🌐 dep: unspecified dependency

Universal Dependencies · 6 key papers · all concepts

Part of: 🗨️ Natural language processing · 🌐 Universal Dependencies

dep (unspecified dependency) is the most general relation label of Universal Dependencies. Annotators or conversion tools use it when a dependency between two words exists but no more specific universal relation can be chosen, often because the text is garbled, fragmentary or unclear.

Also: unspecified dependency, generic dependency, dep

What it is

In Universal Dependencies (UD), every word of a sentence is attached to one head by a typed grammatical relation. Together these attachments form a dependency tree in which content words head function words (Marneffe et al. 2021). The v2 guidelines fix a closed set of 37 universal relation types. Individual languages may refine these with subtypes, but they may not add new basic types (Nivre et al. 2020). The label dep is the escape hatch in this inventory. It states that the dependent belongs to the head, but says nothing about the kind of relation.

The guidelines treat dep as a last resort. It is meant for cases where a parse cannot be determined, for example because of a typo, a cut-off fragment or an opaque string. It is also used where the annotator cannot decide among the specific labels. The head is normally the word to which the unclear element most plausibly belongs, often the root predicate or the nearest noun. The dependent can be of any part of speech.

dep has to be told apart from several neighbouring labels that cover messy input with a definite analysis:

• goeswith joins pieces of a wrongly split word.
• reparandum marks a speech disfluency that is later corrected.
• orphan handles ellipsis.
• list, parataxis and discourse cover loosely joined material whose function is known.

dep has no subtypes defined at the universal level. A treebank-specific refinement would follow the general colon pattern (dep:subtype), but such refinements are rare. If a specific relation is known, it should be used instead.

This label goes back to the earlier Stanford scheme. That scheme also had a generic dep as the top of its relation hierarchy, which carried over into the cross-linguistic redesign (Marneffe et al. 2014).

Examples

In the examples below, the arrow points from the head to the dependent.

1. Web text with a stray string: Call me tomorrow xqzt. Here xqzt is attached to call as dep (call → xqzt), because nothing tells whether it is a name, an error or an abbreviation.
2. A fragment with unclear structure: Prices, the usual, whatever. The annotator cannot decide between appos and parataxis, so whatever may be attached to prices as dep.
3. A foreign insertion: Le rapport, as discussed, arrive demain. The English chunk may hang from rapport or arrive as dep when neither advcl nor a flat foreign analysis fits cleanly.

Across languages

Usage differs by treebank more than by language type, because dep reflects annotation practice and the quality of the source text. It is consistently rare.

In English UD data, the most common pattern is a noun heading a pronoun. Next come noun-headed proper nouns and noun-headed nouns, with numerals and symbols trailing behind. This profile fits list-like web material, names and fragments. French shows a different profile. The leading pattern is verbs heading adpositions, followed by nouns or proper nouns heading X tokens, which are unanalysable or foreign words. Subordinating conjunctions under verbs also appear often. These pairs point to residual uncertainty around prepositions, clause linkers and foreign strings rather than around nominal lists. For background on these languages, see the English and French topics.

In treebanks and research

UD grew from 10 treebanks into a large collection: release v2.7 (November 2020) held 183 treebanks for 104 languages (Marneffe et al. 2021). With many contributors, a fallback label is unavoidable, and its frequency is a rough indicator of how complete an annotation is. Work on annotation consistency examines exactly such irregularities across UD corpora (Marneffe et al. 2017). Validation and tutorial material for UD also treat leftover generic labels as candidates for revision (Zeman 2021).

The motivation for UD was cross-scheme compatibility. Before UD, the dependencies shared between Swedish, Danish and English schemes never exceeded 40% (Nivre et al. 2016). An unspecified label weakens that parallelism, which is why guideline work keeps narrowing its use. Analyses of constructions (Nivre 2025) and of nonprototypical clauses (Nivre et al. 2026) give annotators specific options for structures that might otherwise end up as dep.

In dependency parsing, dep counts in labelled attachment scores like any other label, including in the CoNLL 2017 and 2018 shared tasks built on UD (Marneffe et al. 2021). Because it is rare and heterogeneous, parsers learn it poorly. A parser that predicts a specific label where the gold tree says dep is counted wrong, even if its analysis is arguably better. Automatic converters, such as the constituency-to-UD converter released with Stanford CoreNLP (Schuster & Manning 2016), can also fall back on dep when their rules find no match. Researchers therefore often inspect or exclude dep edges when they study treebanks or cross-lingual transfer.

Key papers

The works this page cites, the most cited first; citations as counted by OpenAlex plus the papers of this library that cite the work, read 2026-10-24.

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-dep/