Mark Johnson (Mark.Johnson@xerox.fr)
Thu, 7 Sep 1995 13:45:56 +0200
Oh dear, I was trying to write something ecumenical, and I wind up being flamed by two of the big guns in the field! Ivan and Geoff seem to have interpreted my posting as an attack on HPSG and support for the Minimalist program, which it was not. It seems as if Geoff in particular interpreted me as saying something quite different to what I actually said. If you read my message, you will see that rather than advocating any particular theory, I suggested a far more conservative stance: we merely try to understand how particular proposals are related to each other. My main point was that there are conceptualizations of grammar in which phrase-structure does not play a crucial role. I made the observation that in theories developed in the 1980s (which I described as ``older'', ``classical'', ``static'' and ``structuralist'') phrase-structure plays a central theoretical role: it is regarded as an entity which is required to satisfy certain constraints (usually involving a kind of universal quantification: e.g., every NP must have a sequence of children of such and such categories, every NP must receive Case, etc.). This conceptualization of phrase structure is in contrast to other theories of grammar. I mentioned Lambek Categorial Grammar (LCG) and the Minimalist program (MP). There are major differences, both in form and content, between LCG and the MP. E.g., work in LCG usually places a lot of emphasis on formalization, while much of the work in MP is done at such an informal level that it can be hard to identify just what is being claimed. (As everyone reading this knows, there has been a debate on the value of formalization in linguistics which I really think is irrelevant to the points I am trying to make: suffice to say that I think that formalization is useful when it claifies or leads to insights into the linguistic problems being studied. It is not clear to me that lack of _precision_ is the greatest problem of modern linguistics; I suspect that lack of _accuracy_ is far more serious.) Returning to LCG and MP; the thing I was interested in that they have in common is the following: neither regards phrase structure as an entity which is required to satisfy constraints involving universal quantification over nodes. Since the work in LCG and MP I am referring to was mainly done in the 1990s, I think it warrants the term ``newer'' in comparison to the work above. The term ``dynamic'' refers to the fact that both LCG and MP can be conceptualized in terms of processes rather than static constraints. This shows up in the LCG proof theory, for example. (LCG also possesses a declarative semantics, e.g., in terms of residuation, although it is not clear if this is linguistically insightful). In MP, it shows up in the way that phrase-structure construction (the process ``merge'') and movement (the process ``move'') are intertwined; there does not seem to be a natural way to conceptualize MP in terms of a single phrase-structure tree (or a tuple of trees) in the way that classical GB could be, for example. (I wish I could do this, since most of my work on parsing assumes such a classical, static perspective). The term ``resource-based'' refers to the fact that in both LCG and MP the processes just mentioned take place in order to satisfy the resource requirements of linguistic entities. This is explicit in LCG; a transitive verb is literally written as something which needs a subject and an object to become an S. This idea also appears in the MP, where the processes ``merge'' and ``move'' can only apply in order to discharge feature requirements on the constituents involved (e.g., an NP moves in order to get Case, etc.). Incidentally, this leads to a very different conceptualization of features. E.g., an NP that needs Case might have a feature -Case, while a Case assigner might have a feature +Case before the movement applies, but the movement operation discharges (i.e., cancels or deletes) these features. Thus the features associated with a constituent change continuously during a derivation (one of the reasons why it is hard to recognize a static phrase structure in MP). One of the things that I find most appealing about the dynamic approaches is that linearity is ``built in'', i.e., it does not need to be stipulated, as in static structuralist approaches. Natural language does seem to exhibit a lot of linear behaviour, e.g., WH-fillers typically bind exactly one gap, predicates take a fixed number of arguments, etc.; this sort of linearity is natural given the resource accounting mechanisms of the dynamic approaches. Non-linearity can originate from lexical items (e.g., a reflexive might be treated as a detransitivizer), but constructions with no obvious lexical source for the non-linearity (e.g., parasitic gaps, relativizer-less relative clauses) might be problematic. (It is true that there is no particular problem coding up mechanisms for doing this resource accounting in classical structuralist theories: in my humble opinion, SUBCAT and SLASH lists are nothing but the ``state'' of such a resource accounting mechanism). As far as I know, currently the classical structuralist approaches have greater linguistic coverage than any of the newer dynamic theories. However, I suspect that most of the intellectual action in the next decade or so will be in the dynamic approaches. This is because we understand fairly well already just what kinds of linguistic phenomena can (and cannot) be analysed using the classical methods, while the field is wide open for the dynamic approaches. Ivan's message refers to another message by Bob Carpenter which I have not seen, so I don't know exactly what claims I am supposed to have made. Never the less, I agree with the reprinted claim attributed to me; let me explain why. Much of my work focusses on developing appropriate constraint languages for expressing linguistic constraints (within what I called above a ``structuralist'' approach). I started with attribute-value (AV) constraints, but it turned out that the computational problems (e.g., satisfiability) for AV constraints are no easier than those for a much more general class of first-order formulae (which can express a kind of set value, tree structure constraints, etc.). I also noticed that there is no particular reason to restrict attention to local-tree constraints. I have worked on methods for transforming or compiling non-local constraints into equivalent (but more complicated) local constraints for parsing purposes; it is in this context that SLASH passing can be seen as the result of compilation of non-local constraints. Of course, the fact that this _can_ be done does not mean that it _should_ be done; it could be that the best description of the non-local filler gap dependencies is in fact in terms of the interaction of purely local dependencies. Ivan's message is exactly to the point here. Now most theories do posit mechanisms for ``marking'' the path of a filler-gap dependency, e.g., subjacency, forcing a COMP to COMP (or in later versions of the theory, Spec of COMP to Spec of COMP) movement. Assuming Spec-Head agreement and feature percolation, it does not seem inconsistent with (classical) GB for a verb's form to vary depending on whether a WH-dependency passes through its complement. But in order to draw meaningful conclusions each case needs to be examined in detail, of course. Well, I hope that this has clarified my original message somewhat; Best, Mark
This archive was generated by hypermail 2.0b3 on Fri Dec 18 1998 - 20:33:03 PST