Social media platforms have recently emerged as the primary space for public conversations, providing a venue for people to share perspectives and for policymakers to promote their decisions and inform the public about them. The massive amount of available opinion data presents tantalizing opportunities to study the perspectives expressed on these platforms. Insights derived from this analysis can help gauge public opinion, inform public policy, and help support human decision making. Realizing these opportunities requires models adapted to the new social media settings, in which linguistic content and its social context cannot be separated. This CAREER project develops novel modeling techniques and learning algorithms for combining these two aspects under a common innovative principle -- creating a socially grounded language representation that views opinion understanding as part of a larger framework of understanding real-world scenarios (such as the implementation of specific policies or the response to an emergency situation) and their participants. This research helps provide the relevant context needed for better understanding social media content and result in highly nuanced analysis, capturing the stances, attitudes and relationships between the different stakeholders of a given real-world scenario.
This project suggests a new way to conceptualize opinionated text analysis, as part of a real-world scenario, reflecting the attitudes-of and relationships-between stakeholders in the scenario from which the text emerges. A major design goal is to avoid the supervision bottleneck, and allow the system to easily adapt to new events and policy issues by using the social information associated with users as a form of indirect supervision over documents they author. This is done by representing documents, authors, referenced entities, their connections and behaviors in a shared neuro-symbolic framework enabling symbolic inference over latent entity representations learned from data. The project addresses three main challenges: (1) constructing a representation language for characterizing opinions, their targets and motivation, and the stances they express, (2) grounding opinion text in real world scenarios by infusing relevant real-world information into a neural language model, and (3) exploiting social information by formulating a unified view of social, behavioral, and textual information. These research efforts help provide nuanced insights from social media content that lacks specificity on its own, while building the computational foundations for jointly processing textual and social information.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
