Andreas Kathol (kathol@let.rug.nl)
Mon, 16 Oct 1995 11:49:14 +0100 (MET)
Steve, Ulli, and everyone else, Ulli Koch wrote: >Your structure > > S > / \ > NP VP > | | > N V > | | > Kim left > >is correct for P&S 87 and P&S 94 (modulo obligatory null >determiners) [...] Steve Wechsler wrote: > Thanks for your reply. It was the only one I received. I would have > thought that more HPSGians would be interested to know that the version of > the theory in P&S 94 fails to generate simple Noun Phrases like [Kim] > (without creating new problems; see below). I had actually meant to say something about Steve's question sooner, but this is more with respect to the verbal part of his example sentence. Am I the only one who feels a deep sense of uneasiness about an analysis that requires a unary expansion whose sole purpose is to license a lexical item as the only daughter of a phrase? Since this situation is brought about by an ID schema which makes explicit reference to signs of sort _word_ ("lexical sign"), let me rephrase my question by asking whether it is necessary and/or desirable to make reference to the word/phrase distinction in syntactic licensing schemata. To be sure, I do not want to question the usefulness of the sortal _phrase/word_ distinction in general. This seems to be necessary for instance in order to appropriately tighten the domain of lexical rules and for grounding the recursive licensing of signs. Hence, a sign S is licensed if: a. either it is of sort _word_, b. or there are two or more objects of sort _sign_ whose combination according to a number of constraints yields S. So, do we need licensing schemata of the sort in b. that mention _word_s, as opposed to _sign_s_? My conjecture is that this is indeed never necessary and that what appear to be _word/phrase_ distinctions are best made in different terms, specifically by looking at the COMBINATORIAL PROPERTIES of a given element. For instance, it seems that we never need to make reference to the _phrase_ status of a (nonfinite) VP complement because that's already given in terms of a specification like: SUBJ <[ ]> COMPS <> which matches both "phrasal" VPs (such as "love Sandy") as well as "lexical" ones (such as "sleep"). If this conjecture is confirmed, we whould never need unary expansions as in the case of VPs consisting of intransitive VPs. So, what about those cases where reference to _word_ has been thought to be necessary? 1. Flat structure of VP, schema 2 (Head-Complement Schema) Assuming that VPs have a flat structure, as opposed to, say, a binary one, yields a straightforward domain for order variation. However, if I see this right, it's actually not crucial to require of the head in schema 2 that it be lexical. Since we also have the requirement that the licensed projection be saturated for all arguments except the subject, it doesn't seem that we incorrectly derive ungrammatical strings (in the same way this does happen with schema 3; see below). But we DO introduce an arbitrary number of analyses for VPs since nonbranching structures are now licensed ad infinitum: (1) . . . VP | schema 2 | VP | schema 2 | VP / | \ schema 2 / | \ V NP NP As an alternative, I would suggest that we assume that licensing schemata OBLIGATORILY introduce syntactic material (by cancelling at least one of the arguments), which of course makes unary expansions impossible. The flipside of this is that schema 1 should not contain a requirement to the effect that subject-taking verbal projections are or sort _phrase_. Otherwise, we either couldn't derive "Kim left" or would be stuck with the kind of nonsolution I discuss below under 5. But as far as I can see, the only way to require obligatory structure building of schemata, allow for order variation, AND avoid spurious ambiguities is to abandon the strict phrase structure-based conception of syntactic licensing that underlies P&S 87/94 and move toward a more flexible system such as Blevins 90 or Kathol 95. 2. Inversion in English, schema 3: I don't want to go through the details of the analysis here, but I don't see any reason why the approach to order variation (in German) of Kathol 95, specifically regarding the placement of finite verbs, should not carry over to English. While this would involve allowing for possibly nonconcatenative operations in the linear component, we also get precisely the same combinatorial structure in both uninverted and inverted sentences. (Of course, this is not a new proposal, see Blevins 90 for similar ideas and some historical background.) (2) a. S [Kim, has, slept] / \ NP VP [has, slept] / / \ Kim V V [slept] | | | has has slept b. S [has, Kim, slept] / \ NP VP [has, slept] / / \ Kim V V [slept] | | | has has slept In other words, do we actually need to have 3 separate schemata for licensing head-argument combinations? (My personal view is that a single binary schema (with obligatory argument cancellation), embedded in a more flexible linearization component, should do the job just fine.) In HPSG, restricting the head to be of sort _word_ is also necessary for avoiding multiple applications of schema 3 resulting in ungrammtical VOS-like structures for English: (3) S / \ schema 3 / \ VP NP /\ / \ schema 3 / \ V[+INV] VP Again, if structure building and linearization are decoupled as in (2), the requirement on _word_-level heads can be dropped, as the combinatorial licensing is no longer involved in issues of order. If desired, I will elaborate this point more fully. 