Dependency Grammar¶

A family of syntactic analyses that makes asymmetric head–dependent relations among lexical units the primary structure of a sentence, without requiring phrase nodes as its basic combinatorial units.

Version: v1 · 2026-08-30 · History
Domain-specific #: 1643
Origin domain: linguistics
Subdomain: syntax
Aliases: Dependency syntax, Dependency-based grammar

Core Idea¶

Dependency grammar is a family of syntactic approaches in which the primary combinatorial structure is a set of asymmetric relations between lexical units: one unit is the head and another is its dependent. A sentence is analyzed by connecting words—or, in some frameworks, morphemes—through these directed relations. Phrasal constituent nodes such as noun phrase and verb phrase need not be inserted as the basic units that hold words together. The defining contrast is therefore not “trees versus no trees.” Both dependency and constituency analyses can be tree-shaped. It is lexical nodes joined directly by head–dependent arcs versus words grouped through phrasal constituents.^[1]^[2]

Dependency grammar is a family resemblance category, not one fully specified theory. Its members disagree about which word is the head of particular constructions, whether function words or content words head their combinations, which dependency labels are primitive, whether every structure must be a single-rooted tree, how ellipsis and coordination are represented, and whether linear order belongs to syntax or a separate component. Those disagreements are real, but they occur inside a common representational commitment. De Marneffe and Nivre identify this commitment as the use of dependency relations between words to represent grammatical structure; their review also emphasizes that different dependency frameworks vary substantially in theoretical and representational choices.^[1]

A familiar basic representation assigns each non-root word one syntactic head and labels its incoming relation. The result is a rooted directed tree. Universal Dependencies (UD), for example, requires its basic representation to be a tree: one word heads the sentence under a notional ROOT, and every other word depends on another word. UD's enhanced layer may add relations and is therefore generally a graph rather than a tree.^[3]^[4] This example demonstrates both the recurring core and an important boundary: single-headed treehood is a widespread realization, not a necessary definition of every dependency grammar.

The abstraction matters because it changes what syntactic analysis foregrounds. Predicate–argument and modifier relations become direct arcs among lexical units. A verb can govern its subject and object; a noun can govern its determiner and modifier; a preposition's status depends on the chosen scheme. The analysis compresses sentence structure into a network whose units are close to the observed word sequence while retaining hierarchy. That design has supported grammatical theory, multilingual treebanks, parsing algorithms, typological comparison, and downstream language-processing applications.^[5]^[6]

Structural Signature¶

The minimal structural signature is:

lexical units → asymmetric head assignment → typed or untyped dependency relations → connected syntactic organization

Five roles make the identity recognizable:

Lexical units. The basic nodes correspond to words or morphemes rather than mandatory phrasal constituents.
Head selection. For a syntactic combination, one unit is structurally privileged as head or governor.
Dependents. Other units are subordinated to heads through directed relations.
Grammatical relation. Each arc represents a syntactic connection; many schemes type it with labels such as subject, object, modifier, or determiner.
Composed structure. Local arcs combine into a representation of a clause or sentence, usually rooted and often tree-shaped.

The invariant is relational lexical organization. If phrase nodes do all primary combining and the word-to-word arcs are only a later conversion or visualization, the originating analysis is constituency-based even if a dependency projection can be extracted from it. Conversely, a dependency grammar may use domains, features, valency frames, or ordering constraints in addition to dependency arcs without ceasing to be dependency-based. The question is which objects and relations carry the central syntactic commitments.

Several often-mentioned properties are contingent rather than defining. Projectivity requires that dependency subtrees occupy contiguous regions of the linear sentence and, graphically, that arcs can be drawn above a sentence without crossing. Some dependency grammars restrict structures to projective trees; others admit nonprojective arcs for discontinuous constructions.^[7] Single-headedness, acyclicity, and one root characterize a common well-formed dependency tree, but enhanced or specialized representations may relax them. Verb centrality is historically important, especially in Tesnière's clause analysis, but it cannot serve as a universal diagnostic across contemporary schemes and construction types.

