iThe rise of auxiliary DO: verb-non-raising or category-strengthening?
The paper contrasts two explanations for Ellegård's statistical data on the rise of auxiliary DO during the 15th and 16th centuries. One is Kroch's explanation in terms of a change on the parameter of verb-raising, which is shown to have a number of serious weaknesses. The other is Warner's explanation in terms of the gradual development of the distinction between auxiliary and full verbs. Though Kroch quotes Ellegård's figures in support of the
Principles-and-Parameters analysis, they actually support Warner's view much better. The paper also considers developments in the auxiliary system since the 16th century and offers a mixture of cognitive and functional explanations for the changes since the 13th century.
1. A qualitative history of the auxiliary DO
One of the most striking characteristics of Modern English is the importance of the auxiliary DO. In the absence of any other auxiliary, it is used in subject-inversion, in negation, in emphatic polarity, in `VP anaphora' and in tag questions, as shown by the following pairs of examples where a sentence containing DO is contrasted with one containing another auxiliary, HAVE. (1) aHave they finished?
b Did they finish?
(2) aThey haven't finished.
b They didn't finish.
(3) a They HAVE finished!
b They DID finish!
(4) a They have.
b They did.
(5) a They have finished, have(n't) they?
b They finished, did(n't) they?
Auxiliary DO is often called `periphrastic' DO because it has no meaning independent of the meaning of the construction concerned; the only reason for using auxiliary DO in Modern English is because the syntax requires an auxiliary and no other auxiliary is needed by the sentence's meaning. DO fills the gaps where non-auxiliary verbs are not allowed and where other auxiliaries are not needed.
In Middle English, in contrast, DO had no special role because auxiliary and non-auxiliary
verbs could be used in much the same ways, as shown in Table 1 (where Middle English word-forms have been modernised). The choice of sentence-types to illustrate the change will be explained below.
[Table 1 about here]
This complementarity between auxiliary DO and other verbs means that any history of DO must also be a history of the whole system of auxiliary and non-auxiliary verbs. In a nutshell, what has happened over the last thousand years is that a range of constructions have developed in which only an auxiliary verb is allowed. The main constructions are exemplified in 0-0. As we shall see, there are other constructions in which auxiliary and non-auxiliary verbs are treated differently, but the ones just mentioned are special because they occur so often in any text, and correspondingly in the experience of any language user or language learner. Any change in these constructions has a high profile in the history of the language. We also know a great deal about the history of DO in these constructions, and can locate the change quite precisely in the fifteenth and sixteenth centuries. During that period, as non-auxiliary verbs waned, so DO waxed; and it makes no sense to try even to describe the history of DO on its own, let alone to explain it. This paper therefore has to present the changes in the use of DO as part of a much larger pattern of change which has affected all verbs.
This much is uncontroversial. The point of contention is the nature of these large-scale changes. I shall compare two very different views which are exemplified by the work of Kroch (e.g. 1989, 1994) and of Warner (especially 1993). I shall try to show that the numerical data which Kroch quotes in support of his account actually support Warner's account much better. In both cases the rise of DO turns out to be just a small corner of a much larger picture of grammatical reorganisation, but the overall pictures are very different. In Kroch's picture the details of change all follow from the resetting of a single underlying `parameter', the verb-raising parameter, during the 15th and 16th centuries. Before the change, English allowed all verbs to raise to the position of the abstract inflectional node in a Barriers-type sentence structure (Chomsky 1986, Pollock 1989); but since the change verb-raising is no longer allowed. In principle this parameter can only be set in one direction, but an individual may have two different grammars, each with a different setting for this parameter. Warner's picture shows a much more gradual development which had nothing to do with verb-raising; in fact, verb-raising is not recognised as a reality at any time in the history of English. Instead, what changed is the meaning of being an auxiliary or non-auxiliary verb. From modest beginnings, this distinction came to be relevant to a larger and larger range of constructions (and other patterns), a development which could extend over many centuries beyond the 15th and 16th centuries - a strikingly different view from Kroch's.
