[developers] Trigger rules with specialization
Woodley Packard
sweaglesw at sweaglesw.org
Sun Dec 11 17:20:46 CET 2011
Hi Francis,
When does the information that distinguishes the two classes of
adjectives become accessible? It sounds like you would like to
trigger "darou-v-cop-lex" when there is a na-adjective present, and
not trigger it when there is not a na-adjective present. If there is
no information distinguishing these two types of adjectives in the
MRS, I don't think my extension would help; it would allow the
specification of constraints on the *triggered* lexeme's AVM, not on
the AVMs of signs that are instantiated to match the context list
(which could be confusing, given the potentially many-to-many mapping
between EPs and signs). It could in principal be used to reduce the
situations in which the trigger rule matched (via failed unification
of those constraints), but I would expect the more useful effect to be
the presence of those additional constraints on that triggered lexeme
in the chart, reducing its rampant combinatorics.
Here's another example in English. When I parse "I think that the dog
barks." and generate from the resulting MRS, the complementizer
"that" (and also "like", "as if" and "as though") are triggered. They
are needed in the chart (obviously) so that they can combine with the
"the dog barks" clause, but there is no way to specify that they can
*only* combine with that clause. As a result, we get edges in the
generator chart that look like "that I think", "that I think the dog
barks", and "that I think that the dog barks" (and many other
combinations). These are wasted computation. With the proposal, the
trigger rule could skolemize the complementizer's
COMPS.FIRST.SYNSEM.LOCAL.CONT.HOOK directly (or something similar).
The combinatorics caused by this example in English are noticeable but
not too bad; I understand it is worse in the Wambaya and German
problem cases.
Woodley
On Dec 11, 2011, at 6:09 AM, Francis Bond wrote:
> G'day,
>
> we have a slightly related issue in Jacy where the semantics doesn't
> constrain things enough. In the most clear case there are two kinds
> of adjectives, one which inflects, and one which needs a separate
> inflecting word. There is nothing in the semantics that distinguishes
> between these two.
> We are currently over-inserting using regexps as in:
>
> darou-cop-lex_gr := arg0e_gr &
> [ CONTEXT [ RELS <! [ PRED "~_a_"] !> ],
> FLAGS.TRIGGER "darou-v-cop-lex" ].
>
> This has the unfortunate effect that it triggers on all adjectives,
> when we really only want one class of adjectives (na-adjectives).
>
> If I understand it correctly, your extension would allow us to trigger
> much more appropriately.
>
> --
> Francis Bond <http://www3.ntu.edu.sg/home/fcbond/>
> Division of Linguistics and Multilingual Studies
> Nanyang Technological University
More information about the developers
mailing list