Anticipating Your Next Question

January 12, 2015

Just 25 years ago, people researching a topic typically started with a visit to their library and a chat with the reference librarian. Today, with the help of an Internet search engine, they can answer almost any question whenever it occurs to them – but only if they can find the right keyword.

As any search engine user knows, finding that keyword can occasionally be frustrating.

“Oftentimes a keyword is strongly associated with a topic that the user is not interested in,” said Hongning Wang, a newly appointed assistant professor of computer science at the University of Virginia. “At other times, it’s simply difficult to reduce an abstract query to a meaningful keyword.”

In his research, Wang applies machine learning and ideas from the social sciences to generate more helpful, personalized search results.

The fundamental problem, Wang said, is the constrained nature of human-computer interactions, which leaves users’ intentions latent, or unexpressed. Keywords are a much more limited form of expression than conversations. While librarians learn the context of a borrower’s requests during free-form, extended discussion and use the borrower’s comments to make informed recommendations, computers currently have no way to do this directly.

Wang is devising mathematical models that computers can use to make appropriate suggestions or ask relevant questions, based on individual user data and the activities of similar users.

For instance, while interning at Microsoft Research, Wang helped develop a task model that enabled Microsoft’s search engine, Bing, to place an individual search within a class of actions – for instance, vacation planning – and then optimize the user’s next search. If you made a plane reservation, Bing could follow up by recommending information about inexpensive car rentals or hotel rooms according to your inferred search intent.

“That’s why big data is so exciting,” Wang said. “Not so long ago, we lacked sufficient data to make these kinds of predictions, but today we have enough to get started. Our challenge is to make sense of the data so that machines can better serve us.”

Wang doesn’t rely on data alone, however. He also draws on knowledge about human behavior from the social sciences to improve the predictive ability of his models. To do so, he takes behavioral rules from sociology and psychology and translates them into mathematical formulas. “My work is grounded in the behavioral as well as the combinative sciences,” he said.

Accordingly, the prospect of collaborating easily across disciplines was one of the inducements that led Wang to join the faculty. “At the University, I have the opportunity to work with domain experts from other areas as well as leading computer scientists,” he said. Wang has begun to make contacts with faculty in the sociology and psychology departments and foresees forging partnerships with the School of Medicine.

Wang’s enthusiasm for more powerful predictive modeling, however, is tempered by his belief that determining users’ intentions should include their preferences for privacy.

“We should have systems that offer the information you feel you need,” he said. “Beyond that, we should be conservative. Our goal is to develop systems that serve users, not annoy or offend them.”

- By Charlie Feigenoff

Media Contact

Josie Pipkin

Director of Communications, School of Engineering and Applied Science