CoNLL Formats

Model: CoNLL-U, CoNLL-X, CoNLL-2003, CoNLL-2005, CoNLL-2009, CoNLL-2012
Origin: Conference on Natural Language Learning shared tasks
Specification: universaldependencies.org/format.html
Key Reference: de Marneffe et al. 2021 (UD); various shared task proceedings

Overview

CoNLL formats are tab-separated column-based annotation formats widely used in NLP shared tasks. CoNLL-U (Universal Dependencies) is the most widely used variant. Each line represents a token with multiple annotation columns. CoNLL formats are flat, single-file representations optimized for machine learning pipelines.

CoNLL-U (Universal Dependencies)

Column Mapping

CoNLL-U Column	Layers Equivalent	Notes
`ID`	`pub.layers.expression` token `tokenIndex`	Token position. CoNLL-U uses 1-based; Layers uses 0-based.
`FORM`	`pub.layers.expression` token `text`	Surface form.
`LEMMA`	`annotationLayer(kind="token-tag", subkind="lemma")` → `annotation.value`	Lemmatization layer.
`UPOS`	`annotationLayer(kind="token-tag", subkind="pos")` → `annotation.label`	Universal POS tag.
`XPOS`	`annotationLayer(kind="token-tag", subkind="xpos")` → `annotation.label`	Language-specific POS tag.
`FEATS`	`annotationLayer(kind="token-tag", subkind="morph")` → `annotation.features`	Morphological features (e.g., `Case=Nom\|Number=Sing`). Each feature key-value pair maps to a `feature` entry.
`HEAD`	`annotationLayer(kind="graph", subkind="dependency")` → `annotation.headIndex`	Governor token index. CoNLL-U's `0` (root) maps to `headIndex` absent or a sentinel.
`DEPREL`	Same dependency layer → `annotation.label`	Dependency relation label.
`DEPS`	`annotationLayer(kind="graph", subkind="enhanced-dependency")`	Enhanced dependencies (multiple heads). Each `head:deprel` pair creates a separate annotation in the enhanced layer.
`MISC`	`pub.layers.defs#featureMap` on the token or annotation	Catch-all for `SpaceAfter=No`, `Translit=...`, etc.

Special Token Types

CoNLL-U Feature	Layers Equivalent	Notes
Multi-word tokens (e.g., `1-2 del`)	`pub.layers.defs#tokenRefSequence`	Multi-word token ranges are represented as a `tokenRefSequence` with `tokenIndexes` covering the component tokens. The surface form and span of the multi-word token are stored in features.
Empty nodes (e.g., `2.1`)	`pub.layers.annotation#annotation` with features	Empty nodes in enhanced UD are represented as annotations (not tokens) in the enhanced dependency layer, with `features` indicating they are empty/null nodes. Their position is tracked via decimal indices stored in features.
Sentence boundaries	`pub.layers.expression` (kind: `sentence`) + `pub.layers.segmentation`	CoNLL-U blank lines between sentences map to sentence boundaries in the segmentation record.
`# text = ...` comment	`pub.layers.expression` sentence `features` or `pub.layers.expression.text`	Sentence-level metadata from comments.
`# sent_id = ...` comment	`pub.layers.expression` sentence `uuid`	Sentence identifier.
`# newpar` / `# newdoc`	`pub.layers.expression` (kind: `section`) boundaries	Paragraph and document boundaries.

CoNLL-2003 (NER)

CoNLL-2003 Column	Layers Equivalent	Notes
Word	`token.text`	Surface form.
POS tag	`annotationLayer(kind="token-tag", subkind="pos")`	POS tag.
Chunk tag	`annotationLayer(kind="token-tag", subkind="chunk")`	IOB chunk tag.
NER tag	`annotationLayer(kind="token-tag", subkind="ner")`	IOB NER tag. Can also be converted to `kind="span", subkind="entity-mention"` with token spans.

CoNLL-2005 / CoNLL-2009 (SRL)

CoNLL-200x Column	Layers Equivalent	Notes
Predicate	`annotationLayer(kind="span", subkind="predicate")`	Predicate identification.
Predicate sense	`annotationLayer(kind="span", subkind="frame", formalism="PropBank")` → `annotation.label`	PropBank sense (roleset ID).
Argument columns	`annotation.arguments[]` with `argumentRef.role`	Each argument column (ARG0, ARG1, ARGM-TMP, etc.) becomes an `argumentRef` on the frame annotation.

CoNLL-2012 (OntoNotes Coreference)

CoNLL-2012 Feature	Layers Equivalent	Notes
Coreference column	`pub.layers.annotation#clusterSet` with `kind="coreference"`	Parenthetical coreference notation (e.g., `(12)`, `(12`, `12)`) maps to cluster membership. Each cluster ID becomes a `cluster` with `memberIds` pointing to span annotations.
Speaker column	`annotationLayer(kind="token-tag", subkind="speaker")`	Speaker diarization.
Named entity spans	`annotationLayer(kind="span", subkind="entity-mention")`	Entity mention spans.

Conversion Notes

CoNLL formats are flat, token-per-line representations. Converting to Layers requires:

Parse the file into tokens, creating a tokenization with kind="whitespace" or appropriate strategy
Group tokens into sentences (blank-line delimited), creating sentence objects
Wrap in a segmentation record bound to an expression
For each annotation column, create a separate annotationLayer with appropriate kind/subkind
IOB/BILOU tags can remain as token-tags or be converted to span annotations — Layers supports both

The reverse conversion (Layers → CoNLL) selects the appropriate annotation layers and serializes them column-by-column.

Overview​

CoNLL-U (Universal Dependencies)​

Column Mapping​

Special Token Types​

CoNLL-2003 (NER)​

CoNLL-2005 / CoNLL-2009 (SRL)​

CoNLL-2012 (OntoNotes Coreference)​

Conversion Notes​