Louis-Philippe Morency spends much of his workday watching YouTube. He logs hours viewing videos of people expounding on everything from presidential politics to peanut butter preferences.
But Morency is no slacker. He’s a scientist at the USC Institute for Creative Technologies (ICT) whose focus is teaching computers to identify and understand the ways people convey emotion, including those occasions when we say one thing and mean the opposite. And in an interesting twist for those who study human communication, it turns out computers are becoming one of the best places to explore how people express themselves.
“There is a growing field of opinion-mining right now where people study Internet posts like Amazon book reviews or other text-based product and movie critiques to find out how people feel about a topic,” said Morency, who is also a research assistant professor of computer science at the USC Viterbi School of Engineering. “We are taking this field one step further by focusing on online videos which provide verbal and nonverbal communication clues beyond just words.”
According to Morency, most people can cite countless cases of misreading either written or body language. For example, consider a tone-deaf email in which a joke can come across as an insult or conversations in which a person’s sentiment can only be understood if a statement is delivered with a smile or a stare.
“By looking at more than just text, we can learn when someone is using sarcasm, for example, saying they love something when their facial expressions and body language indicate that they hate it,” Morency said.
Social scientists have advanced understanding of everything from autism to cross-cultural differences by studying how people use verbal and nonverbal forms of communication. In the past, researchers needed live subjects to carry on their study. But with the increasing volume of videos posted online, the Internet has become an invaluable resource for researchers. For his latest effort — figuring out how to identify when someone is sharing a positive, negative or neutral opinion — YouTube provides a limitless library of likes and loathes.
“There are a lot of people sharing their sentiments on YouTube,” Morency said. “The goal of this work is to see if we can find a way to analyze these millions of videos and accurately assess what kinds of views they are expressing.”
To do this, Morency and his colleagues created a proof-of-concept data set of about 50 YouTube videos that feature people expressing their opinions. The videos were input into a computer program Morency developed that zeroes in on aspects of the speaker’s language, speech patterns and facial expressions to determine the type of opinion being shared.
Morency’s small sample has already identified several advantages to analyzing gestures and speech patterns over analysis of writing alone. People don’t always use obvious polarizing words, such as love and hate, each time they express an opinion. So software programmed to search for these “obvious” occurrences can miss many other valuable posts.
Morency also found that people smile and look at the camera more when sharing a positive view. Their voices take on a higher pitch when they have a positive or negative opinion, and they start to use a lot more pauses when they are neutral.
“These early findings are promising, but we still have a long way to go,” Morency said. “What they tell us is that what you say, how you say it and the gestures you make while speaking all play a role in pinpointing the correct sentiment.”
Morency first demonstrated his YouTube model at the International Conference on Multimodal Interaction in Spain last fall. He has since expanded the data set to include close to 500 videos and will submit results from this larger sample for publication later this year.
The YouTube opinion data is available to other researchers who can contact Morency’s Multimodal Communication and Machine Learning lab at ICT. In the academic community, Morency foresees his research and database serving as resources for scientists striving to understand human nonverbal and verbal communication, helping to identify medical conditions, such as autism or depression, or to build engaging educational systems. And potentially there are marketing surveys for commercial uses.
As for Morency, he plans to continue to view how people behave over computers in order to make computers behave more like people.
And that is an effort worth watching.
More stories about: Research