When people make decisions or form beliefs, they often discuss them with others, seeking out others' opinions and sharing their own. Recently such conversations are occurring online, providing a public source of information of interest to companies, the military, the government, public policy bodies, and educators. Moreover these dialogs occur at scale, allowing researchers in natural language processing access to large-scale dialog datasets for the first time. However, automatically processing such dialogs is challenging, because current tools are targeted to traditional language resources such as newspaper articles. This project develops innovative algorithms for automatically processing and identifying important phenomena in such dialogs including: (a) stance - participants' views on a topic; (b) subjective dialog acts - including sarcasm and humor; and (c) central propositions - core ideas in the dialog, by combining methods of crowd-sourced annotation, bootstrapping and machine learning, and cognitive science. A critical project output is a new corpus, including annotations and dialogic summaries.
Longer term impacts include public policy, providing government and the military with methods to discover what "the man on the street" is saying about current topics. Educators can re-use the corpora and tools to expose children to compelling arguments about important issues. Greater understanding of opinion sharing dialog enables new cognitive experiments and theory: automatically identifying compelling arguments allows political science and social psychology researchers to examine learning and opinion formation. The project trains undergraduate and graduate students in interdisciplinary research combining social media, human computer interaction, computational linguistics and natural language processing.
