Comments on Pullum
Ad Neeleman · Mind & Language · 2013
My comments on Pullum's article will follow after a considerable detour. But for those who want to know where I am headed, let me summarize my criticism in a sentence. In my view, several of the arguments that Pullum presents in favour of a constraint-based model of grammar are weakened considerably by the fact that not enough consideration is given to the distinction between competence and performance. (Found at http://faculty.atu.edu/mfinan/2033/section12.pdf) The central conclusion of Pullum's article is that it is better to conceive of syntactic theory as a set of constraints (MTS) than as a generative procedure (GES). At first sight, this seems to neatly match the model in (4). After all, constraints are static and are therefore most naturally conceived of as belonging to the most abstract level of description. Given that parsing and generation of linguistic utterances must be procedural, one might be inclined to think of procedures as belonging to the intermediate level of description. A: merge with weak inclusiveness and binary branching built in B: merge with strong inclusiveness and binary branching built in C: merge with strong inclusiveness, binary branching and label assignment under directionality built in D: 2n E: merge with strong inclusiveness, binary branching, label assignment under directionality and selection built in However, neither the fact that a constraint-based account of the language faculty must belong to the computational level of description, nor the fact that an account at the algorithmic level of description must be procedural, implies that an account at the computational level must be non-procedural. It is possible that the most insightful description of the formal properties of the way in which language associates sound and meaning is in terms of a generative procedure. All that the model in (4) tells us is that such a procedure would not have to map transparently onto the procedures that are relevant to parsing and generation. We should let the facts decide between these different views. But what facts? It stands to reason that the parser will have evolved to be as fast and as robust as possible. Both properties can give rise to misleading results, in that they may lead to twists in the relationship between grammaticality at the computational level and acceptability as experienced by subjects asked for their intuitions. In all likelihood, the reverse situation also exists. Some perfectly grammatical structures are hard to parse (for instance because they give rise to garden path effects), or in fact impossible to parse. It has been argued that certain cases of multiple centre embedding may fall in the latter category (that is, the parser is held responsible for the unacceptability of examples like the cheese that the mouse that the cat caught ate was imported from France). Thus, hypotheses about the grammar do not directly predict grammaticality judgments. They predict grammaticality judgments in conjunction with a theory of performance about which relatively little is known. For reasons of practicality, linguists must therefore work under some idealization of the influence of the performance systems on grammaticality judgments. The standard assumption is that this influence is negligible (compare the quote in (6)). However, some linguistics have made other assumptions that are usually not explicitly presented as such and that assume some degree of transparency between the descriptions of the language faculty at the computational and descriptive levels. The idea is that, at least in some cases, there is systematic association between the level of experienced unacceptability and the grammatical principle that is violated. For example, in movement theory, violations of Subjacency are taken to give rise to weak unacceptability, as compared to violations of the Empty Category Principle (see Chomsky, 1986, among others). If correct, this could be explained in a model like (4) by saying that there is a repair strategy for (certain) Subjacency violations, but not for ECP violations, leading to subjects having different experiences of two types of examples that are both ungrammatical at the computational level of description. While this is a perfectly reasonable way to proceed, the reality is that we cannot know before hand where the data reflect properties of the grammar (as instantiated in the performance systems) and where they reflect aspects of the performance systems that have to do with efficiency and robustness. In other words, in testing hypotheses about the grammar additional hypotheses about the implementation of the grammar in the performance systems must be made, and what is tested is this constellation of hypotheses, rather than grammatical theory on its own. There is nothing particularly remarkable about this, except perhaps that it is rarely commented on in the literature. It seems to me that the issue explored in the previous sections is particularly relevant to the arguments that Pullum presents in favour of a constraint-based grammar (or a model-theoretic syntax, to be more precise). These arguments are built on general properties of linguistic data that—in the absence of a thorough analysis of individual examples—may not reflect the grammar itself, but could just as well be contributed to limitations of the performance systems as a tool for generating grammaticality judgments. I first consider the argument from gradience of ungrammaticality. The idea is that the standard derivational