What It Is Not¶

Dependency grammar is not the prime Dependency. The prime denotes substrate-neutral directed reliance with a specifiable condition and failure mode. Dependency grammar is a linguistic framework whose technical “dependency” relation organizes syntactic units. A dependent is structurally governed by a head, but the grammar does not reduce to the general claim that one thing relies on another. It additionally specifies lexical nodes, syntactic headedness, grammatical relations, sentence-level composition, and theory-specific well-formedness conditions.

It is not dependency parsing. A grammar or representation says what analyses are licensed and what structures mean; a parser is a procedure that maps an input string to an analysis. Transition-based, graph-based, deterministic, probabilistic, and neural parsers can target dependency structures. Nivre's projective algorithm, for example, guarantees projective acyclic output, but the algorithm is one computational realization, not the identity of dependency grammar.^[8]

It is not Universal Dependencies. UD is a particular cross-linguistic annotation framework and resource community within the larger dependency tradition. Its content-word orientation, relation inventory, morphological layer, basic-tree requirements, enhanced graphs, and language-specific extensions are design decisions, not universal obligations of dependency grammar.^[5]^[6]

It is not merely any directed graph over tokens. Semantic dependency graphs, coreference graphs, discourse-relation graphs, and information-extraction networks may use word nodes and arrows without constituting syntactic dependency analyses. The arcs must carry grammatical organization of lexical combinations.

It is also not simply “a constituency tree with fewer nodes.” Dependency and constituency representations can sometimes be transformed into one another, and typed dependencies can be extracted from phrase-structure parses.^[9] Convertibility does not erase the theoretical distinction: the two analyses choose different primitives, encode certain relations directly or indirectly, and can diverge over heads, discontinuity, and granularity.

Scope of Application¶

In theoretical syntax, dependency grammar offers a framework for statements about headedness, valency, arguments, adjuncts, government, coordination, word order, and discontinuity. Tesnière's structural syntax made dependency and valency central to clause organization; later dependency theories developed distinct formalisms rather than a single lineage with uniform rules.^[10]^[11]

In corpus linguistics, dependency annotation turns sentences into reusable treebank structures. UD combines word segmentation and morphology with a syntactic layer focused on relations among predicates, arguments, and modifiers; its multilingual design supports comparative annotation while permitting language-specific extensions.^[5] In natural-language processing, dependency representations provide targets for parsers and inputs to relation extraction, information extraction, semantic processing, machine translation, and other tasks. Stanford Typed Dependencies was explicitly designed to expose grammatical relations in an accessible, semantically useful representation.^[12]

The abstraction applies to analyses, formalisms, and annotation schemes whose primary syntax has the structural signature. It does not apply metaphorically to software package dependencies, project schedules, causal diagrams, or any graph merely because nodes depend on one another.

Clarity¶

The fastest diagnostic is to ask: What are the basic internal objects of the syntactic representation? If sentence structure is formed by directed grammatical relations whose endpoints are lexical units, the analysis is dependency-based. If it is formed primarily by rewrite rules or dominance relations over phrasal nodes, it is constituency-based. If the arrows instead encode predicate–argument semantics after syntax has been abstracted away, it is a semantic dependency representation rather than dependency grammar.

Consider “The child opened the door.” In a dependency analysis, opened may head child through a subject relation and door through an object relation; child heads the, and door heads its determiner. The analysis directly exposes which lexical item governs each dependent. A constituency analysis instead groups the child as a noun phrase and opened the door as a verb phrase before relating those phrases within a sentence. Both capture hierarchy. Their difference is the ontological inventory and the relations made primitive.

This diagnostic prevents three errors: equating dependency with adjacency, because heads and dependents need not be next to one another; equating dependency with semantics, because syntactic arcs may only approximate semantic relations; and treating every output tree as a grammar, because a parse tree is an analysis produced under a representational scheme, while dependency grammar is the family of schemes that licenses such analyses.

Manages Complexity¶

Dependency grammar manages syntactic complexity by replacing nested phrase inventories with local relational commitments. Each word receives a structural neighborhood: its head, its dependents, and often labels for those arcs. The whole sentence can then be investigated through graph operations—ancestors, descendants, paths, subtrees, roots, crossings—without losing hierarchy.

That compression is especially useful for languages in which surface order is flexible or discontinuous. A dependency relation can remain the same while a dependent appears before or after its head, although actual annotation and grammar still need rules for ordering and nonprojectivity. The representation separates, at least analytically, who depends on whom from where each unit occurs. It does not make free word order trivial; it gives analysts a stable relational layer across different orders.

