Ivan A. Sag (sag@csli.stanford.edu)
Mon, 20 Jan 97 11:20:29 PST
All, As this exchange about adverb extraction has progressed, we seem to be copying more and more people, so we decided to move it to the hpsg list. Bob Levine and I were also wondering why we don't have more discussion of this kind on that list. It's not that people aren't working on issues like this that are of common interest... Happy New Year, Ivan --------------------------------------------------------------------- --------------------------------------------------------------------- Summary: Discussion of adverb extraction -- mid-January, 1997. --------------------------------------------------------------------- --------------------------------------------------------------------- Kathy Baker wrote to Ivan Sag about some issues that arose in implementing the proposal for adverb extraction made in Pollard and Sag (1994), which uses a SLASH-based analysis only for extraction of embedded adjuncts. Ivan Sag wrote back suggesting that one should probably reject the proposal for adverb extraction made in P&S-94, because it appears inconsistent with the results obtained by Hukari and Levine's work (Journal of Linguistics 1995). H&L show that languages that register unbounded dependencies (through verb morphology, the possibility of verb-inversion, tone downstep suppression etc.) do so systematically in matrix adverbial structures, even in sentences analogous to (0): (0) Yesterday, Kim visited Sandy. So it seems plausible that these examples too should be treated as extraction, i.e. in terms of SLASH, which is not the case in the P&S-94 analysis. Ivan suggested the following alternative: (1) Adverb Extraction LR (revised): + + + + |HEAD verb | |HEAD [2] | |SLASH eset | ==> | + XP + | |CONTENT [1]| |SLASH { |MOD [HEAD [2],CONT [1]]| } | + + | |CONT [3] | | | + + | |CONT [3] | + + This would allow any verb to become [SLASH {ADV}], hence allowing a treatment of all adverb dislocations, even (0), in terms of [SLASH {ADV}] specifications. --------------------------------------------------------------------------- Then Bob Levine replied as follows: --------------------------------------------------------------------------- >There's something that has been gnawing at me since you sent us that >message about the new form of the AELR... We have the revised form of >the AELR [in (1)]. And it seems to work as promised. You get the >content of the modification transferred, as it were, to the content of >the verb from which the modifier has been extracted, and you don't >even have to worry about who is the syntactic and who is the semantic >head in this case because the syntactic head contains the content of >the modified VP, unlike the case where the adverb is in situ. The >technical problems we noted in our adjunct extraction paper do indeed >vanish, as you assert. > >But there seems to be a problem when more than one modifier is >involved in such a way that scope considerations become crucial to >truth conditions. Consider the sentences > >(2)a. Robin washed the car frequently rather rarely. > b. Robin washed the car rather rarely frequently. > c. Rather rarely, Robin washed the car frequently. > >(2)a and (2)b have different truth conditions: (1)a is true if Robin >goes through bouts of washing the car twice a day for a week, but only >does so one or two weeks out of a year, while (1)b is true if Robin >washes the car only once a month, and does so ten months out of the >year. In other words, the rightmost adverb has the wider scope. (I'm >ignoring strongly marked prosodic possibilities here that might give >different interpretations here, for the time being). Now as (1)c >shows, the extracted adverb is interpreted so that that its scope is >the whole VP; that is, (2)c is synonymous only with (2)a. But that is >not the result which follows from the revised AELR. The reason is that >when you run the AELR on the verb _wash_, this is what comes out of >it: > >(3) > + > |PHON <wash> + > |HEAD [2][COMPS <[4]NP, [5]NP>] | > | + XP + | > |SLASH { |MOD [HEAD [2],CONT [1]]| } | | > | |CONT [3](rather-rarely'([1]))| | > | + + | > |CONT [3] | > + + > >The revised AELR builds the modification corresponding to the >extraction *into* the content of the verb which is modified by the >extraction. But the problem is that now the in situ adverb will modify >this content, meaning that the functor denoted by the in situ adverb >will have wider scope than the functor denoted by the extracted >adverb, contrary to fact. The interpretation will be determined by the >rule and by the Semantics Principle as follows: > > VP[CONT [3]]/AdjP[4][LOC [1][CONT (rather-rarely'(wash-the-car)')]] > / \ > / \ >(5) VP/[4] AdjP[[3]CONT [frequently'([1])] > [CONT [1]] | > / \ frequently > / \ > V[CONT [1]/[4]] NP > | / \ > wash / __ \ > the car > >And even if you want to argue that the scoping in examples like (2) is >uniformly ambiguous---which I don't think is actually the case---the >problem is that the kind of interpretation that will get cranked out >of the LR *isn't* ambiguous; it's the more marked interpretation, the >one that runs counter to the the scoping the syntax of (2)b, and the >more normal interpretation corresponding to (2)b isn't available. How >do you get scope management to come out right, given this problem? And >notice that with certain combinations of adjuncts, the results of >misscoping can be pernicious; in a question like > >(4) How many times a year does Robin run in the morning? > >the only available interpretation