The Norwegian UD treebank is based on the Bokmål section of the Norwegian Dependency Treebank (NDT), which is a syntactic treebank of Norwegian. The current version of NDT has been automatically converted to the UD scheme by Ingerid Løyning Dale, Per Erik Solberg and Andre Kåsen at the Norwegian Language Bank at the National Library of Norway. This conversion builds to a large extent on previous conversions by Lilja Øvrelid at the University of Oslo.
NDT was developed 2011-2014 at the National Library of Norway in collaboration with the Text Laboratory and the Department of Informatics at the University of Oslo. NDT contains around 300,000 tokens taken from a variety of genres. The treebank texts have been manually annotated for morphosyntactic information. The morphological annotation mainly follows mainly the Oslo-Bergen Tagger. The syntactic annotation follows, to a large extent, the Norwegian Reference Grammar, as well as a dependency annotation scheme formulated at the outset of the annotation project and iteratively refined throughout the construction of the treebank. For more information, see the references below.
In creating the data splits, care has been taken to preserve contiguous texts in the different splits and also to keep a fair balance of genres in each of the splits. Petter Hohle created the splits for the Norwegian UD treebank. The splits were created by concatenating the following files (available with the distribution of NDT):
Training data (15696 sentences, 180 individual files):
- ap001_0000 -- ap012_0002 (53 files)
- bt001_0000 -- bt005_0001 (28 files)
- db001a_0000 -- db013_0004 (42 files)
- kk001_0000 -- kk005_0001 (10 files)
- sp-bm001_0000 -- sp-bm001_0008 (9 files)
- vg001_0000 -- vg002_0003 (8 files)
- blogg-bm001_0000 -- blogg-bm003_0000 (9 files)
- nou001_0000 -- nou004_0000 (10 files)
- st001_0000 -- st005_0000 (11 files)
Development data (2410 sentences, 26 individual files):
- ap012_0003 -- ap014_0002 (7 files)
- bt005_0002 -- bt005_0005 (4 files)
- db013_0005 -- db014_0002 (5 files)
- kk006_00001 -- kk007_0000 (2 files)
- sp-bm002_0000 -- sp-bm002_0001 (2 files)
- vg002_0004 (1 file)
- blogg-bm003_0001 -- blogg-bm003_0002 (2 files)
- nou004_0001 (1 file)
- st005_0001 -- st005_0002 (2 files)
Test data (1939 sentences, 26 individual files):
- ap014_0003 -- ap015_0002 (7 files)
- bt005_0006 -- bt006_0001 (4 files)
- db014_0003 -- db014_0007 (5 files)
- kk007_0001 -- kk008_0000 (2 files)
- sp-bm003_0000 -- sp-bm003_0001 (2 files)
- vg002_0005 (1 file)
- blogg-bm003_0003 -- blogg-bm003_0004 (2 files)
- nou004_0002 (1 file)
- st005_0003 -- st005_0004 (2 files)
Tree count: 20045
Word count: 311277
Token count: 311277
Dep. relations: 35 of which 2 language specific
POS tags: 17
Category=value feature pairs: 31
White space always indicates a token boundary and punctuation constitute separate tokens, except:
- numbers with periods, commas or colons, e.g. 1.3, 0,6, 10:13
- abbreviations, e.g. f.eks., Carl J. Hambro
- URLs, e.g. http://www.ifi.uio.no
The treebank does not contain multiword tokens.
The PoS-tags follow the universal tag set and does not add any language-specific PoS-tags. The morphological features follow the Oslo-Bergen Tagger scheme (Hagen et. al., 2000). PoS-tags and morphological features were converted automatically to the UD scheme.
The syntactic annotation in the Norwegian UD treebank conforms to the
UD guidelines, adding language-specific relations for relative clauses (acl:relcl)
and verb particles (compound:prt). The annotation has been automatically converted to
UD from the original dependency scheme described in Solberg
et. al. (2014) and further described in the NDT guidelines (Kinn
et. al.).
The conversion has not been manually checked. There are a few known discrepancies from UD:
- no mwe analysis in the treebank. This is also information that is not present in the original data.
Kristin Hagen, Janne Bondi Johannessen and Anders Nøklestad: "A Constraint-based Tagger for Norwegian". 2000. Proceedings of the 17th Scandinavian Conference in Linguistics.
Kari Kinn, Per Erik Solberg and Pål Kristian Eriksen. "NDT Guidelines for Morphological Annotation". National Library Tech Report.
Per Erik Solberg, Arne Skjærholt, Lilja Øvrelid, Kristin Hagen and Janne Bondi Johannessen. 2014."The Norwegian Dependency Treebank", Proceedings of LREC 2014, Reykjavik
Lilja Øvrelid & Petter Hohle (2016). "Universal Dependencies for Norwegian", In Proceedings of the Ninth International Conference on Language Resources and Evaluation (LREC'16)
NDT has been automatically converted to the UD scheme by Ingerid Løyning Dale, Per Erik Solberg and Andre Kåsen at the Norwegian Language Bank at the National Library of Norway. This conversion builds to a large extent on previous conversions by Lilja Øvrelid at the University of Oslo. Petter Hohle created the data splits and Fredrik Jørgensen aligned the treebank to the original texts. We thank the annotators of the original NDT: Pål Kristian Eriksen, Kari Kinn and Per Erik Solberg.
--> UD 2.12
- The conversion is completely rewritten using Grew by the Norwegian Language Bank at the National Library of Norway. The conversion is to a large extent based on the guidelines of the previous version.
- som in relative clauses is not longer treated as pronouns, but complementizers with the postag SCONJ and the label mark.
- There is no longer an explicit analysis of verbal particles. The postag has changed from ADP to ADV and the label is advmod.
- The changes in 2.10 and 2.12 are implemented.
UD 1.3 --> UD 1.4
- Added SpaceAfter annotation describing tokenization
- Removed tokens that were introduced during treebanking but are not present in original texts: 1) tokens introduced for paragraph boundaries (|), and 2) extra punctuation introduced following sentence-final abbreviations.
UD 1.2 --> UD 1.3
- In order to improve consistency between the Germanic languages, a list of auxiliary verbs were agreed on. For Norwegian these are: bli 'become', burde 'should', få 'get', ha 'have', kunne 'can', måtte 'must', skulle 'should', tørre 'dare', ville 'will', være 'be'.
- Appositions (
appos) now exclusively follow their head, in compliance with UD guidelines - Earlier
nsubjin cleft constructions are nowdislocated - Adpositions marking a subordinate clause are now
markas opposed to earliercase - In v1.2 a right-headed analysis was used for combinations of proper and common nouns such as Premier League-spillere 'Premier League players'. In v1.3 a
compoundanalysis was chosen for these. - PronType feature was added for all PRON or DET, and VerbType for all VERB and AUX
- General bug fixing