Typed relations further compress recurring grammatical functions. A label inventory lets researchers compare subjects, objects, modifiers, determiners, auxiliaries, and coordination across sentences. But the compression has a cost: choosing heads and labels can force contested theoretical judgments into one arc. UD explicitly notes that its preference for content-word heads conflicts with dependency traditions that make function words heads in many constructions.^[3] A dependency representation is therefore an analytical model, not an unmediated picture of language.

Abstract Reasoning¶

Once an analysis is dependency-based, it licenses several inferences.

Head diagnostic: removing or replacing a proposed head should change the organization or category of the combination more fundamentally than changing an ordinary dependent, subject to the framework's own criteria.
Valency diagnostic: a head's lexical or constructional requirements constrain which dependents may or must occur and how they are interpreted.
Path diagnostic: a syntactic relation between distant words may be decomposed through their path in the dependency structure.
Projectivity diagnostic: if an arc spans material that is not dominated by the same head, a projective analysis fails and the scheme must permit nonprojectivity, revise head choices, or use another device.
Conversion diagnostic: when converting constituency to dependencies, head rules and treatment of function words determine the result; the conversion is not representation-neutral.
Annotation diagnostic: attachment and label disagreements identify different kinds of uncertainty. Correct head with wrong label differs from wrong head, and evaluation should not collapse them.

These operations support predictions without claiming that one dependency scheme is linguistically privileged. For a candidate tree, one can check rootedness, acyclicity, single-headedness, connectedness, and projectivity when those are required. For a grammar comparison, one can hold the sentence constant and examine how head-selection policies alter paths and subtrees. For downstream use, one can ask whether the chosen scheme directly represents the relations the application needs.

Knowledge Transfer¶

Knowledge transfers literally within linguistics and language technology. A linguist's head–dependent analysis can inform a treebank annotation manual; the treebank can train a parser; parser output can supply relations to an extraction system. The representation persists across theory, annotated data, algorithm, and application even though each stage adds its own commitments.

Transfer between languages is more difficult. UD's universal inventory and guidelines seek consistent treatment of comparable constructions, but morphological richness, word segmentation, clitics, ellipsis, serial verbs, and language-specific constructions resist naïve one-to-one mapping.^[6] Good transfer preserves the lexical-relation core while allowing local evidence to change head choices, subtypes, or analysis.

Transfer outside language is analogical only. Software dependencies and organizational reporting lines also form directed graphs, but they lack syntactic categories, grammatical relations, valency, and the word–morpheme boundary. What transfers is the more general prime Dependency or graph reasoning, not dependency grammar as such.

Examples¶

Basic clause. In “Researchers analyzed samples,” analyzed can be the root, Researchers its subject dependent, and samples its object dependent. The lexical units are nodes; the verb is head of two labeled arcs; those arcs compose into the clause's structure. This is a canonical dependency analysis.

Internal nominal structure. In “the old bridge,” a content-word-centered scheme makes bridge the head, with the and old as dependents. Another dependency theory may differ about the determiner. The disagreement does not remove the shared identity because both make word-to-word head assignment primary.

Universal Dependencies treebank. A CoNLL-U record encodes each word's HEAD and DEPREL values. Basic dependencies must form a tree, while the DEPS field can encode enhanced relations that form a graph.^[4] This is an operational instantiation linking the abstraction to a concrete interchange format.

Dependency parser. A parser receives a token sequence and predicts arcs from heads to dependents. The CoNLL shared-task formulation treats words as nodes and arcs as binary head-to-dependent relations, optionally with labels.^[13] This is an application of dependency grammar, not a synonym for it.

Non-example—phrase structure. An analysis that first creates NP and VP nodes through phrase-structure rules and only later derives word-to-word relations remains constituency-based at its primary layer. De Marneffe, MacCartney, and Manning's extraction of typed dependencies from phrase-structure parses demonstrates compatibility between layers without identifying them.^[9]

Non-example—semantic graph. A graph connecting a predicate to semantic arguments may omit function words, introduce implicit concepts, or permit reentrancy for reasons unrelated to syntax. Unless its arcs represent the grammatical organization of lexical units, it is not dependency grammar.

