Graded Decompositional Semantic Prediction

Adam Teichert, Johns Hopkins University
Host: Benjamin Van Durme

Language suggests information about entities and events—real or imagined. We are interested in inferring such information or meaning, such semantics, from text. In this dissertation, we build upon and contribute to a decompositional view of semantic prediction which is inherently (1) structured—multi-dimensional with correlation and possibly constraints among the possible semantic questions, (2) graded—predicted quantities represent magnitudes or probabilities rather than binary or categorical values, and (3) subjective. Combining these aspects leads to interesting opportunities for modeling and annotation and raises important questions about the impact of these practices. Specifically, we propose the first structured model for the task of Semantic Proto-Role Labeling, casting the structured problem as a multi-label prediction task which we related empirically to semantic role labeling. We subsequently propose mathematical models of structured ordinal prediction that allow us to incorporate graded annotation and to jointly model multiple annotators. We investigate the decompositional semantic prediction task of Situation Frame Identification (a flavor of topic identification) and propose a graded model for the binary task. Finally, we address issues in efficient scalar annotation.

Speaker Biography

Adam Teichert is a PhD candidate in the Center for Language and Speech Processing and an Assistant Professor of Software Engineering at Snow College in Ephraim, UT. Before coming to Johns Hopkins, he received a B.S. in Computer Science from Brigham Young University and a MS in Computing from the University of Utah. His research has explored methods for efficient learning and inference in natural language processing with recent focus on structured models and related methods for decompositional semantic labeling and topic identification.