Syntax Tree

Chomsky's "Remarks on Nominalization" (1970) and Ray Jackendoff's X-bar Syntax (1977) generalized phrase structure into a uniform endocentric template. Every phrase XP has a head X (the same category as the phrase), an optional specifier, an optional complement, and possibly adjuncts. NP has N as head; VP has V; PP has P; AP has A. The rules are XP → Spec X', X' → X' Adjunct, X' → X Complement. This eliminated category-specific phrase structure rules and predicted parallel structures across categories. The X-bar schema dominated syntactic theory from the 1970s through the 1990s and remains influential in many frameworks.

Linguists identify constituents — units that form a node in the tree — through several diagnostics. Substitution: a constituent can be replaced by a single word ("the very tall man" → "he"). Movement: only constituents can be fronted ("It was the cat that the dog chased"). Coordination: only constituents conjoin with "and" ("the dog and the cat"). Ellipsis: constituents can be elided ("Mary saw John, and Sue did too"). Pronominalization: pronouns substitute for full constituents. These tests sometimes disagree, generating theoretical debate, but they are the empirical foundation of phrase structure.

Phrase structure (PSG, Chomsky tradition) builds trees with phrasal nodes (NP, VP) representing groups of words. Dependency grammar (Lucien Tesnière, Éléments de syntaxe structurale, 1959) draws arcs between words — each word depends on a head word. "The dog chased the cat" in dependency: chased is the root; dog is its subject; cat is its object; the depends on dog/cat. Modern computational parsing (CoNLL, Universal Dependencies) uses dependency formats. Some claim PSG and dependency are notational variants for surface structure; theoretical frameworks differ on derivational layers and movement.

A treebank is a corpus annotated with syntactic structure. The Penn Treebank (Marcus, Santorini, Marcinkiewicz, 1993) provided phrase-structure annotations for the Wall Street Journal and Brown corpus, totaling about 4.5 million words. It became the training data for nearly all statistical parsers (Collins, Charniak, Klein and Manning, etc.). The Penn Discourse Treebank, Penn Chinese Treebank, and treebanks for many languages followed. Universal Dependencies (Marneffe, Manning, Nivre, 2014) aimed at cross-linguistic dependency annotation; over 100 languages now have UD treebanks. Modern parsers (BERT-based, 2018+) retain treebanks as the gold standard.

A parser computes the syntactic structure of a sentence given a grammar. Top-down parsers (Earley, 1970; CYK, 1965) work from the start symbol; bottom-up parsers work from words. Probabilistic context-free grammars (PCFGs, Manning and Schütze 1999) assign probabilities to rules and select the most likely parse. Lexicalized parsers (Collins 1997, Charniak 2000) condition probabilities on heads. Neural transition-based parsers (Chen and Manning, 2014) and graph-based parsers (Dozat and Manning, 2017) achieve human-level accuracy on Penn Treebank. Modern systems combine pre-trained embeddings (BERT) with parsing-specific decoders.

Movement (or transformation) explains certain dependencies as displacement of constituents. The wh-question "What did Mary see?" is derived from the underlying "Mary saw what" — "what" moves to the front. Chomsky's Aspects of the Theory of Syntax (1965) introduced transformational rules; Government and Binding (1981) reduced them to general operations (Move-α). The Minimalist Program (1995) reduces movement to feature checking. Movement leaves traces (or copies) at the original site, recoverable through binding, reconstruction, and intervention effects. Dependency grammars typically reject movement; HPSG handles long-distance dependencies through feature percolation.

Chomsky's Standard Theory (Aspects, 1965) distinguished deep structure (where semantic interpretation happens) from surface structure (where phonological interpretation happens). Transformations mapped between them. The Minimalist Program (1995) eliminated this dichotomy — derivations involve internal merge (movement) operating on a single representation, with interface conditions at the LF (semantic) and PF (phonological) interpretive levels. Generative semantics (Lakoff, McCawley, Postal 1960s-70s) had argued for deeper representations; the debate split syntacticians and contributed to the rise of cognitive linguistics.