Back to AI Research

AI Research

RAG-RST uses retrieved examples to teach discourse structure without fine-tuning

Key Takeaways

  • The framework supplies annotation conventions through in-context demonstrations, with dynamic retrieval performing best in the reported comparisons.
  • RAG-RST studies discourse parsing: identifying how spans of text relate and which span carries the central content.
  • In Rhetorical Structure Theory, a tree edge includes both a nuclearity label and a relation such as elaboration or background.
  • The authors emphasize that the same pair of clauses can receive different labels depending on rhetorical intent and annotation conventions.
  • Familiarity with language is therefore not the same as knowing the formal scheme used by an evaluator.

The structure behind a paragraph

RAG-RST studies discourse parsing: identifying how spans of text relate and which span carries the central content. In Rhetorical Structure Theory, a tree edge includes both a nuclearity label and a relation such as elaboration or background.
The authors emphasize that the same pair of clauses can receive different labels depending on rhetorical intent and annotation conventions. Familiarity with language is therefore not the same as knowing the formal scheme used by an evaluator. That distinction motivates supplying examples of the scheme rather than assuming a model already applies it consistently.

Retrieval supplies the schema

The framework retrieves examples as in-context demonstrations without updating model parameters. The paper compares a precomputed strategy, which labels discourse-unit pairs before parsing, with a dynamic strategy that retrieves examples at each reduce step. In the dynamic version, the parser's earlier decisions affect which examples reach the generator next.
The authors report nuclearity F1 rising from 0.794 to 0.846 and relation F1 from 0.738 to 0.788 with the supplied schema. Both configurations outperform the compared RST parsers across the reported tree metrics, with the dynamic variant consistently strongest.

More examples are not always better

The study finds that the benefit saturates and then declines beyond 50 retrieved examples. It also depends on structural position. Those findings suggest that example selection and placement matter, rather than retrieval volume alone determining performance.
Franklin's reading is that the paper offers a focused lesson for formal annotation tasks: retrieve demonstrations of the intended convention, then check whether they help at the point of decision. These results concern the evaluated discourse-parsing setup. They do not establish that adding retrieval will improve every language task, or that a larger context automatically compensates for mismatched examples. A reproduction should track retrieval strategy, example count and the separate label metrics.
Read the RAG-RST paper.

Comments