Filling in the Blanks in Understanding Discourse Adverbials: Consistency, Conflict, and Context-Dependence in a Crowdsourced Elicitation Task

摘要

The semantic relationship between a sentence and its context may be marked explicitly, or left to inference. Rohde et al. (2015) showed that, contrary to common assumptions, this isn't exclusive or: a conjunction can often be inferred alongside an explicit discourse adverbial. Here we broaden the investigation to a larger set of 20 discourse adverbials by eliciting ≈28K conjunction completions via crowdsourcing. Our data replicate and extend Rohde et al.'s findings that discourse adverbials do indeed license inferred conjunctions. Further, the diverse patterns observed for the adverbials include cases in which more than one valid connection can be inferred, each one endorsed by a substantial number of participants; such differences in annotation might otherwise be written off as annotator error or bias, or just a low level of inter-annotator agreement. These results will inform future discourse annotation endeavors by revealing where it is necessary to entertain implicit relations and elicit several judgments to fully characterize discourse relationships.