To what extent are the grammatical conventions of human languages shaped by performance? In particular, can rules and principles of the syntax be explained by considerations of processing, and by the temporal sequencing of speech? There has been some discussion of this question in the past, among linguists and psychologists, and the consensus seems to be: only to a limited extent. That is, processing may nibble at the edges of the grammar but does not appear to be a major determinant of syntactic conventions.

I believe that this consensus is mistaken. It comes from looking at insufficient data from too few languages, and from accepting too uncritically certain common assumptions in theories of grammar and in theories of performance. When we expand the data and the language sample, and pursue other theoretical assumptions, we reach a very different conclusion: namely, that grammars are profoundly shaped by processing. I believe that even highly abstract and fundamental properties of syntax can be derived from simple principles of processing that are needed anyway, in order to explain how language is used.

This conclusion has been reached after a close analysis of linear ordering in performance and in grammars. It is further motivated by the correspondences between on-line procedures for recognizing constituent structure and the grammatical devices that make such recognition possible. It is also supported by grammatical constraints on relativization, on movement, and on various other phenomena across languages. As a result, performance now joins the core of an explanatory model of syntax.

In this chapter I shall ask how a grammar could, in principle, respond to processing. Grammars have their own basic format of grammatical categories and principles, as defined within a given theory or model. So when we talk about a grammar responding to a principle of performance, or “grammaticalizing” that principle, we have to view this interaction either within the context of some given format, or (as we shall do here) within the context of general properties of many such formats, if one wishes to remain more theory-neutral. The ultimate explanation for a format or formats presumably reduces to some interaction of the innateness, cognitive-semantic and other explanatory primitives underlying universals that were summarized in ch. 1.5. Performance principles contribute to this format and to the associated grammatical principles, but they do so relative to restrictions that are imposed by the other primitives, and they will, in turn, impose restrictions of their own on other explanatory forces. As a result, the precise manner in which a given performance principle can be expressed in a grammar is complex, and the translation from performance to grammar may not always be one-to-one. I begin this chapter by outlining some general ways in which grammars appear to have responded to processing. This enumeration is not intended to be exhaustive. Rather, it serves merely as a background to the major types of grammatical responses for which I shall argue in this book.

A processing approach to syntax

Since the publication of Chomsky (1957) we have seen the development of a rich tradition of generative research into syntax. This tradition began by incorporating structuralist insights (cf. Saussure 1916; Bloomfield 1933; Harris 1951) into the kinds of grammatical rules that were widely employed in traditional grammars such as Jespersen (1924), thereby making these latter more explicit and more predictive in relation to grammaticality data (cf. Chomsky 1956). It led to a hypothesis about the role of grammatical knowledge within a larger theory of performance (cf. Chomsky 1965). It has contributed fundamental new insights into basic properties of natural language syntax and also into semantics, e.g. “constituent-command” (c-command) (cf. Reinhart 1983). It has elevated the goals of grammatical analysis beyond the facts of individual languages towards the construction of a theory of Universal Grammar (or UG), with built-in parameters of variation (cf. Chomsky 1981, 1982). And it has led to the formulation of numerous alternative models based on competing foundational concepts, whose empirical consequences are currently being worked out (cf. e.g. Aoun 1985, 1986; Gazdar, et al. 1985; Chomsky 1986; Blake 1990; and the papers in Baltin and Kroch 1989).

The philosophy that has been proposed as an explanatory background to, and motivation for, these generative descriptions is the innateness hypothesis (cf. e.g. Chomsky 1968).

In the preceding chapters I have presented evidence and arguments in favor of the following generalization: many fundamental and abstract structural properties of grammars can be explained by simple considerations of processing ease. In other words, grammars are in large measure performance-driven, and we might use the term “Performance-Driven Grammar” to refer to this research program. This conclusion has been motivated by a close analysis of linear ordering in performance and in grammars. It is further motivated by the correspondences between on-line procedures for recognizing constituent structure and the grammatical devices that make such recognition possible. It is also supported by grammatical constraints on relativization, on movement, and on numerous other phenomena across languages.

