Skip to content
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
295 lines (217 sloc) 12.8 KB


The German UD is converted from the content head version of the universal dependency treebank v2.0 (legacy).


The German UD conforms to the UD guidelines, but there are some exceptions.

Morphology in release 2.0

Universal POS tags were assigned manually, while LEMMA and XPOSTAG were predicted by TreeTagger (first for release 1.4; see Changelog below). Morphological features were assigned using rules based on the values of the other columns (UPOSTAG, XPOSTAG, LEMMA, FORM, DEPREL). Gender, number and case of nouns and their det/amod children are based on the (manual) syntactic annotation, e.g. nsubj => nominative. They should have high precision but lower recall because we did not add them where the context did not provide enough clues (morphological analyzer / lexicon was not used at this stage).

Morphology in release 2.2

The XPOS and FEATS have been updated using the mate parser with Tueba-D/Z. The mate parser model was trained on the first 80% of Tueba-D/Z as converted by an updated version of the TuebaUdConverter that improved the morphological tag extraction. Tueba-D/Z tokens containing hyphens were split into multiple tokens to align with the UD tokenization.

During tagging, `` '' and -- were temporarily normalized to " and - and some new spellings were normalized back to old spellings to be compatible with the Tueba-D/Z model. Tueba-D/Z is fairly homogenous and somewhat dated, so the tagger does not perform as well on user-generated content as on older newspaper text. A detailed analysis is below.

After incorporating the automatic mate POSTAG and FEATS, the UD sentences were aligned with Tiger using the mapping in #13 and LEMMA (for content words) and POSTAG and FEATS (for all tokens) were updated with gold values from Tiger. The only exception the feature Voice, whose value is passed through from mate since it's not annotated on the aux in Tiger.

Ordinal numbers split into two tokens in UD (e.g., "3 .") were rejoined as in Tiger and compounds were reanalyzed in order to provide consistent FEATS for all subparts. If any subpart of a compound was tagged as NE by the mate parser, the deprel flat was used, otherwise compound.

Mate model analysis:

The accuracy on Tueba-D/Z test data for XPOSTAG is 98.3% and for FEATS as an unanalyzed string is 91.7%. Analyzing the individual morphological tags, the results are:

       Acc.    Prec. Rec.  F1

Case 0.95580 0.996 0.911 0.952 Definite 0.99653 0.986 0.987 0.986 Foreign 0.99835 0.646 0.714 0.678 Gender 0.96150 0.982 0.929 0.955 Mood 0.99554 0.984 0.959 0.972 NumType 0.99988 0.991 0.998 0.995 Number 0.97852 0.995 0.966 0.980 Person 0.99481 0.986 0.966 0.976 Polarity 0.99998 0.999 0.999 0.999 Poss 0.99993 0.995 0.997 0.996 PronType 0.99734 0.998 0.988 0.993 Reflex 0.99991 0.996 0.992 0.994 Tense 0.99567 0.983 0.962 0.972 VerbForm 0.99492 0.993 0.965 0.978 Voice 0.99934 0.934 0.967 0.950

Case and Gender are not unexpectedly the least accurate of the frequent features. Foreign could potentially be removed, although when inspecting the UD instances the precision seems relatively high.


2018-11-15 Dan Zeman

  • Fixed punctuation using udapy -s ud.FixPunct

2018-04-15 Adriane Boyd

  • Fixed morphology (retagged with Mate, see above), re-synchronized with original Tiger data where applicable.

2017-11-20 Dan Zeman

  • Fixed: copula is AUX.
  • Fixed: capitalization of multi-word tokens at sentence beginning.

2017-04-13 Dan Zeman

Removed duplicate sentences from the training data. They were too long to believe that they were naturally occurring duplicates.

2017-03-01 v2.0

  • Converted to UD v2 guidelines (Dan Zeman)

2016-08-21 Dan Zeman

Added sentence ids. Added LEMMA and XPOSTAG predicted by TreeTagger with a German model supplied with the tagger and available in Treex (, commit 50ad1fe0b9907ac382cbcda0a0f102602abc21a0). The UPOSTAGs from the original data (assigned manually) were not modified. Some features were also added if they could be derived from the information already present. Features that need a lexicon and/or disambiguation, such as Gender, Number and Case, have only been added if they can be deduced from the (manually annotated) dependency structure, plus a few heuristics (e.g. form equal to lemma often but not always means singular).

The work was done mainly using the HamleDT::DE::FixUD block, see

2015-11-08 Wolfgang Seeker

Removed sentences from test due to overlap with dev (sent-no. 6, 8, 79, 80, 88, 108, 109, 118, 152, 154, 164, 167, 190, 191, 195, 206, 215, 220, 229, 247, 295, 346, 451) Removed sentences from dev due to overlap with train (sent-no. 616)

############################################################################### LEGACY README FILE BELOW ###############################################################################

=================================== Universal Dependency Treebanks v2.0

This directory contains treebanks for the following languages: English, German, French, Spanish, Swedish, Korean, Japanese, Indonesian, Brazilian Portuguese, Italian and Finnish.

A description of how the treebanks were generated can be found in:

Universal Dependency Annotation for Multilingual Parsing Ryan McDonald, Joakim Nivre, Yvonne Quirmbach-Brundage, Yoav Goldberg, Dipanjan Das, Kuzman Ganchev, Keith Hall, Slav Petrov, Hao Zhang, Oscar Tackstrom, Claudia Bedini, Nuria Bertomeu Castello and Jungmee Lee Proceedings of ACL 2013

A more detailed description of each relation type in our harmonized scheme is included in the file universal-guidelines.pdf.

Each treebank has been split into training, development and testing files.

