==============================================================================
Czech data for the CoNLL-2009 shared task
"Syntactic and Semantic Dependencies in Multiple Languages"
Version 1.3: January 13, 2012
==============================================================================


(1) LIST OF VERSIONS

  v1.3 [2012/01/13]: OOD data description

  v1.2 [2009/02/07]: missing Y added to FILLPRED at PRED='0'

  v1.1 [2009/02/05]: details on parser added to README.TXT

  v1.0 [2009/01/19]: initial distribution of the trial, train and
    development data, valency lexicon and this README.TXT file.


(2) CONTENTS

* README.TXT
  this file

* CoNLL2009-ST-Czech-development.txt
  development data set for Czech, 5228 sentences

* CoNLL2009-ST-Czech-train.txt
  train data set for Czech, 38727 sentences

* CoNLL2009-ST-Czech-trial.txt
  trial data set (part of train data set) for Czech, 193 sentences

* CoNLL2009-ST-evaluation-Czech.txt
  Czech evaluation data. 4213 sentences

* CoNLL2009-ST-evaluation-Czech-ood.txt
  Czech out-of-domain data. 1184 sentences

* Czech.vallex
  Czech valency lexicon, 14979 entries. See below for details.



(3) ON THE DATA SET

The Czech corpus was converted from PDT 2.0
(http://ufal.mff.cuni.cz/pdt2.0/) following these rules:

- LEMMA contains only the word itself, no details nor explanation.

- Detailed part of speech is part of FEAT (SubPOS).

- HEAD is taken from the analytical layer of PDT 2.0.

- DEPREL is taken from afun attribute of the analytical layer of PDT 2.0.

- Every analytical node referenced by a lexical reference (a/lex.rf)
  from the tectogrammatical layer of PDT 2.0 has a PRED value
  filled. If the refering non-generated tectogrammatical node has a
  valency frame assigned, the value of PRED is the identifier of the
  frame. Otherwise, it is set to the lemma.

- For every tectogrammatical node, corresponding analytical node is
  found:

  1. If the tectogrammatical node is not generated and has a lexical
     reference, the referenced node is taken.

  2. Otherwise, if the tectogrammatical node has a coreference or
     complement reference to a node that has an analytical node
     assigned (by 1. or 2.), the assigned node is taken.

  APRED columns are filled with respect to this correspondence: for a
  tectogrammatical node P and its effective child C with functor F,
  the column for P's corresponding analytical node at the row for C's
  corresponding analytical node is filled with F. (Some nodes can thus
  have several functors in one APRED column, they are separated by a
  vertical bar).

- PLEMMA, PPOS and PFEAT were generated by cross-trained tagger MORCE
  (http://ufal.mff.cuni.cz/morce/) by Jan Raab.

- PHEAD and PDEPREL were generated by cross-trained MST parser
  [McDonald et al., 2005] (Chu-Liu alogorithm) by Drahomira Johanka
  Spoustova.

- The valency lexicon has four columns:
  1. lemma (can occur several times in the lexicon, with different
     frames)
  2. frame identifier
  3. list of space-separated actants and obligatory members of the
     frame
  4. example(s)

The out-of-domain data were created in the same way from a small part
of unfinished Prague Czech-English Dependency Treebank 2.0.


(4) ORGANIZATION

  Jan Hajic, Pavel Stranak, Jan Stepanek
  UFAL MFF UK
  Malostranske namesti 25
  11800 Praha
  Czech Republic
  Phone: +420-221 914 257
  Fax: +420-221 914 309
  E-mail: [lastname][at]ufal.mff.cuni.cz