Structural Tensions¶

Lexical economy versus hidden constituency. Direct word-to-word arcs yield a compact representation, but some constructions appear easier to state with phrases, domains, or multiword units. Adding auxiliary structures may improve coverage while weakening the claim that dependencies alone are sufficient.

Universal comparability versus language fidelity. A common relation inventory enables multilingual research and transfer. The same inventory can flatten distinctions that a language-specific tradition treats as fundamental. Extensions restore detail but reduce comparability.

Content-word heads versus function-word heads. Choosing content words often aligns trees with predicate–argument content and downstream semantics. Choosing auxiliaries, adpositions, or determiners as heads may better reflect distribution, selection, or a theory's conception of syntactic structure. Either policy changes tree shape substantially.

Tree constraints versus descriptive coverage. Single-rooted, acyclic, single-headed trees are computationally tractable and easy to annotate. Coordination, control, ellipsis, shared dependents, and other phenomena can motivate extra heads or nodes. Enhanced dependency graphs improve coverage at the cost of the simple tree invariant.

Projectivity versus discontinuity. Projective trees support efficient algorithms and contiguous subtrees. Natural languages exhibit constructions whose most plausible dependencies cross. Forcing projectivity can distort the analysis; admitting nonprojectivity expands coverage and computational complexity.^[7]

Syntactic directness versus semantic overreading. Dependency arcs often put predicates near arguments and are useful for applications. Yet syntactic head–dependent relations are not automatically semantic roles, entailments, or causal relations. Treating them as such can make an accessible representation look more semantically complete than it is.

Structural–Framed Character¶

Dependency grammar is strongly framed. Its directed relation is structural, but recognition requires the linguistic frame: lexical units, heads, dependents, grammatical relations, clauses, and syntactic well-formedness. The vocabulary travels literally across grammatical theory, treebanking, and NLP because those practices share the language-analysis domain. It does not travel literally to arbitrary networks.

Its identity is also theory-mediated. Head tests, relation inventories, root conventions, and analyses of function words reflect research traditions and annotation purposes. Different schemes can map the same sentence to different legitimate dependency structures. That contestability does not make the abstraction subjective; it shows that the shared core underdetermines a complete analysis.

The abstraction should therefore remain domain-specific. It is broader than one formalism or annotation project, but narrower than the structural prime Dependency and narrower than generic graph representation.

Structural Core vs. Domain Accent¶

The structural core is a directed organization: nodes are connected asymmetrically, local relations compose into a larger rooted or root-like structure, and paths expose hierarchical organization. That skeleton resembles generic dependency graphs.

The domain accent is load-bearing. The nodes are lexical units; headship is syntactic; arc types encode grammatical relations; structures analyze sentences or related linguistic objects; and boundaries are defined against constituency, semantics, morphology, and parsing. Remove those commitments and the result is merely a directed labeled graph. Remove direct lexical head–dependent relations and the result may still be syntax, but it is no longer dependency grammar.

This explains the classification. The candidate has a genuine recurring abstraction across dependency theories, treebanks, and parsers, yet its identity cannot be recognized without linguistic concepts. The portable residue is already represented by primes such as Dependency, Hierarchy, and Parsing; the residual lexical-syntactic package merits its own domain-specific node.

Dependency grammar presupposes Dependency as its closest portable structural relative: syntactic analyses orient an asymmetric relation from a head to a dependent and compose those relations. The match is not an identity. The catalog prime requires directed reliance and a failure-mode commitment across substrates, whereas linguistic dependency is a technical structural relation whose exact interpretation varies by theory.

It also relates to Hierarchy, because dependency paths and rooted structures establish dominance without phrasal nodes; to Parsing, because dependency parsers construct or infer dependency analyses; to Symbolic Representation, because words and labels stand in a formal structure; and to Classification, because annotation schemes assign typed grammatical relations. These are explanatory neighbors, not additional proposed parents. The smallest prospective DAG placement is one proposal-only composition edge to prime:dependency.

Relationships to Other Abstractions¶

Current abstraction Dependency Grammar Domain-specific

Parents (1) — more general patterns this builds on

Dependency Grammar presupposes Dependency Prime

Dependency grammar presupposes Dependency as its closest portable structural relative: syntactic analyses orient an asymmetric relation from a head to a dependent and compose those relations.

