Universal Dependencies (UD)

UD defines a fixed set of 17 universal part-of-speech tags that abstract away from language-specific grammatical traditions. These tags, such as NOUN, VERB, and ADJ, are applied consistently across all treebanks.

• Enables direct cross-lingual parser training
• Reduces the need for language-specific feature engineering
• Forms the foundational input layer for dependency arc prediction

The framework specifies a universal inventory of 37 syntactic relations organized into a taxonomy of clausal, nominal, and modifier dependencies. Core relations include nsubj (nominal subject), obj (direct object), and amod (adjectival modifier).

• Facilitates uniform syntactic analysis across languages
• Allows parsers to share a common output label space
• Supports the development of multilingual evaluation benchmarks

UD prioritizes content words (nouns, verbs, adjectives) as syntactic heads over function words (auxiliaries, adpositions, complementizers). For example, in a prepositional phrase, the noun is the head and the preposition is a dependent marked with the case relation.

• Creates a more semantically motivated tree structure
• Improves consistency in annotation across different language families
• Simplifies the extraction of predicate-argument structures for downstream tasks

Beyond the basic syntactic tree, UD defines an enhanced representation that adds arcs to capture implicit semantic relationships. This includes resolving control structures, shared arguments in coordination, and the subjects of relative clauses.

• Bridges the gap between surface syntax and deep semantics
• Provides a richer graph for relation extraction systems
• Enables more accurate conversion to Abstract Meaning Representation (AMR)

UD provides a layered system for annotating lexical and inflectional features such as Number=Sing, Tense=Past, or Case=Dat. These features are organized into a multi-dimensional attribute-value matrix attached to each token.

• Captures the grammatical properties of morphologically rich languages
• Provides critical signals for disambiguating syntactic attachment
• Enables typological analysis of language structures

All UD treebanks are distributed in the CoNLL-U format, a tab-separated text standard with ten fields per token. Fields include the token ID, lemma, universal POS tag, morphological features, dependency head, and dependency relation.

• Ensures interoperability between different parsing tools and libraries
• Simplifies data loading with standard parsers like Stanza and spaCy
• Contains structured comment lines for metadata and multi-word tokens

The standard tab-separated text format for representing annotated linguistic data within the UD ecosystem. Each token occupies a line with 10 fields, including ID, FORM, LEMMA, UPOS, XPOS, FEATS, HEAD, DEPREL, DEPS, and MISC. Sentences are separated by blank lines, and comments begin with a hash mark. This format ensures interoperability between treebanks and parsers across languages.

A neural graph-based architecture introduced by Dozat and Manning that achieves state-of-the-art accuracy on UD parsing tasks. It uses BiLSTM encoders to generate contextualized word representations, then applies deep biaffine attention to score all possible head-dependent arcs simultaneously. The model decouples arc prediction from label classification, using separate biaffine classifiers for each task. This design efficiently captures higher-order syntactic interactions without explicitly modeling sibling or grandparent features.

The primary evaluation metric for dependency parsers, measuring the percentage of tokens that receive both the correct syntactic head and the correct dependency relation label. LAS is calculated as:

• Correct head + correct label / Total tokens A related metric, Unlabeled Attachment Score (UAS), measures only head correctness. State-of-the-art neural parsers achieve LAS scores above 95% on English UD treebanks, while morphologically rich languages like Czech or Arabic typically score lower due to non-projectivity and complex inflection.

A structural property of dependency trees where no crossing arcs exist when the sentence is drawn linearly above the words. A projective tree can be parsed with a simple stack-based algorithm without reordering. Non-projective trees contain crossing dependencies, which are common in:

• Languages with free word order (Czech, Latin)
• Wh-movement in questions
• Long-distance dependencies The UD framework supports both projective and non-projective annotations, with the Chu-Liu/Edmonds algorithm enabling efficient decoding of non-projective structures in graph-based parsers.

A deterministic parsing paradigm that processes sentences left-to-right using a stack and buffer data structure. The parser applies a sequence of actions—SHIFT, LEFT-ARC, RIGHT-ARC, REDUCE—to incrementally build a dependency tree. Variants include:

• Arc-Standard: Builds dependencies only when both head and dependent are complete
• Arc-Eager: Attaches dependents eagerly before the head is fully processed Transition-based parsers offer linear-time complexity, making them ideal for production systems like spaCy, though they are susceptible to error propagation from greedy decisions.

An extended UD representation that augments basic syntactic trees with additional arcs to capture implicit semantic relationships. Enhancements include:

• Null nodes for elided predicates in gapping constructions
• Propagation of conjuncts to share dependents across coordinated heads
• Control relations linking infinitival complements to their implicit subjects
• Relative clause augmentation to mark shared arguments This representation bridges the gap between surface syntax and deep semantic parsing, serving as input for tasks like Abstract Meaning Representation (AMR) parsing and semantic role labeling.