[Corpora-List] algo for semantic structure

[Linas Vepstas]

2008/9/7 Dan Garrette <dhgarrette at gmail.com>:
> Vrone,
>
> The most common way to turn syntax into semantics is by defining logical
> expressions for each part of speech in a sentence and then composing those
> parts to make the meaning of the whole.  For instance, to find the meaning
> of the sentence "John sees a man" we can start by assigning a logical
> meaning of each word in the sentence:

Surely this is the "least common" way of doing things, as
it completely ignores both traditional work on parsing, as
well as being uninformed by any sort of corpus statistics.
One usually finishes by assigning meaning, not starts.

The most "common" way might be to employ a number
nested, layered techniques, from part of speech taggers
and morphology stemmers, to various phrase structure
or dependency grammars to obtain relations, for example

  subj(John, sees)    # who is seeing?
  obj (a man, sees)  # what are they seeing?

Meaning is provided by both the structure of the sentence,
plus prior knowledge that the word "John" might
be a man, or might be a toilet, and "see" might be "view"
or it might be "accept visitors", so that "John sees a man"
might be a funny way of saying "the toilet is accepting
visitors".

Importantly, one can make rather good progress
by abandoning syntactic structure completely; see
for example Radu Mihalcea's work on word-sense
disambiguation, which, in a nutshell, solves a Markov
chain on word senses.  There's not a drop of grammar
or parsing in it (or first-order logic either).  Its a solid
result which should cause anyone working on
this-n-such theory of grammar to stop, pull their head
out of the sand, and re-evaluate their strategy.

The work described in  http://nltk.org/doc/en/ch11.html
is curious, but I'd think a much stronger approach would be
to assign probabilities to each statement of first-order
logic (and thus obtain, for example, a "Markov logic
network", or a generalization, the PLN) Such probabilities
would be computed from corpus analysis.

I agree that "sense" can be considered to be the set of
predicates and rules that were triggered during a parse.
But the examples at that url also seems to make use of
somewhat olde-fashioned ideas like NP and VP, when
there in fact seem to be a much much richer and broader
set of relationships between words, phrases, colligations, etc.
e.g. the hundreds of dependency links and thousands of
rules in link-grammar, to the thousands of framenet-style
frames.  I just don't see that a first-order logic will ever
successfully capture this -- I'd say that Cyc illustrates what
the limit of that technique is.

--linas

_______________________________________________
Corpora mailing list
Corpora at uib.no
http://mailman.uib.no/listinfo/corpora