Not everyone likes having their beliefs challenged. It makes us feel threatened on a deeply personal level and compels us to seek consistency, so we can continue to function with sanity in this ever-changing, chaotic world.

Take a meat lover, for example, who agrees that animal cruelty is wrong but goes on to include fish and pork ribs in his diet. Or a person who cannot stop smoking even after knowing that it is harmful to their health.

Coping with contradictory ideas — also known as cognitive dissonance — can be mentally stressful, and people tend to find consistency by rationalization (accepting new information), confirmation (sticking with old ideas), or blindly believing whatever they want to believe.

Vasudha Varadarajan and Swanie Juhng, Ph.D. students, along with Prof. Andrew Schwartz, Director of HLAB (Human Language Analysis Beings) at Stony Brook University — realized that even though cognitive dissonance is often expressed in language on social media, detecting it using AI is a rare class problem. Since most AI tasks for language focus on labeling common events — positive or negative — being able to label rare events is difficult, both for the AI and the humans in the loop.

In their paper, titled ‘Transfer and Active Learning for Dissonance Detection: Addressing the Rare-Class Challenge,’ Prof. Schwartz and his team addressed this challenge by prioritizing which examples to label first, and showed benefits both in terms of AI's ability to capture the rare event and in terms of improving the human experience of annotating data in order to train AI. Their work won them the prestigious Outstanding Paper Award at the Association for Computational Linguistics (ACL) conference this year. The award recognizes those who’ve contributed exceptional research in the field.

Vasudha Varadarajan, a fourth-year Ph.D. student working in the areas of natural language processing and the relationships between cognition and language, said, “Pushing for AI to be human-centered by incorporating cognitive and psychological constructs comes with the challenge of making careful choices in low-resource settings. This award is a recognition of our abilities to lead, design, and perform meticulous experiments, and an encouragement to keep making strides in our research.”

“Our AI work building on concepts from social and psychological sciences was often seen as a niche outside the mainstream of the AI community. This award suggests that the view of our work is changing and that human-centered AI is increasingly being valued by our peers.”
-Prof. Andrew Schwartz

The team’s elegant approach, which helps alleviate the “needle in a haystack” problem of annotating rare class data samples, gives rise to notable implications in how language models approach absolute rarity.

Swanie, a Ph.D. candidate focusing on applying NLP and ML to solve problems in psychology, added, “The acknowledgment we received has fueled our motivation to continue doing research. We look forward to making more contributions and connecting the dots in the field of human-centered AI."

Ankita Nagpal
Communications Assistant