News & Events

Cognitive Science Dinner Talk

Benjamin Van Durme

Department of Computer Science, Johns Hopkins University

366 Meliora Hall

Universal Decompositional Semantics

The dominant strategy for capturing a symbolic representation of natural language has focused on categorical annotations that lend themselves to structured multi-class classification. For example, predicting whether a given syntactic subject satisfies the definition of the AGENT thematic role. These 
annotations typically result from professionals coming to mutual agreement on semantic ontologies. The Decompositional Semantics Initiative (decomp.net) is exploring a framework for semantic 
representation utilizing simple statements confirmed by everyday people, e.g., "The [highlighted syntactic subject] was aware of the [eventuality characterized by the salient verb]". This is conducive 
to a piece-wise, incremental, exploratory approach to developing a meaning representation that we can build software systems to target. 

Time permitting, I will relate this work to a sympathetic effort in common sense and natural language inference, two topics of increasingly larger interest within computational linguistics.