Journal of Language and Linguistics Volume 1 Number 2 2002 ISSN 1475 - 8989

1. Theory and practice in the concept of description

1.1. If we agree to use our terms quite broadly, we can define a language to be a general theory of human knowledge and experience, and discourse to be the set of practices for working out the theory (cf. Sapir 1921; Hartmann 1963; Halliday 1994). Language would be a theory - or a whole network of criss-crossing 'theories' - for representing our world and ourselves and each other in the world, and for constructing alternative states of the world or alternative worlds. We understand each other insofar as our theories of our language are similar in principle and get more finely tuned during discourse (Beaugrande 1997a).

1.2. The relations between theory and practice would logically constitute a dialectic, being an interactive cycle wherein two sides guide or control each other. When the dialectic is working smoothly, the practice is theory-driven, and the theory is practice-driven; the theory predicates and accounts for the practice; and the practice specifies and implements the theory (Fig. 1).

The real-life practices of discourse are strongly 'theory-driven' in obliging the participants to 'theorise' about what words mean, what people intend, what makes sense, and so on. Indeed, discourse is the most theoretical practice humans can perform, and also the most efficient and effective in using the least effort for the most goals. In return, language is the most practical theory humans can devise, offering the resources to shape and guide almost any of our practical activities.

1.3. Yet the 'theoreticalness' of language is dexterously concealed from the majority of speakers who practice it. If asked, they would probably describe discourse as a thoroughly practical matter; they would be surprised if we told them they possess a 'theory of their language' that gives them the status of 'theoreticians'. No doubt the theory can be practised so efficiently because many operations function below the level of conscious awareness; in return, the nature and organisation of the theory are difficult to determine or describe by means of introspection alone (but cf. 1.8ff; 3.36f; 4.4).

1.4. Moreover, a language is a unique type of theory. It cannot be conclusively verified or falsified in the conventional manner of a scientific theory, because we cannot adduce some language-independent testing grounds, such as a set of free-standing meanings for which the language could be judged a valid or invalid expression. Instead, language is a theory that partially creates and constitutes what it postulates, and thus tends to confirm itself. For practical purposes, we normally take things to be what our language calls them. When we wish to express them more validly, we can practice our language more elaborately; we cannot suspend its practices and go to meanings or things without it. We cannot get outside language to inspect it.

1.5. By the definitions proposed above, a 'theory of language' expounded in modern linguistics would more precisely be termed a meta-theory, whereas the discourse we produce to expound the theory would manifest our own meta-practices. "The constructs or schemata of linguistics" could thus be described as "language turned back on itself" (Firth 1957 [1950]: 190). This convolution renders linguistics unique among the sciences. We set about formulating an explicit theory of language whilst we already sustain an implicit theory as language; and our formulations are instances of practising the latter theory. Moreover, every explicit theory proposed so far undoubtedly falls far short of the richness and complexity of the implicit theory, though we may not be able to demonstrate just how.

1.6. Modern linguistics might in turn be characterised as a set of projects for rendering explicit the implicit 'theoreticalness' of language. Yet linguistics has been signally undecided about deriving its theories dialectically from the description of the ordinary practices of text and discourse. The most resolute position has been adopted in fieldwork linguistics. Providing descriptions of previously undescribed languages is by necessity practice-driven, since data in and about the language must come from observing the practices of native speakers. In addition, the fieldworker must subject every step in the theorising about the language to practical tests with informants. Achieving a reasonable fluency in the language demonstrates a practical competence that should plausibly enhance the authority of one's theoretical statements.

1.7. Still, fieldwork is theory-driven in its own ways. The linguist holds a general conception about possible types of language, e.g. whether one is "analytic" like Ammanite of Vietnam, or "polysynthetic" like Yana of California (Sapir 1921:142). The type is a high-level meta-theory directing attention to certain classes of features or patterns, such as "reduplication" to "indicate such concepts as distribution, plurality, repetition, customary activity, increase in size" or "intensity" (Sapir 1921:76). But the fieldwork linguist is always stimulated upon discovering some previously unknown feature or aspects, e.g. when Dyirbal of North Queensland was found to have a separate Dyanguy variety or dialect used only in the hearing of taboo relatives like a man's mother-in-law or a woman's father-in-law (Dixon 1968). Such discoveries are also of interest to neighbouring disciplines in the social sciences of sociology, anthropology, and ethnography (cf. 3.8; 3.40).

1.8. The opposite approach commonly goes by the name of 'theoretical linguistics' but might, for the present discussion, be more aptly called homework linguistics.1 It is heavily theory-driven, and presents invented data from well-described languages, notably English, of which the linguists are fluent or native speakers from the start. Instead of deriving the theory of a particular language dialectically by describing its practices, 'homeworkers' derive a theory of language in general by a theoretical bootstrapping that combines their own intuition and introspection with conceptions sporadically borrowed from language philosophy, formal logic, or mathematics (cf. 3.22). The standards of science are to be upheld by 'theorising' the more practical and ordinary qualities out of language. The most scientific statements should describe 'language' in the most abstract and general sense, and ultimately in terms of 'linguistic universals' (cf. 1.16, 20).

1.9. The decisive step in this outlook was to "give priority to introspective evidence" and "intuition" (Chomsky 1965:20). The homework linguist was now said to command an "enormous mass of unquestionable data" merely by virtue of holding the "linguistic intuition of the native speaker"; and precisely for these "data", a "description, and, where possible, an explanation" were to be "constructed" (1965:20). The linguist would apparently become the representative of the "ideal speaker-hearer in a completely homogeneous speech-community, who knows its language perfectly" (Chomsky 1965:4) (1.13). Yet to discredit fieldwork with informants, homework linguists felt impelled to deny that the "speaker of a language", who has "mastered and internalised a generative grammar, is aware of the rules of the grammar or even" "can become aware of them"; and that "his statements about his intuitive knowledge are necessarily accurate", since "a speaker's reports and viewpoints about his behaviour and competence may be in error" (1965:8). These denials should cast serious doubts upon authorising linguists to act as model "speakers", unless their academic training and status grant them super-human powers of introspection (1.12; 3.36). But then they would be patently untypical and unsuited as models of a "completely homogeneous speech-community".

1.10. Such perplexing lines of argument might help to explain why homework linguists have so often used data from a well-described language like English, besides just being native speakers. They could presuppose extensive information about the language and did not have to supply it. They could exploit their own intuition and introspection to swiftly elevate their deliberations up beyond the laborious problems of fieldwork in order to address purely theoretical rather than practical issues: theory becomes meta-theory, or, in the terms proposed here, meta-meta-theory; and their discourse on language manifests not just meta-language but meta-meta-language. So the discussion naturally seeks illustrations in invented data whose status seems so secure as to camouflage the role of the linguist as inventor, e.g.:

(1)	The farmer kills the duckling	(Sapir)
(2)	John ran away	(Bloomfield)
(3)	The man hit the ball	(Chomsky)

Paradoxically, such data were invented to seem incontestable, yet they can be empirically classified as non-authentic insofar as they do not spontaneously occur in ordinary discourse.2 Nonetheless, these same data, accompanied by rather cursory descriptions, have often been adduced to support general statements about the nature of language, e.g., that "word order is unquestionably an abstract entity" (Saussure) or that "grammar is autonomous and independent of meaning" (Chomsky). The essential paradox thus consists of basing a general theory upon special cases by expressly selecting data devoid of special features (cf. 4.2).

1.11. Moreover, non-authentic data represent an unannounced compromise between "langue and parole", or "competence and performance", which homework linguistics has separated by a radical dichotomy. Saussure had roundly asserted that "speech cannot be studied", "for we cannot discover its unity"; it is only a "heterogeneous mass" of "accessory and accidental facts" (1966 [1916]:9, 11) (cf. 1.21f; 3.13; 3.17). In the same vein, Chomsky (1965:4, 201) asserted that the "observed use of language" "surely cannot constitute the subject-matter of linguistics, if this is to be a serious discipline"; "from the standpoint of the theory", "much of the actual speech observed consists of fragments and deviant expressions of a variety of sorts". Such pronouncements suggest that authentic data do not practice theory of a language, but seriously disrupt it. The production of such data would resemble a catastrophic phase transition from the extreme order of language over to the extreme disorder of discourse. The speaker takes order, transforms it into disorder and transmits it to the hearer, who transforms it back into order. Made explicit, this account of the relation between language and discourse is obviously unsustainable.

1.12. In parallel, homework linguists announced that "the concrete entities of language are not directly accessible" (Saussure 1966 [1916]:110); and that "knowledge of the language" is "neither presented for direct observation nor extractable from data by inductive procedures of any known sort" (Chomsky 1965:18). These claims too were meant to discredit fieldwork linguistics. But they also imply an unsustainable account of native-language learning, namely struggling against the grain of what a child can "access and observe" - which is "fragmentary and deviant" anyway. This implication presumably helped to garner support for the universalist notion of an "innate language acquisition device" (Beaugrande 1997b, 1998a).

1.13. Once "actual speech" has been declared "heterogeneous" and "deviant", the linguist can proceed to invent non-authentic data which have been quietly rendered homogeneous and purified of all deviance. Similarly, if language is represented as an abstract, ideal system, then it is most expediently exemplified by idealised data. By implication, homework linguists do not represent ordinary speakers in real life, but rather "ideal" super-speakers who, thanks to their "perfect knowledge", can practice the language with far greater unity and purity (cf. 1.9).

