Computatioal perspective on gfs

[Bruce Mayo]

Responding to the discussion of LFG as a formalism for languages without
grammatical functions:

I'd like to throw in a couple of remarks from the perspective of a
computational implementor on what I think is an often-overlooked side of
the debate.

When you look at the formal organization of LFG -- I mean the factorization
into c-, f-, s- structure etc. -- there is nothing that compells a grammar
writer to declare a set of grammatical functions different from the set of
thematic roles (as J.B. and others have already noted); all that is
necessary are explicit mappings from constituent to syntactic to semantic
structure, which in limiting cases can be trivial, and thus apparently
redundant. But without these separate structures and mappings, you don't
have LFG. So the real question is why would one want to insist on having
f-structure mappings in the descriptions of languages that map one-to-one
from syntactic to semantic arguments (if such languages exist).

Here's a computational argument for having an apparently redundant
f-structure even when, in purely descriptive terms, you don't need it (when
f-structure mapping adds no additional information to the description).

It is possible to write an LFG-grammar with the single, trivial c-structure
rule S -> X+ ^=v, which says that a sentence is just an unstructured string
of unifying words. This appears empty, but it formally acknowledges that
there has to be a distinct process that is looking at and unifying the
incoming words one by one, and that it could -- in another language, or
maybe in some some obscure, overlooked area of the language you're looking
at -- be assigning categorial markers to them. Typologically there is
really no option here: even if a given language doesn't make use of the
opportunity to carry out some analysis while building c-structure, it is
clear that you can't do anything else with a sentence until you first
sequentially read in and start to combine the words. Scrambling doesn't
mean no c-structure. C-structure has to be there, even it doesn't have much
to do, and hiding it in a general unification structure (eg HPSG) doesn't
make it go away.

But why must you build an f-structure,  if it is identical to the semantic
(thematic-argument) structure (or a part of it)? Why not map directly from
c to s-structure if the mapping is clear and unambiguous? Here's why: At
the lowest level of sentence analysis (less so at generation), you're
looking at highly ambiguous data, whether phonetic or visual. Word analysis
and syntactic analysis are not proceeding blithely symbol-by-symbol;
rather, they have to investigate lots of possible combinations of possible
interpretations of the sensory data, which means fetching enormous
quantities of data from lexical storage. If a poorly heard or ambiguously
segmented verb could be ten different lexical items, each of these lexical
entries must be fetched to see which one fits best (unifies) with the (also
ambiguous) sentence environment. But a full lexical entry is a very large
data structure, when you bring in all of the semantic knowledge that can be
attached to each entry. Fetching all of this data for all of the trial
unifications would be very wasteful, because most of it -- for the trials
that don't succeed -- will simply be thrown away. The data pathway to the
lexicon cannot be arbitrarily large, and I would expect that the cognitive
apparatus would try to keep the traffic on this pathway to a minimum at the
early stage of parsing.

The best way to do this would be to provide a kind of compact, simplified
key to the thematic roles and lexical semantic structure. (Data base people
do this all the time.) These keys could look like 'GIVE(^SUBJ, ^ OBJ,
^OBL)', and you could fetch lots of them in a very short time, because they
don't carry with them all of the semantic baggage lying deeper in their
lexical entries. Once syntactic analysis has found out which keys (lexical
forms) match best to the sentence environment, then it would make sense to
fetch the more elaborate semantic data, assign the gramm. function
arguments to semantic arguments, and start working through the
combinatorial possiblities of s-structure.

Thus, whether the argument positions in the lexical forms (marked for
syntax by the gramm. functions) are marked the same way as at the semantic
level is not important to the early phase of analysis; what matters is that
*some* form of simplified lexical representation be made available, and
that there must be a structure in which to carry out trial unifications, in
order to avoid the combinatorial explosions that would arise during a
combined syntactic-semantic analysis. My guess is thus that any language
has to provide lexical forms in its lexicon and an f-structure as a
scratchpad for syntactic analysis, even when a given language appears to
make no use of the mapping possiblities between syntactic and semantic
argument structure. However, once you have a mechanism for learning and
making lexical forms, there is no guarantee that they will remain
one-to-one images of the semantic argument structure, and of course in most
languages, they don't.


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