"just-in-time" sub-grammar extraction

[Karel Oliva]

>>I am looking for relevant work, and ideas, regarding the following
>>practical problem in HPSG parsing:
>
>>Given a large HPSG grammar and a short piece of text (a sentence, a
>>passage, or even a document), efficiently extract a small HPSG sub-grammar
>>(out of the larger one) that gives the same parsing and partial parsing
>>(chart edges) results.  This would improve the parsing speed.
>
>
>I was looking at this problem with respect to parsing it with an
>ordinary (i.e procedural) chart parser, written in C++.  My
>conclusion is that you have to tag the text in a prepass, and call
>a subgrammar which accepts the surrounding tags.


Yes, to use a tagger with 100% recall (even if the precision is relatively
low -  even less than 50%) will speed up the actual parser, since this
helps to discard a lot of partial parses which, however, do not contribute
to a parse spanning the whole input. BUT this does NOT meet Vlado's
condition of providing for the same partial parsing(s) as the full grammar.

However, I wonder what is this condition good for. (???)

If you would be ready to drop it, or reduce it to stg. like "partial parsing
(chart edges) results which possibly participate in building the final
parse" - which seems quite reasonable, at least for all purposes apart from
teaching students chart parsing,
then another strategy would be possible, which I used with a considerable
gain in efficiency in middle- to large-scale parses of Bulgarian and Czech
some 8yrs ago, and namely:
         that you establish cross-links between lexical classes and rules,
with the aim that >>each rule is fired (that is, put into the subgrammar
for parsing the very sentence) only if its lexical trigger is present in
the sentence<<.

To give trivial examples:
1. if there is not a single coordinating conjunction in the input, then the
subgrammar will NOT contain any coordination rule
2. If there is no adjective in the input sentence, then one probably will
not need any rules for left-hand-side modifiers of nouns
3. If there is no preposition in the input, you might forget about
PP-building rules
4. If there is no preposition and no adverb in the input, you might forget
about all rules serving for modifying a V/VP


Well, I know, these examples ARE  oversimplified - but I hope they
illustrate well what I had in mind.

Hoping to have been of service

karel oliva