1.14. The perplexities implied for linguistic description became most virulent in Hjelmslev's "prolegomena to a theory of language".3 Though acknowledging that "the linguist who describes a language" "uses that language in the description", he issued a plea to "rise above the level of mere primitive description to that of a systematic, exact, and generalizing science, in the theory of which all events (possible combinations of elements) are foreseen" (1969 [1943]:9, 121). The "theory" would be "applicable even to texts and languages" that have "never been realised, and some of which will probably never be realised" (1969:17). This startling project would be the linguists' equivalent of a theory of everything, or the grand unification theory currently much sought in physics. "The linguistic theoretician" proceeds to "discover certain properties present in all those objects that people agree to call languages, in order then to generalise those properties and establish them by definition"; by doing so "he decrees to which objects his theory can and cannot be applied" (1969:18). Such a "linguistic theory" "provides the tools for describing" "a given text and language", and "cannot be verified - confirmed or invalidated - by reference to existing texts and languages" (1969:18).

1.15. If these methods were literally adopted, the linguist must examine all the world's "languages" in the ordinary sense (that "people agree" about) and construct the theory solely out of those "properties" that have in fact been "discovered" everywhere. Then, it would trivially, indeed automatically apply to all languages without requiring any "decree", "verification", or "confirmation". Yet the set of properties would undoubtedly be far too small, abstract, and general to "provide tools for describing a text" (4.5). One could only describe the features that the text shares with every other text in every language, including languages that don't exist and never will - an esoteric exercise, to put it mildly.

1.16. When Saussure had earlier counselled "the linguist" to "acquaint himself with the greatest possible number of languages in order to determine what is universal in them", he had surmised that "the diversity of idioms hides a profound unity", and that "all idioms embody certain fixed principles that the linguist meets again and again" (1966 [1916]:23. 99). But he had conceded that "it is very difficult to command scientifically such different languages"; and wryly concluded, with immense understatement, that "the ideal, theoretical form of a science is not always the one imposed upon it by the exigencies of practice" (1966:99). Not so Hjelmslev, who conjured the ideal whereby "mere primitive description" would be replaced by "self-consistent and exhaustive description" (1969:9, 18). To judge from his published work, he never tried to present such a description of any text, and so did not confront its impracticability as a method.

1.17. To include all non-existent, merely "possible" languages, the set of languages to which Hjelmslev's "theory" could apply would be infinite; as a corollary, so too would be the set of "texts" to be "described". If so, the results of describing a text or a set of texts would always seem too restricted to claim genuine significance - just as homework linguists of the generative school would predict anyway (1.20). Yet, again by implication, the processes of comprehending a text would be infinite as well, which is blatantly false. Here, we see how far the requirements placed upon description vastly overreach actual language, even though, as I suggested, the theory falls far short (cf. 1.5). In parallel, the "competence" and "perfect knowledge" of the "ideal speaker-hearer" (1.9) vastly overreach the performance and knowledge of real speakers. Both overreachings render homework linguistics empirically vacuous: striving to describe everything at once and not describing anything.

1.18. I would argue this point just as emphatically for the definition of "language" as an "infinite set of sentences" (e.g. Chomsky 1957:13), presumably calculated to suggest that the description of data was not merely impracticable, but incapable in principle of ever leading to a theory of language (or to a "grammar"). Yet an "infinite set" would contain every conceivable sentence, including the most flagrantly improbable ones offered as counter-examples (like "colourless green ideas sleep furiously"). The paradoxes of the infinite inhabit imaginative prose, such as that of Jorge Luis Borges. In his infinite library:

Moreover, "performance" would require infinite search times. And it would be related to "competence" in purely accidental ways, just as, in the familiar parable, a roomful of chimpanzees with typewriters would, in infinite time, write the complete works of Shakespeare. Such is the proper mathematical meaning of the "infinite", and it cuts a theory of language off from all practices.

1.19. We can accordingly dismiss the reservation that descriptive linguistics is "inadequate" because "the corpus of observed utterances" is "finite" (cf. Chomsky 1957:15; 1965:67). This reservation holds for every set of observations and every set of data in every science. Only the finite can be observed; and data are, both by definition and by etymology, 'the given', and can never be other than finite.

1.20. The justified assessment should be that a language is manifested in a very large but always finite set of data; and that its system provides for indefinitely larger sets, which will also be finite at any time. No such set can ever be completely observed, but due to practical limitations rather than theoretical principles. Like all scientists who work with such large data sets, linguists must manage a trade-off between breadth (how much data a theory can describe) and depth (what degrees of detail and precision the description can achieve) (3.10ff). Now, if a language were an infinite set, then its description would entail an infinite breadth that flattens out our depth to an infinite shallowness, and our description (completed in infinite time, by the way) would capture only infinitesimal details. In practice, homework linguistics evaded its own "infinity" postulate by "assuming that the set of grammatical sentences is somehow given in advance" (e.g. Chomsky 1957:18, 54, 85, 103). Breadth was merely hypothetical, bootstrapped into the theory by invoking "language universals" "stated only in general linguistic theory as part of the definition of the notion 'human language'" (Chomsky 1965:6, 117), Breadth in the practical sense I suggest was left off the agenda, as when "gross coverage of data" was decried because it does not help a linguist "learn anything about the principles" (Chomsky 1982:82f).

1.21. We can also dismiss the reservation that "the corpus of observed utterances" is "accidental". Every science must confront the accidental in its data; the role of theory is not to leave real data aside and invent some data that suits it better, but to stipulate how we can distinguish between accidents and regularities (3.17). And the crucial requirement for doing so is to collect and collate data sets as large as current technologies allow. Of course, the state of technology is itself contingent upon accidents, e.g., whether funds are allotted for super-colliders in physics or for space telescopes in astronomy. But the capacity of technology to produce data has usually been well ahead the capacity of theory to account for those data - and nowhere more so than in linguistics today (3.2).

1.22. Moreover, science can enlist technologies precisely for coping with accidents in our data, most crucially at frontiers where our theories are still struggling to distinguish the accidents from the regularities (3.17). The more significant the potential for accidents, the greater the breadth we should seek, and the more we should deploy those technologies that increase breadth without materially decreasing depth. We may thereby push down the significance of any particular accident (or set of accidents) by reassessing its probability. Conversely, we may discover regularities when we can inspect a large set of data where we saw accidents before (cf. 3.8).

2. Recovering the dialectic

2.1. The issues raised in the foregoing section indicate that mainstream linguistics has not managed to capture the dialectical cycle displayed back in Fig. 1. In descriptive linguistics, the practices have usually run well ahead of the theories. Numerous steps and strategies actually applied in fieldwork research were entirely data-driven, and nowhere accounted for in the sparse linguistic theories of the times. Even Pike's (1967 [originals 1945-1964]) monumental programme to situate language within a "unified theory of the structure of human behavior" was fenced within the confines of behaviorism and 'unified science', which hindered him from expounding a unified theory of meaning (Beaugrande 1991:107-11). More recently, some significant and original phenomena discovered and described in fieldwork, as in Longacre's (1970, 1990) work on "spoken paragraphs" and "storylines", or in Grimes' (1975) work on the "thread of discourse", were nowhere accredited in linguistic theory nor mentioned in conventional linguistics textbooks. Either new terms were coined, such as "staging" and "collateral"; or else accredited terms were assigned unconventional meanings, as for "predicate" and "transformation".

2.2. In generative linguistics, in sharp contrast, the theories have run far ahead of the practices - so far indeed that practices seem to have been left behind altogether (Beaugrande 1998). Descriptive linguistics was sternly rebuked for not being theoretical enough, and, more specifically, for trying to construct theory out of practice, namely through the observation and analysis of data (Chomsky 1957). In respect to fieldwork, the rebuke was patently unfair: no other method can succeed when the linguist has no prior or outside information about the organisation of a language. What emerges is of course a theory about that one particular language, not about the "universal nature" of all languages. But within its modest scope, the theory has been vigorously tested by data (1.6), and can be retested whenever the data undergo a substantive increase.

2.3. In generative linguistics, the construction of theory became independent of the observation and analysis of data; on the contrary, these methods were expressly declared incapable of producing a theory (1.11) They could be bypassed precisely because the linguist as native speaker had so much prior or outside information about the language (1.10). But where then should the theory come from? In the event, it mostly came, impressively recast into more technical terminologies, from traditional grammar-books about that same native language. Thus, the "universality" of "phrase markers" was asserted, yet the accompanying diagrams displayed some obviously English-flavoured grammar-book categories like "definite" and "Article" (e.g. Chomsky 1965:107ff). Long before, Bloomfield (1933:233, 270) had warned against "linguists taking for granted the universal nature" of the "categories" of their own "native language". Now, real prospects arose of "forcing all languages into the mould of English, just as in earlier periods they were forced into that of classical Latin" (Hall 1968:53) (1.10). The relatively rigid word-order of English engendered the theory of "autonomous syntax". The absence of a systematic morphology in English led to morphology being left homeless in generative theory. And so on.

2.4. The dialectical nature of language and discourse was now thoroughly obscured. Language was not regarded as a theory which discourse puts into practice, but as a theory about a theory (a meta-theory about itself) which is independent of practice and indeed disrupted by practice. Paradoxically, these linguists discredited the data produced by ordinary native speakers as "fragmentary and deviant", yet accredited the data invented by themselves on the grounds of their own competence as native speakers (cf. 1.9ff). The data were invented precisely out of the theory - just the reverse of descriptive linguistics. Here, Hjelmslev's vision seems to come alive: a "linguistic theory" that "cannot be confirmed or invalidated" (1.15).

