[matrix] Matrix Coordination Module

[Scott Drellishak]

Hi, all.  I'm working on a coordination module for the Matrix.  I'd like to 
describe my current plans and see if you have either any objections or any 
additional suggestions.

My analysis for coordination is based pretty heavily on the ERG's -- in 
particular, the coordination module will use binary branching structures to 
simulate N-ary branching coordination, and there will be three abstract 
rules from which all part-of-speech-specific rules derive:

 (1) head-marker, which handles the marking of a node as coordinated
 (2) mid-coord, which joins a node on the left with a node marked 
coordinated
     on the right, with the mother node marked coordinated.
 (3) top-coord, which joins a node on the left with a node marked 
coordinated
     on the right, with the mother node *not* marked coordinated.

So we'll have the following structures for the following examples (using 
X-top, X-mid, and X-head as abbreviations for the three rules above, and 
ignoring punctuation):

 (4) fire and steel

     [NP-top [NP fire] [NP-head [CONJ and] [NP steel]]]

 (5) beg, borrow, or steal

     [V-top [V beg] [V-mid [V borrow] [V-head [CONJ or] [V steal]]]]

The new feature SYNSEM.LOCAL.COORD will be a list containing the 
coordination relation specified down in the head-marker rule and all the 
collected handles and indices of the coordinands.  (This is called CONJ in 
the ERG, but since some languages mark coordination without separate lexical 
conjunctions, I'm avoiding using "CONJ" in the general case.)

By varying the definitions of top-coord and mid-coord, we can require 
different strategies for marking coordination.  If the left coordinand in 
both rules must have a non-empty COORD, then we require coordination marking 
(e.g. a conjunction) on each coordinand:

  (6) and A and B and C and D

If the left coordinand must have an empty COORD, then we require 
coordination marking only on the rightmost coordinand:

  (7) A B C and D

Various other patterns are also possible.

Similarly, by varying the head-marker rule, we can get preposed conjunctions 
(as in English), postposed conjunctions, or some kind of non-lexical marking 
(i.e. a suffix, a different verb form, etc).

As described so far, the top-coord rule isn't doing much work -- more or 
less everything could be handled by mid-coord.  However, I intend eventually 
to implement something equivalent to the ERG's LEFT feature, which is a 
two-item list that contains the CONJ values for complex conjunctions like 
"either...or", where "either" is applied by the top rule, and any number of 
"or"s are applied by repeated application of the mid-rule.  (As with CONJ 
above, I think I'm going to rename this feature, perhaps to COORD-STRAT.)

I'm in the middle of the first-pass implementation of this module, which 
will include the abstract rules and also phrase-type-specific rules for NP, 
VP, AP, PP, and probably some others.  Near the end of this month (May), a 
group of students in Emily Bender's Grammar Engineering course who are 
working on grammars of various languages will attempt to integrate my 
implementation of coordination into their grammars.  That should provide 
some welcome testing.

After making any changes based on what I learn from that testing, I'll 
figure out how my rules will be integrated into the module scripts.  That 
will involve structuring the questions asked (e.g. "Does your language 
coordination all phrases with the same coordination strategy?"), and also 
deciding on a file structure.  I'm currently leaning towards one file with 
all the abstract rules, and then one file for each specific factored rule 
(e.g. (a) VP-coordination (b) with morphological marking (c) on each 
coordinated verb, versus (a) NP-coordination (b) marked by conjunctions (c) 
with only one conjunction per coordinated list).

If you have any suggestions, please let me know.  Thanks in advance for your 
time!

Scott Drellishak
"The Coordination Coordinator"
University of Washington Linguistics
sfd at u.washington.edu