🌐 parataxis: parataxis
parataxis is a Universal Dependencies relation that links a word, usually the main predicate of a sentence, to another clause or phrase placed beside it without a coordinating or subordinating marker and without being its argument, as in juxtaposed sentences, parentheticals and some reported speech.
What it is
In Universal Dependencies (UD), clauses can be joined in three broad ways. They can be coordinated, they can be embedded as arguments or modifiers, or they can simply be set side by side. The label parataxis covers the third case. The term comes from traditional grammar, where it is contrasted with hypotaxis, or subordination. In that tradition parataxis also takes in coordination, so the UD label is narrower: it is kept for unmarked juxtaposition, while coordination is annotated as conj.
UD is a lexicalist dependency scheme, meaning that words rather than morphemes are the basic units. Words are linked by typed relations, and content words act as heads (Nivre et al. 2016; Marneffe et al. 2021). Version 2 has 37 universal relation types, and languages may refine them only through subtypes (Nivre et al. 2020). Parataxis belongs to the universal set. Its lineage goes back to the Stanford dependency tradition, which was recast cross-linguistically before being merged into UD (Marneffe et al. 2014; Nivre et al. 2016).
Each part of the arc has a typical role:
- The head is normally the predicate of the first or main unit.
- The dependent is the predicate, or other head word, of the unit that stands beside it.
The relation says that the two units belong together in discourse. It does not say that one fills a grammatical slot of the other.
Parataxis sits among several neighbours that it has to be told apart from:
- conj, for coordination, usually with a conjunction.
- ccomp, for clausal complements of verbs such as say.
- appos, for apposition.
- discourse, for interjections and particles.
- list, for enumerations.
UD 2.0 framed parataxis so widely that annotators struggled to separate it from conj, appos and ccomp. The listed uses of the relation also turned out to have little in common (Ahrenberg 2019).
Some treebanks add language-specific subtypes. Labels such as parataxis:insert (parentheticals), parataxis:discourse, parataxis:rep (repetition) or parataxis:hashtag can be found in individual corpora. They follow the general UD rule that subtypes refine a label without creating a new category (Nivre et al. 2020).
Examples
Below, the head comes before the arrow and the dependent after it.
| Sentence | Arc | Type |
|---|---|---|
| The rain stopped; we went out. | stopped → went | two full sentences joined by a semicolon |
| The plan, Maria thinks, will fail. | fail → thinks | parenthetical reporting clause |
| Take it, it's yours. | Take → yours | asyndetic (conjunction-less) clauses linked only by a comma, the second giving a reason for the first |
In the second example, the reporting clause interrupts the main clause, so it is attached as a parataxis dependent of fail. It is not treated as the governor of an embedded complement.
When the reporting verb comes first, as in She said: "I'm leaving", treebanks diverge. Some use ccomp, with said as head and leaving as dependent. Others use parataxis, either with said as head or the other way round, with leaving as head and said attached to it.
Across languages
The Parallel Universal Dependencies (PUD) treebanks offer a controlled test of these differences. Each holds the same 1000 sentences, most of them originally English and translated into the other languages, so their annotations ought to be close. English PUD was compared with thirteen other PUD treebanks, those in which the label occurs more than ten times. The comparison revealed large differences, both in how often parataxis appears and in where it is placed. Most of these differences came from annotator habits rather than from translation (Ahrenberg 2019).
English PUD contains 97 parataxis arcs. Of these, 44 join side-by-side sentences, 42 mark direct speech, 8 attach parentheticals and 3 fall into other cases (Ahrenberg 2019). The number of instances per treebank ranges from 20 in Czech PUD and 24 in Arabic PUD up to 195 in Russian PUD. A similarity score measures how closely each treebank's placement of the label matches English. It is highest for Portuguese PUD (0.79) and lowest for Czech (0.17) and Arabic (0.24). Swedish PUD kept 86 of the 97 English instances but added many of its own where English has hypotaxis (Ahrenberg 2019). For direct speech:
- German PUD uses ccomp, with the speech verb as head.
- French PUD keeps parataxis but reverses the direction of the arc (Ahrenberg 2019).
In treebanks and research
For dependency parsing, parataxis is a known trouble spot. It joins material without overt markers, so the cues that separate it from conj, ccomp or a sentence boundary often depend on meaning and punctuation. Inconsistent labelling across treebanks also limits both multilingual training and cross-lingual comparison. The consistency of UD corpora has been assessed in a general study that is not limited to this relation (Marneffe et al. 2017). A tutorial on UD principles and tools introduces the guidelines in which these distinctions are drawn (Zeman 2021).
Ahrenberg proposed several revisions to the guidelines (Ahrenberg 2019):
- Keep one general parataxis relation.
- Allow genre- or corpus-specific subtypes, following UD_Italian-Postwita (a treebank of Italian tweets) and UD_French-Spoken (a treebank of spoken French).
- Drop the list relation.
Enhanced representations
Enhanced UD representations add extra edges to the basic tree. Examples include an edge between a subordinate verb and its actual subject, or shared dependents spread across conjuncts. They also refine some labels, for instance by appending the preposition to nmod (Schuster & Manning 2016). The basic tree is kept in full, so the paratactic arc survives. The enhancements described for English target prepositions, conjuncts and controlled subjects, not parataxis (Schuster & Manning 2016). The arc therefore reaches downstream tools for information extraction and discourse analysis much as it stands in the basic tree. Those tools still have to infer for themselves what the juxtaposition means.
What parataxis 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 → VERB | 3,108 |
| English: NOUN → VERB | 704 |
| English: VERB → NOUN | 642 |
| English: NOUN → NOUN | 486 |
| English: ADJ → VERB | 453 |
| French: VERB → VERB | 1,318 |
| French: NOUN → VERB | 298 |
| French: VERB → NOUN | 172 |
| French: ADJ → VERB | 121 |
| French: VERB → PRON | 107 |
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-parataxis/