PDTB-based text enrichment for coherence classification

Determine whether enriching text with Penn Discourse TreeBank relations, rather than Rhetorical Structure Theory or Part-of-Speech information, achieves better coherence-classification results, and assess whether its performance is directly related to the overall adequacy of the text-enrichment technique.

Background

The paper argues that RST enrichment is poorly aligned with the Transformer architecture because RST encodes hierarchical discourse relations, whereas Transformer inputs are linearized sequences. The authors note that the Penn Discourse TreeBank provides a linear discourse annotation framework and is therefore structurally aligned with Transformers. They identify implementing PDTB-based enrichment and repeating the experiments as a future investigation, explicitly framing the expected outcome as a hypothesis.