What about the PRED features?

[Ron Kaplan]

At 01:13 PM 6/1/96 EST, Avery Andrews wrote:

>Another and possibly rather trivial example of this issue is
>PRED-features; many of the proposed semantic theories for LFG have
>the effect of leaving PRED features with no clear function; this is
>the case for the `informational spreading' architecture proposed by
>Chris Manning and I, so we took them out, but it also seems to be
>the case for the linear logic semantics, but they're still there.
>A small point perhaps, but it could generate embarassment if you were
>teaching the material and somebody asked why the PRED features were
>there, and you couldn't answer them (the practical effects of this could
>be bad, if it happened to me while teaching the LFG course at the
>upcoming ALI in July).

Here is how I would avoid embarassment:

When we were designing the LFG syntactic formalism, the semantic-form PRED
values were introduced as a compact, encapsulated encoding that included the
semantic information that we thought would have some interaction with
syntactic constraints.  We wanted to provide syntacticians a way of
expressing (and observing) the syntactic/semantic interactions without at
the same time requiring them to have or develop a complete theory of
semantic specification--an important but somewhat independent research
enterprise. The semantic forms thus serve as a relatively opaque token in
syntax for a separate full-fledged semantic theory.  
I think that that particular purpose can still be served by the semantic
forms and PRED values, even though our understanding of what a full-fledged
semantic theory might be has advanced since the late 70's and even if we are
willing to commit to one of the alternatives (e.g. the linear logic
approach) that have emerged in the last 20 years.  Much of the information
carried by a semantic form is indeed superfluous when the syntactic system
is in fact coupled with a full-fledged semantic theory.  Even so, it may
still be a convenient formal device to allow syntactic investigations to
proceed in appropriate semantic ignorance.

I don't think it is embarassing to present this motivation and also to
explain why some of the information in a semantic form overlaps the
constraints that a real semantic theory would provide.

A full-fledged semantic theory can be used to explain more precisely the
different kinds of information that a semantic form encapsulates.  Thus, a
semantic form carries 4 kinds of information:

1.  The assignment of grammatical functions (subcategorization)
2.  The mapping of grammatical functions to semantic arguments
3.  The individuation of semantic entities (by the instantiation property)
4.  The name of the semantic relation.

Items 2, 3, and 4 would clearly be covered by any real, elaborated semantic
theory.  Even with a simple semantic (sigma) projection formulated as an
attribute value matrix, we can explain what a semantic form is doing.
Consider the semantic form 
        (^ PRED)='kick<(^ SUBJ)(^ OBJ)>'
The mapping and relation information here is redundant with (using the Ascii
notation of the Xerox LFG workbench):
        (sigma::^ REL)=kick
        (sigma::^ ARG1)=sigma::(^ SUBJ)
        (sigma::^ ARG2)=sigma::(^ OBJ)
If these codescriptive statements were associated with the lexical entry,
there would of course be no need to also include the argument mapping and
relation separately in the semantic form.  We can assume also that the
semantics itself will worry about individuation, so that that property of
the semantic form can also be dispensed with.

What remains is the subcategorization property, the specification of the
governable functions that the predicate allows.  In some notes that I wrote
several years ago (and perhaps can put on our ftp server), I suggested that
this information could be specified in a different way, simply by adding the
set of Governable Grammatical Functions directly as a new formal property of
an f-structure.  The lexical entry for "kick" could assert, for example,
that GGF(^)={SUBJ, OBJ}, and this would drive the completeness and coherence
checking.  Having done this as well, there is nothing left for the PRED
value to convey.

Thus, with a reasonable semantic theory at hand, we can formally eliminate
the semantic form PRED values entirely if we also introduce the independent
GGF idea.  On the other hand, I can think of 2 reasons why we might still
want to decorate our f-structures with PRED semantic-forms:

(1)  The original motivation of making syntactic research and
representations somewhat independent of the details of a semantic theory
while still making crucial interactions apparent.  We would know that these
specifications would be replaced by their equivalent semantic formulations
when syntax is embedded in a more comprehensive theory of language.

(2)  Knowing what devices a particular semantic theory uses to characterize
items 2,3,4 above, we can retain the semantic forms in the lexicon and
grammar but treat them as an abbreviatory notation for the more detailed
specifications that the semantic theory requires.  Thus, in the example
above, we would think of the PRED semantic form as a short-hand for the 3
sigma equations.  Or, on the linear logic view, we would think of the
semantic form as short-hand for premises such as "if (^ SUBJ) means X and (^
OBJ) means Y, then ^ means kick(X,Y)".

Either way, it seems to me that the semantic forms might still serve as
useful entities in the representations we use to guide our purely syntactic
research--nothing to be embarassed about.

--Ron

Until June 30:  kaplan at nias.knaw.nl
Permanent:  kaplan at parc.xerox.com