🌐 flat: flat expression

Universal Dependencies · 6 key papers · all concepts

Part of: 🗨️ Natural language processing · 🌐 Universal Dependencies

flat is the Universal Dependencies relation for a flat expression: a multiword unit, such as a personal name, a date or a stretch of foreign text, that has no clear internal syntactic head. By convention, its first word heads it and every later word attaches to that first word.

Also: flat expression, flat:name, flat:foreign, name (UD v1), foreign (UD v1)

What it is

Universal Dependencies (UD) builds a dependency tree over syntactic words. Each word attaches to a head by a typed relation, and content words normally serve as heads (Marneffe et al. 2021). Some word sequences resist this treatment. In Marie Curie, neither word governs the other. In a quoted Latin phrase, the parts follow a grammar the host language does not analyse. The flat relation covers such cases.

The guideline is a convention, not a grammatical claim. The leftmost word of the expression becomes its head, and each following word depends on that head through flat. The result is a one-level "bush" rather than a chain. The whole expression then attaches to the rest of the sentence through its first word, by an ordinary relation such as nsubj or obj.

UD v2 keeps 37 universal relation types. Languages may refine these with subtypes but may not add new basic types (Nivre et al. 2020). As a matter of textbook history, v2 folded the v1 relations name and foreign into flat. Many treebanks preserve that distinction as two subtypes:

• flat:name, for proper names
• flat:foreign, for foreign material

The relation descends from the cross-linguistic Stanford dependency inventory that UD absorbed (Marneffe et al. 2014; Nivre et al. 2016).

Neighbouring relations

The table separates flat from its neighbours by two questions. Compound has an internal head, while flat does not. Fixed also lacks an internal head, but it is reserved for function-like idioms. A name is open and productive, so it is not fixed.

A related flat-structure idea appeared in enhanced English UD. There, multiword prepositions and quantifier phrases were given a flat internal analysis so that content words could be linked directly (Schuster & Manning 2016). That analysis concerns multiword expressions of the fixed kind, not flat in the v2 sense.

Examples

The arrow points from head to dependent:

• Marie Curie won twice. Here Marie → Curie is flat (or flat:name), and Marie is the nsubj of won.
• He shrugged and said "que sera sera". Here que heads the quotation, and sera and sera each attach to it by flat:foreign.
• French Monsieur Dupont est arrivé. Here the common noun Monsieur heads, and Dupont attaches to it by flat:name.

A reduplicated interjection such as ha ha is often annotated the same way, with the second ha as a flat dependent of the first.

Across languages

Use of flat varies with each language's naming habits and with each treebank team's choices. In the English UD treebanks, the relation is overwhelmingly proper noun to proper noun. Smaller groups follow: a name followed by a number (World War 2), foreign X-tagged strings, a common noun followed by a number, and repeated interjections.

French UD also uses flat mostly between proper nouns. Unlike English, a sizeable share joins a common-noun title or category to a name, as in the Monsieur Dupont pattern. Foreign strings and number-to-number sequences come next. Overall, French uses flat somewhat more often relative to its size than English does.

Some differences reflect annotation policy rather than grammar:

• Some treebanks treat titles as heads, others treat them as modifiers.
• Some split dates into nummod structures, others keep them flat.
• Some use the subtypes, others only bare flat.

In treebanks and research

Because flat is defined by convention, treebanks have applied it unevenly. Its consistency has been checked alongside that of other relations in assessments of the UD corpora (Marneffe et al. 2017). Joakim Nivre and colleagues present the subtype mechanism as the way UD reconciles universal categories with language-particular distinctions (Nivre et al. 2020). The subtype mechanism is also central to tutorials on UD principles (Zeman 2021).

Later work on how UD encodes constructions also touches the boundary between headed and headless structures. One example is the treatment of strategies for expressing a construction (Nivre 2025). Another is the analysis of nonpredicational clauses (Nivre et al. 2026).

For dependency parsing, flat edges tend to be short and head-initial. Parsers trained on the CoNLL 2017 and 2018 shared-task data learn them alongside all other relations (Marneffe et al. 2021). Errors usually concern the boundary of the expression or the confusion of flat with compound or appos.

Flat spans also overlap closely with named entities. Downstream systems can therefore read flat:name structures as cheap entity candidates, while the basic tree still records how the name functions in its clause (Marneffe et al. 2021).

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-flat/