AI Glossary / Natural language and speech / Intermediate

EDU

Definition

EDU stands for elementary discourse unit: a minimal segment used as a building block in a discourse analysis, especially Rhetorical Structure Theory. EDUs are combined into larger spans connected by discourse relations. They often correspond to clauses, but the exact boundaries depend on the annotation guidelines rather than on sentence punctuation alone.

What it means

A sentence can contain more than one communicative unit. “The server restarted because the update required it” contains a main statement and an explanatory clause that a suitable discourse analysis can represent separately.
The word “elementary” means minimal for the chosen discourse representation. It does not mean a single word, a fixed number of tokens, or a complete standalone sentence in every case.
An EDU is also different from an arbitrary document chunk used for storage or retrieval. A chunking system may split text by length; EDU segmentation aims to identify units relevant to discourse structure. Confusing these purposes can lead to misleading descriptions of a system's output.

How it works

An annotator or segmentation model locates boundaries according to a defined scheme. The scheme specifies how to handle clauses, coordination, quotations, relative clauses, and other constructions.
Once the units are established, discourse analysis connects them and larger spans through relations such as explanation or contrast. In a typical RST tree, EDUs form the leaves, while higher nodes represent combinations of units.
Segmentation affects everything that follows. If an explanation is merged into the wrong unit or a single construction is split inappropriately, downstream relation predictions are evaluated over the wrong pieces. Reporting whether a parser receives gold-standard EDUs or predicts its own boundaries is therefore essential.

Three practical examples

• A discourse annotation team segments a report into EDUs before assigning relationships between its parts.
• A parser evaluation compares predicted boundary positions with a reference segmentation under the same guidelines.
• A summarization study examines which combinations of discourse units preserve the main claim and its necessary qualifications.

Think of it this way

Think of EDUs as the pieces chosen for a particular analytical jigsaw. Larger structures are built by relating those pieces. The analogy captures composition, but language does not arrive with universally agreed cut lines: segmentation is guided by a theoretical and annotation framework, and some boundaries require judgment.

Worked example

Consider this invented passage: “Although the prototype passed its basic tests, the team extended the trial. Several users still needed onboarding.”
For a simplified clause-oriented segmentation, mark U1 = “Although the prototype passed its basic tests,” U2 = “the team extended the trial,” and U3 = “Several users still needed onboarding.” The first sentence contains two units, while the second contains one.
The segmentation alone does not assign relations. A later analysis might connect U1 and U2 through concession and interpret U3 as explaining the extension. Those are additional decisions, not automatic consequences of having three units.
Now consider “The team believed that onboarding was incomplete.” Some guidelines, including the cited Stede, Taboada, and Das scheme, keep this kind of cognitive-verb complement together rather than treating every embedded clause as a separate EDU.
The lesson is that “split at every clause” is too crude. A project must state its segmentation policy before treating the resulting units as reference answers. These examples illustrate the selected distinctions rather than prescribing one universal segmentation for all corpora.

Illustrative case study

Fictional case: DiscourseDesk builds an annotation tool for short business reports. Its first version splits text at every comma and labels the pieces as EDUs. Reviewers find list items incorrectly separated and meaningful clause boundaries missed when no comma appears.
The team adopts an explicit segmentation guide and adds examples for coordination, embedded complements, and explanatory clauses. It records disagreements for adjudication rather than assuming punctuation always decides the boundary.
The resulting dataset documents its conventions alongside the annotations. This fictional case illustrates a quality-control process, not a claim that one guide is universally best or that segmentation alone makes the reports understandable to an AI system.

Common misconceptions and mistakes

An EDU is not necessarily a sentence, and a sentence is not necessarily one EDU.
Do not equate discourse segmentation with splitting text into fixed token windows. The objectives differ.
Avoid assuming that every syntactic clause must be separated under every scheme. Guidelines differ, especially for embedded and attributed material. Also distinguish a boundary error from a relation-labeling error; combining them without explanation can hide the source of a parser's failure.

Business use

The term matters when teams use discourse annotations, evaluate parsers, or design structured text-analysis datasets. It helps explain exactly what units the system is connecting.
A practical review should inspect the segmentation guide, a sample of difficult cases, and the conditions under which performance was measured. If evaluation supplies correct EDUs in advance, the reported result describes a narrower task than processing raw documents end to end. That distinction affects expectations about implementation work and downstream reliability.

People also ask

Is every EDU a complete sentence?

No. Many EDUs are clauses or other guideline-defined segments within a sentence.

Are EDU boundaries universal across datasets?

No. Annotation schemes can make different choices, so comparisons require compatible definitions.

Does identifying EDUs also identify discourse relations?

No. Segmentation establishes the units; linking them and labeling their relationships are further parts of discourse analysis.

Check your understanding

A system splits a report into blocks of exactly 100 tokens and calls each block an EDU. Is the fixed length enough to justify that label?

Answer: No. EDU status depends on discourse segmentation conventions, not an arbitrary token count. A length-based block may cut through a unit or combine several units.

Related terms

• Discourse parsing
• Rhetorical structure theory
• discourse segmentation
• discourse span
• nucleus
• satellite
• coherence relation
• sentence boundary

Learn next

Segment a short passage using a stated RST guide, then compare it with punctuation-only splitting. Mark which differences follow from the guide and which need judgment. Only afterward assign relations between the resulting units.

More in Natural language and speech

• Dynamic programming - Dynamic programming is a method for solving a problem by combining solutions to smaller subproblems and reusing their results.
• Dynamic time warping - Dynamic time warping, or DTW, aligns two sequences while allowing differences in their local timing.
• Edge-factored - Edge-factored describes a structured model whose score for a graph or tree decomposes into contributions from its individual edges.
• Effect size - Effect size is a numerical description of how large a difference or association is.
• Elman Networks - Elman networks are simple recurrent neural networks that feed the previous hidden state into the computation of the next hidden state.
• Emission probabilities - Emission probabilities describe how likely particular observations are given a hidden state in a probabilistic model such as a hidden Markov model, or HMM.

Sources

Sources: Stede, Taboada, and Das, Annotation Guidelines for Rhetorical Structure, segmentation section: https://www.sfu.ca/~mtaboada/docs/research/RST_Annotation_Guidelines.pdf ; Jurafsky and Martin, Speech and Language Processing, Discourse Coherence: https://web.stanford.edu/~jurafsky/slp3/25.pdf Independently checked the segmentation guidelines, including their specific treatment of cognitive complements and adjunct clauses.

Previous: Edmonds-Karp Next: Effect size