Whereas Kroch locates his explanation firmly in the transformational paradigm, Warner's
belongs explicitly to the cognitive tradition with its stress on structured categories (such as word-classes). The debate involves serious questions about the nature of grammars and sentence structure; in particular, is the verb-raising transformation real in any languages? Indeed, are any
transformations real? If Kroch's account is successful then it must count as serious support for the transformational view of sentence structure. If Warner is right, on the other hand, the diachronic facts are equally compatible with monostratal theories of syntactic structure such as Head-driven Phrase Structure Grammar (Pollard and Sag 1994), Lexical Functional Grammar (Bresnan 1982) or Word Grammar (Hudson 1990).
The facts for which these two views are alternative explanations are not in dispute. As mentioned earlier, Warner recognises a broader range of facts as relevant, but we can start by focussing on the patterns which Kroch discusses. These are the five constructions shown in Table 1 - negative and interrogative sentences of various kinds. These are the most relevant to Kroch's account because they are the ones where (on his analysis) verb-raising once applied but no longer does. However they are also relevant because we have useful statistical data which we shall consider in the next section. We shall consider Warner's broader view in section 4.
2. A quantitative history
The historical facts relevant to negative and interrogative sentences can be summarised as follows (I follow the account in Denison (1993:255f)). The `periphrastic' (semantically empty) DO appeared first in the 13th century, but at that time it had no special connections with negation or inversion; nor, indeed, were these restricted by rule to auxiliary verbs. The link between auxiliary-hood and these sentence-types appeared first in the 15th century, and had almost reached its modern solidity (with some differences of detail) by the end of the 16th. The changes shown in Table 1 were completed within two centuries, and our challenge is to explain not only why they happened at all, but also the way they spread through both the grammar and the population.
First, the data. For each of the sentence-types shown in Table 1, Kroch quotes both the total number of textual examples counted in which DO might have occurred (N) and the
iipercentage of these in which DO was used. His raw numbers (based on Ellegård 1953:159) are
repeated in Table 2.
[Table 2 about here]
As can be seen from the second column, the data-points are not distributed evenly through time. It will be helpful to present the figures graphically, so we can reorganise them in equal periods of 25 years by collapsing the figures from 1526-1550, and by splitting those for 1426-1475 (with some averaging to smooth transitions where appropriate). The result is the rather complex graph in Fig. 1, which appears to confirm Kroch's Constant Rate Effect - that `the
rate at which the newer option replaces the older one is the same in all contexts' (1994:181). According to Kroch, a statistical analysis in terms of logistic regressions shows that the rate of slope is the same in all contexts (ibid:182). More generally, he says that the same pattern, the Constant Rate Effect, is found repeatedly in empirical investigations. He offers an explanation for the change which we shall consider in the next section.
[Fig 1 about here]
One of the most striking features of Fig. 1 is that the five lines appear to form two clusters. The lines for negative questions and transitive (positive) questions are consistently higher than the other three, and by the mid-16th century the two clusters are very clearly separated. I shall call the two clusters the `high-do' and `low-do' constructions. How seriously should we take the apparent difference shown in this graph? One way to test its significance is by testing the statistical significance of differences between the constructions at each point in time. For example, in 1400 11.7% of negative questions contained DO, whereas 0% of transitive questions did. This percentage difference is enough to separate the two lines in Fig. 1, but when we look at the raw figures we find that they are based on a total of only 17 negative questions and three transitive questions - i.e. 2/17 negative questions had DO, compared with 0/3 transitive questions. It is easy to test the statistical significance of this difference (by the Chi-square test). It is not at all significant - the probability (p) is over 0.5, meaning that such a difference is likely to be found purely by chance in over 50% of random samples. It is normal to set a threshold of 5% (p < 0.05) for
statistical significance, so the question is which of the differences shown in Fig. 1 are significant in this technical sense.
In a nutshell, the answer is that (with a handful of exceptions which we shall consider below) the only significant differences are those between our high-do and low-do groups. More precisely, at least one high-do construction (transitive question) is almost always significantly different from at least one low-do construction (intransitive question); the two high-do constructions are almost never significantly different from each other; nor are the three low-do constructions. Fig. 2 shows the much simpler picture which results if we ignore all the differences which are statistically below the threshold of significance. Clearly, we are dealing with two groups of concomitant changes, rather than with a single one; so any explanation must be able to explain why the changes are grouped in this way.
[Fig. 2 about here]
As mentioned above, Fig. 2 is actually too simple. The statistical analysis reveals two exceptions.
