Syntax
Dependency Grammar
Tesnière's alternative — words connected by directed asymmetric relations, not nested phrases
Dependency Grammar (DG) is a syntactic framework that represents sentence structure as a set of directed binary relations between words, rather than as hierarchical phrase constituents. Each word (except the root) depends on exactly one other word. In "John reads books quickly", "reads" is the root; "John" is its subject, "books" its object, "quickly" its modifier. Lucien Tesnière's posthumous 1959 Éléments de syntaxe structurale founded the modern theory. DG has roots in Pāṇini's Sanskrit grammar (~500 BCE) and the Arab grammarians. Today dominant in computational linguistics — Universal Dependencies (de Marneffe et al. 2014) annotates 100+ languages in DG format. Competes with constituency-based approaches.
- FounderLucien Tesnière (1959, posthumous) — Éléments de syntaxe structurale
- Older rootsPāṇini (~500 BCE), Arab grammarians, medieval modistae
- Core unitWord, not phrase
- DependencyAsymmetric relation: head ↔ dependent
- Universal Dependencies100+ languages annotated in unified DG (since 2014)
- Computational dominanceMost modern parsers use DG (Stanford, spaCy, Stanza)
Interactive visualization
Press play, or step through manually. The visualization is yours to drive — try it before reading on.
Watch the 60-second explainer
A condensed visual walkthrough — narrated, captioned, under a minute.
Why dependency grammar matters
- Computational parsing. Modern NLP pipelines (Stanza, spaCy) output dependency trees by default.
- Universal Dependencies. Standardized cross-lingual annotation enables comparative typology.
- Free-word-order languages. Latin, Russian, Czech are easier to analyze in DG than constituency.
- Information extraction. Dependency paths between named entities support relation extraction.
- Machine translation. Tree-to-tree MT has been built on dependency representations.
- Linguistic typology. Head-directionality and valency patterns visible directly in dependency structure.
- Pedagogy. Many traditional grammars (Latin, Russian) implicitly use dependency-style analysis.
Common misconceptions
- Dependency grammar has no hierarchy. The head-dependent chains form a tree; hierarchy is real, just word-based.
- DG and constituency are interconvertible without loss. Conversion is approximate; some phenomena fit one framework better than the other.
- DG is recent. Pāṇini's grammar (~500 BCE) and Arab grammarians used dependency-like analysis millennia ago.
- All trees are projective. Many languages exhibit non-projective dependencies (crossing arcs); DG can handle them.
- Dependency labels are universal across languages. Universal Dependencies aim for cross-linguistic consistency, but language-specific extensions exist.
- Modern syntax has chosen one framework. Both constituency and dependency thrive; framework choice depends on goals.
Frequently asked questions
How does dependency grammar differ from constituency?
Constituency groups words into nested phrases (NP, VP, PP). Dependency directly connects each word to one head, producing a tree of words rather than a tree of phrases. The two are partially convertible — dependency representations can be derived from constituency trees by selecting heads. But DG is more economical (fewer nodes) and naturally handles free-word-order languages.
Who was Lucien Tesnière?
French linguist (1893-1954) at Strasbourg and Montpellier. His magnum opus Éléments de syntaxe structurale was published posthumously in 1959. He developed valency theory — the idea that verbs have a fixed number of "actant" slots (1 for intransitive, 2 for transitive, 3 for ditransitive). His "stemma" diagrams pre-figured modern dependency trees. Underappreciated in Anglophone linguistics until computational linguistics revived the framework.
What is valency?
A verb's argument structure — how many participants it requires. "Sleep" is monovalent (one argument: subject). "Hit" is bivalent (subject and object). "Give" is trivalent (subject, indirect object, direct object). Tesnière's valency theory anticipates modern argument-structure work (Levin, Rappaport-Hovav). Dependency grammar makes valency central by directly representing each argument as a dependent of the verb.
What are Universal Dependencies?
Universal Dependencies (UD) is a project (de Marneffe, Manning, Nivre, others, since 2014) standardizing dependency annotations across languages. Same labels (nsubj, obj, amod, advmod) for English, Korean, Arabic, etc. Now covers 100+ languages with treebanks. Enables cross-lingual NLP research, multilingual parsing, typological studies. Dominant data format in modern dependency parsing.
Why is DG popular in computational linguistics?
Several reasons. Dependency trees are flatter and faster to parse. Labels are typically more interpretable (subject, object, modifier). Free-word-order languages (Czech, Russian, Latin) handle gracefully without empty categories. Cross-linguistic comparison is easier with shared labels. Modern neural parsers (Stanza, spaCy) achieve high accuracy with dependencies. Constituency parsing is mostly used in linguistic research now.
Can DG handle long-distance dependencies?
Yes, with extensions. Pure projective DG (no crossing arcs) struggles with phenomena like wh-movement and topicalization. Non-projective DG allows crossing arcs and handles flexible word order. Czech and Latin treebanks have substantial non-projectivity. Mel'čuk's Meaning-Text Theory adds multiple levels (semantic, deep-syntactic, surface-syntactic) to handle complex phenomena.
Is constituency or dependency more correct?
Both capture aspects of syntactic structure. Constituency emphasizes hierarchical grouping and movement; dependency emphasizes head-modifier relations. They are partly inter-derivable. Many modern frameworks (HPSG, LFG) blend both. Question is empirical and theoretical: which framework predicts what data better. For computational use, dependency dominates; for theoretical work, both are widely used.