Children (1) — more specific cases that build on this

Functional Generative Description Domain-specific is a kind of Dependency Grammar

Dependency Grammar is the proposed immediate parent.

Hierarchy path (1) — routes to 1 parentless root

Dependency Grammar → Dependency

Neighborhood in Abstraction Space¶

Dependency Grammar sits in a sparse region of the domain-specific corpus (85^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Dependency (prime): substrate-neutral directed reliance; it lacks lexical units, heads, grammatical relations, and sentence analysis.
Phrase Structure Rules: rules that build phrasal constituents. They can be converted to dependencies under head rules but choose different primitives.
Parsing (prime): analysis of an input under a formal system. Dependency parsing is one application; dependency grammar is the target representation and theory family.
Abstract Syntax Tree: a programming-language representation that omits surface syntax and organizes code constructs, not a grammatical theory of natural-language lexical dependency.
Valency theory: the study of a head's combinatorial capacity, especially a predicate's arguments. It is historically and theoretically close but does not exhaust dependency grammar.
Universal Dependencies: one multilingual dependency-based annotation framework, not the whole family.
Semantic dependency graph: a predicate–argument or meaning representation whose arcs need not be syntactic headship.
Dependency tree: one concrete structural product; dependency grammar is the family of principles that licenses such products.
Dependency parsing: an algorithmic task that predicts dependency structures.
Link grammar: a related lexical-relation formalism with its own representational commitments; similarity of word links does not make every link-based system an unrestricted synonym.

References¶

[1] Marie-Catherine de Marneffe and Joakim Nivre, “Dependency Grammar,” Annual Review of Linguistics 5 (2019), 197–218. https://doi.org/10.1146/annurev-linguistics-011718-011842 registry ↩a ↩b

[2] Joakim Nivre, “Dependency Grammar and Dependency Parsing,” technical survey (2005). https://citeseerx.ist.psu.edu/document?doi=0734a44075c1b43726014146430cb329e8b34c9d registry ↩

[3] Universal Dependencies, “Syntax: General Principles,” version 2 documentation. https://universaldependencies.org/u/overview/syntax.html registry ↩a ↩b

[4] Universal Dependencies, “CoNLL-U Format,” current format documentation. https://universaldependencies.org/format.html registry ↩a ↩b

[5] Joakim Nivre et al., “Universal Dependencies v2: An Evergrowing Multilingual Treebank Collection,” LREC 2020, 4034–4043. https://aclanthology.org/2020.lrec-1.497/ registry ↩a ↩b ↩c

[6] Marie-Catherine de Marneffe et al., “Universal Dependencies,” Computational Linguistics 47.2 (2021), 255–308. https://doi.org/10.1162/coli_a_00402 registry ↩a ↩b ↩c

[7] Joakim Nivre, “Constraints on Non-Projective Dependency Parsing,” EACL 2006, 73–80. https://aclanthology.org/E06-1010/ registry ↩a ↩b

[8] Joakim Nivre, “An Efficient Algorithm for Projective Dependency Parsing,” IWPT 2003, 149–160. https://aclanthology.org/W03-3017/ registry ↩

[9] Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning, “Generating Typed Dependency Parses from Phrase Structure Parses,” LREC 2006. https://aclanthology.org/L06-1260/ registry ↩a ↩b

[10] Lucien Tesnière, Elements of Structural Syntax, trans. Timothy Osborne and Sylvain Kahane (John Benjamins, 2015). https://doi.org/10.1075/z.185 registry ↩

[11] Klaus Fischer, “Dependency Grammar and Valency Theory,” in The Oxford Handbook of Linguistic Analysis (2012), 223–256. https://doi.org/10.1093/oxfordhb/9780199544004.013.0010 registry ↩

[12] Marie-Catherine de Marneffe and Christopher D. Manning, “The Stanford Typed Dependencies Representation,” COLING 2008 workshop, 1–8. https://aclanthology.org/W08-1301/ registry ↩

[13] Joakim Nivre et al., “The CoNLL 2007 Shared Task on Dependency Parsing,” EMNLP-CoNLL 2007, 915–932. https://aclanthology.org/D07-1096/ registry ↩