Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Phrase Detectives Corpus 2.1.4

Jon Chamberlain, Silviu Paun, Juntao Yu, Udo Kruschwitz and Massimo Poesio Release date: January 2019

This is a corpus of anaphorically annotated documents, consisting of all the documents completely annotated as of 12th October 2018 using the Phrase Detectives game-with-a-purpose:


The corpus is avaliable at this link.

The Documents

A total of 542 documents are included, for a total of 408K tokens and 108K markables. The documents are divided in:

  • a 'gold' subset of 45 documents (23K tokens, 6K markables) annotated both by Phrase Detectives players and by two experts. This subset is a cleaned-up version of the Phrase Detectives 1.0 corpus released in 2016. In addition to the gold annotation, a 'silver' annotation is included, extracted using the Mention Pair Annotation (MPA) probabilistic aggregation method for coreference discussed in the paper:

    Silviu Paun, Jon Chamberlain, Udo Kruschwitz, Juntao Yu and Massimo Poesio, 2018. A Probabilistic Annotation Model for Crowdsourcing Coreference. In Proc. of EMNLP.

  • a 'silver' subset of 497 documents (384 tokens, 101K markables) for which only the silver MPA annotation is provided.

The silver subset consists of documents from two separate genres:

  • Wikipedia pages (350 documents, 218K tokens, 57678 markables)
  • fiction from the Project Gutenberg collection (145 documents, 158739 tokens, 41898 markables)

The gold subset consists of documents from three genres:

  • Wikipedia pages (35 docs, 15287 tokens, 3957 markables)
  • fiction from the Project Gutenberg collection (5 documents, 7536 tokens, 1947 markables)
  • art history texts from the GNOME corpus (5 documents, 989 tokens, 274 markables)

A document is considered `complete' is all its markables have at least 8 annotations, and each interpretation has been validated by at least 4 players. On average, 20.6 annotations per markable are present.

Markup formats

The documents are provided in three different formats: MAS-XML, CONLL, and CRAC.


This is the format used for the original release of the corpus. It is an inline XML format, in which each markable is identified with an <ne> element, and anaphoric annotation is stored using <PDante> elements containing in turn <anchor> elements specifiying annotations and validations (all annotations and validations are provided) and <comment> elements with additional information provided by the players. Details for <PDante> and <anchor> are as follows:

id = id of the markable
visibility = whether the administrator marked this to be "hidden"
start_chr = the markable span was edited and the new start character is this
end_chr = the markable span was edited and the new end character is this

timestamp = time the annotation was created
interface = interface the annotation was created on (0 = standalone PD, 1 = Facebook PD)
mode = mode of annotation (a = annotation, v = validation, e = expert)
agree = was this annotation in agreement with the interpretation
annotation_time = time to create the annotation response from page load
user_rating = rating of the user
user_id = encrypted user_id
type = interpretation type (DN = Discourse New, DO = Discourse Old, NR = Non-referring, PR = Property, SE = same entitiy)
ante = antecedent markable id
favoured = the expert indicates whether this is the best interpretation (y) or another might be better

type/type_id = the type of error detected:

  1. not_selectable
  2. out_of_context_window
  3. parse_error
  4. discourse_diexis
  5. ambiguous
  6. non_referring
  7. nearest_mention_embedding
  8. bridge
  9. quantifier


This is the format used in the CONLL 2011 and 2012 Shared Tasks on Coreference. Unlike the MAS-XML format, where all interpretations produced by players are included, only one interpretation is provided for each markable in a doc in this format. For the docs in the silver subset, this is the silver interpretation. For the docs in the gold subset, both the gold and the silver interpretations are included, in separate versions of each document.


This is the format used in the CRAC 2018 Shared Task on Anaphora Resolution with the ARRAU corpus. As in the case of the CONLL format, only one interpretation is provided for each markable.

Annotation scheme

Markables are annotated according to a scheme including the following labels:

  • NR: non-referring. Used for expletives, as in '[It] rains'.
  • PR: used for predicatives NPs, as in 'John is [a policeman]'
  • DN: Used for discourse-new mentions.
  • DO: Used for discourse-old mentions.

Split-antecedent plurals are marked.

Folder structure

The gold and silver subsets are contained in subfolders called 'gold' and 'silver', respectively.

Document Guide

Source files are included in this corpus release including files appended:

No appending or -no_headder = The original file
.filtered = Coverted to an input format
-masxml_xsd = Original MASXML file
-masxml = File converted to MASXML format for the game
-sgf = File converted to SGF format to imported into the game

Export files are appended -game.


Full documentation and preliminary analysis of the data has been published in the following NAACL paper:

Poesio, M., Chamberlain, J., Paun, S., Yu, J., Uma, A., and Kruschwitz, U. (2019). A Crowdsourced Corpus of Multiple Judgments and Disagreement on Anaphoric Interpretation. In Proceedings of the 2019 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2019).

Earlier papers on the corpus include:

Chamberlain, J., and Poesio, M., and Kruschwitz, U. (2016). Phrase Detectives Corpus 1.0: Crowdsourced Anaphoric Coreference. In Proceedings of the 2016 Language Resources and Evaluation Conference (LREC'16).

Poesio, M., Chamberlain, J., Kruschwitz, U., Robaldo, L., and Ducceschi, L. (2013). Phrase Detectives: Utilizing collective intelligence for internet-scale language resource creation. ACM Transactions on Interactive Intelligent Systems.


Please report any errors or issues to

Change Log

January 2019: Version 2.1 created.


No releases published


No packages published