How researchers trained one AI system to start asking its own questions

A deep learning system from Cornell, developed by asking human teachers to generate questions from Wikipedia posts, could be a boon for educators. Here's what it does, and the limitations it faces.

Video: How to tell the difference between AI, machine learning, and deep learning

We normally use artificial intelligence (AI) to answer questions. But what if it could ask us the questions instead?

That's the subject of new research by Xinya Du, a first year student at Cornell University. Du studies natural language processing and machine learning, focusing on deep learning for natural language generation and question answering. Du first worked on question-answering systems, when it occurred to him to see if the AI could also generate those same questions. "It seemed to work," he said.

Du's AI system generates questions by studying a set of more than 500 Wikipedia articles and 100,000 questions about those articles, all composed by human teachers. The teachers were asked to create specific questions regarding some aspect of the information presented in the Wikipedia articles.

"Then, we utilize that data set to learn from how people ask questions regarding those paragraphs of those Wikipedia articles," said Du. "Such as subtype questions regarding location, time, and maybe of the 'would' or 'what' type questions."

For instance, if the Wikipedia article mentions a species dying during a natural disaster, the AI could generate the question, "Why weren't there birds during this period?"

"It's like our system infers that birds belongs to the species," said Du.

There are many potential uses for a system like this, especially in an educational setting or a chatbot realm. The system could aid chatbots, potentially, by generating real questions and linking them to responses.

Although there are occasional errors, the current system is great at generating fact-oriented questions, according to Du. Questions that require contextual knowledge, on the other hand, are more difficult for the AI to produce. For instance, the system could easily ask a question relating to the dates of a certain historical period when income levels fell. But any question of interpretation, like asking about how class distinctions were affected, would be more difficult for the system to generate.

AI experts see this as a potential contribution in education. Roman Yampolskiy, director of the Cybersecurity Lab at the University of Louisville, suggests that it "could be incorporated into toys to interact with children while improving their critical thinking skills."

Other AI experts expressed more skepticism.

Marie desJardins, AI professor at the University of Maryland, Baltimore County and former chair of AAAI (the National Conference of the Association for the Advancement of Artificial Intelligence), thinks the research is "interesting in a limited sort of way." According to desJardins, "all of the questions it generates are very fact-based and concrete." Instead, she said, doing real question initiation would involve "modeling information that isn't in a text, and some kind of exploration model about pushing a dialog beyond what previous interactions have explored." Still, she said she sees importance in moving research in this direction.

Vincent Conitzer, professor of computer science at Duke University, said he sees limitations as well. He also added that AI initiating conversations is not new. "In fact, many chatbots seem to try to take charge of the conversation, introducing new topics, to avoid questions on other topics that will expose their lack of understanding," Conitzer said.

Regardless of its current limitations, Du said he believes that it's worth trying to generate these kind of context-related questions by introducing an outside knowledge base that can challenge the AI.

Still, he said, "I don't even know if all human teachers do that."

Also see...


Pepper, the humanoid robot developed by Aldebaran, specializes in customer service interactions.

Image: Aldebaran