claiming to have the new benchmark in conversational models, only to be disappointed with the results. As it stands, there’s just no incentive for other researchers to piggyback on these benchmarks to add value to the technology.
Weston explained that while ParlAI does not claim to be the bridge that connects all of these separate data sets, it does aim to leapfrog dialog reinforcement involving chatbots. Think of this machine-learning technology as a baby—the more people talk to it, the more it learns and anticipates. However, the technology will take time.
“That will take a while before those things are general enough that they can take care of all the things of a human assistant,” LeCun said, before adding, “We’re talking decades.”