carp@research.bell-labs.com
Thu, 17 Jul 1997 14:59:05 -0400 (EDT)
Stefan Mueller writes: > >From some reviewers I got comments of the kind: "Where is the problem? > Just take Ingria's approach!" I am a little bit surprised that people I don't know how many reviews you got, but I wrote one of them and I sure didn't suggest you take Ingria's approach. I think you're right in abandoning subsumption. The logic is wrong (in particular, the failure of conjunction to distribute over generalizations, so that (A & (B generalize C)) =\= (A&B) generalize (A&C)). Here's what I wrote for the review of the abstract in question (I also wrote that the empirical issues were interesting and that the paper should definitely be accepted): > Relational analyses in HPSG can be pretty ad hoc ... > The solution['s] ... use of disjunction is reminiscent of the > approach used for CG by Sam Bayer... I'm of the opinion that just building the right logic of disjunction into HPSG (rather than always treating it as non-determinism at the highest possible level of program scope), would be the right way to go. In case you want to see more of my own reasoning, check out my previous post to this list on the topic (and the thread around it): ftp://ling.ohio-state.edu/pub/HPSG/Log/MailArchives/Archive4/0131.html > My question now is: Did I miss something? Is there actually somebody > using subsumption tests in an implementation? This I don't know the answer to. But even if someone was using it, we know why it won't work. I know that Ivan Sag and Jay Blevins were recently looking into similar issues involving coordination (in addition to Ivan's latest take on relatives), so you may want to contact them to see if they've got anything to say that might be useful to you. - Bob
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:36:08 PST