Improving Computational Models of Semantics Using Discourse

Michael Roth

Univeristy of Edinburgh

Tuesday, April 26, 2016
4:00 p.m., ICSI Lecture Hall

In this talk, I will discuss how models of lexical and event-level semantics can be improved by taking into account knowledge or context on the discourse level. The talk will begin with motivation based on earlier linguistic studies. I will then present two recent lines of work by myself and colleagues that showcase the potential of using discourse knowledge.

The first task concerns distinguishing between paradigmatic relations such as synonymy, antonymy and hypernymy. Within this application, we explore how far discourse relations can serve as an alternative or complementary set of features to traditional lexico-syntactic patterns. For the second task, semantic role labeling, we define a novel set of features that take into account sentence and discourse context. We experimentally demonstrate that such features lead to significant improvements of a state-of-the-art model.

The main conclusion of the talk will be that discourse matters and that we should keep this in mind when designing/learning computational models for word-level and sentence-level tasks.

Bio:

Michael Roth is a DFG Research Fellow and visiting researcher at the University of Edinburgh. He holds a PhD in Computational Linguistics from Heidelberg University (2013) and a MSc in Language Science and Technology from Saarland University (2008). His research focusses on developing computational models of language that can facilitate automatic text understanding beyond the sentence level.