Extraction of Force-dynamic Image Schemas for Event Structure Representation of Procedural Text
Speaker: Michael Regan (University of New Mexico)
Date and Time: November 19 (Tuesday)
Event structure decomposition is integral to machine reading comprehension, dialog state tracking, and other natural language understanding tasks where models of entity states and temporality as well as of event-event, participant-event, and participant-participant relations are desirable. Previous work in event representation has yet to show how organizing conceptual structures into well-defined, flexible units representing both background knowledge of the world and how that knowledge is expressed in a certain language might improve AI reasoning capabilities. In preparation for my dissertation research into this question, in this talk I will argue for a fine-grained approach to event representation based on theoretical work done in cognitive semantics: force dynamics (Talmy 1988; Croft 2012), a image schematic model of causation that my research may show can support common-sense reasoning applications. Event representation is often done at the macro-level, with a focus on sequences of events in narratives, procedures, etc.; in contrast, at a micro-level we examine subevents of clausal events by employing the entity-centric, fine granularity of a force-dynamic approach to characterize participant interaction as a function of time, qualitative change, and transmission of force. As a proof of concept linking micro- and macro-level analyses, I propose designing, implementing, and evaluating a computational model for the extraction of participant histories as storylines from scientific, procedural text. One task will be to examine the use of dynamic knowledge graphs as a representation for evolving fine-grained storyline event structure and the tracking of entity states. A second task will be to compare how well force-dynamic event structure can be extracted using a symbolic approach based on a model of argument structure and verb classes versus a neural approach using pre-trained language models. A third task will be to examine the applicability of image schemas for effect prediction via a generative process with force-dynamic structures as priors. The primary hypotheses of the study will be examined along with a tentative timeline for how my research may progress.
Michael Regan is a PhD student in Linguistics and MS student in Computer Science at the University of New Mexico, and a Professional Research Assistant in Computer Science at the University of Colorado Boulder. His research interests include cognitive semantics, event structure representation, multilingual NLP, and representation learning.