[erg] Puzzle about IND and DIV

[Dan Flickinger]

Hi David -

Sorry for the long delay in replying.  The re-entrancy of LOCAL.AGR and LOCAL.CONT.HOOK.INDEX for ordinary nouns is asserted in the type `noun_synsem' in lextypes.tdl, and this means that the two DIV values you observed will also have to unify (or fail to do so).  

Your strategy of simply removing the DIV constraint on the normal bare-NP rule has the potential disadvantage that you can no longer tell whether a given input is an ordinary well-formed sentence, or instead slightly ill-formed and dependent on the constraint relaxation you introduced.  This may not seem to matter, but I think it may cause trouble with parse selection, where that bare-NP treatment of a singular count noun will appear as a candidate when you don't want it.  You might consider making a copy of the rule that is just for singular count nouns, so the parse selection machinery can tell it apart from the normal one, and adopt different likelihoods for each.  If you adopt this strategy, you may find of interest the additional mal-rules that are defined in the erg/educ subdirectory, and you can compare your rule with the 'mal_bare_np_sg_phrase' defined in erg/educ/lextypes-epgy.tdl.

 Dan

----- Original Message -----
From: "David Mott" <mottdh at googlemail.com>
To: erg at delph-in.net
Sent: Wednesday, October 9, 2013 9:41:39 AM
Subject: [erg] Puzzle about IND and DIV

BACKGROUND

I am writing a "mal rule" that will allow sentences like "male called a
woman". (IE this is an ungrammatical version of "a male called a woman").
The standard ERG (1111) will not parse that because male does not act as a
mass noun, and so will not allow creation of a bare np.

I have decided that my desired parse will be the application of hdn_bnp_c
to n_sg_ilr to male_n1. This fails at the composition with hdn_bnp_c. After
experimenting with the definition of hdn_bnp_c, I have found that the
constraint:

SYNSEM.LOCAL.AGR.DIV +

seems to be the problem. If I construct a mal-rule version of hdn_bnp_c
without this constraint, then the sentences parse fine, eg:

derivation[1] (0):male called a woman

 (805 sb-hd_mc_c 0 0 4 [root_informal]
   (787 mal_hdn_bnp_c 0 0 1
     (293 n_sg_ilr 0 0 1
       (22 male_n1/n_-_c-nocnh_le 0 0 1 []
         (20 "male" 0 0 1 <0:4>))))
   (578 hd-cmp_u_c 0 1 4
     (228 v_pst_olr 0 1 2
       (62 call_v1/v_np*_le -0.2128 1 2 [v_pst_olr]
         (17 "called" 0 1 2 <5:11>)))
     (501 sp-hd_n_c 0 2 4
       (181 a_det/d_-_sg-nmd_le -0.4307 2 3 []
         (18 "a" 0 2 3 <12:13>))
       (212 n_sg_ilr 0 3 4
         (201 woman1/n_-_c_le 0 3 4 []
           (19 "woman" 0 3 4 <14:19>))))))

So I have achieved my mal-rule. But....

THE PUZZLE

I want to understand exactly why this constraint causes the problem.
Intuitively this seems right, in that "male" is not divisible, so I would
expect a clash somewhere between + and - on "INDEX.DIV". It also makes
sense that n_sg_ilr would create DIV to be -, which it does.

BUT I have traced the TFS for male_n1, n_sg_ilr and hdn_bmp_c and I cannot
see the linkage. I have found on hdn_bmp_c:

SYNSEM.LOCAL.AGR.DIV +

and I have also found on hdn_bmp_c (after composition with n_sg_ilr):

C-CONT.HOOK.INDEX.DIV -
SYNSEM. --SIND.DIV -

but no link between the ARG.DIV structure and anything else at all. So how
could the clash between + and - on DIV occur?

Maybe my analysis is wrong? Maybe there is a special out-of-band (not in
the TFS system) linkage between AGR and other stuff? Maybe there is a
completely different explanation?

I would be grateful for any thoughts on this.

BTW I am interested to find out why I have failed to see the
inconsistencies in the constraints, rather than other ways to do what I am
trying to do. I am happy with my mal-rule, its just I don't understand why!

best wishes

David Mott