3. Structure of Verb Cluster in German: Let me now turn to a case where I believe that viewing syntactic construal in terms of the _word/phrase_ distinction actually makes the WRONG predictions. Kiss 92 (don't take this personally, Tibor!) presents a right-branching analysis for canonical verbal clusters in German (cf. (4)) along the lines in (5): (4) dass Peter das Buch lesen koennen wird. that Peter the book read can will `that Peter will be able to read the book.' (5) V Complement / \ Head / \ V V | Comp. / \ Head | / \ lesen V V [LEX+] | | koennen wird [LEX+] This analysis ia argued to be preferable to the left-branching analysis, cf. (6), because it relies on "lexeme-selection", which seems prima facie to pick out a natural class of possible complements and allows a rather appealing bifurcation of complementation types parallel to the sortal _phrase/word_ distinction. The problem with the right-branching analyis, the way I see it, is that it has absolutely NOTHING of insight to say about a related set of facts that FALL OUT on a left-branching analysis like the one in (6): (6) V / \ Head / \ V V / \ Head | / \ wird V V | | lesen koennen Among the facts that the type of analysis shown in (6) can account for rather straightforwardly are Partial VP Fronting constructions, cf. (7), and Aux-Flip constructions, cf. (8): (see also Hinrichs and Nakazawa 1994). (7) [lesen koennen] wird Peter das Buch sicherlich. read can will Peter the book certainly (8) dass Peter das Buch [wird [lesen koennen]]. that Peter the book will read can To my mind, this suggests that the kind of syntactic unit that serves as the complement for governing verbs in German is related to the notion of _word_ only in an indirect manner--certainly the sortal _word/phrase_ distinction is irrelevant for this phenomenon. 4. Other phenomena I'm really curious to find out if there are any other phenomena that seem to require reference to _word_. One example that comes to my mind is the restriction in pseudocleft constructions, where the initial element of the wh-clause must not be complex: (9) a. [[What] Sandy likes] is for Kim to go away. b. *[[For what] Sandy longs] is for Kim to go away. c. [[What] Sandy longs for] is for Kim to go away. But it's anything but clear to me whether this is an indication that this first element is of sort _word_ (e.g.: why do we then have dislocation of a _word_-level element in (9c)?), or whether the operative constraints are to be located elsewhere; e.g. in terms of internal complexity of the initial element which can be captured by other means. 5. Word/phrase and the lexicon Coming back to the problem of unary expansions for VPs in sentences like "Kim left", one possible response may be that maybe we need to specify an element like _left_ as being of sort _sign_ in the lexicon and have the syntactic environment resolve this sort to _word_ or _phrase_. Besides introducing a strange partition on words in the lexicon (i.e. elements that are ONLY words and those that are words or phrases), it doesn't actually solve the problem. This is because once we have an environment where _sleeps_ needs to be resolved to _phrase_ in the P&S 94 system (as in the combination with the subject _Kim_), we immediately have a DTRS attribute appropriate for that type of linguistic object. There will have two substructures: one given as the value of HEAD-DTR (the _word_-level instantiation of the intransitive) and an empty list as the value of the COMPS-DTRS attribut. Hence, there doesn't seem to be a way to avoid unary syntactic expansions by underspecification in the lexicon so long as schema 1 requires the predicate to be of sort _phrase_. 6. Other frameworks To the best of my knowledge, unary expansions are never necessary in Categorial Grammar, which is what one should expect since there is only one way of classifying linguistic objects, namely according to their combinatorial properties and NOT via some "external" categorization as with the HPSG _word/phrase_ distinction. I always found this parsimony appealing. Similarly, there seems to be a methodological imperative in much HPSG work to avoid unecessary proliferation of structure. Mostly this has manifested itself in a high degree of skepticism towards empty categories. But I suppose one may ask if this principle shouldn't also be applied to vestiges of 80s-style X-bar syntax, which, incidentally, seem to have been abandoned by Chomsky as well. I find the existence of unary expansions somewhat ironic as HPSG and LFG used to make a big deal of the fact that they are lexicon-driven in a way that pre-"Bare Phrase Structure"-TG was not. I'm sure I missed something. --Andreas @phdthesis{ author = "James Blevins", year = {1990}, title = "Syntactic Complexity: Evidence for Discontinuity and Multidomination", school = {University of Massachusetts, Amherst}} @incollection{ author = "Erhard Hinrichs and Tsuneko Nakazawa", year = "1994", title = "{Linearizing Finite {AUX} in {G}erman Verbal Complexes}", booktitle = "German in Head-Driven Phrase Structure Grammar", editor = "John Nerbonne and Klaus Netter and Carl Pollard", pages = "11--38", publisher = "Stanford: CSLI Publications"} @phdthesis{ author = "Andreas Kathol", title = "Linearization-Based {G}erman Syntax", school = "Ohio State University", year = "1995"} @phdthesis{ author = "Tibor Kiss", year = "1992", title = "Infinite Komplementation: Neue Studien zum deutschen Verbum infinitum", school = "Universit{\"a}t/Gesamthochschule Wuppertal"}
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:33:04 PST