Facebook has released a platform that could combine different advances in artificial intelligence and make machines a lot more articulate.
The framework, called ParlAI, offers researchers a simpler way to build conversational AI systems, and to combine different approaches to machine dialogue. The framework should make it easier for developers to build chatbots that aren’t so easily stumped by an unexpected question. A common criticism of the chatbots released to date, including those available via Facebook, is that they are too narrowly focused and too easily confused.
The long-term hope is that ParlAI will help advance the state of the art in natural language research by reducing the amount of work required to develop and benchmark different approaches.
ParlAI comes with more than 20 different natural language data sets built in. These include question-and-answer examples from Stanford, Microsoft, and Facebook. And it provides ways to use several popular machine-learning libraries. Creating algorithms capable of learning to perform multiple different things—such as answering various types of questions—is seen as an important challenge in AI research.
The system is also integrated with Amazon’s Mechanical Turk, a platform for outsourcing small tasks. This means researchers will be able to ask humans to help train their dialogue systems, something that many see as vital to the development of smarter conversational agents.
Creating a machine capable of holding a decent conversation remains one of the holy grails of artificial intelligence (see “AI’s Language Problem”). It’s possible to have a computer use conversational ticks and tricks that provide the appearance of intelligence for a few minutes. But the complexity of language, and the way it taps into learning and common-sense knowledge, causes this to quickly unravel.
“This is a problem that goes beyond simply getting machines to understand language or generate speech,” says Yann LeCun, Facebook’s director of AI and a major figure in the field.
Facebook and others also believe that better dialogue systems could have myriad commercial applications. The social network has created tools to allow outside developers to build chatbots on its platform that perform useful tasks, like looking up information or ordering products. Facebook has also poured significant resources into its own conversational assistant, called M, which is being tested by a select group of users (see “Facebooks Perfect, Impossible Chatbot”).
ParlAI represents recognition of the challenge that remains, but it could spur progress.
“A complete question-answering system requires a lot of different components, which this framework looks to provide,” says Richard Socher, chief scientist at Salesforce and a well-respected expert on machine learning and computer dialogue. “The community will benefit immensely from a larger data set testing platform like this.”
Oren Etzioni, head of the Allen Institute for AI in Seattle, says ParlAI should be welcomed by anyone working on natural language understanding. “It isn’t a breakthrough, but it should be a helpful enabling technology,” Etzioni says.
Etzioni adds that there has been relatively little focus on dialogue systems in recent years. He also says that new machine approaches such as reinforcement learning, which mimics the way animals learn through positive feedback, could help make conversational machines way smarter.
“I think you will see real progress in the next five years,” he says.