2.5. Perhaps the most far-reaching implication of this approach is that the very term "language" no longer refers to what most people, including most scientists, consider a language. Instead, it refers to a construct of linguistic theory so strenuously idealised it might ironically qualify as a Hjelmslevian "language that has never been realised and will probably never be realised" (1.14). How and why such a construct should promote the description of the languages that are being realised all around the world has never been convincingly expounded. Indeed, we could predict some compelling obstacles against description.

2.6. One obstacle lies in the terminology. The purely virtual status of "language" as a non-realised system at the centre of linguistics spreads out into the more specific terms. Such seems to have occurred with "syntax" as a formal system of rules which determine the word-order of all "grammatical sentences" in a language. Because real speakers put words in order for many motives quite unrelated to formal rules, this "syntax" does not exist in real language (Beaugrande 2000a). Still less does a "semantics" exist which assumes a fully stable, deterministic meaning for each expression of a language, whether based upon "meaning postulates" or "semantic features" (Beaugrande 1984). The virtual, non-existent status of these two "levels" or "components" of language makes them unsuitable in principle for the description of authentic data, whence the unquestioned substitution of non-authentic invented data (cf. 1.10ff).

2.7. The second obstacle is the peculiar meaning allotted to the term "description". When the operation of "assigning a structural description to a sentence" was equated with "generating the sentence" (Chomsky 1965:9), the formal analysis of data was equated with the original production of data, despite Chomsky's denials of doing so. Yet the categories of that same analysis are utterly insufficient for production, e.g., in taking no account of meaning during the "generative" stage. In effect, this "description" strips the sentence of most of its operational features and leaves a mere trace - not even a blueprint for the design, let alone a record of the implementation of the design.

2.8. Evidently, replacing "language" with a virtual construct leads to replacing "description" with a virtual operation. Here too is a motive for preferring non-authentic data: they are most amenable to just such an operation. A "transformational grammar" needs only the descriptive categories for converting the sentence into another more essential and general structure ("kernel", "deep structure" etc.). This operation does not even describe given the sentence itself, but analyses it away, and presents once again a structure which requires no confirmation because the theory had introduced it in the status of an axiom. So the description is effectively circular in the manner of a foregone conclusion.

2.9. If linguistics is to reinstate language as an empirical object of study, we must reassert its descriptive heritage and recover the dialectical interaction between language as theory and discourse as practice. These two sides must be seen to constitute a dynamic cycle between two distinct but closely co-ordinated modes of order. The order of language must be practice-driven and expressly designed to support the theory-driven order of discourse without fully predetermining it. So far, a large grey area persists between these two orders, comprising a host of constraints that are more specific or local than a language yet more general or global than a discourse (Beaugrande 2000b) (cf. 4.2).

3. The impact of very large corpora

3.1. For practical reasons, much corpus research based upon fieldwork in the past has had to be content with relatively small amounts of data. I can discover in that work no theories stipulating just how large a corpus ought to be; nor would such a theory be particularly relevant or interesting as long as the fieldworker may have to confront bizarre, fortuitous circumstances to get data. In his fieldwork on Cantonese in the 1940s, Halliday made speech recordings on cumbersome wire spools, and the breaking of wires would frequently damage or destroy his data.4 Improved technologies have reduced such mechanical dangers, but not the labours of transcribing and interpreting the data. Voice recognition by computer, now finally achieved, will help us only for transcribing data in those languages that have already been described extensively enough to configure the program; and transcribing data is just a partial step in analysing or interpreting it.

3.2. Today, corpus research has access to very large corpora of authentic data for a number of languages, and may confidently foresee many more in the near future. We now face a daunting decision about whether some established theory and practice of linguistic description will be reapplied to corpus studies; or whether the foundations of linguistics will be revised in light of corpus studies (Tognini Bonelli 1996; Sinclair 1999). As we know from the work on "scientific revolutions" in the philosophy of science since Kuhn (1970), a theory is not displaced by data alone but only by another theory which handles more data and extracts new and important insights from data. My own experiences in corpus research lead me to predict that linguistics should brace itself for a major scientific revolution or paradigm shift similar to those ensuing upon the introduction of such technologies as the telescope in astronomy or the microscope in biology (Sinclair 1994, 1999). As with other technologies, this one wields the capacity to produce data far ahead of the capacity of our theories to account for those data (1.21). To extend the analogy: we are 'seeing' phenomena in language which only become visible through the technology.

3.3. However, the technology also renders visible some far-reaching problems. These problems do not, as has sometimes been argued (e.g. Widdowson 1991), arise from weaknesses inherent in corpora. Rather, the problems have been inherent in language research all along but would hardly be addressed when data were either restricted by the practices of fieldwork linguistics or else marginalised by the theories of homework linguistics. Now, corpus research confronts us with principled questions like these:

What size should a corpus have in order to represent a language? What is the ratio between quantity and quality of data? What is the ratio between breadth and depth of description? What is the ratio between the uniformity and diversity of data? What is the ratio between regularities and accidents in data? What is the ratio between grammar and lexicon in a language? What is the ratio between manifest and underlying organisation of language?

These questions are so intricately related to each other that discussing any one of them by itself is an uneasy task. Even so, corpus research should eventually lead us toward some worthwhile answers through the aid of technology itself (cf. 4.6).

3.4. So our first question concerns the representative size of a corpus. The notion of an entire language having a quantifiable size at all hardly seems to figure in modern linguistics.5 It would of course be moot if language is defined as an "infinite set of sentences"; but I have tried to show why this definition is invalid (1.18).

3.5. Once a language is defined as a finite though very large set of data, and also a system providing for indefinitely larger sets (1.20), then our question concerns the ratio between the actual size of a corpus and its potential size. Actual size has been mainly dominated by practical factors. In early corpus research on computers, when the technology of memory and programming were rather limited, a million words seemed an ambitious size. When the technology advanced, practical motives were again dominant in bumping up the size to 20 million and then to 200 million in the Collins Birmingham University International Database (COBUILD) - familiarly called the "Bank of English" (BoE) - namely, for compiling a new type of data-driven dictionary that soon became the market standard. Then the corpora themselves were offered on commercial markets, such as COBUILD on CD-ROM (5 million) and the British National Corpus (BNC) from Oxford University Press (100 million).

3.6. This dominance of the practical side was to be expected. Lexicography has traditionally been a practical enterprise; and theoretical linguistics has focussed far more upon grammar than on the lexicon (cf. 3.11; 3.23). Even so, practical advances are still needed for more friendly technology at the users' end. Direct access to a corpus via the Internet is subject to multiple disturbances, such as lines being overloaded, busy, or periodically cut off in mid-operation. A corpus on a single CD-ROM (like the COBUILD's) can only hold a modest data set and do simple searches and calculations. For larger sizes and more complex searches like the BNC, users work with several CDs on ponderous operating systems like UNIX or LINUX, and require technical training in mastering systems like the "Corpus Data Interchange Format" based on "Standard Generalised Markup Language" (Aston and Burnard 1998).

3.7. But viewed from inside linguistics, theory is the side where advances are pressingly called for now. There, size leads us to the further question of the ratio between quantity and quality of the data. The null hypothesis would be that beyond some threshold (say, a million words), increases in size just multiply out in a mechanical proportionality: an item or pattern appearing once at 1 million words will appear 20 times at 20 million words and 200 times at 200 million words. But this hypothesis could hold only if a language were so uniform a system that its output hits a definite information ceiling and its features go asymptotic. Beyond that, quantity would rise whilst quality remained constant.

3.8. Corpus research, on the contrary, suggests a dialectical ratio whereby a major rise in quantity brings a rise in quality; so the language system must be far more diverse than the null hypothesis stipulates. New data can reveal previously undetected constraints upon an apparently unconstrained regularity. For example, most grammars of English, including the COBUILD Grammar based on a 20-million-word corpus, present the pattern of Definite Article plus Adjective for referring to a whole class of people, and declare it "possible to use almost any Adjective this way" (Sinclair et al. 1990:21f). But Sinclair (1998:86) recently reported "attitudinal biases and selectional restrictions" in the corpus at 336 million words: the pattern is mainly reserved for "unfortunate" people, such as the elderly, the injured, the unemployed, the sick, the aged, the poor, and the handicapped, as in (4). Fortunate people occurred mainly by way of contrast with the unfortunate, as in (5-6).

(4)	On services to the mentally ill, the elderly and the handicapped, Mr Cook pledged that Labour would appoint a minister for community care.	(newspaper)
(5)	This is a system in which the rich are cared for and the poor are left to suffer in silence.	(newspaper)
(6)	The appeal, especially in Latin countries, is rather to envy the fortunate than to pity the unfortunate.	(Bertrand Russell)

This "attitudinal bias" might be explained from the effect of depersonalising by omitting a Noun for the Adjective to modify. Such explanations may not be foreseen or admissible in established linguistic theories, but could be helpful for fieldwork research as well as ethnography (1.7), and also in the teaching of English (4.4).

3.9. New insights are both reassuring and disturbing. Just because linguists are stimulated when new regularities are discovered (1.7), we are troubled by the prospect of stopping the advance of theory by freezing the size of a corpus for practical or technological motives. This fate may befall when a dictionary or reference work arrives on the market, and the funding agent terminates support. Linguistics should therefore provide the public in general and user groups in particular with enough theoretical and practical knowledge to appreciate the dialectical ratio between quantity and quality. Only then will commercial markets be impelled to build larger corpora as grounds to claim better products.

