This is the readme file for the Turkish part of the CONLL-X Shared Task.

Version: $Id: README,v 1.2 2006/01/08 02:24:04 sabine Exp $


Note: this file contains Unicode characters


1. Preamble

    1.1 Source

        METU-Sabanci Turkish Treebank, in the corrected version
        received from Gülşen Eryiğit on 29 December 2005.

	Note: METU is the English abbreviation for the Middle East
	Technical University. ODTÜ is the Turkish abbreviation, so in
	Turkish, the Treebank is called ODTÜ-Sabancı.

    1.2 License

        See metu_sabanci_license.html


2. Documentation

    2.1 Data format

        Data adheres to the following rules:

        * Data files contain one or more sentences separated by a
          blank line.

        * A sentence consists of one or tokens, each one starting on a
          new line.

        * A token consists of ten fields described in the list
          below. Fields are separated by one tab.

        * All data files will contains these ten fields, although only
          the ID, FORM, CPOSTAG, POSTAG, HEAD and DEPREL columns are
          guaranteed to contain non-underscore values for all
          languages.

        * Data files are UTF-8 encoded (Unicode).

        Field 1: ID         

          Token counter, starting at 1 for each new sentence.

        Field 2: FORM

          Word form or punctuation mark.

          In the CoNLL-X data, the '_' in the FORM column are for
          non-last IGs in a word, see below, section 2.3 Conversion

          !!! Tokens with FORM '_' WILL NOT BE SCORED UPON !!!

        Field 3: LEMMA         

          The lemma of the FORM. Non-first IGs in a word have '_' as
          the lemma.

        Field 4: CPOSTAG 

          Coarse-grained part-of-speech tag. See the Appendix of
          ttbankkl.pdf for a full list ("Major Parts of Speech").

        Field 5: POSTAG         

          Fine-grained part-of-speech tag. See the Appendix of
          ttbankkl.pdf for a full list ("Minor Parts of Speech"). In
          most cases, the "Minor Part of Speech" fully determines the
          "Major Part of Speech". In those cases, POSTAG simply
          contains the "Minor Part of Speech". However, the "Minor
          Parts of Speech" 'PastPart' and 'FutPart' occur with nouns
          as well as adjectives. We have therefore prefixed them (and
          'Inf' and 'PresPart' for consistency as well) with 'N' and
          'A' respectively.
	  If no "Minor Part of Speech" is given in the treebank,
          POSTAG is identical to CPOSTAG.

        Field 6: FEATS         

          List of set-valued syntactic and/or morphological
          features. See the Appendix of ttbankkl.pdf for a full list
          with explanations.

        Fields 7: HEAD         

          Head of current token, which is either a value of ID or zero ('0').
          A value of zero means the token attaches to the virtual root node.
          The dependency structure resulting from the HEAD information can be
          non-projective. 
        
        Field 8: DEPREL         

          Dependency relation to the HEAD. As the labels are in
          English, they should be relatively obvious. In addition, see
          user_guide.pdf (section 2.1 YAPISAL ROLLER) for example
          sentences in graphical form.
          During the conversion, we introduced three additional
          DEPREL's: 'ROOT' for a token with HEAD=0 (unless a different
          DEPREL is available from the treebank), 'PUNC' for
          punctuation and 'DERIV' for non-last inflectional groups
          (IGs), see below, section 2.3 Conversion
         
        Field 9: PHEAD         

          Projective head of current token, which is always an
          underscore because it is not available from the Turkish
          treebank.

        Field 10: PDEPREL 

          Dependency relation to projective head, which is always an
          underscore, because it is not available from the Turkish
          treebank.

    2.2 Text

        The texts in the METU-Sabanci Turkish Treebank come from the
        balanced METU-Sabanci Corpus. They are extracts of novels,
        news, etc. See user_guide.pdf (in Turkish) for an overview of
        proportions of the genres. 

    2.3 Conversion

        The Turkish treebank annotates dependencies not beween words
        (which are quite complex in Turkish because it's an
        agglutinative language) but between inflectional groups (IGs),
        see ttbankkl.pdf. Each word consists of one or more inflection
        groups. An inflection group consists of either a stem or a
        derivational suffix, and all the inflectional suffixes
        belonging to that stem/derivational suffix. 

        It's a bit like saying that in "the usefulness of X for Y",
        "for Y" links to "use-" and not to "usefulness". Only that in
        Turkish, "use", "full" and "ness" each could have their own
        inflectional suffixes attached to them.

        In the CoNLL-X data, the '_' in the FORM column are for
        non-last IGs in a word. Those IGs will always link to the next
        IG and have deprel 'DERIV', so they should be very easy for
        machine learners. 

        !!! Tokens with FORM '_' WILL NOT BE SCORED UPON !!!

        Ideally, derivational suffixes should have a form and a stem
        as well (unless it's zero derivation) but that information is
        not yet available in the treebank. It might be in future
        versions.

        Most punctuation is unattached in the original
        treebank (or attached to the virtual root node in an
        alternative interpretation). However, some punctuation
        characters are attached and have children. These mainly occur
        with coordination and indirect speech. In addition, the main
        verb is a child of the end-of-sentence punctuation.

	For CoNLL-X, the dependency structure was changed to so that
	punctuation does not have any more children. This is feasible
	as normally punctuation should have only one child, or at
	least one special child with the function 'SENTENCE'. Other
	children are then attached to the special child, while the
	only or special child is attached to its original grandparent
	(i.e. punctuation's head). If there are several children and
	not a single special one, all are attached to their original
	grandparent.

        Unattached punctuation is attached as "high" as possible
	without creating a new crossing dependency or attaching it to
	the virtual root node.

    3. Acknowledgements
        
       Bilge Say and Kemal Oflazer for granting the license for
       CoNLL-X and, especially Bilge, for answering questions. I hope
       you get well soon!

       Gülşen (Cebiroğlu) Eryiğit for making lots of corrections to
       the treebank and discussing some aspects of the conversion.