model of grammar cannot deal with the fact that some sentences are neither fully acceptable, nor fully unacceptable. This is because a generative procedure, by its very nature, can either produce a string or not produce a string. There is no middle way. A constraint-based grammar, however, allows strings that violate certain constrains, but not others. If the observable data were direct expressions of competence grammar, this would be a very strong argument. But as argued above, the data are not. A subject's experience of a test sentence does of course depend on whether that sentence is grammatical (that is, whether in parsing it can be covered by a single connected structure). However, to repeat, it also depends on how easy it is to find a covering structure (if one exists) and how easy errors are detected (if the sentence is ungrammatical). If no covering structure is found, the subject's experience will further be affected by whether or not the substructures the parser builds up can be integrated though discourse mechanisms and whether or not there are repair mechanisms that can fix the problem. Given this range of factors, we expect, irrespective of the nature of the competence grammar, to find variation in grammaticality judgments. In particular, gradience of judgments is perfectly compatible with a description of knowledge of language in terms of a generative procedure. As the example of Subjacency violations versus ECP violations in the previous section demonstrated, this is true, even if the variation in judgments is taken to reflect violations of specific grammatical principles. The same general issue affects Pullum's other arguments. Take the fact that people are able to give grammaticality judgments of sentence fragments. Admittedly, this fact cannot be accounted for by a competence grammar that employs a generative procedure that must start with, or terminate in, an S-symbol. But that is not a problem, as the relevant fact will have to be accounted for in any case by any description of the language faculty at the algorithmic level. It is well known that parsing is incremental (see Gorrell, 1995 and references mentioned there). People are able to recover an interpretation for incoming language word by word. This implies that the performance mechanisms, while designed to operate in accordance with the rules of grammar, must be able to deal with incomplete structures. But if an account is available (or must be developed) for the interpretability (and acceptability) of incomplete structures at the algorithmic level, we are not required to provide an additional account at the computational level (that is, as a part of our competence grammar). Similarly, there is no problem for a conception of the syntax as a generative procedure arising from the lexical independence of grammaticality judgments. Lexical dependence is incorrectly predicted to exist by models of the language faculty in which the grammar is seen as a generative procedure and is held directly responsible for parsing. But there is no need to make this combination of assumptions. The problem dissolves if parser and grammar are taken to be descriptions of the language faculty at different levels. The robustness of the parser makes it likely that the parsing process will not terminate if a terminal is identified whose phonology does not match an existing word. In fact, this must be so in view of the phoneme restoration effect. In case the input contains a phonological form that is not part of a speaker's permanent lexicon, a good strategy might be to store the relevant form in a ‘temporary lexicon’, and to try and identify a meaning for it (presumably, this is how new words are learned). Again, this can and should be dealt with at the algorithmic level; there is no need to burden the competence grammar with it. The idea behind the argument is that a generative procedure either produces a well-formed output or ‘crashes’ and therefore produces no output at all. In (11) there is no grammatical way to express the intended content and yet alternatives do not seem fully ungrammatical either. Quandaries of this type, Pullum suggests, are therefore incompatible with a conception of the grammar as employing a generative procedure. But it seems to me that this argument is really not very different from the argument from gradience of ungrammaticality, and my reaction is therefore not very different either. As long as the grammar is taken to be a characterization of the language faculty at the computational level, it is possible that there are cases in which a given semantic content cannot be expressed by a structure of a particular form. But in view of the robust nature of parsing, it is entirely possible that in performance some of the ungrammatical candidate structures can be associated with the intended interpretation through repair mechanisms of various types, giving rise to an experience in subjects that is neither one of grammaticality nor one of ungrammaticality. None of this means that competence grammar is not constraint-based. It just shows that we need better arguments to decide the issue. One kind of argument that I personally find intriguing is frequently used in Optimality Theory. In this constraint-based theory, the evaluation procedure for candidate structures is defined in such a way that a set of constraints used to describe a given language automatically generates a language typology, thus increasing testability. Such typological predictions of course do not follow from any known generative procedure.