Formalism & Theory

[Avery Andrews]

Some of my thoughts on the foundational issues.

1.  I disagree with Paul King about the status of generative grammar as
a true deductive science.  The basic problem is that in empirically
relevant cases, we can't prove much of anything that is actually useful.
So I can't prove that classic 1982 LFG (without inside-out devices)
doesn't contain satisfactory analyses of case-stacking etc. in languages
like Kayardild and Martuthunira.  What can do is show that Jane Simpson's
analysis of case-stacking in Warlpiri doesn't carry over in any obvious
way (since these languages don't split up NP's in the way that Warlpiri
does), and that some other obvious moves don't work, but this is just an
argument, not a proof.  So no deductive science.

2.  But I very much agree with him that restrictiveness per se is a
false goal.  For at least three reasons:

  A)  We don't know what's not out there (in the Amazon, for example)

  B)  Even if we did, we wouldn't know whether its absence was
      accidental (if it were not for the languages in the geographically
      restricted area of Southern Africa, we would surely be boasting
      about how our theories predicted the impossibility of clicks being
      speech-sounds, in spite of their prevalence in para-language.

  C) And even if we had full knowledge of the facts of A and B, we still
     wouldn't know whether things were non-occurring because of
     properties of the language faculty per se (and so should be
     explained by our generative theory), or for other reasons, such as
     functional considerations, or the environment that the language-
     faculty is functioning in.

3.  So what to about instead?  I think that what's important is what I
would call `appropriate descriptive power', which has several facets.

  a) the ability to frame clear & explicit descriptions of the kinds of
     basic linguistic patterns that people notice (the basic LFG scheme
     of annotated PS rules, grammatical functions,  agreement-via-
     unification etc. does pretty well in this regard, tho I find lots
     of rough edges).

  b) automatic prediction of interactions that aren't so easy to notice.
     for example in LFG, if you've done predicate adjectives and
    `subject raising' constructions correctly in a language like
    Icelandic or Ancient Greek, then long-distance agreement works out
    automatically, for `for free'.  Not as a deductive prediction, since
    we can't prove that there isn't some equally nice or nicer grammar
    for the basic facts in which LDA doesn't work, but as a sort of
    practical, `empirical' one.

 c) suggestiveness.  This is basically, as Bruce Hayes put it in his
    paper on Grounded Phonology, the capacity of a theory to serve
    as an amplifier for the pattern-detection abilities of linguists.
    So if you look at Dyirbal with good ole' TG in your toolkit,
    you have a good chance of discovering syntactic ergativity, and
    if you do the same for Modern Icelandic, you have a good chance
    of discovering non-nominative subjects.  I don't think this is
    really independent of (a, b), but it is certainly independent of
    even basic coherence of a theory w.r.t. deductiveness.  So PPA
    has served very well in helping people discover and organize facts
    about word-order variations and the relationships that these might
    or might not have w.r.t. inflection in European languages.

Perhaps a unifying theme to (a-c) is that appropriate descriptive
power means that analyses that make correct predictions are easy
(for linguists) to find.

The advantages of looking in this direction rather than towards
restrictiveness is that we can tell whether a given analysis of some
data is clean and nice & captures generalizations, or whether it is
a mess.  And I think that if we do concentrate on improving `ADP',
we'll make progress towards understanding why language is learnable,
which amounts to finding a discovery procedure (basically because
ADP I think to a large degree is a matter of making the right analyses
easier for linguists to find than the wrong ones, which should
contribute to making them easier to find for algorithms).  This more
modest goal is sort of like Chomsky's original proposal of going
for an `evaluation metric' rather than a `discovery procedure', though
it's not really the same.

And so now we get to formalism and architecture: details of formalism
and archtecture have a big influence on ADP, but these influences are
subtle and complex, and there are no proveable results. All you can do
is note that some proposals do support nice analyses of specific
empirical phenomena, and others have not yet been observed to.

For example LFG's correspondence-based architecture supports a principle of
`morphological blocking' that makes it quite easy to describe cases
where the presence of an inflection inhibits the use of a free pronoun.
This is much harder to work out in a derivation based theory like TG,
which causes these theories to have problems with basic descriptive
details of languages such as Irish (and also lots of Oceanic ones, which
have similar effects).

On the other hand, classic LFG f-structure-based architecture seems to
have considerable difficulties in dealing with the basic facts about
complex predicates that Miriam Butt, Alex Alsina and others have been
documenting for the last 10 years or so.  Chris Manning and I discuss
and propose a solution to this in our recent book, but our architectural
proposal has the unfortunately property of `breaking' analysis that
depend on inside-out functional uncertainty:  I've current got a
nasty-looking fix for that, and am trying to find a nice one.

Something else I'm presently interested in is the introduction of
resource-based ideas into syntax; it seems to me that these ideas
might be able to replace the rather awkward apparatus of `Formal'
Completeness and Incoherence (which apply to FORM features), with
a more uniform approach where by lexical stems all introduce a form
feature, and these are consumed to produce meanings.  `Regular'
compositional interpretations arise when a single FORM feature
is consumed to produce a meaning-element, idiomatic/non-compositional
ones when 2 or more are.  If something like this can be worked out,
it might help a lot in analysing serial verbs, where noncompositionality
is very common (and I can't imagine anyone who does descriptive
work on SVCs taking the FORM feature proposal seriously), and also
for polysynthetic languages, where non-composiional interpretation of
stem and various affix-slot combinations is very widespread.  A
circumfix, for example, can be treated as a pair of affixes whose
FORM resources provide no semantic intepretation individually, but
only together, via a 2-premise deduction (a circumfix would then be
basically just an extreme case of noncompositionality).

So while I won't quibble with Mark Johnsona about whether the
formalism is the `central' part of a theory or not, I would insist
that it's the deepest and most fundamental:: changing it may not have obvious
deductive consequences  (basically because we're not smart enough to
find them), but it can have all sorts of effects on what you can
describe easily, which is what I think is the most important attribute
of a linguistic theory.  So RLFG is a waste of time if it has no effect on
this, a mistake if it irretrievably breaks things that we really need,
and a success if it doesn't break important stuff, and makes other
things easier to do.

And as someone who has spent almost all of his professional career
surrounded by descriptivists whose problems he usually can't help
with in any substantial way, I think generative theory has a long way
to go in attaining appropriate descriptive power.


  - Avery Andrews