CoNLL-U Format

Every token occupies a single line with 10 tab-separated fields. Empty lines mark sentence boundaries, and lines starting with # store metadata comments. The rigid column structure ensures consistent parsing across all UD treebanks.

• ID: Token index (integer) or range for multiword tokens (e.g., 1-2)
• FORM: The surface form as it appears in the text
• LEMMA: Base or dictionary form of the word
• UPOS: Universal part-of-speech tag (e.g., NOUN, VERB)
• XPOS: Language-specific POS tag (optional, _ if absent)
• FEATS: Morphological features as Key=Value pairs separated by |
• HEAD: ID of the syntactic head (0 for root)
• DEPREL: Universal dependency relation label (e.g., nsubj, obj)
• DEPS: Enhanced dependency graph as HEAD:REL pairs
• MISC: Any other annotation as Key=Value pairs

The format handles syntactic words that span multiple orthographic tokens (e.g., Spanish 'dámelo' = 'da' + 'me' + 'lo'). A multiword token line uses a range ID (e.g., 1-3) followed by individual token lines with dashed IDs (1-1, 1-2, 1-3). This preserves both the original text segmentation and the syntactic analysis.

• Range ID line contains the full surface form
• Sub-token lines carry the morphological and syntactic annotations
• Enables accurate alignment between raw text and linguistic analysis

The FEATS column uses a structured Key=Value notation to encode inflectional and derivational morphology. Multiple features are separated by the pipe character (|), and the underscore (_) indicates no morphological annotation is available.

• Example: Gender=Fem|Number=Sing|Case=Acc
• Universal features defined by UD include Animacy, Aspect, Case, Definite, Degree, and 20+ others
• Enables cross-linguistic queries for specific grammatical phenomena
• Language-specific extensions can be added as needed

The DEPS column extends basic syntax with additional arcs capturing implicit predicates, shared arguments, and control relationships. Each entry is a HEAD:REL pair, with multiple pairs separated by |. This bridges surface syntax and deeper semantic interpretation.

• Example: 3:obl:agent|5:nsubj:xsubj
• Captures null elements and elided material
• Supports downstream tasks like semantic role labeling and AMR parsing
• The _ placeholder indicates no enhanced annotation is present

Lines beginning with # store structured metadata as key-value pairs. Standard comment keys include sent_id for unique sentence identifiers and text for the raw sentence string. Transliterations, translations, and corpus-specific annotations are also stored here.

• # sent_id = weblog-blogspot.com_norvig_20060911092800_ENG_20060911_092800-0001
• # text = What if Google Morphed Into GoogleOS?
• # translit = ... for scripts requiring romanization
• Comments precede the token lines they describe

The format supports empty syntactic nodes to represent elided or gapped material in coordinated structures. Empty nodes use an ID with a decimal (e.g., 5.1) and carry the _ placeholder in the FORM and LEMMA columns while retaining full morphological and dependency annotations.

• Example: 'Mary bought a book and John a newspaper' — an empty node represents the elided verb 'bought' in the second clause
• Enables complete syntactic representation of elliptical constructions
• Critical for languages with frequent argument dropping

The syntactic analysis task whose output is stored in CoNLL-U files. A dependency parser analyzes the grammatical structure of a sentence to determine head-dependent relationships between words.

• Constructs a directed graph where nodes are tokens and edges are typed dependencies
• The HEAD field in CoNLL-U (column 7) stores the index of the syntactic governor
• The DEPREL field (column 8) stores the specific grammatical relation

The foundational NLP task that populates the UPOS and XPOS fields (columns 4 and 5) in CoNLL-U. Taggers assign grammatical categories to each token.

• UPOS: Universal part-of-speech tag from the UD inventory
• XPOS: Language-specific part-of-speech tag (may be empty with an underscore)
• Serves as a critical input feature for downstream dependency parsing and entity recognition

The FEATS field (column 6) in CoNLL-U encodes inflectional properties of a token using a key-value pair notation. This structured representation captures grammatical categories like tense, number, case, and gender.

• Format: Feature=Value pairs separated by pipes (e.g., Number=Sing|Case=Nom)
• Enables precise linguistic analysis across morphologically rich languages
• An underscore (_) indicates the feature is not applicable or unspecified

An extended representation in Universal Dependencies that augments the basic syntactic tree in the DEPS field (column 9). Enhanced dependencies add extra arcs to capture implicit predicates, shared arguments, and control relationships.

• Represents semantic roles beyond surface syntax
• Critical for tasks like semantic role labeling and coreference resolution
• Stored as a list of head-relation pairs separated by pipes in CoNLL-U

CoNLL-U handles complex tokenization through its multiword token lines (lines 1-2) and empty node lines (lines 3-4). This structure represents cases where syntactic words do not align with orthographic words.

• Multiword tokens: A single orthographic token spanning multiple syntactic words (e.g., French 'du' = 'de' + 'le')
• Empty nodes: Syntactic words with no surface realization (e.g., elided verbs in gapping constructions)
• The ID field uses dash notation (e.g., 1-2) for multiword spans and decimal notation (e.g., 2.1) for empty nodes