Re: OT: THEORY Fusion Grammar

--- Patrick Littell <puchitao@...> wrote:

> On 7/14/06, Gary Shannon <fiziwig@...> wrote:>
> > Hypothesis: For any natural language, related
> elements
> > are always immediately adjacent and there exists a
> > complete fusion grammar for that language.
> >
> > Comments? Counterexamples?
> >
>
> To the extent that a language lacks constituency, it
> will be a
> counterexample; the extreme examples of
> nonconfigurationality
> (Warlpiri, etc.) would lead to nearly insurmountable
> problems.  (In
> Warlpiri you could, for example, put "dog" and the
> beginning and "old"
> at the end and "ugly" in the middle.)
Interesting. I'd love to take a look at some sample
sentences in Warlpiri. What is it that links these
distant elements syntactically or semantically?

> About "Fusion Grammar" in general: Aside from the
> issues above, some
> modification may be needed to prevent circularity.
> (Or I'm just
> misreading things, in which case I apologize in
> advance.)  In order to
> parse the sentence above, it appears you already
> have to *know* its
> meaning.  Take "Mary's ugly dog".  You have to know
> "ugly" is related
> to "dog" (rather than Mary) in order to "fuse" them.
<snip>

In this case (where Mary's might be a possesive or it
might be a contraction of "Mary is") both cases are
developed in parallel. The 's would be expanded as
both "Mary is" and "Mary OWNEROF" (where "OWNEROF" is
a sort of internal possesive particle).

Then both phrases would be parsed, and ultimately, by
the time the whole sentence has been parsed from both
starting points, one of the parses will not be capable
of completion while the other will. In case the
sentence is inherently ambiguous then both parses will
be completed and the ambiguity will be revealed by the
existence of more than one valid parse.

At the intermediate stages of parsing there may well
be dozens of copies of the sentence being worked on
concurrently until one of them is revealed to be
completable at the end, at which time the others are
discarded.

Thus: Mary's ugly dog barked.

1a: Mary (is ugly) dog barked.
1b: (Mary OWNEROF) (ugly dog) barked.

2a: (Mary (is ugly)).SV (dog barked).SV
2b: ((Mary OWNEROF)(ugly dog)) barked.

3a: No SV+SV rule. Cannot be completed.
3b: (((Mary OWNEROF)(ugly dog)) barked).SV

3b has an SV fusion as the last step while 3a has no
applicable fusion rule and is discarded, thus
resolving the ambiguity. Since the virtual element
"START" that occurs invisibly at the beginning of each
sentence has a rule for fusing with type SV to form a
sentence, (simultaneously consuming the last element)
the parse is complete.

Note also that "tags" might also encode semantic
information. Rather than just being tagged as a noun,
for example, something might be tagged as a "liquid
(or pourable sunstance like flour, sugar, etc.)", or
as an "animate being". Thus the tags could be more
specific than simple parts of speech, and verbs that
apply to liquids (pour, drink, spill) might have rules
that fuse them with nouns specifically capable of the
action described by the verb. Thus "Who is the
president of the United States." could never be
mistaken for the assertion that the rock band "Who"
holds that office.

It also occurs to me that fusion rules might include
specific words rather than just tags. The word "the"
for example, might be included in a rule as the
literal word rather than some tag type. I'm not sure
if that's useful or not yet. I'll need to get a little
further along in formalizing the rules before I can
tell.

--gary

>
> Roughly, though, the way you're building up the
> phrase structure isn't
> so different from the sort of Bare Phrase Structure
> that's popular in
> the Minimalist program.
>
> -- Pat
>

Reply