; In 1500-25, transitive and intransitive questions fall together, each leaving the group that it belongs to at other periods - i.e. DO is significantly rarer in transitive questions than in the other high-do construction, and significantly more frequent in intransitive questions than in the other
; In 1525-50, Wh-object questions become significantly rarer than the other low-do constructions.
These two aberrations are shown in Fig. 3. If we use statistical analysis at all, we must respect it in all cases. A complete explanation for the historical changes that led eventually to the relatively simple two-way split in 1550-75 will have to explain these details, but I admit immediately that I have no such explanation.
[Fig. 3 about here]
To judge by these graphs, the high-do contexts had a growth-spurt around 1500 which was not matched by the low-do contexts until much later - in fact we can only assume the existence of a later spurt during the next century after the period covered by Kroch's data. We must obviously also assume that the spurt for the high-scorers levelled off during this period, as the figures approached 100%. In short, we seem to have evidence here for two separate S-shaped curves whose spurts cover about a century, but which (according to Kroch's statistical tests) have the same slope and therefore could illustrate different applications of the same change. The challenge, therefore, is to explain first, what this change consisted in, and second, to explain why it affected negative and transitive questions before all the other sentence-types.
3. An explanation in terms of the verb-raising parameter
Kroch takes the statistical linkage between the various changes as evidence for a fundamental linkage in the speakers' competence:
We take its general validity to indicate that what changes in frequency in the course of time
during a syntactic change is language users' overall tendency to choose one abstract
grammatical option over another in their language production. ... The unity of the change
is defined at the level of the grammar, not at the level of the surface contexts. This unity he finds in a single Chomskian parameter which has many ramifications, so that the transition period covered by his data involved competition between two distinct grammars.
The parameter in question is the presence or absence of verb-raising (or, in more recent terms, whether or not verb-raising is forced to apply before the structure is `spelled out' phonetically). Fig. 4 shows the basic sentence-structure which he assumes.
[Fig. 4 about here]
The assumption is that the node `I' (for Inflection or INFL) is crucial for subject inversion and negation, so if a verb is raised to I both of these possibilities are open to it. Without verb-raising, neither inversion nor negation is possible, so DO (which can occur at I) must be supplied, whereas (according to Kroch (ibid:192)), auxiliary verbs are exempt from the loss of verb-raising, which is why they do not need DO.
One of the attractions of the verb-raising analysis is that it also explains a change in the position of adverbs. Kroch quotes the following example to illustrate the change: (6) a (Middle English) Queen Ester looked never with swich an eye.
b (Modern English) Queen Ester never looked with such an eye.
If we assume that adverbs such as never are generated as adjuncts on the left edge of the VP, then
verb-raising would move looked past never; so the loss of verb-raising would explain why never
cannot follow the verb in Modern English. Kroch quotes further figures from Ellegård (1953) which confirm that this change took place at roughly the same time as the rise of auxiliary DO, and that it spread at roughly the same rate as the other changes. More precisely, the pre-verbal position of never arrived about a generation before the use of DO in high-do contexts as can be seen from Fig. 5.
[Fig. 5 about here]
Most of Kroch's paper deals with the mechanism, called Grammar Competition, which drove our linguistic ancestors to sort out the conflict between their two grammars. He suggests that:
the historical evolution of competing variants in syntactic change is similar to the evolution of
morphological doublets. In both cases, the coexistence of the variant forms is
diachronically unstable: one form tends to drive the other out of use and so out of the
He suggests that all syntactic change may be located in the requirements of individual formatives (184), and Grammar Competition eliminates alternative syntax just as it eliminates synonyms (except when these are supported by sociolinguistic differences). Unfortunately his general discussion of Grammar Competition does not deal directly with verb-raising, but he presumably sees this too as something which is triggered by the formative INFL, so the competition is between an INFL which does trigger raising and one which does not.
