Robert Levine (levine@ling.ohio-state.edu)
Thu, 23 Jan 97 15:42:27 EST
Gosse's inference about the significance of filler/gap pathway morphosyntax as sketched in my previous note is absolutely correct, with one important caveat that I want to get to later. He's also correct that if you amalgamate SLASH on the head in languages like Chamorro, Irish and the rest, in *all* cases---including adjuncts---then the issue of whether or not there are traces is essentially beside the point. Of course, in order to do that, given the way the traceless approach in Pollard and Sag 1994 and Ivan and Janet's WCCFL paper is set out, the only way to ensure that SLASH will show up on the head when it's an adjunct that's extracted is either to have a special lexical rule such as Ivan was contemplating earlier or to put adjuncts on valence lists so they can be converted into SLASH specifications by the CELR. The first way is, apparently, very problematic, so it looks like you have to go with the second, and you have to do it in such a way that slashed categories show up on ARG-ST if you want to get any explanatory milage from this approach vis-a-vis the binding domain facts. The significance of the coreference facts that I noted for English, and which Tibor says are very similar for German, is that they seem to run counter to this conclusion. In this connection, Gosse notes that Dutch and Japanese provide examples which look problematic for the form of binding theory that I was appealing to; presumably the problem with (1) Eva{i} zag mij plotseling voor zich{i} staan Eva saw me suddenly before ZICH standing is that _zich_ is not accessible to condition A on the assumption that the adjunct _voor zich staan_ is unselected, so you can't force the coreference that you do get. But isn't it the case that within the VP complement of _voor_ you have a SUBJ specification which is structure shared with the synsem of _Ewa_ by whatever mechanism you appeal to get subject-controlled adjuncts to be subject controlled? If so, then there's no need to make the adjunct selected by the head, any more than there is in, e. g., (2) The twins{i} do crazy things without ever getting themselves{i} in trouble. Gosse further observes that one could imagina a `mixed' approach to the problem of adjuncts (similar in fact to what AGKMS propose for French), where sometimes adjuncts are selected and sometimes not, but that if the two different modesof selection have such far reaching consequences for binding and for extraction, it should be relatively easy to check whether this is a coherent position. (I.e. the cataphora effects observed by Bob should only be possible with unselected, and therefore non-extractable, adjuncts.) Something like this contrast was what I was getting at in the examples in one of my earlier notes; cf the following: (3)a. We only buy presents for him{i} on Robin's{i} BIRTDAY. b. *I told him{i} about Robin's{i} BIRTHDAY. c. On whose birthday are we going to drink that bottle of champagne? So for English, at least, there is a serious misprediction that follows if your approach entails that cataphora and extractability are complementary. But it's not really the binding domain facts that seem to me central. There's a more general point at stake, namely, the question of whether you want to be forced to say that if two things behave in the same way in certain respects, they must be in some respect instances of the same thing. We have a traditional distinction between adjunct and complement which is based on a number of behavioural distinctions between the two classes: (i) cooccurrence between heads and complements is lexically idiosyncratic, but cooccurrence between heads and adjuncts is limited only by semantic coherence (this gives rise to `unicity' effects and the like); (ii) complement order is positionally far more constrained than adjunct order with respect to either complements or other adjuncts; (iii) extraction from within adjunct clauses is typically dicier than extraction from complement clauses; and various other difference that have been noted in the literature; the binding theory differences that I've been harping on should be counted here as well. The problem is that to extract adjuncts on the traceless approach, you've got to treat these differences as in some sense secondary to a fundamental identity between complements and adjuncts: both are `selected' by heads. It seems to me that this outcome reflects the fact that HPSG has no way to characterize the notion of `dependent', where both complements and ad-categories are dependents of a head in the way intended by dependency grammars or certain work of Carl's and Eun-Jun's. The generalization is that dependents of a head can in general extract, and that in binding domain languages the head records extraction of/from (one of) its dependents. Because the only way to capture this notion of dependency in HPSG is by means of valence list, you're forced to say that adjuncts are on valence lists, which is an instance of the modern cliche `if your only tool is a hammer, everything looks like a nail'. If the theory is forcing you to assimilate adjuncts and complements to the same thing, in spite of fairly far-reaching differences between them, then maybe the theory needs to be amended in some way to capture their commonality without forcing that commonality to be one of lexical selection. Otherwise the notion of selection is in danger of becoming contentless. Bob
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:36:03 PST