The Danish UD treebank is a conversion of the Danish Dependency Treebank.
The Danish UD treebank has been converted from the Danish Dependency Treebank (Buch-Kromman,2003) into Universal Dependencies (UD). It consists of 5,512 sentences (100k words). The Danish source texts and the Danish part-of-speech tags were created by the PAROLE-DK project (Keson 1998) by the Danish Society for Language and Literature.
In the DDT formalism, determiners head nouns, and auxiliaries head verbs. In order to promote content words to heads, we have applied a cascade of graph transformations that make function words (determiners, auxiliaries, conjunctions, etc) leaves in the dependency tree, instead of intermediate elements between content heads.
The part-of-speech tags and labels from the original treebank have been partially converted using mappings, and partially using the new calculated tree structure as a reference to assign labels.
The Danish Dependency Treebank was released under the GNU GPL license, hence that license can be used for UD_Danish as well. However, since GPL is more suitable for programs than for data (see UniversalDependencies/docs#296 for a discussion), we asked for and Matthias Buch-Kromann was kind enough to grant the permission to use the Creative Commons license as an alternative.
Contributors (in order of last names)
- Anders Johannsen
- Héctor Martínez Alonso
- Barbara Plank
-
Johannsen, Anders, Martínez Alonso, Héctor and Plank, Barbara. "Universal Dependencies for Danish". TLT14, 2015.
-
Buch-Kromann, Matthias T., Line Mikkelsen, and Stine Kern Lynge. "Danish dependency treebank.". TLT. 2003.
-
Keson, Britt (1998). Documentation of The Danish Morpho-syntactically Tagged PAROLE Corpus. Technical report, DSL
GNU Public License, version 2 OR CC BY-SA 4.0
Tree count: 5512 Word count: 100733 Token count: 100733 Dep. relations: 39 of which 5 language specific POS tags: 17 Category=value feature pairs: 46
- reattached nonprojective punctuation using udapy ud.FixPunct
- replaced goeswith relations that were used wrongly between number and %
- fixed left-headed appos relations
- fixed numeral-like ADJ to NUM
- fixed: clausal (VERB) adverbial modifiers should be advcl, not advmod
- fixed: nominal (NOUN|PROPN|PRON|DET|NUM) adverbial modifiers should be obl, not advmod
- fixed non-leaf case and mark
- fixed: pronouns cannot be mark
- adverbs attach as advmod, not obl
- obl:lmod instead of obl:loc (compatibility with other languages)
Changes for this version include:
- CCONJ attach to the second conjunct
- harmonized POS and labels to v2.0
- added obl / nmod distinction
- resolved punctuation bugs
- included ~4K words from the trailing train section into test and dev, marked with dev2 and test2 in the sentence comments
next release should improve the labels of subordinate clauses, and resolve the attachment of relative clauses dependending on pronouns like "det".
From v1.1 to v1.2, many improvements were made and integrated in the conversion scripts. In particular, we split multiword units, fixed issues with AUX and DET (some words that are not closed-class items were labeled as such), changed the headness of infinitive phrases (for at gøre), and added heuristics to introduce clausal dependency relations. Further smaller modifications include the change of POS for "som" and "der" from PART to "PRON", and fixing the lemma of quotes from HTML-escaping to plain quotes (" => "). Release v1.1 contained many "dep" relations for constructions that were not yet labeled in the first release (e.g. clausal dependents). The current release reduces the number of "dep" rels significantly, from over 7k in v1.1 to only 77 remaining cases in v1.2.
This first release contains the original tokenization of DDT, where multiword units like i_dag (today) or selv_om (even though) have been underscored together. These units are made up of more than one syntactic word and need to be split, with a subsequent reattaching of the dependency tree, for the treebank to be fully compliant with UD tokenization.
Relations across clauses are not labeled in this scheme, i.e., the current conversion does not assign labels like csubj.
Dear Dan,
The GPL seemed to be the best (and possibly only) license around when we created the DDT a long time ago. But I certainly don't mind double-licensing the DDT under a CC-BY license as well. Is there anything you want me to do, other than giving you this permission in an email as I have done now?
Best regards, Matthias
On Fri, Sep 9, 2016 at 10:58 PM, Dan Zeman [email protected] wrote:
Dear Matthias,
I am writing to you regarding the license of the Danish Dependency Treebank. We have been playing with this data, converting it first to the Prague style in the HamleDT project (http://ufal.mff.cuni.cz/hamledt), then Barbara Plank and others converted it to Universal Dependencies ( http://universaldependencies.org/).
All this happened with the GPL license in mind, which means we also attach the same license to any conversions we create. However, the vast majority of treebanks in UD are under one of the CreativeCommons licenses, and there is now a discussion about how much GPL is actually suitable for data (as opposed to software code). The discussion is here:
UniversalDependencies/docs#296 (comment) 245569728
Hence I was wondering whether you would mind allowing a CreativeCommons license as an alternative to GPL?
Thanks and all the best,
Dan Zeman
=== Machine-readable metadata (DO NOT REMOVE!) ================================ Data available since: UD v1.1 License: CC BY-SA 4.0 Includes text: yes Genre: news fiction spoken nonfiction Lemmas: converted from manual UPOS: converted from manual XPOS: not available Features: converted from manual Relations: converted from manual Contributors: Johannsen, Anders; Martínez Alonso, Héctor; Plank, Barbara Contributing: here Contact: [email protected] ===============================================================================