Confronted with election data that suggests 60% of counted ballots belong to the Democratic Party, would you say that the Democrats “will probably win the election,” or would you be more cautious and say they “might win the election”? Uncertainty expressions — words like “might” and “probably” — are spoken more or less cautiously in different contexts. So how do people’s beliefs affect their interpretation of uncertain data?
Anzi Wang ’25 intends to find out. In her independent research project, “Studying Confirmation Bias through (Adaptation to) Uncertainty Expressions,” Wang uses experimental and computational analysis to help improve the computational model used for psycholinguistic analysis — which seeks to understand why people say what they say.
Her research over the past three years at ߲ݴý earned her an honorable mention from the Computing Research Association, an organization that supports researchers across industries and recognizes those “who demonstrate strong research capabilities and a commitment to advancing the field.”
Wang’s project began in March 2024, when she was nominated by her academic and research adviser, Professor Grusha Prasad, and selected to be one of two Mind, Brain, and Behavior Scholars at ߲ݴý.
Wang’s experiment compared two probability scenarios and the language people used to describe them. The first depicted a gumball machine with orange and purple gumballs. Participants were asked whether the first gumball that came out would not be, might be, or would probably be a certain color. The second scenario offered a hypothetical election prediction from an unknown country. Based on the data, which was stated to be from a reliable source, participants were asked whether a certain (unidentified) party would not win, might win, or would probably win.
When it’s as simple as gumballs, people’s language about uncertainty matches their literal meaning. Elections are not so straightforward. The word might suggests that there is some probability that an event will happen. When asked how they would relate election information to another person, participants only said that a party might win when the polls indicated a 50% chance. In the gumball scenario, participants were much more liberal with their use of the word might, attributing it frequently to probabilities below 50% — and staying true to the meaning of the word.
Why the difference? Wang suggests that individuals go through an additional interpretative step when reasoning about complex events like elections. They think about the messy contexts of other elections and consider their beliefs and perceptions, rather than accepting the truth of the poll results. As she expands this study for her thesis project, Wang hopes to learn more about biases and preferences in the language used to describe uncertain events. She can use this data to develop an improved, more accurate language model for use by future researchers.
Wang and Prasad have been collaborators since February 2023. Wang worked alongside another student to set up an eye-tracking lab and design an experiment using the new tech. During summer 2023, Prasad’s research team designed a new curriculum for language models. Now, Prasad is mentoring Wang as she applies to graduate schools.
While at ߲ݴý, Wang has made the most of her interdisciplinary interests. A computer science and philosophy double major, she has been both a computer science teaching assistant and a conversationalist at the “Ask-A-Philosopher” booth at the Hamilton Farmers Market. She worked at the ALANA Cultural Center and served as a core member of Lambda. After graduation, Wang plans to obtain her PhD in computational psycholinguistics.