[developers] LOGON: error `input fix-up ambiguity: 2 outputs'

[Stephan Oepen]

yes, exactly: the ‘fixup’ mechanism in the generator, i.e. the optional
transfer grammar invoked on generator inputs prior to lexical lookup,
is defined to be ambiguity-free, i.e. must not contain optional rules.

this step is not really intended for transfer, but rather to make minor
adjustments to one generator input, resulting in one generator input
that has a higher chance of success, for example making sure that
all instance variables are quantified (if a grammar requires that) or
that labels are linked up right between a quantifier and its entity.

in case you need a full-fledged transfer step, i think that should be
separated from generation, such that it can fan-out appropriately.

i realize the explanation offered by woodley (and seconded by me
now) seems at odds with your observation that it does not matter
whether or not the rule in question is activated.  but maybe there
are other optional rules in the ‘fixup’ transfer grammar?

best, oe

On Sat, Mar 1, 2014 at 6:27 AM, Woodley Packard <sweaglesw at sweaglesw.org> wrote:
> Hi Emily,
>
> The message means that the translation rules (acm.mtr) produce multiple
> possible outputs.  This is because the pro-drop transfer rule is marked as
> optional (type "monotonic_omtr" as opposed to "monotonic_mtr").  I'm not
> sure if that was intentional or?  I tried translating eng2eng with ACE and
> got a related message (I assume you're using the LKB for translation - ACE
> issues the warning but proceeds with generation with just the first transfer
> result).
>
> In general it is legal to have multiple transfer results, but when transfer
> is done as part of the "fixup" stage, there is no opportunity for
> disambiguation, so the LKB and ACE both (apparently) assume there will only
> be one.  If you want the ambiguity (e.g. if you want to try generating both
> ways), perhaps explicitly separating the transfer step from the generation
> step is called for.
>
> You say that the error appears whether or not the pro-drop rule is enabled
> as part of the grammar or not.  That is not the case on my end (admittedly a
> different setup -- eng2eng instead of frr2eng, and ACE instead of LKB), so
> I'm not sure what to suggest there.
>
> Good luck,
> -Woodley
>
> On Feb 28, 2014, at 7:54 PM, "Emily M. Bender" <ebender at uw.edu> wrote:
>
> Hi,
>
> I'm trying to get the LOGON MT set up ready for my students to work with,
> and I've hit an error message that's new to me (in the message subject).
> This comes about when I try to translate from Frisian (frr) to English (eng)
> for any sentence with a pronoun in it.  Sentences without pronouns work
> fine, and eng2frr works, even with pronouns.
>
> Perhaps relatedly, the pro-drop rule I've defined (for other output
> languages) doesn't fire for frr inputs though it does for eng inputs, though
> I've been unable to spot the difference in the (relevant portions of) the
> MRSs.   (And the `input fix-up ambiguity' error appears whether or not that
> rule is enabled as part of the grammar.)
>
> I'm attaching the grammars in case anyone is curious, but what I'm really
> looking for is any leads on what the error indicates.
>
> I suspect the difference has to do with the fact that the frr grammar was
> built last year, while eng was (re)built this year, so the Matrix versions
> are not the same.  But I haven't yet been able to pinpoint a relevant
> difference.
>
> (This is all with Ubuntu+LKB 17, 64-bit version.)
>
> Thanks for any leads,
> Emily
>
> --
> Emily M. Bender
> Associate Professor
> Department of Linguistics
> Check out CLMS on facebook! http://www.facebook.com/uwclma
> <eng.tgz><frr.tgz>
>
>