AI Solutions for Social Good: Ph.D. Student Enlists LLM Assistants ...

Driven by her commitment to community service, Jaspreet Ranjit is leveraging the power of large language models to tackle pressing social issues like homelessness.

Homelessness is one of the most prevalent problems in California. In 2020, there were 161,548 homeless people living in California alone, which accounts for 28% of the homeless population in the United States. 

For the state’s most populous city, Los Angeles, the issue feels even more dire. In 2023, an estimated 75,518 individuals were unhoused on any given night in Los Angeles County, marking a 9% increase from 2022. According to the Los Angeles Homeless Services Authority, 46,260 of them resided in the City of Los Angeles. A recent Supreme Court decision allowing cities to ban homeless people from sleeping outside in public places paved the way for California Gov. Gavin Newsom to issue an order encouraging cities to dismantle homeless encampments. Just last week, Newsom even participated in cleaning up such a campsite in L.A. and threatened to withhold state money from counties that don’t demonstrate improvement on the issue.

In 2020, there were 161,548 homeless people living in California alone, which accounts for 28% of the homeless population in the United States. 

Ranjit, an incoming third-year Ph.D. student in the Thomas Lord Department of Computer Science, became interested in harnessing large language models to study public attitudes towards homelessness on social media. Her project – OATH-Frames: Characterizing Online Attitudes Towards Homelessness via LLM Assistants – won the Best Student Poster award in ShowCais 2024, a yearly symposium hosted by the USC Center for AI in Society that highlights the works of students and faculty that promote the use of artificial intelligence for social good.

“The state of homelessness in the U.S. is largely influenced by social and political factors, which sparks a diverse spectrum of attitudes on social media,” Ranjit said. “We characterize this discourse at a large scale by using LLMs as assistants to domain experts.” 

Using a subset of 2.4 million tweets collected from the social media platform X, formerly Twitter, Ranjit and her collaborators first developed the typology – Online Attitudes Towards Homelessness (OATH-Frames) – which consists of 9 frames capturing perceptions, critiques and responses on homelessness. The tweets collected for the project span between 2021 to 2023.

Given the mental and emotional challenges associated with manually annotating sensitive data, the researchers began to investigate the ability of LLMs as assistants to speed up the annotation process. The utilization of LLMs with domain experts in social work resulted in a 6.5x speed up in annotation time. A total of 8,000 posts were annotated to train a Flan-T5 model which was used to annotate the remaining collected tweets.

The annotated posts allowed the researchers to generate a large-scale analysis across U.S. states and time periods, revealing changing trends in attitudes towards homelessness across key socio-political events. The researchers hope this analysis would assist organizations in driving advocacy efforts.

OATH-Frames took about a year to complete with collaborations between the USC Viterbi School of Engineering and the Suzanne Dworak-Peck School of Social Work.

“Jaspreet has been interested in making a societal impact with AI right from day one.”  

“Jaspreet has been interested in making a societal impact with AI right from day one,” said Swabha Swayamdipta, a WiSE Gabilan assistant professor of computer science who is Ranjit’s advisor. “She spearheaded this effort, coordinating an interdisciplinary team of researchers in computer science and in social work – no small feat for anyone, let alone a junior researcher.”

From an early age, Ranjit has been heavily involved in community service. She has volunteered with a number of organizations since middle school and has started her own volunteering group in her hometown of Northern Virginia. Currently, she works alongside School on Wheels, an organization offering free tutoring services to children who are experiencing homelessness in the Greater Los Angeles area. 

After completing her bachelor’s and master’s education at the University of Virginia in computer science, Ranjit wanted to explore avenues where she could merge her interests in engineering and community service. This led her to pursue a doctoral degree at USC.

“I was interested in exploring ways I could combine my interests in engineering and serving the community, and I thought that pursuing a Ph.D. was the best way to do so,” Ranjit said.

Prior to starting her doctoral degree, Ranjit specialized in studying social and gender biases in visual recognition models. Previously, she interned at Vimeo, where she created a framework that quantifies gender biases within Vimeo’s video based search system. Additionally, her prior research as a Master’s student revealed that gender biases are amplified after fine tuning a model. 

Ranjit said their analysis using OATH-Frames identifies harmful generalizations towards people experiencing homelessness, which are often mislabeled by popular sentiment and toxicity classifiers. She hopes OATH-Frames serves as a resource to accelerate future work on utilizing LLMs to understand discourse on complex social issues. 

In the present day, LLMs are being deployed as annotation assistants within the social sciences. However, questions on their utility still remain open ended. Ranjit remains optimistic about the use of LLMs for social work, believing that despite the challenges, LLMs could be leveraged as powerful tools to tackle social problems.

“I feel excited about the potential uses of LLMs to serve as assistants to domain experts who are working head-to-head on society’s most pressing issues,” Ranjit said. “We hope our research serves as an avenue for future work on further investigating the utility of LLMs in the social sciences.”

Published on September 16th, 2024

Last updated on September 16th, 2024