3.10. Our next and closely related question concerns the ratio between breadth and depth of description (1.20). Whereas fieldwork research managed a balance by sheer practical diligence in describing authentic recorded data, homework research sought to appropriate "infinite" breadth and "universal" depth by sheer theoretical bootstrapping with handfuls of non-authentic invented data (cf. 1.7f; 1.20). So whereas breadth and depth were slowly achieved by fieldworkers through an arduous progress of small steps, they were swiftly built right into the theory of "language" by homeworkers.

3.11. Today, the very large corpus makes unprecedented breadth accessible but not necessarily achievable. The computer resembles a long ladder on which we are still learning the skills for scaling the higher levels in language description. Here too, much depends on how uniform or diverse a language system might be. For a highly uniform system, a description would have favourable chances to be both complete (total breadth) and precise (total depth). The closest approximation in actual language research is in the companion sciences of phonology and phonetics, sharing theory and practice in impressive accord. But their uniformity is a straightforward projection from the human vocal apparatus and the phonetic alphabet. In grammar, uniformity was brightly postulated in theory but never demonstrated in practice. And in the lexicon, the undeniable diversity has kept many linguists from undertaking research at all (cf. 3.23)

3.12. Breadth becomes a virulent issue when we get access to vast quantities of data. Depth becomes virulent when we must choose among sources for those data. Most descriptions produced in modern linguistics have been aimed at an entire language, e.g. at the "set of grammatical sentences somehow given in advance" (1.20). Data sources were not acknowledged to constitute a problematic factor, least of all when the data were invented by the linguists. The same depth of description would be appropriate everywhere, as would the methods for achieving it. In corpus research, this optimism soon breaks down. A language itself is by no means uniformly deep; the Number of Nouns is less deep than Definiteness: Polar Auxiliaries are less deep than Modal Auxiliaries. Reaching one depth is likely to open a view of still further depths, as when an analysis of the Agency of Verbs leads to the discovery of constraints on Pronouns as Subjects or Objects (cf. 3.32ff; 3.44). And the breadth of a deep description, once achieved, may be hard to determine, e.g. how many Verbs share constraints on their Agency (3.34).

3.13. By now we are in the midst of probing the ratio between uniformity and diversity in a language. Here too, linguistic theory has often inclined to a sharp dualism. Total uniformity was attributed to language, witness Chomsky's "completely homogeneous speech-community" (1.9); yet total diversity was attributed to discourse, witness Saussure "heterogeneous mass of accidental facts" (1.11). And theory nowhere explained how so extreme a dualism of order and disorder could inhabit the same system (1.11).

3.14. No doubt the heavy emphasis upon uniformity was intended to accommodate commonplace notions of science, but failed to recognise the uniqueness of language as an object of scientific investigation. There, uniformity and diversity constitute a dynamic dialectic, parallel though not identical to the dialectic between language and discourse. Every aspect of uniformity in a language must be designed to sustain diversity (cf. 3.41). In phonology, the uniformity of phonemes as shared targets underwrites enormous diversity among acts of pronunciation due to such factors as the age, gender, and emotional state of speakers, and their regional or educational background. In grammar, the functions of uniformity are different in modality due to their more complex and multimodal needs for expressing multiple modes of meaning. And the lexicon of English - in contrast to many languages - affords fairly modest and sporadic uniformity, due to its historical and cultural overlayering of extrinsic or specialised approaches to word-composition, e.g., borrowing roots from Latin and Greek.

3.15. Corpus research is now beginning to reveal the significance of the dialectic between uniformity and diversity. Language is found to be less uniform, and discourse less diverse, than linguistic theory is wont to assume. The uniformity of language is designed to generate diversity on-line; and the diversity of discourse continually refers back to and renews the uniformity of language (cf. 3.41).

3.16. In terms of corpus practice, uniformity may actually be a drawback. If we are compiling what Sinclair (1999) calls a 'generic or reference corpus' to cover the English language as broadly as possible, then we must consider how far the newly arriving data appear uniform or diverse alongside our already acquired data. The information value of a corpus would not rise significantly from increasing the store of uniform data of the same type. This problem applies especially to mass media, such as the plentiful newspapers conveniently posted on the Internet or made available by direct electronic transmission, like the Sunday Times. There, the diversity of the data is restricted in being authored by a relatively small, well-trained group of writers, and being edited by an even smaller group. I would also point out the massive ballooning of frequencies like I found in the BoE in July 1994 of key-words such as violence (19,226), kill (51,746), death (31,013), murder (18,383), rape (5,890), and assault (4,055),6 reflecting the morbid, voyeuristic interests of mass media more than the frequencies of authentic English at large.

3.17. Similar factors bear upon the ratio between regularities and accidents. Once again, linguistic theory has been largely dualistic: language constituted by regularities insofar as it can be an object of science; and discourse littered with accidents and therefore no fit object of science. Before very large corpora became available, projects for actually demonstrating regularities by means of statistic frequencies and probability measures were rare and laborious (e.g. Kucera and Francis 1967). Linguists gave reassurances that "a linguistic observer can describe the speech habits of the community without resorting to statistics" because "the forms of language" are "rigidly standardized" (Bloomfield 1933:37); or that when neither "sentences nor any part of them have ever occurred in any English discourse" or in "the linguistic experience of a speaker", they are "statistically" all "equally remote" (Chomsky 1957:17). These two reassurances flatly contradicted each other - data being all highly probable or highly improbable. But neither could be tested without powerful technology for measuring the ratio between regular and accidental (1.22).

3.18. The potential roles for statistics and probabilities are surely due for reassessment now that we have very large corpora (Halliday 1991, 1992). Finding and counting manifest items is most tractable, yet least informative. The frequencies of items in a corpus may give no reliable indication of their functional load in the language system. Finding exactly 6000 occurrences for of in the 5-million-word COBUILD Corpus on CD-ROM is not helpful; we need to know the proportions for each of its multiple functions in combinations. And combinations too are subject to the ballooning effects I noted a moment ago in news media. Among the 20,569 occurrences of sex returned by the BoE in July 1994, I found Sex Pistols (at 707), sex appeal (at 762) oral sex (at 203), and sex discrimination (at 209). Such frequencies are not meaningful unless we can determine how far the occurrences entail the 'same' item at all.

3.19. The frequency of manifest combinations is thus less tractable, but more informative. Corpus research has devoted much exploration to the typical lexical combinations called collocations, and the typical grammatical combinations called colligations.7 Yet typicality is not readily explained in terms of frequency alone. In my own 62.5-million-word corpora of British and American Writers, which I shall be citing further on, among 339 occurrences of the Verb fled were only 3 of the collocation fled the country. To my intuition, this combination seems typical even if its frequency and statistical probability are negligible. It also occurred just once among 99 uses of fled in the COBUILD on CD-ROM:

(7)

... after the collapse of Tsarist authority, opportunists declared an independent democracy, then a military junta that fled the country.

(book)

But I can draw some confirmation where the Verb fled took country names as Direct Objects: France, Iraq, Kuwait, Croatia, Germany.

3.20. We now come to a truly daunting question: the ratio between manifest and underlying organisation of language. Modern linguistics has been postulating an "underlying" organisation of language all along (e.g. Saussure 1966[1916]:56; Sapir 1921:144; Bloomfield 1933:225f; Hjelmslev 1969 [1943]:9f; Chomsky 1965:4f, 10, 18, 22). Among the grandest prospects was that the "descriptive grammars of diverse languages" will "some day" enable us to "read from them the great underlying ground plans" (Sapir 1921:144). Presumably, such "plans" are the goal of work on "linguistic universals", but most of that work lacks a secured base in descriptive grammars.

3.21. Moreover, linguistics has remained disturbingly evasive about how we can derive the "underlying" organisation from the manifest organisation. Thus, Chomsky's provision that "actual data of linguistic performance" would provide "evidence for determining the correctness of hypotheses about underlying structure" conflicted with his insistence that "surface structure" is "unrevealing" and "irrelevant" and "hides underlying distinctions" (1965:18, 24). With surprising candour, he conceded that his proposed "grammar does not, in itself, provide any sensible procedure for finding a deep structure of a given sentence"; and he evaded the whole issue by operating on the "simplifying and contrary to fact assumption that the underlying basic string is the sentence" (1965:141, 18).

3.22. Such evasions readily follow from the already noted tendencies to attribute to language highly idealised modes of order and to transpose the concept of language from the particular instance over to a universal abstraction (cf. 1.8, 13, 16, 20; 2.5). Doing so naturally fosters a readiness to see disorder in manifest data, and hence a reluctance to exploit them in the search for underlying order (cf. 1.11; 3.12). Instead, artificial modes of order get borrowed from sources like formal logic or mathematics, which only intensifies the idealised and abstract nature of "language" (1.8).

3.23. Here, we can highlight the ratio between grammar and lexicon. Linguistic theory has long regarded "grammar" as the epicentre of uniformity and regularity for an entire language and as a home for linguistic universals (compare Saussure 1966 [1916]:133, 152; Sapir 1921:38; Bloomfield 1933:163; Chomsky 1957:56). In exchange, linguists have long concurred that the lexicon is a mere "list of basic irregularities" (Bloomfield 1933:274; cf. Sweet 1913:31; Saussure 1966 [1916]:133; Chomsky 1965:86f, 142, 214, 216). On a smaller scale, this dichotomy re-enacts the dichotomy between the order of language and the disorder of discourse (1.11), and again linguistics has chosen order: much work on grammar, little on lexicon (3.6). Eventually, a homework linguist can baldly announce that "linguistics is not about language; it is about grammar" (Smith 1984).

