Determine the effective form of retrieved knowledge for RST schema induction

Determine what form retrieved knowledge must take at inference time to supply an effective RST annotation schema, including how the number, consistency, and structural relevance of labeled neighboring examples affect nuclearity and relation prediction.

Background

RAG-RST addresses the difficulty that foundation models may recognize discourse phenomena while lacking the annotation convention that maps those phenomena to specific RST nuclearity and relation categories. Retrieval-augmented generation can provide labeled examples of that convention at inference time, but the effective design of the retrieved context is not established.

The unresolved issue concerns the properties of the retrieved demonstrations: how many examples should be supplied, how consistently they should reflect the annotation scheme, and how their usefulness varies with the target pair’s position in the discourse structure. The paper evaluates retrieval size and structural conditioning, but frames the broader question of the required form of supplied knowledge as remaining open.

References

What form the supplied knowledge must take, however, remains open: labeled neighbors function as demonstrations of an annotation scheme, and their utility depends on how many are shown, how consistent they are, and where in the discourse structure the target pair sits \citep{mei2025survey}.

— RAG-RST: Retrieved Exemplars as In-Context Discourse Schema for RST Parsing  (2610.22150 - Sediqin et al., 15 Oct 2026) in Section 1, Introduction