Dataset Card for DaCoref

Dataset Summary

This dataset contains coreference annotations of part of the Copenhagen Dependency Treebank.

Supported Tasks and Leaderboards

This dataset is meant to train coreference resolution models.

Languages

The dataset is available in Danish (da).

Dataset Structure

Data Instances

• Size of downloaded dataset files: 569 KB
• Size of the generated dataset: 1099 KB
• Total amount of disk used: 1668 KB

split	samples
train	2,686
val	332
test	385

An example from the dataset looks as follows.

{
  'sent_id': 'train-v2-0',
  'doc_id': 'nw/0442',
  'text': 'På fredag har SID inviteret til reception i SID-huset i anledning af at formanden Kjeld Christensen går ind i de glade tressere.',
  'tokens': ['På',
    'fredag',
    'har',
    'SID',
    'inviteret',
    'til',
    'reception',
    'i',
    'SID-huset',
    'i',
    'anledning',
    'af',
    'at',
    'formanden',
    'Kjeld',
    'Christensen',
    'går',
    'ind',
    'i',
    'de',
    'glade',
    'tressere',
    '.'
  ],
 'clusters': [[13, 14, 15]]
}

Data Fields

The data fields are the same among all splits.

• sent_id (string): The sentence ID from the Universal Dependencies.
• doc_id (string): The document ID from the Copenhagen Dependency Treebank.
• text (string): The document.
• tokens (list[str]): The tokens appearing in the document.
• clusters (list[list[int]]): The coreference clusters in the document, where the integers refer to the indices in tokens.

Dataset Creation

This coreference dataset was originally annotated as part of the Copenhagen Dependency Treebank (CDT) project but never finished. The incomplete annotation can be downloaded from the project github.

The CDT documentation contains description of the coreference classes as well as inter-annotator agreement and confusion matrices.

For this resource, we used the annotation files from the annotator "Lotte" along with the UD syntax which is an automatic conversion of the CDT syntax annotation by Johansen et al. (2015). We provide the sentence ID from the UD resource as well as the document ID from CDT. The document ID has been prepended with a two letter domain code compatible with the domain codes of the Ontonotes corpus. This is a manually mapping of the sources listed in the CDT. Only nw (newswire), mz (magazine), and bn (broadcast news) were present:

• 299 nw documents
• 41 mz documents
• 1 bn

For the CDT, only the core node of each span was annotated and one annotator manually propagated the label to the entire span. A few systematic errors were corrected in this process, the most important being that plural pronouns "we" and "they" can be coreferent with company names if they refer to the employee group of this company.

For this resource we have merged the following labels to form uniquely numbered clusters: coref, coref-evol, coref-iden, coref-iden.sb, coref-var, and ref. Coref-res and coref-res.prg are also included as clusters but not merged with any other label, nor each other.

Some notes about the annotation, but see also the CDT documentation: If conjunctions of entities are only referred to as a group, they are marked as one span. (e.g. if "Lise, Lone og Birthe" are only referred to as a group, e.g. by the plural pronoun "de"), "Line, Lone og Birthe" is marked as one span. The spans are generally as long as possible. Example: Det sidste gik ud over politikerne, da de i sin tid præsenterede [det første forslag til den milliard-dyre vandmiljøplan].

Singletons are not annotated. The annotation does not label attributative noun phrases that are connected through copula verbs such as to be. Name-initual appositive constructions are part of the same mention as the name. Generic pronouns (mainly "man" and "du") are not clustered unless they are part of a cluster, e.g. with a reflexive or possesive pronoun.

Additional Information

Dataset Curators

The work was conducted by Maria Jung Barrett, and has been uploaded to the Hugging Face Hub by Dan Saattrup Nielsen from the The Alexandra Institute.

Licensing Information

The dataset is licensed under the CC BY 4.0 license.