Dr. Hila Gonen brings trustworthy and reliable AI to UBC Computer Science
New Assistant Professor Hila Gonen to lead research group on reliability and multilinguality of language models
Growing up, Hila Gonen loved all things analytical — from math to basic sciences to linguistics. She knew that for many fields, coding and computer science would be useful, hence her decision to major in computer science.
While she enjoyed coding and being able to create programs, it wasn’t until her third year of university when she attended a seminar on natural language processing that she found her calling.
“I was fascinated because I’ve always loved languages and learning new languages. It made a lot of sense to me to combine my love of languages with computer science,” Dr. Gonen says.
She took a short detour studying theory of computer science during her master’s program at the Hebrew University but ultimately felt the pull to go back to studying language models for her Ph.D. at Bar Ilan University.
At the time, researchers in the field started to notice certain biases in language models. For example, certain professions were associated more with certain genders, such as women and nurses or men and doctors. While scientists tried to find ways to remove these types of biases, Dr. Gonen’s research found that these biases are heavily ingrained into the models and are more difficult to remove than researchers initially realized.
“This work was important in making people realize that it’s a real challenge to remove these biases and it’s something that we need to invest a lot of effort into,” she says.
According to Dr. Gonen, these biases are still present in language models today. For example, if you asked about salary expectations for a friend with a traditionally female name, it tends to suggest a lower salary than if you asked for a friend with a traditionally male name.
For her postdoctoral fellowship at the University of Washington, she shifted her focus to understanding language models and controlling what the models do. She also looked into how models represent different languages and their ability to translate concepts from different languages.
Dr. Gonen hopes to continue her work as a new Assistant Professor at UBC’s Department of Computer Science, studying reliable and responsible language models and collaborating with the larger UBC Natural Language Processing Group.
“We know that today’s language models are not very consistent — the same question phrased differently might get different answers, which hurts our trust in these models,” she says. “I want to focus on making models more consistent and reliable.”
In addition, these models sometimes provide inaccurate information, such as telling us about an event that never happened. Dr. Gonen’s goal is to understand how language models come up with these answers and help them choose to abstain from answering if they don’t have all the information to provide facts.
For Dr. Gonen, developing more trustworthy and reliable language models is crucial if these applications are to be deployed in healthcare and medicine. Now at UBC, she hopes to collaborate with researchers in health and medicine to develop medical applications that can potentially save time for clinicians and patients and help with early diagnosis of health conditions.
As she reflects on her work and journey into the field, she recommends that computer science students be skeptical and try to find “failure modes,” or cases in which the models or systems fail.
“A lot of what I liked about my work started by finding problems with today’s models and trying to fix them,” Dr. Gonen says. “Try to always be skeptical. Make sure that things actually work in practice and try to find places where things don’t work as expected. This is where you can make a change.”