I am trying to understand why violating the locality principle is bad.
Pollard and Sag restrict their SUBCAT list to only have access to complements’ SYNSEMs, in order to maintain certain “locality” restrictions. They observe: “Just like there are no verbs in any language we know of that select a sentential complement whose verb phrase is headed by a transitive rather than intransitive verb, there are no verbs in any language that assign roles to a complement within a complement that they select” (P&S87, p. 143).
Do I understand right that what they are after here is the restrictiveness of the theory, meaning they want to make sure that I am unable to specify such a verb because they believe it does not exist, but that this isn’t about there being specific examples in form of ungrammatical sentences of how that would be bad?
In other words: is what is going to happen, should I violate this principle, that I would have a lexicon which is less theoretically sound but not that I will have ungrammatical sentences generated?
I think from a linguistic standpoint your reading is correct. There is also the practical angle: all processors discard the contents of the daughter’s AVM (by deleting the ARGS attribute, and a few others specified in the config file) after rule application. This is pretty important for efficiency: without it you will get zero packing of ambiguity. This means a rule or lexeme that references a granddaughter or ?niece? directly through the ARGS feature path won’t work. This practical concern can of course be sidestepped by passing whatever information you are interested in from the daughter up to a feature on the mother, as we commonly do with almost all of the features of the head daughter (but generally very few, other than appending RELS, from non-head daughters).
Sorry for being late to the party. I think your understanding is basically right — Ivan liked to insist on the importance of locality as a kind of prediction that the theory makes. Of course, it’s not a 100% strict prediction because any information we need to expose we can, we just have to deliberately pass the information along through the feature structures.
As Woodley says, the daughters of rules aren’t directly accessible, but that doesn’t mean you can make the info accessible via feature copying. For example (and as we were discussing in person yesterday) a complementizer can expose the FORM value of its complement by making it its own FORM value. Then that info is passed to the CP mother via the HFP. Similarly, if we want information about complement’s complements, then we can do a non-cancellation analysis of the valence features (like I do for the Wambaya grammar, but actually for other reasons).
No – ‘non-cancellation’ refers to not shortening the valence lists when an argument is picked up. That contrasts with the usual approach to ‘cancelling’ such requirements as they are satisfied, which in turn is an important way in which locality is maintained.