spaCy Dependency Parser

Uses a shift-reduce algorithm with an arc-eager strategy to build dependency trees incrementally. The parser processes tokens left-to-right, maintaining a stack and buffer to decide between SHIFT, LEFT-ARC, and RIGHT-ARC actions. This deterministic approach yields linear O(n) time complexity, making it ideal for high-volume text processing pipelines.

The default model uses a linear classifier with millions of sparse, hand-crafted features extracted from the parser state. Features include:

• Token attributes: word form, lemma, part-of-speech tag
• Contextual features: words and tags of stack/buffer items
• Structural features: existing dependency relations and arc labels This design enables fast training and inference without GPU requirements.

spaCy supports CNN-based feature extraction via its Tok2Vec component, which can replace or augment the linear model. When neural networks are enabled, tokens are encoded into dense vector representations using hash embeddings and residual convolutions, capturing semantic similarity that sparse features miss. This hybrid approach balances speed and accuracy.

Handles non-projective dependencies—crossing arcs common in languages with free word order like German, Dutch, and Czech. The parser uses a pseudo-projective approach: it parses with projective constraints, then applies learned transformations to recover non-projective arcs. This maintains linear-time complexity while accommodating linguistic reality.

The dependency parser operates as a pipeline component ('parser') that depends on the tokenizer and tagger outputs. It consumes part-of-speech tags and produces a Doc object with:

• token.dep_: dependency relation label (e.g., nsubj, dobj)
• token.head: syntactic governor token
• token.children: iterable of dependent tokens This seamless integration enables end-to-end NLP workflows.

Models are trained on Universal Dependencies (UD) treebanks, outputting the standardized UD v2 relation set. This ensures cross-linguistic consistency—the nsubj label means the same thing in English, Spanish, and Japanese. spaCy provides pretrained models for 20+ languages with UD-compliant dependency annotations.

The primary evaluation metric for dependency parsers. LAS measures the percentage of tokens that are assigned both the correct syntactic head and the correct dependency relation label (e.g., nsubj, dobj). A high LAS indicates the parser accurately understands both the structure and the grammatical function of each word in a sentence.

A cross-linguistically consistent framework for grammatical annotation. UD defines a universal set of part-of-speech tags and dependency relations to facilitate multilingual parser development. spaCy's models are trained on UD treebanks, ensuring consistent syntactic analysis across languages.

Projectivity is a property where a dependency tree has no crossing arcs when drawn linearly above the sentence. Non-projective parses contain crossing dependencies, common in languages with free word order or long-distance questions. spaCy's transition-based parser handles non-projectivity through specific parsing actions.

An extension beyond syntactic structure that identifies predicate-argument structures and semantic relations capturing the meaning of a sentence. While spaCy's core parser handles syntax, its output serves as a critical input feature for downstream semantic role labeling and Abstract Meaning Representation (AMR) tasks.

The foundational NLP task of assigning a grammatical category (noun, verb, adjective) to each token. This is a critical prerequisite for dependency parsing, as the parser's feature model heavily relies on POS tags to predict valid arcs. spaCy's pipeline performs POS tagging before parsing.