We use a revised version of the CoNLL-X format called CoNLL-U. Annotations are encoded in plain text files (UTF-8, using only the LF character as line break) with three types of lines:
- Word lines containing the annotation of a word/token in 10 fields separated by single tab characters; see below.
- Blank lines marking sentence boundaries.
- Comment lines starting with hash (#).
Sentences consist of one or more word lines, and word lines contain the following fields:
- ID: Word index, integer starting at 1 for each new sentence; may be a range for multiword tokens; may be a decimal number for empty nodes.
- FORM: Word form or punctuation symbol.
- LEMMA: Lemma or stem of word form.
- UPOSTAG: Universal part-of-speech tag.
- XPOSTAG: Language-specific part-of-speech tag; underscore if not available.
- FEATS: List of morphological features from the universal feature inventory or from a defined language-specific extension; underscore if not available.
- HEAD: Head of the current word, which is either a value of ID or zero (0).
- DEPREL: Universal dependency relation to the HEAD (root iff HEAD = 0) or a defined language-specific subtype of one.
- DEPS: Enhanced dependency graph in the form of a list of head-deprel pairs.
- MISC: Any other annotation.
The fields DEPS and MISC replace the obsolete fields PHEAD and PDEPREL of the CoNLL-X format. In addition, we have modified the usage of the ID, FORM, LEMMA, XPOSTAG, FEATS and HEAD fields as explained below.
The fields must additionally meet the following constraints:
- Fields must not be empty.
- Fields other than FORM and LEMMA must not contain space characters.
- Underscore (_) is used to denote unspecified values in all fields except ID. Note that no format-level distinction is made for the rare cases where the FORM or LEMMA is the literal underscore – processing in such cases is application-dependent. Further, in UD treebanks the UPOSTAG, HEAD, and DEPREL columns are not allowed to be left unspecified.
Words, Tokens and Empty Nodes
To accommodate the representation of both words and (multiword) tokens (see Tokenization and word segmentation), we adopt an extension of the original CoNLL-X token indexing scheme, where words are indexed with integers 1, 2, 3, …, while (multiword) tokens are indexed with integer ranges like 1-2 or 3-5. Lines representing such tokens are inserted before the first word in the range. These ranges must be nonempty and must not overlap. They have a FORM value – the string that occurs in the sentence – but have an underscore in all the remaining fields (because the token represents multiple words, each with its own lemma, part-of-speech tag, syntactic head, and so on). This is illustrated in the following annotation snippet, showing only the first three fields for the Spanish sentence vámonos al mar (let’s go to the sea):
1-2 vámonos _ 1 vamos ir 2 nos nosotros 3-4 al _ 3 a a 4 el el 5 mar mar
We extract the word sequence by skipping all range IDs:
1 vamos ir 2 nos nosotros 3 a a 4 el el 5 mar mar
We extract the raw token sequence by skipping all integer IDs that are included in a preceding range ID:
1-2 vámonos _ 3-4 al _ 5 mar mar
To accommodate the use of empty nodes for the analysis of ellipsis in the enhanced dependency representation, we adopt a further extension of the indexing scheme from v2. It is possible to insert one or more empty nodes indexed i.1, i.2, etc. immediately after a word with index i (where i = 0 for sentence-initial empty nodes). Here is an example showing the use of an empty node in the analysis of the sentence Sue likes coffee and Bill tea:
1 Sue Sue 2 likes like 3 coffee coffee 4 and and 5 Bill Bill 5.1 likes like 6 tea tea
Empty nodes must have non-empty ID and DEPS fields and empty fields (i.e. underscores) for HEAD and DEPREL, because they are only part of the enhanced dependency graph. All other fields may contain either underscores or values: for example, they can optionally have a word form and lemma as in the example above.
The UPOSTAG field contains a part-of-speech tag from the universal POS tag set, while the XPOSTAG optionally contains a language-specific part-of-speech tag, normally from a traditional, more fine-grained tagset. If the XPOSTAG field is used, the treebank-specific documentation should define a mapping from XPOSTAG to UPOSTAG values (which may be context-sensitive and refer to other fields as well). If no language-specific tags are available, the XPOSTAG field should contain an underscore for all words.
The FEATS field contains a list
of morphological features, with vertical bar (|) as list separator and with underscore to represent the empty list.
All features should be represented as attribute-value pairs, with an equals sign (=) separating the attribute from the value. In addition, features should as far as possible be selected from the universal feature inventory and be sorted alphabetically by attribute names. It is possible to declare that a feature has two or more values for a given word:
Case=Acc,Dat. In this case, the values are sorted alphabetically. In sorting, uppercase letters are considered identical to their lowercase counterparts. Feature names must have the form
[A-Z0-9][A-Z0-9a-z]*(\[[a-z0-9]+\])? and feature values must have the form
Here is an example, showing only the first five fields for the Swedish sentence Då var han elva år (Then he was eleven years old):
1 Då då ADV AB _ 2 var vara VERB VB.PRET.ACT Tense=Past|Voice=Act 3 han han PRON PN.UTR.SIN.DEF.NOM Case=Nom|Definite=Def|Gender=Com|Number=Sing 4 elva elva NUM RG.NOM Case=Nom|NumType=Card 5 år år NOUN NN.NEU.PLU.IND.NOM Case=Nom|Definite=Ind|Gender=Neut|Number=Plur 6 . . PUNCT DL.MAD _
Morphological annotation is only provided for words. Tokens that are not words have an underscore in the UPOSTAG, XPOSTAG and FEATS fields.
