Word sense and semantic role analysis together form a "who does what to whom" analysis of free text, identifying and labeling events and their participants. I present a system for automatic word sense and semantic role analysis that uses FrameNet as a resource for sense and semantic role labels. I argue in favor of a modular toolchain for semantic analysis, with exchangeable components, with word sense disambiguation and semantic role labeling as two building blocks.

I then focus on the question of word sense, taking a closer look at data from a manual annotation effort where we annotated a German newspaper corpus with FrameNet-style information. In the construction of word sense inventories, and in the task of word sense disambiguation, it is usually assumed that different word senses are disjoint. But there are intriguing cases of word sense overlap that question this assumption. I sketch the road towards a fundamentally graded representation of word sense and speculate on suitable word sense disambiguation models.

About the speaker:

Katrin Erk is an assistant professor in the Department of Linguistics at the University of Texas at Austin. She completed her dissertation on tree description languages at Saarland University in 2002, advised by Gert Smolka. From 2002 to 2006, she held a researcher position in Saarbruecken working with Manfred Pinkal. Her current work includes research on machine learning methods for semantic analysis, the acquisition of lexical information from corpora, manual semantic annotation, the detection of multiword expressions, and computational models for word sense.