Summary

The Serbian UD treebank is based on the SETimes-SR corpus and additional news documents from the Serbian web.

Annotation Source

• Lemmas: automatic pretagging + manual correction (there's no format/annotation style for this)
• UPOS, features: converted from MULTEXT-East
• Dependency realations: automatic preprocessing + manual correction, all directly in UD.

Data Split

Training data: 80% of pseudorandom documents from SETimes-SR (sentence ids set*) and 13 web news documents (sentence ids news*), comprising 3497 sentences (77,334 tokens).

Development data: 10% of pseudorandom documents from SETimes-SR (sentence ids set*) and 13 web news documents (sentence ids news*), comprising 476 sentences (11,460 tokens).

Test data: 10% of pseudorandom documents from SETimes-SR (sentence ids set*) and 13 web news documents (sentence ids news*), comprising 411 sentences (8,879 tokens).

The corpus is parallel with a subset of UD Croatian-SET. In release 2.4, the Serbian and Croatian treebanks were re-split so that training, development and test sets are compatible (corresponding documents are in the same section in both languages).

Changelog

• 2019-04-30 v2.4

  • New data split
  • 13 web news documents added

• 2018-04-15 v2.2

  • Repository renamed from UD_Serbian to UD_Serbian-SET.

=== Machine-readable metadata =================================================
Data available since: UD v2.1
License: CC BY-SA 4.0
Includes text: yes
Genre: news
Lemmas: manual native
UPOS: converted from manual
XPOS: not available
Features: converted from manual
Relations: manual native
Contributors: Samardžić, Tanja; Ljubešić, Nikola
Contributing: elsewhere
Contact: tanja.samardzic@uzh.ch
===============================================================================