Joe:
[...]
> > > > > What kind of evidence would it take to change your mind?
> [snip!]
> > If the parts are well-formed then the whole is well-formed,
> > irrespective of how frequent it is. If we found that some ill-formed
> > wholes consisted entirely of well-formed parts then it would be
> > easier to swallow the idea that grammar is basically a mirror of
> > usage.
>
> I know that you're not going accept this but...
>
> - * I'm leaving if here comes my bus. (Lakoff 1987: 473)
>
> In terms of constituency, "here comes my bus" is well-formed but
> "if here comes my bus" isn't. So, you might be able to say that
> you have two well-formed parts ("if" + "here comes my bus") that
> give rise to an ill-formed composite structure.
>
> In terms of dependency, I guess that (a) a verb like "come" is
> going to have to know about subject delay and locative inversion
> and (b) "if" is going to have to access this information. In
> principle, it can be done (if it hasn't already).

You're right; I'm not going to accept it. IIRC locative inversion
can't occur in any subordinate clauses; in this respect it is
like wh-inversion. OTOH, other inversions, such as there-inversion,
be-inversion, neg-inversion can. I think we need a syntactic
explanation for this; usage effects will not suffice.

> In any event, I'm not sure if your proposal is really very
> revealing when it comes to DG: all the individual parts
> (= words) can be well-formed yet give rise to an anomalous
> whole (= sentence).

Individual parts = dependencies.

> > But in fact it's obvious to almost everyone that grammar is
> > a productive system and generates novel structures that do not
> > have to have been sanctioned by accumulated prior usage.
>
> Yes, but I hope that this is not leading to a Poverty-of-Stimulus
> argument!

No. To a langue/parole distinction. Even from a mentalist perspective,
it seems that you can't get away from a productive system that
operates independently from usage. Quite probably we learn the system
inductively from usage, of course. I'm not saying anything here that
a good Word Grammarian would not...

--And.