[developers] Trigger rules still being requested
Emily M. Bender
ebender at uw.edu
Thu Feb 18 01:53:07 CET 2016
Yeah, unfortunately at least one case does lead to semantically empty
things combining fairly freely with just about anything: the argument
Agreed that trigger rules aren't fun, but they can be useful, even in
I've sent Dan's patch along to the students, so we're at least not seeing
On Mon, Feb 15, 2016 at 11:20 AM, Ann Copestake <aac10 at cl.cam.ac.uk> wrote:
> It's an expanded version of the mrscomp grammar which I've just checked in
> to the LKB svn repo. I created it to mirror the dmrscomp grammar and just
> thought I'd check that it's still possible to avoid trigger rules since I
> wasn't sure whether the code still allowed it.
> In terms of whether or not you need trigger rules for a particular
> grammar, I would suggest just trying it. The additional edges will only
> explode the search space if the lexical items with null semantics can
> combine relatively freely with some of the semantically licensed edges. Of
> course this can happen even if you have trigger rules, if they can't be
> defined tightly. But I think of trigger rules as a particular form of
> nastiness that I would be happy not to inflict on grammar writers unless
> it's essential.
> oe has offered to fix the warning messages
> On 15/02/16 19:03, Emily M. Bender wrote:
> Hi Ann,
> What kind of small grammar though? There are a few things (argument
> auxiliaries, free word order) that can drastically expand the search space
> for the generator.
> I haven't tested this, but I wouldn't be surprised if even right out of
> the customization
> system box I could come up with a place where trigger rules are needed.
> But at any
> rate, I'd be happy if the warnings that did show up were legitimate ones
> (and will
> try applying Dan's patch).
> On Mon, Feb 15, 2016 at 12:57 PM, Ann Copestake <aac10 at cl.cam.ac.uk>
>> I just want to confirm that with a small grammar with a small number of
>> null semantics entries (five of them), trigger rules are not required to
>> generate with the LKB. So I think that no warnings about the `missing'
>> trigger rules should be produced by default with small grammars, though I
>> don't know what the criteria for switching on the warnings should be.
>> On 11/02/16 00:05, Ann Copestake wrote:
>> The comment about construction from corpora was really about big grammars
>> - i looked at the ERG trigger rules about a year ago, and was a bit
>> horrified ...
>> For small grammars, is it really hopelessly inefficient to just run in
>> the mode where all the empty semantics items are added to the chart? Does
>> it simply not work if you have no trigger rules? (I admit not having tried
>> this for years.)
>> However, if not, it seems to me that the extract from a corpus technique
>> could work with small grammars assuming you've got a suitable set of test
>> sentences before you start to generate. i.e., you'd parse those and
>> accumulate trigger rules automatically from the parsed set, repeating as
>> you expand the grammar. It doesn't matter if the trigger rules are very
>> liberal, after all.
>> On 10/02/2016 23:14, Emily M. Bender wrote:
>> The tdl output is helpful to the extent that it shows people what the
>> value of TRIGGER should
>> be, but that's about it. It's particularly unhelpful when the trigger
>> rules are actually there... Also,
>> learning trigger rules from parsed/treebanked data sounds interesting,
>> but won't help at all
>> in the case of starter grammars, which we *do* want to be able to
>> On Wed, Feb 10, 2016 at 3:00 PM, Ann Copestake < <aac10 at cam.ac.uk>
>> aac10 at cam.ac.uk> wrote:
>>> Hmm - does anyone find the TDL output for the trigger rules helpful?
>>> because, having looked at the code, I would prefer to remove that facility
>>> completely unless it has support, since it's hardwiring type and feature
>>> names into the code. i.e., I would rather go back to the original
>>> behaviour which meant that warning messages were output (if there were no
>>> appropriate existing trigger rules) without the TDL being suggested. I
>>> believe oe may feel the same way.
>>> The trigger rule mechanism is a mess, in fact. Learning trigger rules
>>> from parsed/treebanked data is feasible, but too big a project for me to
>>> take on.
>>> On 10/02/2016 18:47, Dan Flickinger wrote:
>>> Hi Emily -
>>> I ran into this behavior too when doing the SWB-oriented grammars for
>>> the grammar engineering course I taught with Tom last quarter. The culprit
>>> turns out to be the function index-lexicon(), which right at its end
>>> cheerfully emits those warning/advice messages for all semantically empty
>>> entries regardless of whether trigger rules have been defined. Since the
>>> same shortcoming is repeated in the separate function reindex-lexicon(), it
>>> proved to be easier for me to correct this by changing the function
>>> make-trigger-tdl() which they both call, so that the global
>>> mt::*transfer-triggers* is checked before emitting the warning for each
>>> semantically empty lexical entry. Here is the patched version of this
>>> function analogous to what I added to my grammars' user-fns.lsp file:
>>> (defun make-trigger-tdl (empty-semantics-lexical-entries)
>>> (let ((empty-no-trigger-entries
>>> (loop for id in empty-semantics-lexical-entries
>>> unless (or #+:mt
>>> (gethash id mt::*transfer-triggers*)
>>> collect id)))
>>> (loop for x in empty-no-trigger-entries
>>> (format t "~%~%~a_gr := generator_rule &
>>> [ CONTEXT.RELS <! [ PRED \"non_existing_rel\" ] !>,
>>> FLAGS.TRIGGER \"~a\" ]."
>>> (string-downcase x)
>>> (string-downcase x)))))
>>> I expect that Ann or Stephan could see a more elegant way to achieve
>>> this effect, but this patch might be a useful temporary expedient.
>>> *From:* developers-bounces at emmtee.net <developers-bounces at emmtee.net>
>>> <developers-bounces at emmtee.net> on behalf of Emily M. Bender
>>> <ebender at uw.edu><ebender at uw.edu> <ebender at uw.edu>
>>> *Sent:* Wednesday, February 10, 2016 7:43 AM
>>> *To:* developers
>>> *Subject:* [developers] Trigger rules still being requested
>>> Dear all,
>>> In the Matrix derived grammars, we have the following in lkb/script:
>>> ;;; SSH 2012-03-27 This invokes trigger rules for semantically empty
>>> ;;; lexical entries. The trigger rules themselves should be in
>>> ;;; trigger.mtr in the grammar directory. The documentation on trigger
>>> ;;; rules: <http://moin.delph-in.net/LkbGeneration>
>>> ;;; SSH 2013-05-23 The semi.vpm file needs to be loaded *before* the
>>> ;;; trigger.mtr file in lkb/script for correct interaction.
>>> (lkb-pathname (parent-directory) "trigger.mtr"))
>>> "Generation trigger rules"
>>> :filter nil :task :trigger :recurse nil :edges 200 :subsume nil)
>>> ;;; Matrix-derived grammars often have stable enough semantic
>>> ;;; representations that the grammar can usefully be tested
>>> ;;; by generating from them. The function index-for-generator()
>>> ;;; must be called for generation to be enabled. In grammars with
>>> ;;; small lexica, this takes a trivial amount of time. If
>>> ;;; generation is not being used regularly, and the indexing
>>> ;;; is taking too long, comment out the following. index-for-generator()
>>> ;;; can also be run from the expanded LKB top menu, under
>>> ;;; Generate.
>>> It is my understanding that (index-for-generator) is what is behind the
>>> suggesting trigger rules for semantically empty lexical entries, but
>>> even though
>>> we call it after loading the trigger rules, we still get messages such
>>> as the following:
>>> Is there something else we're supposed to do so that we only see those
>>> when an entry truly lacks a trigger rule?
>>> 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>
> 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...
More information about the developers