CHS: Small: Collaborative Research: Human-Centered Semantic Relatedness

Northwestern University, Evanston IL

Investigators

Abstract

This work addresses key open questions about semantic relatedness (SR) measures, a family of algorithms used throughout computer science and related fields that help computers replicate human assessments of the relatedness between two concepts. Decades of research and development have transformed SR measures into a critical component of a wide swath of intelligent technologies in areas ranging from information retrieval to human-computer interaction to spatial computing. However, despite the importance and ubiquity of SR, researchers have only recently begun to examine it from a human-centered perspective. These human-centered studies have problematized key assumptions underlying the entire SR literature, e.g. that people from all cultural contexts agree on a single relatedness value between any two concepts. This project addresses long-overdue open questions in the SR literature that will move the field towards a more human-centered approach to SR. To do so, this research will collect datasets of relatedness judgments, mine patterns in Wikipedia and other content, perform statistical analyses, create and evaluate algorithms, develop software, and conduct large-scale user studies. First, this research will complete four threads of work that redefine semantic relatedness to address human-centered concerns raised in the SR literature: (1) investigate the role of culture in SR and use these results to redefine SR to incorporate cultural context, (2) study SR among the low-notability concepts that are critical to end users but entirely ignored by the SR literature, (3) address the need for SR measures that explain their relatedness estimates to users, and (4) develop more robust human-centered SR evaluation procedures and support their adoption through easy-to-use software. Second, this research will develop new conceptual representations for SR measures that accommodate differing cultural perspectives and create compact contextual SR models that empower applications with tractable human-centered SR algorithms. Finally, the research will demonstrate the power of human-centered SR approaches through their application in improved recommender systems, enhanced Wikipedia reader experiences, and novel information discovery tools.