Donkey Anaphora is a name for cases where covariance of an antecedent and (what looks like) a bound variable occurs in the absence of c-command, as in the classic (if not entirely pleasant) example in (1):
(1) Every farmer who owns a donkey beats it.
In (the most salient reading of) ex. (1), the expression it is understood to covary with the expression donkey. This, despite the fact that neither c-commands the other.
What I’d like to point out here is that, in cases of Donkey Anaphora, the covarying expressions must match even in features that are considered “semantically inert” – e.g. grammatical gender on inanimates, or number on plurale tantum. So, for example:
(2) Every linguist who has purple pants looks silly in them/*it.
The same seems to hold of grammatical gender on inanimates, for languages that have it. (Towards the end of this handout, you’ll find some Hebrew examples corresponding to (2) – but using grammatical gender, not number, as the operative feature – and involving inanimate antecedents, just like in (2).)
This is proof positive that phi-features such as uninterpreted number (in plurale tantum) and grammatical gender (even on inanimates) can be transmitted by a non-syntactic mechanism. That’s because a syntax that can relate two expressions in the absence of any c-command or locality holding between the two is no syntax at all. (Notice that the antecedent, here, is located inside a relative-clause Complex NP island, inside a Subject Island, and does not come close to c-commanding the pronoun.)
Similar claims were made by Dowty & Jacobson (1988), but the force of many of their arguments rested on a really shoddy theory of Control, and so not much remains of those particular arguments, I don’t think. (Note that this is not true of every single one of their arguments; e.g. the argument from unanteceded deixis of inanimates is not plagued by this problem.)
The Donkey Anaphora case is rather different. For one thing, it involves covariance – so, not accidental coreference or deixis. Second, this is rather different from the Control cases, because there is really no conceivable syntactic maneuver that could help here. There is effectively no structure that the two covarying expressions in (2) cannot be embedded in, while retaining both their understood covariance and their phi-feature matching.
I believe this observation is rife with really interesting consequences. For one thing, anyone who wants syntactic phi-feature agreement to play any role in binding (of anaphors or pronouns) bears an enormous burden of proof. We know that there must be a mechanism, call it M, that ensures that binders and bindees match in phi-features, even in those phi-features which seem to be semantically inert, and does so non-syntactically. Given the serious problems with theories of binding that are based on syntactic phi-feature agreement (see the handout linked above for a survey of these problems), we’d better have a really good reason to keep using that, and not the independently-needed M, to do the job.
Second, let me say a little bit about what M might be. I think M is just semantics/pragmatics. That is to say, I am “going the Dowty & Jacobson route” (in a sense; see below for an extremely important caveat), and assuming that the semantic-pragmatic system has access to information about grammatical gender, number on pluralia tantum, etc. This might fly in the face of some people’s intuitions about what “semantics” and “pragmatics” even mean, but, as they say, tough luck: there is no coherent notion of syntax that could do M’s job, which leaves (as far as I can tell) only this possibility. The matching between pants and them in (2) (and similarly for the linked-to examples involving grammatical gender on inanimates) must be achieved by M.
Now, it is tempting at this juncture to ask: given these results, do we need syntactic phi-feature agreement at all? Can’t we do, say, predicate-argument agreement (my favorite empirical playground) via M, as well? (The rhetoric in D&J seems to suggest this in places, though in other places they claim to be talking only about agreement in Control configurations.)
If you think this could work, think again: predicate-argument agreement really is syntactic. Here’s how we know. In a configuration like (3), where ‘>>’ represents c-command and H is some functional head, H cannot agree with DP2 across DP1:
(3) H >> DP1 >> DP2
There are superficially similar effects in binding, and even in association with focus, but this case is quite different. As has been shown elsewhere, the intervention effect in (3) is purely structural: it cares about case; it’s blind to thematic roles or any other semantic properties; it can be fixed by A-moving DP1 out of the way; etc. etc. – see ch. 8 of my book for a survey of some of the relevant literature. (This is not to be confused with the phenomenon of dative intervention in A-movement, which Benjamin Bruening and, more recently, Kenyon Branan, have challenged the very existence of. I’m talking about cases where DP2 is not a candidate to A-move, regardless of whether DP1 is there or not, and only agreement between H and DP2 is at stake, e.g. in Basque or Icelandic. Neither Bruening nor Branan have anything to say about such cases.)
Binding, on the other hand, doesn’t care about case (in fact, the ability of Icelandic dative subjects to bind subject-oriented anaphors is one of the most striking pieces of evidence that they are indeed subjects). Its minimality violations cannot be repaired by movement – A-moving the downstairs subject out of the way does not allow the experiencer in (4) to bind the downstairs reflexive, even though the experiencer is generally able to bind into the infinitive, as the disjoint-reference effect in (5) attests.
(4) John seems to Mary to like her/*herself.
(5) John seems to her to like Mary. (coreference impossible)
And, most importantly, predicate-argument agreement never does things like (2). Verbs (well, the functional heads that verbs typically occur with) can agree with DPs that are not their arguments; they can agree with DPs in other (lower) clauses; but they can’t blatantly ignore the contours of syntactic structure in the manner shown in (2).
So, what have we learned? That just because two expressions must match in phi-features – even semantically inert phi-features – it is not an argument that the phi-features on the two are related by syntax. Not even close.