Re: A syntactical database of the NT

From: Daniel Ria–o (danielrr@mad.servicom.es)
Date: Tue Dec 08 1998 - 12:04:30 EST


        Here are a few preliminary thoughts about a syntactical database of
the NT, based in personal experience for the most part. Some of the things
I'll say are a matter of controversy while other may appear pretty obvious,
but because of the fact that there is some confusion even about the last
kind of questions, I beg your indulgence if I insist too much in such
matters.
        (This mail is OFFLIST. I am forwarding this mail to C. Conrad and
the people that seem to be most interested in this thread, but maybe it
could be better to write it to the whole list, and keep open this thread
using always the same "Subject" line).

        a) First of all it is necessary to distinguish between "parsing"
and "diagramming". "Parsing" does not require diagramming; diagramming does
require parsing. I can parse a sentence just indicating verbatim the
relations among the members of a sentence and the function of every
syntagm, without any diagramming. Therefore I will always distinguish
between "diagramming", which is a way of represent graphically a
syntactical parsing, and the "parsing" itself.
        On the other hand, "diagramming" can be understood in a restrictive
sense, to denote exclusively the representation schemes that use graphics
(lines, boxes, etc), and in a broader sense, to comprehend every way to
represent graphically (over paper, or a computer screen, etc) a syntactical
parsing (here graphically stands in opposition lo logically). Many people
wouldn't call "bracketing" a kind of "diagramming", but for our purposes,
we can call "diagramming" to all the possibilities to represent a parsing
as a graphic in two spatial dimensions.
        Once a sentence is parsed in a way which is both coherent and
machine-readable, it is very easy for a computer to produce different
representations of the same syntactical description (the parsing).
Therefore, the first aim of a project to develop a syntactical description
of a text corpus is to produce and employ a robust schema of syntactical
annotation. You can left the details of the syntactical representation for
a second phase.

        b) Syntactic functions are referred to syntagms, not to words [I
don't know the exact translation of Spanish "sintagma": it is a concept
very often used in linguistic in Spain, France, Germany etc, but I don't
know if there is an exact translation into English: It covers much (but not
overlapes with) the notion of Immediate constituent, or better, with the
linguistic substance that can be analysed in immediate constituents]. As
the nucleus of a syntagm, there will always be a word (or al least a lexia,
or special combinations, see the "Drew Barrymore" example infra), but a
given word can be the nucleus of many syntagms, each of them with its
function. Therefore, any robust parsing scheme must indicate the
function(s) of every syntagms and their nucleus.

        c) There is not a one to one relation between syntagms and
functions. Some syntagms can simultaneously play more than one function.
The clearest example in Greek are relatives and predicatives. It belongs to
the very definition of a relative the notion that it is a subordinative
element (function= subordinative, like a conjunction or a preposition) and
it plays a another syntactical role in the subordinate sentence (function =
subject, object, etc). A robust parsing scheme should be able to indicate
the different functions of such syntagms.

        Some starting points

        Here I would like to introduce the eight levels of syntactical
representation proposed by the ÇEAGLESÈ standard, that can be considered a
good starting point, but note that it omits any reference to text
transmission, and, in general, to any item considered not linguistic but
"philological").
        1. Bracketing of segments.
        2. Labelling of segments.
        3. Showing dependency relations.
        4. Indicating functional labels.
        5. Marking subclassification of syntactic segments.
        6. Deep or ÇlogicalÈ information.
        7. Information about the rank of a syntactic unit.
        8. Special syntactic characteristics of spoken language.

        Now allow me the pedantry to quote myself. I suggested in a paper
the list of "desiderata" of an annotated corpus (and to a large extent, to
the schema for the syntactical annotation employed):

        - NON-CLOSENESS: possibility of adding new annotations to the corpus.
        - CONSENSUALITY: deploy a scheme appreciated by experts in the field..
        - DIVISIBILITY: Codes must reflect the pertinence of a sub-category
to its category.
        - EXHAUSTIVITY of the analysis.
        - RECOVERABILITY: original text and all the linguistic data must be
recoverable with the pertinent software.
        - ELASTICITY: the same corpus should be able to be used by
different people in different studies.
        - SISTEMATICITY during the annotation process.
        - SIMPLICITY: the annotation scheme must facilitate the automatic
searches.
        - TRANSPARENCY: the meaning of the annotation tags should be
self-evident.
        - ELEGANCE:

        The problem with this decalogue of desiderata is that unlike other
decalogues, we must sacrifice some of the criteria of the list to obtain
some others. I'll need to refer in this and next postings to the points of
this decalogue.

        In spite of what I said at the beginning of this post, in most of
the existing annotated corpora of modern languages there is a deliberate
correspondence between the scheme of annotation and a particular scheme of
representation. That I consider a handicap of such models.
        To start with a simple model, here's an example of annotation from
ÇThe American Printing House for the BlindÈ (APHB), where the tagging
method corresponds to a particular kind of bracketing (It is always
possible, however, to use the data for another kind of diagramming).

B0100111 v
[N Mr._NNSB1 Utterson_NP1 [N the_AT lawyer_NN1 N]N][V was_VBDZ [N a_AT1
man_NN1 [P of_IO [N a_AT1 rugged_JJ countenance_NN1 ,_, [Fr that_CST [V
was_VBDZ never_RR lighted_VVN [P by_II [N a_AT1 smile_NN1 N]P]V]Fr]N]P]N]V]
;_; [[J cold_JJ ,_, scanty_JJ and_CC embarrassed_JJ [P in_II [N
discourse_NN1 N]P]J] ;_; [ backward_RR [P in_II [N sentiment_NN1 N]P]] ;_;
[J[ lean_JJ ] ,_, [ long_JJ ] ,_, [ dusty_JJ ] ,_, [ dreary_JJ ] ,_, and_CC
yet_RR [ somehow_RR lovable_JJ ]J]] ._.

        The advantages of such formalism are evident: it allows the
software to easily tag the words of the text. Here simplicity, transparency
y consensuality are close to the optimum, while other criteria, such as
exhaustivity and divisibility are neglected: if we want to use brackets to
indicate further relations between syntagms, we must use a more complicated
model, like the following examples from ÇThe Penn TreebankÈ do.

a. Chris knew yesterday that Terry would catch the ball
(S (NP-SBJ Chris)
   (VP knew
       (SBAR *ICH*-1)
       (NP-TMP yesterday)
       (SBAR-1 that
               (S (NP-SBJ Terry)
                  (VP would
                      (VP catch
                          (NP the ball)))))))

Everyone seems to dislike Drew Barrymore
      (S (NP-SBJ-3 Everyone)
         (VP seems
            (S (NP-SBJ *-3)
               (VP to
                  (VP dislike
                      (NP Drew Barrymore))))))

        I'd like to hear your commentaries about this very few but verbose
ideas, before I continue exposing my ideas about this matter and present
you my schema for syntactical annotation.

___________________________________________________________
Daniel Ria–o Rufilanchas
c. Santa Engracia 52, 7 dcha.
28010-Madrid, Espa–a
___________________________________________________________



This archive was generated by hypermail 2.1.4 : Sat Apr 20 2002 - 15:40:09 EDT