Rhetorical Structure Theory (RST) models discourse as a tree, where each edge carries two labels: a nuclearity label indicating which span holds the central content, and a relation drawn from a closed inventory of functional categories. Getting these labels right is a core challenge in discourse parsing, as they capture how clauses and sentences connect rhetorically.
The paper introduces RAG-RST, a retrieval-augmented approach that pulls in exemplars and uses them as in-context discourse schema for parsing. Rather than relying solely on a fixed model, the parser is guided by relevant examples retrieved at inference time, which may help it apply the relation inventory more consistently.
The abstract does not report quantitative results, so the contribution here is the method itself. As a single source, this summary reflects only the paper's framing; no independent comparison or follow-up work is discussed.