Causal Relations in Discourse for Pragmatic Inference

Angela Cao

Date: 2022-01-28 / 4:00 ~ 5:00 PM
Location: https://emory.zoom.us/j/99364825782


Abstract

we aim to create a corpus of newspaper articles, short stories, and Reddit threads manually annotated for causal relations from Wolff and Song’s (2003) categories of CAUSE, ENABLE, and PREVENT, which are based on Talmy’s (1988) force dynamics theory of causation. Previous work in computational linguistics (Mostafazadeh et al., 2016; Mirza et al., 2014; Dunietz, 2018) have implemented this categorization in created corpora annotated for causal relations, but have been constrained to limited appearances of such. In our project, we adapt a pre-identified bank of causal connectives (the Constructicon) from Dunietz, which we use as triggers for annotation instances. The goals of this project are thus threefold: (1) to support Wolff and Song’s causal concepts through high inter-annotator agreement of distinguishing between instances of such, (2) to build upon previous annotation work that aim to embed this theory of causation, and (3) to extract frequent causal relations to build commonsense inference. By achieving high agreement, our corpus will provide high quality data for understanding textual causality.

Link

Presentation