3.24. Here again, linguistic theory should replace the dichotomy with a dialectical relation, this one co-ordinating grammar and lexicon and constituting the interactive lexicogrammar, the "semogenic powerhouse of language" () The two sides differ not in kind, but in degrees of delicacy: lower toward the grammatical side and higher toward the lexical side. Perhaps the lexicon could be regarded for some purposes as "most delicate grammar" (Halliday 1961:256; Hasan 1987:184; Cross 1993:199).8

3.25. The interactions of grammar and lexicon are readily evident from corpus research on colligations and collocations in the sense of 3.19. Since these are defined as typical combinations, they continually draw our attention toward plausible motives of speakers or writers for coordinating multiple selections. For example, the English Verb brook meaning "accept, tolerate" usually requires a Negative element (Sinclair 1994), as in:

(8)	Johnson could not brook appearing to be worsted in argument.	(Life)
(9)	Bouille rides, with thoughts that do not brook speech.	(French)
(10)	His work was of a sort that would brook no negligence.	(Lady)

This Verb is infrequently used, and preferentially in solemn language about some weighty business, as in Shakespearean drama:

(11)	This weighty business will not brook delay.	(Henry VI)
(12)	My business cannot brook this dalliance.	(Comedy of Errors)
(13)	False king, why hast thou broken faith with me, Knowing how hardly I can brook abuse?	(Henry VI)

This second constraint is more delicate than the one requiring a Negative, yet more difficult to define in terms of manifest lexical choices. The weighty business might be the assassination of a Duke (11), or just the collection of a debt (12). The weightiness comes in part simply from using brook rather than, say allow or tolerate.

3.26. Such data from the lexicogrammar of English point us toward the immense task of accounting for multiple parameters of variation in a language: genre, register, and style. In terms of theory, these constitute intermediary control systems between the language and the discourse (Beaugrande 1997eh?). Their design must be such that when one of them is activated, the activation level is raised for appropriate options and lowered for inappropriate ones (Kintsch 1988; Rumelhart et al. 1986). In terms of practice, they obviously affect the selections and combinations we can expect to find in authentic discourse data; but how to describe those effects is far from clear at this stage.

3.27. Here, we might pursue a strategy of dialectical resolution: building sub-corpora where we predict systematic distinctions in quality; and then using our findings to test and refine our predictions and to assess the typicality of specified data inventories as indicators of some genre or style (cf. 4.5f). For a brief demonstration, I shall draw upon my own British and American Writers Corpus, dating mainly between 1750 and 1920 and together, and including: (a) two corpora of literature, one by British authors (e.g., Austin, Dickens, Wilde) and one by American authors (e.g., Hawthorne, Mark Twain, Willa Cather), (b) two corpora of academic and civic writers, again including British (e.g., Charles Darwin, J.S. Mill, Mary Wollstonecraft) and Americans (e.g., Thomas Jefferson, Jane Addams, W.E.B. DuBois); and (c) a corpus of popular writers of adventures, mysteries, and children's literature (e.g. Conan Doyle, J.M. Barrie, Kate Douglas Wiggin). Theses of corpora, totalling all together 62.5 million words I compiled myself to run on WordPilot©, a resource program developed by John Milton at the Hong Kong University of Science and Technology (Milton 1999). My compiling too faced fortuitous practical restrictions: I had to rely on texts which are in public domain and can be downloaded from Internet sites.

3.28. In sources (a) and (b), the pattern of Definite Article plus Adjective was found to be more balanced than in the COBUILD data reported in 3.8. The highest frequency appeared among academic and civic writers, who are logically prone to classify people. Alongside the contrasts like those noted by Sinclair, e.g. (14-15), I found many where the fortunate people occurred alone, although sometimes with the intriguing ironic twist of not being secure in their good fortune (16-17).

(14)	Smile with the simple and feed with the poor? […]; let me smile with the wise, and feed with the rich	(Boswell, quoting Samuel Johnson)
(15)	None know the unfortunate, and the fortunate do not know themselves	(Poor Richard's Almanack)
(16)	There is always some levelling circumstance that puts down the overbearing, the strong, the rich, the fortunate, substantially on the same ground with all others	(Emerson)
(17)	the educated see a menace in his [the black man's upward development	(W.E.B. DuBois)

If grammar-books describe the pattern as being more general than is confirmed by contemporary usage in the COBUILD, then perhaps by intuitively taking academic discourse to be a model of English usage at large.

3.29. On the face of it, dialectical resolution might look circular: using the type to identify the features of interest, whilst using those features to identify the type. But text types cannot in theory be defined through rigorous proof, since in practice most types are defined through intuitive heuristics by language users. Besides, types are frequently mixed, as in:

(18)

A wedding is a time for merriment and an apt occasion to showcase age-old traditions in an age where modernity is eroding important aspects of yesteryear. This much-privy glimpse of Arabia was a re-enacted wedding ceremony of the indigenous people, reflecting the timeless beauty and simplicity of Arabia's life-styles, customs and unique identity until the '70s oil-boom brought in dramatic socio-economic development.

(Khaleej Times)

Such discourse briskly mixes the styles of solemnity (merriment, yesteryear), social science (modernity, indigenous, identity, socio-economic development), and tourism (age-old, timeless beauty and simplicity, life-styles), along with the occasional solecism (much-privy glimpse). The mix reflects multiple goals, such as disguising a tourist trap as a cultural site whilst flattering the readers' command of an educated variety of English here in Gulf States.

3.30. Another strategy might be for us to create local regions of substantial depth by describing narrow data sets with some thoroughness. The resulting insights might then be projected across broader sets and guide our selection of aspects and features to investigate. For example, the COBUILD data at 20 million words showed a Verb like elude being used only in the Active (cf. Sinclair et al. 1990:407), e.g.:

(19)	Newer techniques, such as bone-scanning and ultrasound, have enabled us to find more of the causes of back-pain, but a large number still elude us	(magazine)
(20)	Sylvie Guillem as Nikiya gave us her faultless technique and musicality, although the spirituality of the role so far eludes her	(newspaper)

In my literary and academic corpora I found elude in the Passive just six times, as in:

(21)	My importunities would not now be eluded	(Wieland)
(22)	they lessen the consumption; the collection is eluded; and the product to the treasury is not so great	(FedPap)

The meaning for data like (19-20) is roughly: some knowledge or skill would be fitting but is not found. The meaning for data like (21-22) is more like: some people finding ways of avoiding something. The Passive does seem to me intuitively old-fashioned; and Passive versions of these Actives seem utterly improbable:

(19a)	? ?	we are eluded by a large number of the causes of back-pain
(20a)	? ?	Sylvie Guillem is eluded so far by the spirituality of the role

3.31. Now, to increase the depth of our analysis of elude, we can examine some typical collocations and colligations. Among the Nouns as Direct Objects, the collocations noticeably clustered around vigilance, which occurred 9 uses, e.g. (23), along with associates like observation (24), eyes (25), and glance (25).

(23)	Nelson feared the more that this Frenchman might get out and elude his vigilance	(Nelson)
(24)	I had not neglected precautions to secure my personal safety, if I could only elude observation.	(Eyre)
(25)	That I could elude Rima's keener eyes I doubted	(Mansions)
(26)	Hare's fateful glance, impossible to elude	(Desert)

Other typical collocates included grasp (6 uses), e.g. (27), and pursuit (4 uses), e.g. (28).

(27)	the maiden eluded the grasp of the savage	(Last)
(28)	I stopped at one or two stands of coaches to elude pursuit	(Wrongs)

The meanings of all these collocations involve two opposing agencies, one of them seeking to elude the other and the potential consequences.

3.32. Among the colligations, the most striking one by far was a marked preference for Personal Pronouns as Direct Objects. Of the 17 occurrences in COBUILD data, 13 showed this colligation, as in (19-20). Other examples included:

(29)	he defines his essential position, as a man in permanent search of a God who eludes him.	(newspaper)
(30)	they were artists of considerable distinction, but man-in-the-street recognition has eluded them	(newspaper)
(31)	'River Lane,' said Shields. 'Clarke, of course!' That was what had been eluding him.	(book)

Here, a further meaning concerns the lack of some insight or knowledge; the Active Transitivity shifts the Agency in this lack from the person over to the knowledge.

3.33. The proportions among the colligations in my other corpora were less striking but still suggestive: out of 76 occurrences, 22 with Personal Pronoun Objects. Alongside an idea (32) or a fact (33), concrete agents like a person (34) or animal (35) did the eluding.

(32)	He spoke like one who was trying to keep hold of an idea that eluded him.	(Time)
(33)	Something seemed to give way in Jimmy's brain. The simple fact which had eluded him till now sprang into his mind.	(Damsel)
(34)	Although Sam haunted lobby and stairway and halls half the night, the fugitives eluded him	(Whirl)
(35)	All four boats gave chase again; but the whale eluded them	(Moby)

Only two eluded Agents appeared in Direct Objects as Nouns rather than Pronouns:

(36)	this whale eludes both hunters and philosophers.	(Moby)
(37)	often the Captain darted out of the shop to elude imaginary MacStingers [his landlady]	(Domb)

I am aware of no reference, in the linguistic literature on Pronouns, to classes of Verbs which colligate with Pronoun Objects, let alone any prospective theoretical account. Provisionally, we might describe such Verbs as expressions of Agent-Opposing Processes, which are usually accompanied by some preparatory background identifying the Agents. In some contexts, both Agents are persons (or animals), the Subject doing something and the Object eluding it. In other contexts, the Subject is not a person and hence a Pseudo-Agent, but some knowledge or skill that is lacking, and the Object is an Agent who does not have the initiative. In either type of context, the eluded Agent is often clear and can be designated by a Pronoun.