that the new AELR allows is one >where _how many times a year_ has narrow scope and _in the morning_ >has wide scope, making (4) an insane question instead of the perfectly >sensible one it in fact is. > >I've propably missed something quite basic, but I can't see what. Have >I really, or is there indeed a problem for your new adjunct LR? --------------------------------------------------------------------------- Ivan then replied: --------------------------------------------------------------------------- >Bob is absolutely right, of course, as long as the only adverbs being >selected by heads are the extracted ones. But, as Rob Malouf points >out, you get a better result if, as assumed for decades by >dependency grammarians (cf. also current discussion on the categorial >grammar mailing list), heads *IN GENERAL* select for adjuncts. > >So if we replace the AELR by the following Adverb Addition LR, we >get a different result. > >Adverb Addition LR (AALR): > >+ + + + >|HEAD verb | |CONT [3] | >|CONTENT [1]| ==> | AdvP | >|COMPS [2] | |COMPS [2] (+) < |MOD [CONT [1]] | > | >+ + | |CONT [3] | | > + + > >Now this rule can apply to its own output, so it allows the >possibility of two AdvPs in the COMPS list, the later one outscoping >the earlier one. So, assuming now that this LR feeds CELR (or the >equivalent constraint in the ruleless alternative suggested by Gosse >Bouma), then the AdvP that ends up in the SLASH set can be either the >first AdvP on the COMPS list or the second. This means that either >scope will be derived for Bob's examples. > >>(2)a. Robin washed the car frequently rather rarely. >> b. Robin washed the car rather rarely frequently. >> c. Rather rarely, Robin washed the car frequently. > >This prediction might not be a problem, for, as Rob also points out, >there are strong pragmatic grounds for disfavoring the reading of (2)c >where frequently outscopes rarely. However, as Bob points out, the >effect is pretty strong! > >But even assuming that the grammar must impose an absolute restriction >against the unwanted scoping, we might be able to tune this analysis to >take care of business. In particular, Postal has been arguing for >years that complements are right-shifted before being extracted. The >analogoue to this here would be introducing a way of adjusting COMPS >order (say to handle Heavy XP Shift) that enables an analysis where >only COMPS-final elements can be extracted. This would accommodate >Postal's claims, and predict wide scope for the extracted adverbs as >well. > >I'm not sure whether the scope constraint should be part of the >grammar. Are any of you? > >The resulting analysis has the nice property, though, that it posits >no extraction-specific mechanisms and it reconciles the SLASH-based >analysis with the Hukari-Levine observations about adverb extraction, >which can now (assuming lexically induced SLASH amalgamation) be >analyzed in terms of whether an element has a SLASHed ARG-ST element >or not. > ------------------------------------------------------------------------- Bob then replied: Date: Sat, 18 Jan 97 14:16:43 PST From: "Ivan A. Sag" <sag@csli.Stanford.EDU> To: sag@skol.Stanford.EDU ------------------------------------------------------------------------- >Ivan's new rule does avoid the problem of enforced narrow scoping >that caught my eye in the last version of the rule: once you've added >the adjunct synsem value to the COMPS list, you can then run the >output of the rule through the CELR, and since the output of the >adjunct LR will be a head whose CONT reflects widest scope for that >added adjunct, the output of the CELR will allow you to extract that >adjunct while yielding a head whose CONT also reflects widest scope >for that (now extracted) adjunct. And it also avoids all the other >technical problems. But there's something else about it... isn't this >essentially the same rule that's presented in the van Noord & Bouma >paper on adjuncts as complements that came out a few years ago in one >of the computational journals? I'm *sure* it's basically that same >rule! ... >The issue of whether English adjuncts should be put on COMPS lists or >not seems to me inextricably linked with how adjuncts should be >treated in general; what do we know about their properties and what >solutions plausibly account for their behavior in other languages? In >this case we have a substantial of work on French adverbs that needs >to be taken into account in looking at the English facts. The series >of papers cowritten by various subsets of Abeille, Godard, Kim, Miller >and Sag (AGKMS, pronounced `ackems' or something like that, I suppose) >examines a variety of kinds of French adverbs and makes various >proposals about their configurational possibilities, some of which >supersede earlier ones. The earlier AGKMS papers (by SK and AG) >presented some cogent arguments that so-called strictly negative >adverbs in French should be added by a lexical rule to the COMPS list >of finite verbs, but that these same adverbs are *adjoined* to >infinitives. > >(i) Part of the reason for treating negation in finite VPs as >complementation rather than adjunction was in fact the contrast in >iterability of these adverbs in finite and infinitive cases >respectively: you can iterate negative adverbs like (ne) pas in >infinitives whereas you can't in finite clauses, so that in the latter >case the adverbs display a `unicity' property characteristic of >complements. > >(ii) A second piece of evidence for the complementhood of strictly >negative adverbs