How Rhetorical Structure Theory Guides Discourse Parsing AI

This article explores how Rhetorical Structure Theory (RST) serves as the foundational framework for discourse parsing artificial intelligence. By breaking down complex documents into basic functional units and organizing them into hierarchical tree structures, RST allows AI systems to identify and map coherence relations—such as contrast, elaboration, and cause—across entire texts. This structural mapping transforms raw, unstructured text into a logical map of meaning, enabling machines to understand not just individual sentences, but how ideas connect to form a coherent argument or narrative.

Breaking Text into Elementary Discourse Units

Before an AI can map how ideas relate to one another, it must first divide a document into manageable segments. Under the guidance of RST, discourse parsing AI begins by segmenting text into Elementary Discourse Units (EDUs). EDUs are typically clause-like structures that serve as the minimal building blocks of discourse.

For example, the sentence "Because it was raining, the match was canceled" is split into two distinct EDUs:

  1. [Because it was raining,]
  2. [the match was canceled.]

Machine learning models use syntactic cues, punctuation, and lexical markers to accurately identify the boundaries of these EDUs. Once the text is segmented, the AI has the base components required to start mapping relationships.

Identifying Nucleus and Satellite Dynamics

A core principle of RST is asymmetry in text. Not all clauses carry equal weight. RST categorizes EDUs into two main roles: the Nucleus and the Satellite.

  • The Nucleus contains the central, essential piece of information. If it were removed, the surrounding text would lose its core meaning.
  • The Satellite contains supplementary information that supports, explains, or modifies the nucleus.

Discourse parsing AI uses this distinction to determine the hierarchy of importance within a document. In our previous example, "[the match was canceled]" is the nucleus, while "[Because it was raining]" acts as the satellite. In rare cases where two clauses hold equal weight—such as a list or a balanced contrast—RST allows the AI to classify the relationship as multinuclear.

Mapping Specific Coherence Relations

Once the AI identifies the nuclei and satellites, RST provides a strict taxonomy of coherence relations to label how these units interact. AI models look for discourse markers (words like "however," "therefore," or "furthermore") as well as implicit semantic cues to assign these labels.

Here is how RST guides AI to map the three most critical types of relations:

1. Cause (and Result) Causal relations explain why something happened. RST guides the AI to look for cause-and-effect dynamics. If the AI detects a satellite that explains the reason behind the action in the nucleus, it maps a "Cause" or "Reason" relation. This is heavily reliant on recognizing both explicit conjunctions ("because," "since") and implicit chronological or logical sequences.

2. Elaboration Elaboration occurs when a satellite provides additional detail about a concept introduced in the nucleus. Because elaboration often lacks explicit connecting words, AI relies on RST's structural rules alongside semantic similarity models (like neural embeddings). If a subsequent clause drills down into a specific attribute, example, or step related to the previous clause, the AI maps an "Elaboration" relation, linking the detailed satellite back to the broader nucleus.

3. Contrast Contrast relations highlight differences or contradictions between two segments. Depending on the context, RST helps the AI classify this as either a nucleus-satellite relation (such as "Concession," where one point yields to a stronger point) or a multinuclear "Contrast" relation (where two equal points are compared, e.g., "John likes apples, whereas Mary likes oranges"). The AI detects these through adversative markers ("but," "although," "on the other hand") and antonymic semantic patterns.

Constructing the Hierarchical Discourse Tree

RST stipulates that coherence relations are recursive. This means that once two EDUs are linked together, they form a larger span of text that can then act as a single nucleus or satellite to an even larger section of the document.

Discourse parsing AI uses transition-based (shift-reduce) algorithms or graph-based neural networks to build these connections from the bottom up, or the top down.

  1. Local Level: The AI links adjacent clauses into sentences.
  2. Paragraph Level: It links sentences together to determine the main point of a paragraph and how the surrounding sentences support it.
  3. Document Level: It links paragraphs and sections to map the overarching thesis of the entire text.

The result is a single, comprehensive tree structure that spans the entire document. Through this RST-guided tree, the AI achieves a complete map of the document's logical flow, allowing downstream applications—such as text summarization, sentiment analysis, and automated question-answering—to extract core truths (nuclei) while understanding the context (satellites) that shapes them.