Over the past decade, AI has reshaped our lives in numerous ways --- in how we communicate with one another, how we consume information, and how we form our political opinions, to name a few. Despite this tremendous social impact, AI techniques tend to have dangerously little, if any, awareness of the social contexts in which they operate in the real world. This is a cause of great concern, especially with the increasing influence these algorithms have in our daily lives.

My research seeks to build Socially Cognizant Artificial Intelligence by finding efficient and effective ways to incorporate social intelligence in AI algorithms, as well as by bringing these algorithms to new and socially beneficial application domains such as racial disparities in policing, power and gender in workplace, and abusive language online. I embrace truly interdisciplinary research that is driven by the challenges in the real world and is informed by social science theories through close collaborations outside of Computer Science, with researchers and practitioners from Linguistics, Psychology, Education, and Journalism.

• 

  NLP for Social Good (Law Enforcement)

  Can NLP help analyze police body camera footage to improve police community relationships?

  Collaborators

  Prof. Dan Jurafsky (Stanford; Linguistics and CS)
  Prof. Jennifer Eberhardt (Stanford; Psychology)
  Prof. Nelson Morgan (UC Berkeley; EECS)
  Prof. Benoit Monin (Stanford; Statistics)

  Projects

  Institutional Act Tagging for Traffic Stops: We build end-to-end machine learning models to automatically detect the recurring sequences of institution specific dialog acts during traffic stops (such as greetings, giving the reason for the stop, asking for documentation, issuing the sanction, etc.). Institutional acts combine the aspects of dialog acts and those of topical segments, all conditioned by the institutional context. We then study systamatic variations in the presence and sequence of these acts depending on factors such as type of stop and race of the driver, to uncover potential implicit bias.

  Automatic Traffic Narrative Mining: We build machine learning models to obtain actionable isnights from traffic narratives written by police officers describing the circumstances that led to a traffic stop and the events happened during the stop. In our first model, we built models to extract the discretion officers had in making the stop (severity of stops), which when applied to about 20,000 stops from the whole year, revealed that high discretion (less severe) stops disproportionally affect minorities.

  Computational Models for Officer Respect: We build computational models for police officer respect during traffic stops. Applied to one month of traffic stops, our model revealied significant racial disparities in officer respect.

  
  
• 

  NLP to Foster Civility in Interactions

  Can we build machine learning models to foster civility in online environments? Specifically, can we build computational models of subtle forms of incivility such as condescension?

  Collaborators

  Prof. Yulia Tsetkov (CMU; CS)
  Prof. David Jurgens (Univ. of Michigan; School of Information)
  Prof. Dan Jurafsky (Stanford; Linguistics and CS)
  Prof. Chris Potts (Stanford; Linguistics)

  Projects

  Coming Soon!

• 

  Computational Analysis of Social Power

  How is power manifested in language? Can we computationally detect powerful participants in interactions?

  Collaborators

  Dr. Owen Rambow (Columbia and Elemental Cognition; CS)

  Projects

  Coming Soon!

• 

  NLP for Social Good (Science)

  Can NLP help analyze decades worth of scientific literature to gain insights about the underlying social processes that guide scientific intellectual movements? Can we build predictive models that influence funding decisions and science policy?

  Collaborators

  Prof. Dan McFarland (Stanford; School of Education)
  Prof. Dan Jurafsky (Stanford; Linguistics and CS)
  Prof. Jure Leskovec (Stanford; CS)

  Projects

  Coming Soon!