On the Relation between Implicitness and Predictability: Combining Annotations of Evoked Questions and Discourse Relations

This thesis investigates how readers’ evoked questions align with discourse relations and how that alignment governs the omission of explicit connectives. We compared two annotation schemes on the same data, one capturing discourse senses (TED-MDB) and the other capturing evoked questions (TED-Q). This was done by annotating the evoked questions from TED-Q using the same PDTB sense hierarchy as TED-MDB. We show that when a reader’s anticipated sense matches the actual relation, authors are more likely to leave the relation implicit. We further demonstrate that causal and resultative relations attract the strongest alignment, indicating that readers’ evoked questions predict these relations most reliably. Finally, by comparing sense alignment with the extent to which questions are answered, we find that alignment is a more powerful predictor of implicitness than answered-score. These findings support several linguistic hypotheses surrounding predictability and suggest features for...Show moreThis thesis investigates how readers’ evoked questions align with discourse relations and how that alignment governs the omission of explicit connectives. We compared two annotation schemes on the same data, one capturing discourse senses (TED-MDB) and the other capturing evoked questions (TED-Q). This was done by annotating the evoked questions from TED-Q using the same PDTB sense hierarchy as TED-MDB. We show that when a reader’s anticipated sense matches the actual relation, authors are more likely to leave the relation implicit. We further demonstrate that causal and resultative relations attract the strongest alignment, indicating that readers’ evoked questions predict these relations most reliably. Finally, by comparing sense alignment with the extent to which questions are answered, we find that alignment is a more powerful predictor of implicitness than answered-score. These findings support several linguistic hypotheses surrounding predictability and suggest features for various natural language processing applications.Show less

Leiden University Student Repository

Documents

In Collections

On the Relation between Implicitness and Predictability: Combining Annotations of Evoked Questions and Discourse Relations