3.34. The next and much harder problem would be to explore how broad this locally detected constraint might be. Since a brute-force query of Verb + Personal Pronoun Object in a large corpus would be explosive, we can tap our intuition to suggest plausible candidate Verbs. By this means, my queries brought to light the Verbs rebuke colligating in the Active with Personal Pronoun Objects in 24 out of 51 occurrences; beseech in 94 out of 126; and thank in 121 out of 185. (Also, thank had a fair quota of Personal Pronoun Subjects, namely in 84 occurrences.) Similar measures of Personal Pronoun Objects were found with the Pseudo-Agent Verbs behove in 14 out of 19 occurrences; and befall in 108 out of 189. The data for befall showed a distinct and ominous attitudinal bias for choices of the Pseudo-Agent Subjects: the primary collocates were misfortune (at 26), accident (at 23), calamity (at 19), and disaster (at 10).

3.35. Using intuition in this way is far from proclaiming it to supply the "enormous mass of unquestionable data" invoked by homework linguists (1.9). Intuitions are always questionable, and the corpus makes the questioning easy. For example, my intuition suggested Verbs that the corpora did not display in the colligation patterns of rebuke in any significant proportions, such as reprimand (6 out of 27) and rebuff (3 out of 34).

3.36. Corpus research recasts the linguist: not in the role of the "ideal speaker-hearer in a completely homogeneous speech-community, who knows its language perfectly", but in the role of an ordinary speaker-hearer (and writer-reader) in a heterogeneous community, who knows its language only partially and actively seeks access to the knowledge of others. We claim authority for our statements not from harbouring super-human powers of introspection (1.9), but from examining large sets of authentic data produced by a community that puts their implicit theories of the language into a wide range of practices (cf. 1.3). And our statements are not about "language" as some "universal" abstraction, but about those data in one language and often about only one genre, register, or style (3.25). Such statements can easily be "confirmed or invalidated" by more or other data - a normal effect of the dialectic of quantity and quality (3.7) - but either step confirms once again the vitality of using authentic data.

3.37. Intuition and introspection are thus largely heuristic and opportunistic. They suggest things to try or watch for, and they help us determine status and meaning after the fact once authentic data are put before us (Francis and Sinclair 1994:194). They are not too reliable as sources of data, and still less as sources of information about the proportions among selections and combinations of data.

3.38. Allow me to demonstrate this point with one final data set. In July 1994, I found 515 occurrences of couldn't help and could not help in the Bank of English, then at 225 million words. My intuition led me to predict a fair quantity of data colligating with a Direct Object Noun for some Target person who could not be given assistance, but I found just four, not even 1% of the total. Here I encountered another phenomenon pointed out by Sinclair (1991:493f): the presumably basic stand-alone meaning listed in first place by conventional dictionaries not being at all the most frequent in corpus data. The meaning of help as 'give assistance to' is listed first in Webster's Seventh Collegiate (p. 387) whereas the meaning of 'refrain from' or 'avoid doing' is listed in seventh place. The design of such a dictionary would hardly admit a separate definition for not help or could not help, even though the meaning is demonstrably distinct.

3.39. The leading colligations by far in the COBUILD data were with Verbs: either a Present Participle (e.g. couldn't help admiring) or else with but + Infinitive (e.g. couldn't help but laugh). This I could have predicted, but not my finding that no Adverb ever came in between (e.g. couldn't help deeply admiring her) - a fully grammatical option, but not found (but cf. 3.45). In return, I found two less grammatical mixed patterns (couldn't help but thinking and couldn't help from crying) - the second one by the distraught Mary Wells, Tornado Victim.

3.40. Still less could my intuition have predicted the proportions among the collocations. Almost half of the total (at 234) collocated with one out of a set of just four Verbs; could you predict which ones? They were feel (at 68), notice (at 58), think (at 59), and wonder (at 49). Still, if I could not predict, I might 'retrodict' after the fact by noting that these Verbs represent Processes which might well be judged not properly subject to conscious control: they might lead into emotions, perceptions, and thoughts where it seems fitting to remark that someone couldn't help it. The pattern might therefore be termed a Face-Saving Auxiliary: an expression which attenuates the Agency of Process Verbs in order to save face after some Action that might be interpreted as hasty or inappropriate. Such an explanation may again not be foreseen or admissible in the theories of mainstream linguistics, but might be useful for ethnographers (3.8).

3.41. Moreover, these same frequent Verbs could also provide useful Headwords for most of the more delicate collocations, indicating one important way that uniformity is designed to support a diversity (cf. 3.14) The top-ranked feeling could be the Headword for attested collocations with crying, laughing/chuckling, smiling/grinning, blushing, fearing, liking, loving, marvelling, sympathising, wincing, worrying, plus nearly all the delicate collocates in colligation with being or be: touched, charmed, impressed, moved, emotionally involved, fascinated, struck, carried away, swept along, amused, jealous, puzzled, nervous, frightened, surprised, shocked, offended. Emotions might plausibly render you self-conscious, whether pleasant or unpleasant, witness also the list of Direct Objects or Modifiers collocating with the Verb feel in the data: the pleasant ones enthusiasm, passion, thrill, pleased, impressed, vindicated, and the unpleasant ones envy, guilty, ashamed, sorry, miffed, apprehensive, alarmed.

3.42. The slightly less frequent noticing could provide a Headword for seeing, looking at, glancing, hearing, overhearing, remembering, being consciously aware. Thinking could be the Headword for knowing, considering, reflecting, imagining, and could subsume the frequent wondering, where uncertainty rather than emotion might be making you self-conscious.

3.43. One group of collocates formed a cluster with no frequent Headword: speaking, saying, telling, commenting, pointing out, remarking, declaring, suggesting, responding, agreeing, objecting, reminding, congratulating, blurting out. Here we might pick the Headword by its generality rather its frequency: speaking being involved in all the others but not vice-versa (proverbially, one can speak without saying anything).

3.44. The colligating Subjects were evenly divided between Nouns and Pronouns. Yet the proportions among the Pronouns were dramatically uneven. I logged in far ahead at 150 occurrences, followed after a large gap by she (48) and he (45), and then after another gap by you (15), we (7), and they (6), plus the Impersonal one (11) - for a total of 282 Pronoun Subjects (55% of the total data). Here we may have evidence for constraints upon what we could call Multi-Process Agency, such that the identity of the Agent is established for one (or more than one) Process before saying that Agent couldn't help it.

3.45. The data in my two literary corpora gave a more delicate picture of these constraints. There, I registered 147 occurrences of couldn't help and 320 with could not help, for a total of 467. Also, those 320 constituted 86% of the 370 occurrences of not help. The frequency is highly significant if we consider that these corpora, at a total of just 8.7 million words, are about 25 times smaller than the COBUILD at 225 million, which returned 515. The most plausible explanation I can find - again not a "linguistic" one in any established sense - is the useful function for framing Events in literary discourse so as to communicate to the reader a character's own perspective, such as what someone was feeling or thinking, perhaps with no manifest Action, as in:

(38)	Connie stuck to him passionately. But she could not help feeling how little connexion he really had with people.	(Chatter)
(39)	Mrs Tulliver's imagination was not easily acted on, but she could not help thinking that her case was a hard one	(Floss)

The literary style might account for the attestation of inserted Adverbs, which never appeared in COBUILD data (3.39), such as:

(40)	She could not help frequently glancing her eye at Mr. Darcy	(Pride)
(41)	she could not help secretly advising her father not to let her go.	(Pride)
(42)	Florence could not help sometimes comparing the bright house with the faded dreary place	(Domb)

In some such data, there is no other reasonable place to put the Adverb.

3.46. The personal and internal quality might also help explain the tremendous frequencies, similar to those noted in COBUILD data, of First and Third Person Singular Pronouns as Subjects: I (151), he (75), and she (85), for a total of 311 (67% of all my data). The Plurals were rare -we (6) and they (5) - probably because a feeling or a thought normally belongs to just one Agent. The Second Person Pronoun you was rare too (4), doubtless because of the low probability of telling somebody else to their face what they couldn't help.

3.47. At still greater delicacy, I found that choice of the Contraction couldn't made a difference here. Whereas she and he were about half as frequent as for could not, I was more than twice as frequent:

	couldn't help (total of 147)	could not help (total of 320)
I	73 (49%)	78 (24%)
she	17 (11.5%)	68 (21%)
he	19 (13%)	78 (24%)

I checked all the data to see if the Contraction was preferred for spoken discourse. And in fact, only 14 out of 73 uses with couldn't did not occur in direct speech like (43), but in the narrator's voice of first-person narratives like The Adventures of Huckleberry Finn, e.g. (44); this last work alone contributed 7 uses, but then Huck never says could not in any context. Conversely, only 4 out of 78 uses with could not appeared in direct speech like (45); all the rest were in the narrator's voice, like (46).

