CRII: RI: Explainable Recognition of Social Relationships from People's Linguistic Interactions

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

Social relationships between people influence not only what is said but how a message is said through linguistic cues for aspects like respect, politeness, or solidarity. These stylistic aspects not only reflect a particular social context, but can also affect how the recipient interprets the message itself. These relationships are often multifaceted and come with different linguistic expectations. For example, while parents, professors, and police may all choose the language of authority, in their respective interactions with children, students, and the community, their language reflects a broader set of linguistic expectations in how they relate to others. This project aims to study social relationships and their effect on language by building new technology that recognizes the linguistic cues of relationships resulting in significant improvements in everyday interactive language technologies such as voice assistants, translation, and summarization and lead to more natural human-machine interactions by improving the ability of machines to understand social aspects of communication that cannot be obtained from the words alone. This research will also support an interdisciplinary cohort of PhD and undergraduate students and the development of a graduate-level course on Computational Sociolinguistics with open access materials. This CISE Research Initiation Initiative (CRII) project investigates social relationships and their impact on language by developing new algorithms, methods, and datasets, building upon sociolinguistics theory to create a new framework for studying the interplay of language and relationships in real-world social environments. This project contains two main phases. The first phase develops new deep learning methods to accurately infer people's relationship from linguistic cues in their conversation. To support this effort, it also constructs a new, massive corpus of interpersonal conversations from real-world settings and from fictional settings in movies, television, and literature. The second phase builds interpretable computational models of communication that explain how a relationship is signaled through linguistic choices in communication, while also disentangling the linguistic influences of the conversational partners' identities. These models enable testing and extending social science theories about relationship and social organization. Further, this project also introduces a new natural language understanding task for identifying messages that are offensive only in certain contexts, where the linguistic cues violate social norms for a relationship, helping to further improve technologies to better the public's online experiences. 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.