[developers] Fwd: stack overflow on generation

[Emily M. Bender]

Thanks, Woodley.  That is quite helpful!

On Tue, Dec 9, 2014 at 1:06 PM, Woodley Packard <sweaglesw at sweaglesw.org>
wrote:

> Hi Emily and Antariksh,
>
> For ACE, the primary problem was that the trigger rules in trigger.mtr did
> not have unique names, so "kii" was not getting added to the chart.  ACE
> was printing some warning messages when compiling the grammar about them
> being redefined.  When I give them unique names (crucially the "kii"
> rules), I am able to generate from this sentence.  I get 6 results:
>
> Ravi vicaar-at kii tsor zhop-t-o kaa
> Ravi vicaar-t-o kii tsor zhop-t-o
> Ravi vicaar-t-o kii tsor zhop-t-o kaa
> Ravi vicaar-t-o asa kii tsor zhop-t-o
> Ravi vicaar-at kii tsor zhop-t-o
> Ravi vicaar-t-o asa kii tsor zhop-t-o kaa
>
> However, ACE still isn't very happy.  It says the "head-comp" rule is
> loopy, i.e. can apply to its own input somehow (it doesn't apply during
> parsing because it takes 2 inputs and you run out of words eventually, but
> in generation the vacuous "kii" can keep being added to a sentence
> apparently).  I don't think that necessarily implies it's specifically that
> rule that's got a bug in it, but something in the chain of applications
> that leads to that cycle is buggy.  If you run the generator with packing
> turned off (-p), then ACE fails to notice the loop and spins for a while
> before failing.  If you also add a couple of -v's (so -epvv), then you'll
> see the strings it's generating:
>
> kii kii kii kii kii kii zhop-t-o
> vicaar-t-o asa kii kii kii kii kii zhop-t-o
> kii kii kii kii vicaar-t-o kii kii kii kii kii zhop-t-o
>
> etc etc.  I assume some of these are ungrammatical :-).  You should also
> be able to see rule names and trace back what the bad chain was with a
> little patience.  Or maybe you can already guess at this point :-)
>
> I would guess that this cyclic application is also what's throwing the LKB
> off (and maybe the LKB is more relaxed about multiple trigger rules having
> the same name), though I don't know for sure.  Hope that helps,
> -Woodley
>
> On Dec 9, 2014, at 12:47 PM, "Emily M. Bender" <ebender at uw.edu> wrote:
>
> Dear all,
>
> Antariksh and I have hit a generation debugging mystery, that I'm hoping
> for some help with.  Our grammar (of Marathi) can generate for some strings
> (with both ace and lkb) but for this particular item:
>
> *ravi vicaar-t-o kii tsor zhop-t-o kaa*
> (Ravi asks if the thief sleeps)
>
> it gives a "Stack overflow (signal 1000)" error on the lkb and just no
> results with ace.  Turning packing off doesn't affect things.  I played
> around a bit and found that if I set *maximum-number-of-edges* to 717, I
> run out of edges before the stack overflow, but with 718 or more it's back
> to that. (I tried diffing the output of (pprint *gen-chart*) in those two
> conditions, but that wasn't informative ... too many spurious differences.)
> batch-check-morphology() ran happily for a while and then returned nil,
> which I assume means no problems.  Setting *gen-packing-p* to nil doesn't
> change the behavior.
>
> This is clearly a problem with the grammar, but I'm at a loss as to how to
> localize it.  Are there any other tools or strategies I should be aware of?
>
> Thanks,
> Emily
>
> In case it's helpful, the grammar can be found here:
> http://faculty.washington.edu/ebender/mar.tgz
>
> ace output:
>
> ubuntu at UbuntuLKB:~/Downloads/mar$ ace -g mar.dat -Tf1 | ace -g mar.dat -e
> ravi vicaar-t-o kii tsor zhop-t-o kaa
> NOTE: 1 readings, added 653 / 401 edges to chart (76 fully instantiated,
> 84 actives used, 49 passives used)    RAM: 3766k
> NOTE: 122 passive, 620 active edges in final generation chart; built 305
> passives total. [0 results]
>
>
> --
> Emily M. Bender
> Professor, Department of Linguistics
> Check out CLMS on facebook! http://www.facebook.com/uwclma
>
>
>


-- 
Emily M. Bender
Professor, Department of Linguistics
Check out CLMS on facebook! http://www.facebook.com/uwclma
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.delph-in.net/archives/developers/attachments/20141209/80e6963b/attachment-0001.html>