Establishing a Mechanism-Based Framework for the Corpus-Informed Analysis of Multi-Word Discourse Markers