The HEAD and DEPREL fields are used to encode a dependency tree over words. The DEPREL value should be a universal dependency relation or a language-specific subtype of such a relation (defined in the language-specific documentation). As in the case of morphology, syntactic annotation is only provided for words, and tokens that are not words have an underscore in both the HEAD and DEPREL fields.
The HEAD and DEPREL values define the basic dependencies which must be strictly a tree. However, in addition to these basic dependencies, treebanks may optionally provide an enhanced dependency representation that specifies additional dependency relations, for example, when dependencies propagate over coordinate structures. The enhanced dependency representation, which in general is a graph and not a tree, is specified in the DEPS field, using a list of head-relation pairs. We use colon (:) to separate the head and relation and (as usual) vertical bar (|) to separate list items and underscore for the empty list. The list is to be sorted by the index of the head:
Note that the DEPS field should be a self-contained representation of the enhanced dependency graph, which means that dependencies that are shared between the basic and the enhanced dependency representations must be repeated in the DEPS field. Here is an example, showing the first nine fields for the English sentence They buy and sell books:
1 They they PRON PRP Case=Nom|Number=Plur 2 nsubj 2:nsubj|4:nsubj 2 buy buy VERB VBP Number=Plur|Person=3|Tense=Pres 0 root 0:root 3 and and CONJ CC _ 4 cc 4:cc 4 sell sell VERB VBP Number=Plur|Person=3|Tense=Pres 2 conj 0:root|2:conj 5 books book NOUN NNS Number=Plur 2 obj 2:obj|4:obj 6 . . PUNCT . _ 2 punct 2:punct
The dependency relations in both DEPREL and DEPS must have the form
and use of the dash (
-) character instead of underscore (
_) is encouraged.
The final MISC field is for storing any additional information that does not fit into any of the other fields, such as language-specific annotation, any information about other linguistic levels such as discourse, or projective heads and dependency relations (cf. the old PHEAD and PDEPREL fields of the CoNLL-X format). The exact format used in this field should be specified in the treebank-specific documentation, but it has to be formatted as a list that can be split on the bar character (|) without special escaping. If the MISC field is not used, it should contain an underscore.
To facilitate reconstruction of original (pre-tokenization) text, the information on original word segmentation should be kept if available. Every token after which there was no space in the original text should contain
SpaceAfter=No in its MISC field. Note that this feature applies to the token level, not to the word level. Syntactic words that are just part of surface tokens will be ignored during detokenization and thus do not need the feature. In the example below, the line indexed 3 does not contain the
SpaceAfter feature even though there was no space between für and das in the underlying sentence. However, if there were no space between fürs and the following token, the 3-4 line would have
Note that columns 5 to 9 are collapsed in the following example.
# text = Er arbeitet fürs FBI (deutsch etwa: „Bundesamt für Ermittlung“). # text_en = He works for the FBI (German approx: “Bundesamt für Ermittlung”). 1 Er er PRON … _ 2 arbeitet arbeiten VERB … _ 3-4 fürs _ _ … _ 3 für für ADP … _ 4 das der DET … _ 5 FBI FBI PROPN … _ 6 ( ( PUNCT … SpaceAfter=No 7 deutsch deutsch ADV … _ 8 etwa etwa ADV … SpaceAfter=No 9 : : PUNCT … _ 10 „ „ PUNCT … SpaceAfter=No 11 Bundesamt Bundesamt NOUN … _ 12 für für ADP … _ 13 Ermittlung Ermittlung NOUN … SpaceAfter=No 14 “ “ PUNCT … SpaceAfter=No 15 ) ) PUNCT … SpaceAfter=No 16 . . PUNCT … _
Sentence Boundaries and Comments
There must be exactly one blank line after every sentence, including the last sentence in the file. Empty sentences are not allowed.
Lines starting with the
# character and preceding a sentence are considered as carrying metadata relevant to the following sentence. These lines are an integral part of the format as they give the ability to embed metadata together with the sentences. Consequently, any tools compatible with the CoNLL-U format should carry these lines over into their output (unless specifically designed to process them in some way). The contents of the comments and metadata is basically unrestricted and will vary depending on the application, but from v2 every sentence must have a unique sentence id, formatted as in the examples below. In sentence ids, the slash character (“/”) is reserved for specialized downstream use and should be avoided in UD treebanks. Comments used to specify the unannotated sentence as a single string should also be formatted as below. Comment and metadata lines inside sentences (i.e., between the token lines) are disallowed.
# sent_id = 1 # text = They buy and sell books. 1 They they PRON PRP Case=Nom|Number=Plur 2 nsubj 2:nsubj|4:nsubj _ 2 buy buy VERB VBP Number=Plur|Person=3|Tense=Pres 0 root 0:root _ 3 and and CONJ CC _ 4 cc 4:cc _ 4 sell sell VERB VBP Number=Plur|Person=3|Tense=Pres 2 conj 0:root|2:conj _ 5 books book NOUN NNS Number=Plur 2 obj 2:obj|4:obj SpaceAfter=No 6 . . PUNCT . _ 2 punct 2:punct _ # sent_id = 2 # text = I have no clue. 1 I I PRON PRP Case=Nom|Number=Sing|Person=1 2 nsubj _ _ 2 have have VERB VBP Number=Sing|Person=1|Tense=Pres 0 root _ _ 3 no no DET DT PronType=Neg 4 det _ _ 4 clue clue NOUN NN Number=Sing 2 obj _ SpaceAfter=No 5 . . PUNCT . _ 2 punct _ _