Annotating causal language using corpus lexicography of constructions

Abstract

Detecting and analyzing causal language is essential to extracting semantic relationships. To that end, we present an annotation scheme for English causal language (not metaphysical causality), and discuss two methodologies for annotation. The first uses only a coding manual to train annotators in distinguishing causal from non-causal language. To address low inter-coder agreement, we adopted a second methodology, in which we first created a causal language constructicon based on corpus analysis, then required annotators only to annotate instances based on the constructicon. (This resembles the methodology used for annotating the FrameNet and PropBank corpora.) Our contributions, in addition to the annotation scheme itself, are methodological: we discuss when constructicon-based methodology is appropriate, and address the validity of annotation schemes that require expert-level metalinguistic awareness.