🌐 discourse: discourse element
discourse is the Universal Dependencies relation that attaches interjections, fillers, discourse markers and similar expressive elements to the clause they occur in. Such elements carry no grammatical role within the clause's argument or modifier structure.
What it is
In Universal Dependencies, every word of a sentence hangs from exactly one head under a typed label. The inventory of these labels is fixed at 37 universal relation types, and individual languages may refine them with subtypes but not add new basic categories (Nivre et al. 2020). The label set grew out of an earlier cross-linguistic typology of Stanford dependencies (Marneffe et al. 2014). It was fixed for the first multilingual release in 2014–2015 (Nivre et al. 2016).
The guidelines reserve discourse for words that relate to the sentence only expressively or interactionally. Typical cases are interjections (oh, wow), hesitation fillers (um, uh), discourse markers (well, like, you know), sentence-final particles in languages such as Chinese, and emoticons. The usual head is the main predicate of the clause, since UD treats content words as heads (Marneffe et al. 2021). Where an element clearly comments on one phrase, that phrase's head can take it instead. The dependent is most often tagged INTJ, but particles and adverbs also occur.
Neighbouring relations
Several relations sit close to discourse:
- vocative covers expressions that name the addressee (John, sir).
- advmod covers adverbs that genuinely modify the predicate (She spoke well).
- cc covers coordinating conjunctions, including a sentence-initial and.
- parataxis covers loosely juxtaposed clauses.
- reparandum covers disfluent material that is later corrected.
Discourse also differs from discourse relations in the sense of discourse analysis. Those relations hold between clauses or sentences, such as contrast and cause, and are often unsignalled (Marcu & Echihabi 2002). The UD label, by contrast, marks a single word inside one tree.
Subtypes
The relation has no universal subtypes. Some treebanks add their own, such as discourse:sp for Chinese sentence particles, discourse:emo for emoticons and emoji, and discourse:filler for hesitations. These are language-specific refinements of the kind the scheme allows (Nivre et al. 2020).
Examples
Arrows run from head to dependent.
| Sentence | Arc |
|---|---|
| Well, I think she left. | think → Well (discourse) |
| Oh, that is great. | great → Oh (discourse) |
| I, um, forgot the keys :) | forgot → um (discourse); forgot → :) (discourse) |
In the second sentence the copula is is not the head, because UD makes the predicate adjective great the head of the clause (Marneffe et al. 2021). An interjection that stands alone, as in Wow!, is the root itself. Such clauses without a prototypical predicate are discussed by (Nivre et al. 2026).
Across languages
The relation is universal, but how much it is used depends on the language, the genre and treebank policy. In the English treebanks of release 2.18, interjections attached to verbs form by far the largest group, followed by interjections on nouns, adjectives and pronouns. Speech-like and web data make English comparatively rich in discourse arcs.
In French the relation is about half as frequent relative to corpus size. The head is again mostly a verb, but adverbs play a larger part both as heads and as dependents. Markers such as bon or alors tagged ADV are one example.
Languages with grammaticalised sentence particles, such as Mandarin or Japanese, bring a further class of dependents into the relation. Whether these particles count as discourse or as another function is decided treebank by treebank. Choices of this kind, where a language-specific strategy is mapped onto a universal construction, are the subject of (Nivre 2025).
In treebanks and research
Because discourse elements are syntactically loose, they are a known source of inconsistency. The boundaries with advmod, cc and parataxis are applied unevenly across corpora, and annotation consistency in UD has been assessed with automatic checks (Marneffe et al. 2017). Tutorials on UD principles and tools cover exactly such borderline label decisions (Zeman 2021).
For dependency parsing, the label is usually easy when the dependent is a clear interjection. Accuracy depends on correct part-of-speech tagging and on choosing the right attachment site. UD-based parsers trained on raw text, as in the CoNLL 2017 and 2018 shared tasks, report it among all labels (Marneffe et al. 2021). Enhanced UD graphs leave discourse arcs untouched, since they add or refine argument-related edges (Schuster & Manning 2016).
The relation is most useful in spoken and conversational treebanks, and in studies of speech, pragmatics and dialogue. There, markers like well or so carry interactional meaning. They also supply cue words that discourse-level studies of explicit and implicit connectives rely on (Pitler et al. 2009).
What discourse joins: the most frequent head tag → dependent tag
The five most frequent in each language's treebanks, by tokens.
The numbers
| Pair | Tokens |
|---|---|
| English: VERB → INTJ | 5,726 |
| English: NOUN → INTJ | 1,758 |
| English: ADJ → INTJ | 754 |
| English: INTJ → INTJ | 371 |
| English: PRON → INTJ | 370 |
| French: VERB → INTJ | 1,663 |
| French: NOUN → INTJ | 947 |
| French: ADV → INTJ | 406 |
| French: ADJ → INTJ | 293 |
| French: VERB → ADV | 290 |
Key papers
Related concepts
- Universal Dependencies — Universal Dependencies is a framework and an open collection of treebanks that annotates the grammar of many languages with one shared…
- Dependency parsing — Dependency parsing is the task of recovering a sentence's syntactic structure as directed, labelled links between its words. Each word is…
More papers
Topics: 🌐 Universal Dependencies · 🌳 Dependency parsing
Shelves: 🗣️ Language and Linguistics
This page in Markdown → https://md.research.elamur.ai/free/en/concepts/ud-discourse/