Semantics and processing

[Bruce Mayo]

Avery Andrews writes (24 May 96)
>One thing that disturbs me about the `linear logic' approach in the
>DLPS paper and several others is that the idea of basing the
>theoretical architecture on monotonicity and a correspondence
>relation seems to be gone; . . .
>But if a powerful `glue language' is going to be used to assemble
>meanings, why not try to do more with it, such as assemble the
>c-structures as well as the meanings?  The potential existence of this
>kind of possiblity makes LFG + `Semantic Constructors' look like an
>inherently unstable theoretical combination, and difficult from a
>pedagogical point of view:

My understanding of these issues is superficial, at best, but I would like
to pose a related question that nags at me, actually more of an engineering
question but one that must eventually be addressed by any theory that makes
a claim to psychological plausibility (as LFG always has). The problem is
this: the real input to a speech or optical character recognition system is
inherently very ambiguous, both in the segmentation and in the
identification of the segments. Add to this that the syntax and semantics
of a real-world grammar are going to be ambiguous, giving many possible
parses and semantic analyses for each of the many possible interpretations
of the input, you get a computational nightmare. How do you keep the number
of possible (though mostly improbable) interpretations of a given utterance
to a managable number?

The approach that has been used in practical speech and optical text
recognition systems (some of which actually work in very limited domains)
is to divide the processing into several stages, each of which is
restricted to a domain whose principles of well-formedness can be described
independently of the other domains. Everything in each domain is
statistically weighted, so that within each domain you can do the
following: 1) multiply out all of the possible combinations (e.g., ways you
can recognize the sounds or shapes as known phonemes or letters, ways you
can recognize these as words, ways you can recognize the words as
sentences, etc.); 2) then sort the enormous heap of results within a domain
on the basis of the overall weightings, giving a rank-ordering of all the
possible analyses, 3) throw away all but the very best (say top 10 percent)
analyses. Doing this, you reduce the input to the next domain to an amount
that is again computationally tractable.

One thing that always appealed to me in the classical LFG architecture
(Bresnan 1982 and Halvorsen 1983) was that it seemed to offer a series of
well-formedness domains that could be independently computed: you can
compute lots of possible c-structures without worrying about anything else,
thow away most of those, use this small number number of best c-structures
to compute (possibly) lots of f-structures, thow away most, compute lots of
s(emantic)-structures, thow away most and get to a small number of
most-plausible results.

I don't think this is simply a computational issue - my bet is that people,
too, use their knowledge of well-formedness principles within various
linguistic modules or domains to sort out recognition ambiguities, and that
these domain-specific principles are part of the lingustic knowledge they
have to acquire in order to use their language.

But if you conflate semantics and f-structure (let alone c-structure!), you
lose the notion of what a well-formed f-structure by itself is (this is my
impression of the recent linear-logic proposals), and then you lose the
possibility of ranking, sorting and throwing away f-structures *before* you
go on to the next stage of processing. (I seem to remember hearing that in
linear logic semantics, for example, the independent testing of f-structure
for coherence and completeness is to be folded into the semantics, which
means you can't throw away all those ill-formed structures beforehand, and
sematics has to deal with them.) Or am I mistaken in this impression?


Bruce Mayo, FG Sprachwissenschaft
Universitaet Konstanz
Postfach 5560 D 177
D-78434 Konstanz
eMail:  bruce.mayo at popserver.uni-konstanz.de
Tel.:   (+49) 7531-88-3576