(43)	'she took all the grit out o' him. I couldn't help feelin' sorry for him sometimes'.	(Fauntle)
(44)	I had to skip around a bit, and jump up and crack my heels a few times - I couldn't help it	(Finn)
(45)	'He was a very good man, sir; I could not help liking him'.	(Eyre)
(46)	For my part, I could not help thinking this lawyer was not such an invalid as he pretended to be.	(Clink)

3.48. Related constraints applied to the occurrences of the Pronoun it as Direct Object. The data with the Contraction logged 70 instances (47%), the data with could not a mere 19 (6%). Here also, the context tends to establish Identities: not for Agents and Targets, as for the Subject (3.44), but for Actions and States. The tiny frequencies of the third Person Pronouns her (1), him (1), and them (2) as Direct Objects again documents the rarity of the sense of help as 'give assistance to' (3.38). The few Nouns as Direct Objects were also expressions for Actions, not Agents, as in:

(47)	Connie could not help a sudden snort of astonished laughter	(Chatter)
(48)	'I couldn't help the interruption, but I made up for it afterward by working until two'	(Carrie)

I accordingly found a modest scatter of pairs with the same Action as Noun or as Verb, as in:

(49)	With such a possibility impending he could not help watchfulness.	(Caster)
(50)	Catherine, though not allowing herself to suspect her friend, could not help watching her closely	(Abbey)

The colligation with a Verb in the Present Participle was quite conspicuous with could not: 256 out of 320 (exactly 80%). For couldn't help, this colligation logged in at 61 out of 147 (41%), having to compete there with it at 70. Some authors used couldn't help exclusively with it, such as Mark Twain, Harriet Beecher Stowe, and Theodore Dreiser.

3.49. The matter of authors' preferences as compared to linguistic regularities is a puzzling one in corpus research. We might contend that my corpora are far too small, which is doubtless perfectly true, the more so given the sheer size of some single texts, such as Joyce's Ulysses at over 266,000 words. However, differences in size among sample texts is an important empirical given, especially when the public is expected to read the whole text. Besides, we cannot determine in advance how far an author or a text might be internally consistent enough to skew our measurements in one direction - Ulysses surely is not. The colligation depend upon it (meaning 'you may be sure') appears 55 times in my corpora, of which 28 come from Jane Austen; yet her usage was typical of whole sample, where fully 46 are Imperatives and 8 more colligate as you may depend upon it in the same meaning. The typicality was confirmed by data in my corpora of British and American academic and civic writers. There, depend upon it appears 23 times again as Imperative or with you may. 14 of them were uttered by Dr. Johnson in Boswell's Life, whose following item Sir in 12 occurrences can be safely charged to a personal idiosyncrasy.

3.50. At least as puzzling is the matter of translators' preferences as compared to linguistic regularities of multiple languages. The English colligation couldn't help plus Verb (51-52) does not show regular correlates in the German (51a-52a) or Spanish (51b-52b) versions of Alice in Wonderland, whereas French makes do with ne pouvoir s'empêcher (51c-52c). But the colligation couldn't help it has a separate correlate in all three versions (53-53c).

(51)	Alice was very nearly getting up and saying, 'Thank you, sir, for your interesting story', but she could not help thinking there MUST be more to come
(51a)	Alice war nahe daran, aufzustehen und zu sagen: 'Besten Dank für deine wirklich interessante Lebensgeschichte', aber dann sagte sie sich, daß doch noch einfach etwas kommen mußte
(51b)	Alicia estaba dispuesta a levantarse y decir: 'Gracias, señora, por su interesante historia', pero no pudo dejar de pensar que algo más iba a decir la Tortuga
(51c)	Alice fut sur le point de se lever en disant: 'Je vous remercie, madame, de votre intéressante histoire', mais elle ne put s'empêcher de penser qu'il devait sûrement y avoir une suite
(52)	it WOULD twist itself round and look up in her face, with such a puzzled expression that she could not help bursting out laughing
(52a)	hatte das Tier eine Art, sich umzudrehen und ihr mit einem so verwunderten Ausdruck ins Gesicht zu sehen, daß sie laut herauslachen mußte
(52b)	el ava de pronto se giraba, mirándole a la cara con tan perpleja expressión que Alicia no podía contener la risa
(52c)	le flamant ne manquait pas de se retourner et de la regarder bien en face d'un air si intrigué qu'elle ne pouvait s'empêcher de rire
(53)	'Look out now, Five! Don't go splashing paint over me like that!' 'I couldn't help it', said Five
(53a)	'Paß doch auf, Fünf. Du spritzt mich ja überall voll mit deiner Farbe!' 'Dafür kann ich nichts', sagte Fünf
(53b)	'¡Ten cuidado, Cinco! ¡Me estás salpicando todo de pintura!' 'Fue sin querer'- dijo Cinco
(53c)	'Fais donc attention, Cinq! ne m'éclabousse pas de peinture comme ça!' -'Je ne l'ai pas fait exprès', répondit l'autre

Here looms a vast field of research for translation studies with parallel text corpora (cf. King and Woolls 1996). Correlated expressions that collocate and colligate the same way in two or more languages will probably prove to be rare indeed.

4. Into the millennium

4.1. I would be content if the present discussion has managed to raise some vital issues in the changing picture of language and discourse under the impact of large corpus data. The impact seems sufficiently radical that a major scientific revolution or paradigm shift could be predicted. In the past, linguistics has tended to cultivate a large supply of abstract theories whilst postponing and marginalizing description of practices. Today we confront a far larger supply of concrete practices, which must be described before we can even define what a "language" is. I do not advocate that theory-building should be shelved, even temporarily; but rather that theory-building should finally and definitively cease to run so far ahead of practice, and cease to devise arguments why theory cannot be derived or tested from practice.

4.2. As a corollary, unquestioned scientific priority would no longer be allotted to abstract and general statements. These may prove the hardest to demonstrate with authentic data. And we may incur the paradox of trying to base a general theory upon special cases by selecting data devoid of special features (cf. 1.10). How general or specific a description deserves to be should be decided by our data and by the purposes of our research. Concrete and specific statements may prove more realistic, and for some purposes, such as language teaching, more useful. Moreover, data-driven descriptions are by nature specific in the incipient stages, and gradually gain generality as our picture improves of what to examine. A substantial range of constraints should turn out to be more specific than a discourse yet less general than the whole language (2.9).

4.3. As a further corollary, we should no longer displace real data with invented data, or convert data into formal representations. Instead, we should work to get as far as we can using real data to represent themselves. Even our description of the underlying organisation of data should be as data-driven as possible, rather than expressed in some purely theory-driven "deep structure" comprising "universal categories", which I hold least suitable to "provide tools for describing a text" (cf. 1.14-15) To judge from past experience, 'universals' tend to be indirectly extrapolated from particular languages after all, especially English (2.3). The latter's dominance in linguistic theory can only be effectively transcended by much resolute work on large corpora in as many languages as possible, each treated on its own terms.

4.4. Meanwhile, the well-described languages like English could be used by corpus researchers not to hasten above and beyond the data (as homework linguists did, 1.10), but to present data to wide audiences of specialists and non-specialists to test and discuss. By broadening our audience base, we can most safely offset personal biases in our own intuition and introspection. And the chances for productive applications will improve, such as language teaching.

4.5. My own prediction would be that progress will evolve out of the process I have called dialectical resolution. (3.27): the corpora that confront us with problems will provide vital support in solving those problems. If authentic data confront us with diversity, then we should keep building sub-corpora until each of them displays signally enhanced internal uniformity. Then we can compare these sub-corpora to identify and investigate which parameters and constraints are more general or more specific. My own work on text types indicates that types are often untidy and fuzzily defined, due especially to differences between insiders and outsiders, e.g., between academic journals and learner textbooks (Beaugrande 2001). Much academic writing is strenuously and gratuitously technical and actually impedes communication; but effective strategies to improve efficiency require corpus data for describing current practices.

4.6. Again by dialectical resolution, a large corpus can increase breadth without flattening depth if the technology itself is enlisted in the operations of description. Doing so requires sophisticated software for 'tagging' and 'parsing' the data; the description of 'open text' with no such preparation is still not genuinely operational (Sinclair 1999). The more secure categories like "Article", "Preposition", or "Auxiliary Verb" are by no means delicate enough. The more innovative ones, like "staging" and "collateral" in fieldwork (2.1), or "Agent-Opposing Process" and "Face-Saving Auxiliary" proposed here (3.33; 3.40), are not secure. At this stage, the categories of our description can only be heuristic, not formalised. Certainly, we have no sound reason to junk our established terms, nor to reintroduce them in technical guises; instead, corpus data should enable us to render them more applicable and precise as tools of description. We could for example retain the terms "Noun" and "Verb" whilst exploiting corpus data to make their meanings more delicate, e.g., by determining whether the "Nominal" or the "Verbal" formation from the same stem can be regarded as more basic; or whether the two might have evolved apart into quite distinct ranges of colligation and collocation.

4.7. If the dialectic of language and discourse can be restored to the centre of linguistic description, then the prospects for dialectical resolution should be favourable in the long run. For the present, the imperative would be to sustain a spirit of renewal and openness for new phenomena, new methods, and new discoveries stretching out into a new millennium.

Endnotes

1 Fillmore's (1992) term of 'armchair linguist' is not quite accurate when a computer terminal is in front of the chair.
2 For the record, none of these three had appeared in the Bank of English, the world's largest data corpus, as of July 1994.
3 The noted Danish linguist Jacob Mey (personal communication) tells me that the term "prolegomena" with its Kantian echoes in the English title was entirely on the initiative of Hjelmslev's translator, the late Francis Whitfield. Hjelmslev "was never a Kantian, not even a Neo- one"; Mey "would in hindsight characterise him as a purebred Neopositivist".
4 Reported to me by Halliday in conversation in Beijing in July 1995.
5 Except by Halliday, who told me in 1994 that he had written a paper entitled "How big is a language?", but it was not published. Some of its core ideas were taken over into Halliday (1996).
6 These totals include inflected forms too.
7 The terms were introduced by J.R. Firth (1957 [1934-51], 1968 [1952-59]), but gained little substance until corpus data arrived.
8 Actually, the term used in this work is not "lexicon" but "lexis", defined to cover "the resources of the vocabulary" and the "process of lexical choice" (Cross 1993:196). But we might retain the old term in this newer meaning.

