[developers] Newer 64-bit builds cause Non-structure argument error

[Woodley Packard]

I ran into an MRS with that type of LNK information a few weeks ago and ended up making ACE a bit less picky about what's between the < and >.  There are about 4 formats discussed in a comment in the relevant LKB code, including the empty '<>' discussed in this thread, but Stephan's reply seemed to me to suggest that the '<>' is a bug rather than a feature.  For the sake of future clarity, could someone please enumerate what formats are actually considered valid?  The wiki "MrsRfc" page only gives characterization, e.g. '<3:10>', and if there are other acceptable formats it should be extended.

Woodley

On Aug 12, 2012, at 3:05 AM, Stephan Oepen wrote:

> hi again, mike,
> 
> i just checked in new (32- and 64-bit) LOGON binaries.
> 
> the MRS comparison problem you reported was caused
> by bogus LNK information (`<>', no content) on predicate
> symbols.  it seems this was introduced by a code change
> (from february 2012, though the first builds that included
> it probably date to mid-may) i had made.  i have changed
> the LNK creation to avoid the bogus outputs; i also made
> LNK reading more robust, and further wrapped the MRS
> comparison calls from [incr tsdb()], to catch conditions
> thrown from the [incr tsdb()]-external code.
> 
> could you please see whether the latest builds work for
> your purposes?  i am about to commit a revision to the
> [incr tsdb()] database schema (to better support forest-
> based treebanking), and it would be nice to know there
> is a functional revision in SVN before releasing a not
> fully backwards-compatible change ...
> 
> best, oe
> 
> 
> On Sat, Aug 11, 2012 at 6:07 AM, Michael Wayne Goodman
> <goodmami at u.washington.edu> wrote:
>> On Sat, Aug 11, 2012 at 3:39 AM, Stephan Oepen <oe at ifi.uio.no> wrote:
>>> it looks as if you parsed using the LKB for this run?
>> 
>> Yes, our regression testing framework still parses with the LKB.
>> 
>>> the MRSs in current/tiniest are not well-formed, due
>>> to bogus characterization, e.g. "_dog_n_rel"<>.
>> 
>> Ah yes, I see. I also confirm that I get these empty characterizations
>> when parsing interactively in the LKB. I do not get this problem when
>> parsing with cheap.
>> 
>>> i can't quite guess how you arrive at this, but would
>>> a (tiniest and self-contained, if possible) grammar to
>>> debug this further.  do you have a suitable grammar
>>> (where, i guess, just browsing the MRS in the LKB
>>> should reveal the above problem on your side too)?
>> 
>> Sure. I attach the "tiniest" grammar. Note that "tiniest" is not the
>> only grammar showing the problem (indeed, I think all are), but I use
>> it for debugging because of its small size. I thought that perhaps our
>> REPP was malformed, since we started using one for tokenization
>> recently, but that doesn't explain why different versions of the LOGON
>> tree lead to different behavior. Anyway, I hope you have some better
>> ideas than I do.
>> 
>> Note that tiniest is not English. You can use the following sentences
>> to test parsing:
>> 
>> dog slept
>> dog cat chased
>> 
>>> cheers, oe
>>> 
>>> 
>>> On Tue, Aug 7, 2012 at 11:20 AM, Michael Wayne Goodman
>>> <goodmami at u.washington.edu> wrote:
>>>> On Tue, Aug 7, 2012 at 5:08 PM, Stephan Oepen <oe at ifi.uio.no> wrote:
>>>>> i would
>>>>> like to think you should then also see this issue in the
>>>>> interactive environment?
>>>> 
>>>> You are correct; I now see the error when using the podium. I didn't
>>>> think to turn on the MRS field.
>>>> 
>>>>> if so, could you just tar(1) up the two profiles that you
>>>>> compare in detail and send them to me?
>>>> 
>>>> Certainly. I attach the gold and parsed profiles of the "tiniest" grammar.
>>>> 
>>>> Many thanks,
>>>> 
>>>> --
>>>> -Michael Wayne Goodman
>> 
>> 
>> 
>> --
>> -Michael Wayne Goodman