[developers] overgenerating pronouns

Michael Wayne Goodman goodmami at u.washington.edu
Wed Jan 7 06:52:34 CET 2009


Francis and I, in our continued quest to improve Jacy's generation,
are seeing issues with overgeneration of pronouns. For example, we
parse a sentence with a 2nd-person pronoun (such as あなた), confirm that
the relation for said pronoun has "PERS: 2" in the MRS, then are taken
by surprise when we see 1st and 3rd pronouns (eg. 私 or 彼) coming out
in generated sentences.

We have tried the following lkb variable settings:

(setf mrs::*mrs-results-check* t)
(setf mrs::*mrs-results-check* nil)

(setf *bypass-equality-check* nil)
(setf *bypass-equality-check* :filter)

...but no combination of the above prevented the incorrect pronouns
from coming through. We thought that at least one of those options
would ensure that the MRS going into and coming out of generation are
compatible. Did we miss something?

Just for further clarity, here are some sample MRS bits for the pronoun rels:

input (あなた):
[ pron_rel<0:3>
  LBL: h3
  ARG0: x4 [ x PERS:2 ]]

output (私達):
[ pron_rel<@5652>
  LBL: h3
  ARG0: x4 [ x PERS: 1 NUM: PL ] ]

As always, a preemptive thanks!
-Michael Wayne Goodman

More information about the developers mailing list