About the Author

The central theme of Robert de Beaugrande's work has always been the search for a dialectical relation between theory and practice, which has been pervasively unbalanced in modern society and its institutions, including the science of language. Most recently, this search has been directed to the exploration of very large corpora, with some surprising results to be presented in the volume A New Introduction to the Study of Text and Discourse, now in preparation. A listing of published works is
available at www.beaugrande.com, where some 50 or so can be downloaded.

E-mail: beaugrande69@hotmail.com

References

Aston G. and L. Burnard. 1998. The BNC Handbook: Exploring the British National Corpus with SARA. Edinburgh: Edinburgh University Press.

Beaugrande, R. de. 1984. 'Linguistics as discourse: A case study from semantics.' WORD 35.15-57

Beaugrande, R. de. 1991. Linguistic theory: The discourse of fundamental works. London: Longman.

Beaugrande, R. de. 1997a. New foundations for a science of text and discourse. Stamford, CT: Ablex.

Beaugrande, R. de. 1997b. 'Theory and practice in applied linguistics: Disconnection, conflict, or dialectic?' Applied Linguistics 18/3.279-313.

Beaugrande, R. de. 1998. 'Performative speech acts in linguistic theory: The rationality of Noam Chomsky.' Journal of Pragmatics 29.1-39.

Beaugrande, R. de. 2000a. 'There is no such thing as syntax - And it's a good thing too!' Festschrift in honour of Jan Firbas. Ed. Josef Hladky et al. Amsterdam: Benjamins.

Beaugrande, R. de. 2000b. 'Text linguistics at the millennium: Corpus data and missing links.' Text 21.

Beaugrande, R. de. 2001. 'Cognition and technology in education: Knowledge and information - language and discourse.' Cognition and Technology 1/2.

Bloomfield, L. 1933. Language. New York: Holt.

Borges, J.L. 1964. Labyrinths. New York: New Directions.

Chomsky, N. 1957. Syntactic structures. The Hague: Mouton.

Chomsky, N. 1965. Aspects of the theory of syntax. Cambridge: MIT.

Chomsky, N. 1982. The generative enterprise. Dordrecht: Foris.

Cross, M. 1993. 'Collocation in computer modelling of lexis as most delicate grammar.' Register Analysis. Ed. M. Ghadessy. London: Pinter. pp 196-220.

Dixon, R.M.W. 1968. The Dyirbal language of North Queensland. London: University of London PhD Thesis.

Firth, J.R. 1957. Papers in linguistics 1934-1951. London: Oxford University Press.

Firth, J.R. 1968. Selected papers of J.R. Firth 1952-1959. Ed. F.R. Palmer. London: Longman.

Francis, G. & J.McH. Sinclair. 1994. 'I bet he drinks Carling Black Label: A riposte to Owen on corpus grammar.' Applied Linguistics 15.190-200.

Grimes, J. 1975. The thread of discourse. The Hague: Mouton.

Hall, R.A. Jr. 1968. An essay on language. Philadelphia: Chilton Books.

Halliday, M.A.K. 1961. 'Categories of a theory of grammar'. WORD 17.241-292.

Halliday, M.A.K. 1991. 'Corpus studies and probabilistic grammar.' English Corpus Linguistics. Ed. K. and B. Alterberg. London: Longman. pp. 30-43.

Halliday, M.A.K. 1992. 'Language as system and language as instance: The corpus as a theoretical construct'. Directions in corpus linguistics. Ed. J. Svartvik. Berlin: Mouton de Gruyter. pp. 61-77.

Halliday, M.A.K. 1994. Language in a changing world. Sidney: Australian Association of Applied Linguistics.

Halliday, M.A.K. 1996. 'Grammar and grammatics.' Functional descriptions. Ed. R. Hasan, C. Cloran, and D.G Butt. Amsterdam: Benjamins. pp. 1-13.

Halliday M.A.K. 1997. 'Linguistics as metaphor.' Reconnecting language. Ed. A.M Simon-Vandenbergen, K. Davidse, and D. Noël. Amsterdam: Benjamins. pp. 3-27.

Hartmann, P. 1963. Theorie der Sprachwissenchaft. Assen: van Gorcum.

Hasan, R. 1987. 'The grammarian's dream: Lexis as most delicate grammar.' New developments in systemic linguistics. Ed. M.A.K. Halliday and R. Fawcett. London: Pinter. pp. 184-211.

Hjelmslev, L. 1969 [orig. 1943]. Prolegomena to a theory of language. Madison: University of Wisconsin Press.

King, P. & D. Woolls. 1996. 'Creating and using a multilingual parallel concordancer.' Translation and Meaning 4.459-66.

Kintsch, W. 1988. 'The role of knowledge in discourse comprehension: A 'construction-integration model'.' Psychological Review 95/2.163-82.

Kucera, H. & W.N. Francis. 1967. Computational analysis of present-day American English. Providence: Brown University Press.

Kuhn, T. 1970. The structure of scientific revolutions. Chicago: Chicago University Press.

Longacre, R. 1970. Discourse, paragraph, and sentence structures in selected Philippine languages. Santa Ana: Summer Institute of Linguistics.

Longacre, R. et al. 1990. Storyline concerns and word order typology in East and West Africa. Studies in African Linguistics, Supplement 10. Los Angeles: UCLA Dept. of Linguistics.

Milton, J. 1999. 'Lexical thickets and electronic gateways: Making text accessible by novice writers.' Writing: Texts, processes and practices. Ed. C. Candlin and K. Hyland. London: Longman. pp. 221-243.

Pike, K.L. 1967 [orig. 1945-64]. Language in relation to a unified theory of the structure of human behavior. The Hague: Mouton.

Rumelhart, D., et al. 1986. Distributed parallel processing: Explorations in the microstructures of cognition. Cambridge, MA: MIT Press.

Sapir, E. 1921. Language. New York: Harcourt, Brace, & World.

Saussure, F. de. 1966 [orig. 1916]. Course in general linguistics. Transl. Wade Baskin. New York: McGraw-Hill.

Sinclair, J.McH.1991. 'Shared knowledge.' Georgetown University Round Table on languages and linguistics 1991. Ed. J. Alatis. Washington, DC: Georgetown University Press. pp. 489-500.

Sinclair, J.McH. 1994. Large corpora are here to stay. Lecture at the University of Vienna, June 1994.

Sinclair, J.McH. 1998. 'Large corpus research and foreign language teaching.' Language Policy and Language Education in Emerging Nations: Focus on Slovenia and Croatia. Ed. R. de Beaugrande, M. Grosman, and B. Seidlhofer. Stamford, CT: Ablex. Pp. 79-86.

Sinclair, J.McH. 1999. 'New roles for language centres: The mayonnaise problem.' Language centres: Innovation through integration. Ed. D. Bickerton and M. Gotti. Plymouth: CercleS. pp. 31-50.

Sinclair, J.McH. et al. 1990. Collins COBUILD English grammar. London: Harper Collins.

Smith N.V. 1983. Speculative linguistics: An inaugural lecture. London University College.

Sweet, H. 1913 [1875-76]. 'Word, logic, and grammar.' Collected papers of Henry Sweet. Oxford: Clarendon. pp. 1-33.

Tognini Bonelli, E. 1996. Corpus theory and practice. Pescia: Tuscan Word Centre.

Widdowson, H.G. 1991. 'The description and prescription of language.' Georgetown University Round Table on Languages and Linguistics. Ed. J. Alatis. Washington, D.C.: Georgetown University Press. pp. 11-24.

Appendix 1. Key to Abbreviations

Abbey: Jane Austen Northanger Abbey
Carrie: Theodore Dreiser Sister Carrie
Caster: Thomas Hardy Mayor of Casterbridge
Chatter: D.H. Lawrence Lady Chatterley's Lover
Clink: Tobias Smollett The Expedition of Humphry Clinker
Damsel: Pelham Grenville Wodehouse A Damsel in Distress
Desert Zane Grey The Heritage of the Desert
Domb: Charles Dickens Dombey and Son
Emerson: Ralph Waldo Emerson Essays
Eyre: Charlotte Brontë Jane Eyre
Fauntle: Frances Hodgson Burnett Little Lord Fauntleroy
FedPap: Alexander Hamilton, John Jay, and James Madison The Federalist Papers
Finn: Mark Twain, The Adventures of Huckleberry Finn
Floss: George Eliot Mill on the Floss
French: Thomas Carlyle The French Revolution
Lady: Henry James The Portrait of a Lady
Last: James Fenimore Cooper The Last of the Mohicans
Life: James Boswell The Life of Samuel Johnson, LL.D.
Mansions: W. H. Hudson Green Mansions
Moby: Herman Melville Moby Dick
Nelson: Robert Southey The Life of Horatio Lord Nelson
Poor Richard: Benjamin Franklin Poor Richard's Almanack
Pride: Jane Austen Pride and Prejudice
Time: H.G. Wells The Time Machine
Whirl: O. Henry Whirligigs
Wieland: Charles Brockden Brown, Wieland
Wrongs: Mary Wollstonecraft, Maria, or The Wrongs of Woman