in finite VPs is the fact that the lexical properties >of verbs showing up with _pas_, for example, are different from the >properties such verbs show without _pas_. Clitic order in imperatives >changes, and selectional properties are different as well. If _pas_ is >part of the COMPS list in the cases where it appears with finite Vs, >then these Vs will be different lexical items from those where no >negation is involved and other differences in lexical properties, such >as clitic order and selection, might then be expected. > >Interestingly, the earlier work on French adverbs, such as the AG >paper `The syntax of French negative adverbs', treat many lexical >adverbs such as _bien_ as complements as well, and specify two lexical >rules adding adverbs to COMPS lists, one for strictly negative adverbs >and one for all others, including some with a negative flavor which >are distinguished from the first group for various reasons. At the >same time, provision is made for adverbs which are not complements to >still appear as sisters to the verb (a suggestion that was first made, >as Iatridou points out in her critique of Pollock, by Lisa Travis in a >1988 McGill WPL). So now we have three possibilities for adverbs: (1) >they may be adjoined at various phrasal levels; (2) they may be >complements of the verb; (3) they may be unselected sisters of the >verb. But in a later paper on the status of the feature [LEX] that is >central to AG's account of linear ordering properties, they withdraw >their proposal that non-negative adverbs are complements and treat >virtually all such cases as instances of either (1) or, for [+LEX] >adverbs, (3). In restricting the treatment of adverbs by >lexical rules making them complements to strictly negative adverbs, >they state quite explicitly that their previous analysis of >non-negative adverbs as complements was based on consideratins of >`systematicity' but really had no empirical support. And it's true >that no comparable arguments from unicity and differential lexical >properties were ever given in the earlier paper for non-negiative >adverbs as complements; that treatment more or less rode the coattails >of the analysis given for the negative adverbs. > >So it turns out (at the moment anyway) that AGKMS seem to be requiring >at least unicity and some evidence of variation in lexical properties >on the head in order to qualify an adverb as a complement. Evidence of >sisterhood to the head is insufficient, since that can be accomodated >without positing complementhood. More to the point, AG evidently do >not regard extractability of adverbs as evidence of complementhood; >French adverbs extract just as English ones do, but that apparently >wasn't taken by AG as evidence in favor of complementhood. Was this an >oversight, or was it based on the sense that the most important >consideration was the lack of evidence for a `split' in a verb's >lexical entry when a non-negative adverb is present as vs. when it >it's not? If it's the latter, then the approach that Gosse and Gertjen >first proposed in their paper and that Ivan and Rob rediscovered seems >to run into the same problem as the French non-negative adverbs: the >LR rule in question coins a new avatar of the verb, but where is the >evidence that you really have a separate lexical item? > >That's really the main point I wanted to get at, but there are others >that might be pursued. Tom [Hukari - IAS] and I have been working for a >while now on the problem of getting the binding theory to work out right >for adjunct clauses (including those contained in fillers, giving rise to >the so-called antireconstruction effect among a subset of speakers); >one thing that seems to emerge is that if you treat adjunct *CLAUSES* >as complements then you both lose the possibility of an account of the >antireconstruction cases and, less exotically, you lose a good story >about the contrast between (1)a and b: > >(1)a. *You can't tell them{i} that the twins{i} are being offensive. > b. You can't say anything to them{i} without the twins{i} getting offended. > >and similarly for (2): > >(2)a. *I told them{i} about the twins'(i} BIRTHDAY. > b. I only get them{i} presents on the twins'{i} BIRTHDAY. > >(The contrast in (2) seems a little less sharp than in (1) but is to >my ears still striking. The cases in (3) seem quite parallel, though: > >(3)a. We'll wait to start criticising them{i} till after > the twins{i} depart. > b. We'll wait to start criticising them{i} till after > the twins'{i} departure.) > >You might want to say that sure, clausal adjuncts of the sort in (1) >aren't complements; they really do adjoin, but perhaps they're not >extractable either, so it doesn't matter. On the other hand, it's a bit less >plausible to say this in the case of (2) or (3)b, since you can extract such >adverbial PPs: `On whose birthday did....' etc. But this is getting >into somewhat different terrain than I wanted to; the main issue I >wanted to raise is this question of the English adjunct situation >vis-a-vis the one Ivan and company have already explored for French, >and if, modulo extractability itself, there's independent reason in >English to posit complementhood for adjuncts when the AGKMS group seem >to have abandoned complementhood for the corresponding cases in >French. > [NOTE FROM IVAN: I edited all this only in minor ways. I hope I haven't distorted anything. If you've managed to get this far, you are up to date in our discussion.]
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:36:03 PST