Re: GLR ambiguity

[Hans Aberg]

On 16 Jun 2007, at 08:15, Paul Hilfinger wrote:

> > On 14 Jun 2007, at 12:48, Alessandro Di Marco wrote:
> > I was trying to create a GLR grammar for natural languages
> > ...when I stuck on the
> > following s/r ambiguity.
>
> > text:
> >           /* empty */
> >         | text sentence
> >         ;
> >
> > sentence:
> >           WORD EOL
> >         | DOUBLEQ WORD EOL
> >         | DOUBLEQ WORD EOL DOUBLEQ
> >         ;

> GLR does not resolve grammar conflicts statically.  Bison will
> continue to report conflicts, and these reports really don't tell you
> much.  Since natural languages ARE ambiguous, what you must use GLR
> for is to gather the possible interpretations.  That is the purpose of
> %merge, which allows you, on encountering two different parses of the
> same phrase, to collect the interpretations (syntax trees, or
> whatever semantic values you are using) and return this collection
> (represented however you choose) as the value of the ambiguous
> construct.  %merge also allows you to reject some interpretations on
> context-sensitive grounds.  When I say "allows you" I don't mean that
> it provides specific facilities to do any of this, but rather that it
> gives a parser structure that allows YOU to write the necessary  
> actions.

We did not get to know how much of actual language process that was  
intended - perhaps the intent was only to filter out quotations?!

But it would sure be interesting if somebody took up the quest of  
doing a natural language grammar. There is, in fact, a Flex/Bison  
grammar for the constructed language lojban, which is made not not be  
ambiguous:
  http://www.lojban.org/tiki/tiki-index.php?page=Home+Page&bl
It might give inputs on how to do it for natural languages.

  Hans Aberg