Each file is formatted according to the CoNLL 2006/2007 guidelines:

The treebank annotations use basic Stanford Style dependencies, modified minimally to be sufficient for each language and be maximally consistent across languages. The original English Stanford guidelines can be found here:

============================================================================== Version 2.0 - What is new

  1. Added more data for German, Spanish and French.
  2. Added Portuguese Brazilian, Indonesian, Japanese, Italian and Finnish.
  3. New content-head versions for 5 languages (see below).
  4. A number of bug fixes in the harmonization process.

===================== Standard dependencies

In release 2.0 we include two sets of dependencies. The first is standard Stanford dependencies, which correspond roughly to the output of the Stanford converter for English with the copula as head set to true. In general, these are content-head dependency representations with two major exceptions: 1) adpositions are the head in adpositional phrases, and 2) copular verbs are the head in copluar constructions.

This data is in the std/ directory and contains all languages but Finnish.

Version 1.0 of the data is only standard.

========================== Content head dependencies

In order to converge to a more uniform multilingual standard, in particular for morphologically rich languages, this release also includes a beta version of content-head dependencies for five languages: German, Spanish, Finnish, French and Swedish. Here the content word is always the head of a phrase.

============================================================================= Language Specific Information

==================== English dependencies

Note that the English dependencies are based on the original Penn Treebank data automatically converted with the Stanford Dependency Converter. Instructions for how to do this with corresponding scripts are included in the English directory.

==================== Finnish dependencies

Finnish data is in the ch/fi directory and was produced by researchers at the University of Turku. In that directory there are specific README and LICENSE files for that data. Two things to note. First, the Finnish data is only content-head. This is due to difficulties in automatically converting the data to standard format from its original annotations. Second, we have included a test set in the release, but this is not the real test set, just a subset of the training. The true test set for this data is blind (as per the wishes of the researchers at Turku). The unannotated test data is included as well as instructions for obtaining scores on predictions.

============================================================================= Other Information

================================ Fine-grained part-of-speech tags

In the CoNLL file format there is a coarse part-of-speech tag field (4) and a fine-grained part-of-speech tag field (5). In this data release, we use the coarse field to store the normalized universal part-of-speech tags that are consistent across languages. The fine-grained field contains potentially richer part-of-speech information depending on the language, e.g., a richer tag representation for clitics.

========================= Licenses and terms-of-use

For the following languages

German, Spanish, French, Indonesian, Italian, Japanese, Korean and Brazilian Portuguese

we will distinguish between two portions of the data.

  1. The underlying text for sentences that were annotated. This data Google asserts no ownership over and no copyright over. Some or all of these sentences may be copyrighted in some jurisdictions. Where copyrighted, Google collected these sentences under exceptions to copyright or implied license rights. GOOGLE MAKES THEM AVAILABLE TO YOU 'AS IS', WITHOUT ANY WARRANTY OF ANY KIND, WHETHER EXPRESS OR IMPLIED.

  2. The annotations -- part-of-speech tags and dependency annotations. These are made available under a CC BY-NC-SA 3.0 non commercial license. GOOGLE MAKES THEM AVAILABLE TO YOU 'AS IS', WITHOUT ANY WARRANTY OF ANY KIND, WHETHER EXPRESS OR IMPLIED. See attached LICENSE file for the text of CC BY-NC-SA.

Portions of the German data were sampled from the CoNLL 2006 Tiger Treebank data. Hans Uszkoreit graciously gave permission to use the underlying sentences in this data as part of this release.

For English, Italian, Finnish and Swedish, please see licences included in these directories or the following sources.

Finnish - Swedish - Italian -

We are greatful to researchers at those institutes who provided us data, in particular:

Maria Simi and company from the University of Pisa. Converting Italian Treebanks: Towards an Italian Stanford Dependency Treebank Bosco, Cristina and Montemagni, Simonetta and Simi, Maria Proceedings of LAW VII & ID

Filip Ginter and company from the University of Turku. Building the essential resources for Finnish: the Turku Dependency Treebank Haverinen, Katri and Nyblom, Jenna and Viljanen, Timo and Laippala, Veronika and Kohonen, Samuel and Missil{"a}, Anna and Ojala, Stina and Salakoski, Tapio and Ginter, Filip Language Resources and Evaluation, 2013

Joakim Nivre and company from Uppsala University.

And Chris Manning and Marie-Catherine de Marneffe from Stanford and Ohio. Generating typed dependency parses from phrase structure parses MC De Marneffe, B MacCartney, CD Manning, Proceedings of LREC, 2006

Any use of the data should reference the above plus:

Universal Dependency Annotation for Multilingual Parsing Ryan McDonald, Joakim Nivre, Yvonne Quirmbach-Brundage, Yoav Goldberg, Dipanjan Das, Kuzman Ganchev, Keith Hall, Slav Petrov, Hao Zhang, Oscar Tackstrom, Claudia Bedini, Nuria Bertomeu Castello and Jungmee Lee Proceedings of ACL 2013

======= Contact

=== Machine-readable metadata (DO NOT REMOVE!) ================================ Data available since: UD v1.0 License: CC BY-NC-SA 3.0 US Includes text: yes Genre: news reviews wiki Lemmas: automatic UPOS: converted from manual XPOS: automatic Features: automatic Relations: converted from manual Contributors: Petrov, Slav; Seeker, Wolfgang; McDonald, Ryan; Nivre, Joakim; Zeman, Daniel; Boyd, Adriane Contributing: here Contact:

(Original treebank contributors: Quirmbach-Brundage, Yvonne; LaMontagne, Adam; Souček, Milan; Järvinen, Timo; Radici, Alessandra)

You can’t perform that action at this time.