Must every relation have a satellite?

No; multinuclear relations connect nuclei.

Can a relation connect larger spans?

Yes, including spans with internal structure.

Does Evidence mean verified proof?

No; it describes a rhetorical role.

AI Glossary / Natural language and speech / Advanced

RST

Definition

RST stands for Rhetorical Structure Theory, a framework for describing how parts of a text relate to one another. It represents relationships such as evidence, elaboration, and contrast, often distinguishing a more central span, the nucleus, from a supporting span, the satellite. Some relations connect equally central nuclei.

What it means

RST asks what role a passage plays in text organization. A sentence might supply evidence for a claim, explain a circumstance, or add detail to an earlier statement. These relationships help describe why the text is more than an unordered collection of sentences.
Centrality concerns the writer's communicative purpose in the analysis. It does not mean that the nucleus is always the first sentence, the longest passage, or the only information a reader needs. A satellite can contain crucial qualifications or support.
RST provides an analysis of discourse organization, not a truth detector. Calling a passage Evidence describes its proposed rhetorical relationship to another passage. It does not independently establish that the evidence is accurate, sufficient, or persuasive to every reader.

How it works

An analysis begins by dividing the text into elementary discourse units, often clauses or other short spans under specified annotation conventions. Relations then connect spans, including spans that already contain smaller structures. This creates a hierarchical account of the text.
In a nucleus-satellite relation, the two spans have different roles. For Evidence, supporting material is interpreted as increasing the reader's belief in the nucleus. For Elaboration, a satellite adds detail to something presented or accessible in the nucleus. Multinuclear relations instead connect spans with comparable status; Contrast is one example.
An automated discourse parser must identify boundaries, structure, nuclearity, and relation labels. Errors at one stage can affect later decisions. A correct label on the wrong pair of spans is not a correct overall analysis.
Annotation schemes differ in their relation inventories and conventions. Evaluation therefore needs a specified scheme and reference analysis. Words such as because or although can be useful signals, but a connective alone does not determine every relation, and a relationship can be understood without an explicit connective.

Three practical examples

A summarization researcher examines which passages carry the main claims and which provide supporting details, while checking that compression preserves qualifications.
A writing-support tool highlights a claim whose intended supporting paragraph appears disconnected from it.
A document-analysis project annotates how explanations and comparisons connect across paragraphs before training a discourse parser.

Think of it this way

Think of a presentation with main points, supporting slides, and equally emphasized comparisons. RST describes comparable organization in text. That organization does not prove the claims.

Worked example

Consider three fictional discourse units: U1, “The new packing method reduced breakage in this trial.” U2, “The old method produced twelve broken items out of one hundred.” U3, “The new method produced four broken items out of one hundred.”
Assume the writer intends U2 and U3 as a balanced comparison supporting U1. One plausible teaching analysis joins U2 and U3 through a multinuclear Contrast relation. Their combined span then serves as the satellite in an Evidence relation whose nucleus is U1.
This analysis has three leaf units and two internal grouping nodes. The nodes cover U2 - U3 and U1 - U3. It contains one multinuclear relation and one nucleus-satellite relation. The larger satellite contains two local nuclei, showing that a span can be central within a smaller grouping while serving a supporting role in a larger grouping.
The supplied figures give breakage proportions of 12/100=12% and 4/100=4%. The observed difference is eight percentage points. Relative to the old method's proportion, the reduction is 8/12, approximately 66.7%. These arithmetic facts help explain the intended evidence relationship, but the passage supplies no study-design details that would establish a general causal conclusion.
Now suppose a parser preserves both groupings and the Evidence direction, but labels the U2 - U3 relation as Sequence. Under a deliberately simple exercise score over the two fixed relation labels, it gets one of two correct, or 50%. That number is not a standard complete RST benchmark score: the exercise assumes the unit boundaries and grouping structure are already correct.
A second parser labels the top relation Evidence but reverses nucleus and satellite. Matching the word Evidence alone would miss the error. The proposed support and the supported claim now have the wrong roles, even though the relation name looks right.
Finally, keeping only U1 would preserve the main claim under this analysis but remove its numerical basis. Keeping only U2 and U3 would preserve the comparison but omit the writer's explicit conclusion. Choosing a useful summary therefore requires a purpose and a fidelity check, rather than mechanically deleting all satellites.
This invented analysis is illustrative; other interpretations may be possible.

Illustrative case study

Fictional example: BriefBuilder produces short summaries of internal trial reports. Its team experiments with a rule that favors nuclear spans. Reviewers find that some summaries retain confident conclusions while dropping the passages that explain their limited scope.
The team keeps the discourse analysis as one input but adds checks for qualifications and supporting figures. It compares summaries with the source passages before accepting the rule. This fictional scenario illustrates evaluation design.

Common misconceptions and mistakes

RST is not a sentence-syntax tree with different labels. Its relations concern how text spans function together. A nucleus is also not automatically true, and a satellite is not automatically dispensable.
Another mistake is assuming one fixed universal label set or one certain interpretation for every passage. Annotation guidelines and communicative context matter. Do not score only relation names when the span boundaries, attachment, or nuclearity can be wrong. Nor should a predicted structure be treated as direct access to the writer's private intentions.

Business use

Discourse analysis can support document review and summarization experiments when teams define a clear use case. Keep the original passages available for inspection. Evaluate whether outputs preserve claims, qualifications, and evidence. A parsing score alone does not establish summary quality.

Check your understanding

Which unit is the top-level nucleus in the example?

Answer: U1.

Can U2 and U3 be local nuclei inside a larger satellite?

Answer: Yes.

What is the observed difference in breakage rates?

Answer: Eight percentage points.

Related terms

Discourse parsing
Argument mining
Discourse coherence
elementary discourse unit
nucleus
satellite
rhetorical relation
text summarization

Learn next

Annotate a short passage twice, recording its unit boundaries, grouping, nuclearity, and relations separately. Compare disagreements before interpreting a single overall score.

Sources

Jurafsky and Martin, Speech and Language Processing, chapter 25, discourse coherence and RST: https://web.stanford.edu/~jurafsky/slp3/25.pdf RST website, relation definitions, including Evidence, Elaboration, and multinuclear Contrast: https://www.sfu.ca/rst/01intro/definitions.html RST website, introduction to the framework: https://www.sfu.ca/rst/01intro/intro.html

Previous: Rprop Next: Rubin causal model

RST