Social media notifications flood our phones daily. Our feeds are filled with news, family updates, and ads, but these messages have untapped information that a Mason Engineering professor is mining using the power of technology and artificial intelligence (AI) to tackle social problems.
Hemant Purohit, assistant professor of information sciences and technology, has been working with Bonnie Stabile from the Schar School of Policy and Government for the last two years on their multidisciplinary research project to use AI techniques to look at the connection between social media and its impact on policy and law and the way people perceive misleading or misguided information on the web.
“We address this problem of how policies to support women empowerment are undermined due to the way women, who are considered weak by the power in a society, are socially constructed negatively in online social spaces by people with potentially malicious intent,” says Purohit. “Like the Stanford swimmer sexual assault case where some people took a stand for the accused, and they were maliciously creating uncertainty about the survivor perhaps due to prior biased beliefs.”
Purohit and Stabile wanted to see if one could recognize this social construction on the web and find the intent of the people doing it by translating this challenge into a computational problem of intent mining and employing state-of-the-art deep learning (AI) techniques.
Purohit and his PhD student Rahul Pandey in collaboration with Stabile and her PhD student Aubrey Grant created a computational framework to collect, process, and analyze social media messages discussing rape or sexual assault topics using natural language understanding methods of AI. They then categorized their data set of tweets into four categories based on intent: accusational, validational, sensational, and other, says Purohit.
“We found that nearly half of 100,000 tweets referencing key terms associated with ‘rape’ and ‘deception’ were accusational, meaning that they blamed, disparaged, and disbelieved women reporting rape or sexual assault,” says Stabile.
In comparison, only 12 percent of these tweets validated the victim, making accusational and negative tweets almost three times as prevalent, which Stabile and Purohit believe directly reflects an unfortunate reality of how social construction in online public conversations could influence the implementation and actual impact of policy and laws, given there are several strong policies and laws to curb gender violence and sexual assault.
“We believe that this both reflects damaging portrayals of those who experience sexual assault and can serve to perpetuate such portrayals,” says Stabile. “Such evidence is critical in bringing attention to how women, the predominant victims of sexual assault and harassment, can be mischaracterized in ways that can disadvantage them when policy is made.”
Purohit and Stabile have been working to combine their theoretical and computational expertise to tackle social issues and their work has been recognized at international forums.
Their paper led by PhD student Rahul Pandey, “Distributional Semantics Approach to Detect Intent in Twitter Conversations on Sexual Assaults,” won the best paper award at the 2018 IEEE/WIC/ACM International Conference on Web Intelligence hosted in Santiago, Chile. They recently published a journal article entitled, “She Lied: Social Construction, Rape Myth Prevalence in Social Media and Sexual Assault Policy.”
“We will continue to explore other application areas of policy and laws where we could take our multidisciplinary research approach to study the online social media and web spaces with theoretically-inspired computational methods,” says Purohit. “We are currently exploring hate speech on social media platforms and how to detect and mitigate that, again given the fact that there are already many policies and laws to prevent this phenomenon."