Architecturally compatible RST enrichment for coherence classification

Investigate whether enriching text with Rhetorical Structure Theory information yields better coherence-classification results when coupled with architectural modifications that account for hierarchical structures, such as Poincaré embeddings or Graph-aware Isomorphic Attention.

Background

The paper finds that inserting RST-derived special tokens into XLM-RoBERTa Longformer inputs reduces coherence-classification accuracy relative to the plain-text pipeline. The authors attribute this result to a structural mismatch between hierarchical RST representations and the Transformer architecture, which flattens the discourse structure. They therefore propose investigating architectures that explicitly model hierarchy or graph structure, while presenting the expected benefit as an unverified hypothesis.