This conclusion places performance at the core of an explanatory model of syntax, and departs from a long research tradition in which performance has been viewed as playing either no role, or only a peripheral role, in the core grammar. The role of an innate UG is correspondingly reduced, while that of (ultimately innate) processing mechanisms is increased. The processing mechanisms in question are those that make it possible for humans to recognize grammatical structure in a rapid and efficient manner. They can be seen as the syntactic counterparts of the highly efficient (and again innate) mechanisms for sound and word recognition that have been experimentally validated in numerous studies.

In ch. 3.6 I formulated three sets of predictions made by EIC for grammatical conventions involving linear order: predictions for basic orders (3.6.1); for grammatical rearrangements of basic orders (3.6.2); and for implicational hierarchies of linear orderings (3.6.3). These predictions are derived from the assumption that there is a precise correlation between the data of language performance and the grammatical conventions of the world's languages. The explanation that has been proposed for this correlation is processing complexity. In the area of linear ordering, the metric of complexity is that of EIC. More generally, I argued in ch. 3 that the notion of a Constituent Recognition Domain (CRD), upon which the linear ordering metric is defined, is just one instance of a Structural Domain (SD) of structurally related nodes, whose relative size and complexity determines the applicability of numerous grammatical operations, such as relativization and extraction, pronoun retention, and so on. Whenever there are differences in structural complexity between different points in a syntactic tree, or between alternative structural renderings of one and the same semantic proposition, I claimed that we should see a “performance-grammar correlation” (cf. ch. 2.3): grammars will conventionalize the easier structures before they conventionalize the more difficult ones (hence the cut-off points on various hierarchies), assign more surface coding to arguments in complex structural positions than to those in simpler positions (in order to make semantic and syntactic interpretation easier in performance), and grammaticalize those linear orderings that result in faster and more efficient structural recognition in performance.

In this chapter I shall test the predictions made by EIC for language performance formulated in ch. 3.5. Some preliminary evidence taken from the published literature and involving psycholinguistic experiments, acceptability intuitions, and text frequency counts has already been summarized in ch. 3.3. This evidence did not test EIC directly, however, and was limited in the range of languages considered, just English and Japanese. In this chapter performance data will be considered from the following ten languages: English, German, Greek, Polish, Rumanian; Finnish, Hungarian; Turkish; Japanese and Korean. These languages are typologically diverse and exemplify all the structural possibilities set out in table 3.2. They also represent four distinct genetic stocks (in the classification of Ruhlen 1991: 380): the first five are Indo-European and cover four subfamilies (Germanic, Greek, Slavic, and Romance); Finnish and Hungarian are from the Finno-Ugric branch of Uralic; Turkish is from the Turkic branch of Altaic; and Japanese and Korean belong to the Japanese-Korean-Ainu grouping which Ruhlen now considers separate from Altaic. These languages exhibit considerable freedom of word order, though they also have many grammaticalized orders as well, especially English.

The EIC was explicitly tested on nine of these languages (all but Greek), both by myself and by collaborators at the University of Southern California and in the Constituent Order Group of the European Science Foundation Programme in Language Typology.

In ch. 3.2 I defined two parsing principles, Mother Node Construction (MNC) and Immediate Constituent Attachment (ICA). The former is a construction principle: it defines a method for recognizing abstract phrasal groupings on the basis of terminal elements within a parse string. Without such a principle, constituent structure would not be recognizable on-line, there would be no such thing as a Constituent Recognition Domain (CRD), and so on. Clearly, the construction of abstract syntactic nodes, and the correct assignment of words to their appropriate dominating nodes, is a vital prerequisite for the everyday use of human language.

This chapter will consider the theory of node construction in more detail. MNC is an extremely general and all-pervasive principle which has been largely sufficient to illustrate the wide range of ElC-derived word order predictions presented in chs. 4 and 5. But the construction of a higher node by an immediate daughter is not the only way of adding non-terminal structure to the terminal categories of the parse string, and I have already alluded to the existence of some additional construction principles in previous chapters. The purpose of this chapter is to define these principles and to consider the predictions they make for variation in both category selection and ordering across languages. A further goal is to consider how grammars have responded to the parser's need to recognize higher constituent structure in a rapid and efficient manner.