Kroch's statistics do seem to support his claim that DO spread through the various sentence-types at the same speed, and that this spread could be linked to the change in adverb position. However there are serious weaknesses in his theoretical interpretation. ; If the change involved nothing but the verb-raising requirements of INFL, why did it affect the different sentence-types at different times? Kroch's analysis predicts that DO and adverb-preposing should have grown not only at the same speed but also at the same time, with no difference at all across sentence-types; and similarly for the various sentence-types which were affected. Kroch recognises this problem, but offers only a partial explanation: .. the approach taken here implies directly that the frequency differences in different contexts of
a change must be due to factors orthogonal to the grammatical change itself .. and that
these orthogonal factors are responsible for the differences in the .. parameter values in
[our tables]. Such factors are not well-understood but must involve psycholinguistic and
information processing preferences, which, in usage, favor one form or the other
differentially in different linguistic contexts whenever a language, for any reason, happens
to allow more than one option for expressing a given linguistic content. (183) No doubt processing factors are relevant, and I myself shall invoke them below, but Kroch's account makes them bear a remarkably heavy burden in the explanation. The problem is as follows. Whatever the influence of processing may be, this influence is presumably constant across time; so the only possible source of variation through time is the number of speakers who have the alternative grammars. Accordingly, the model predicts that the relationships among the sentence-types should stay constant through time; but this is not what we find.
For example, take the relationship between preposed adverbs and negative declaratives with DO. In 1525 the respective figures are 89% and 8%, but 25 years later it was about 90% and 43% - a very different relationship. Putting this another way, let's assume that the order adverb-verb presupposes a non-raising grammar; therefore anyone who uses this order must already have such a grammar. Given the figure of 89% adverbs preposed in 1525, at least this proportion of the population must have had non-raising grammars by that date. Therefore the much lower figure of 8% for negative declaratives with DO must be due to some processing consideration which makes verb-raising in this sentence-type about ten times as likely as adverb-preposing. But if this processing consideration is constant it should still have roughly the same effect 25 years later. At this time the number of people with a non-raising grammar is roughly similar to the previous period (when almost everyone already had one); therefore the proportion of negative declaratives with DO should also be roughly the same. But it is not - it is five times greater.
The same problem arises with any pair of syntactic constructions. If their differences are due to processing factors, they should stay constant; but they don't. What we observe is much more suggestive of a change which spreads through the grammar construction by construction, with one group of constructions setting the pace and other constructions following after. As with lexical diffusion, the change spreads through both the grammar and the population at the same time, so the figures that emerge from a corpus of texts produced by different people reflect both these changes.
; The assumptions of Principles-and-Parameters theory force Kroch to consider the old and new patterns as distinct grammatical systems (183):
The options in question .. are not alternating realizations within a single grammar, like
extraposed versus non-extraposed constituents. Rather they seem always to involve
opposed grammatical choices not consistent with the postulation of a single unitary
analysis. In the present case, for example, contemporary accounts of verb-movement to
INFL all agree that it is forced by the morphosyntactic contents of functional heads and
cannot be optional. Because the variants in the syntactic changes we have studied are not
susceptible of integration into a single grammatical analysis, the variation does not stabilize
and join the ranks of a language's syntactic alternations. Instead, the languages always
evolve further in such a way that one or the other variant becomes extinct.
In other words, at least some of the variability is due to code-switching. But the codes concerned are grammars which are identical except for one parameter, the presence or absence of verb-raising. How can we distinguish empirically between the predictions of this claim and of one in which the trigger of verb-raising is optional? And if we can distinguish the two, how plausible is Kroch's view? At present its only support is its compatibility with Principles-and-Parameters theory, so those of us who question that theory need more empirical evidence.
; The explanation depends crucially on the the verb-raising analysis, but the empirical and theoretical underpinnings for this analysis are weak. Some of these weaknesses have been pointed out by others (notably Kim and Sag 1996), so I shall not try to repeat the exercise. The devil lies in the detail of such analyses, and one of the problems of verb-raising lies in dealing with all the details. Kroch discusses a lot of details in the paper quoted (1994), but there appear to be some important inconsistencies. First, if auxiliary verbs (HAVE, BE and the modals) are verbs which continue to be raisable, how can verb-raising be controlled only by INFL (as Kroch claims)? According to this analysis, INFL either raises all verbs, or it raises none of them, so there is no way to accommodate lexical exceptions. And second, Kroch suggests (195) that verbs and nouns may always be raised to a higher functional head before surface structure. This suggestion clearly conflicts with the claim that modern English verbs do not raise (to INFL or anywhere else) at surface structure.
; According to Kroch verb-raising had disappeared from English by the end of the 16th century (`.. 1575, the date at which V-to-I movement is definitively lost' - ibid:181). This assumption flies in the face of a variety of facts, not least the figures which he himself quotes showing that most low-do constructions still managed without DO in 1575. Furthermore, according to Warner (1993:66) subject inversion was still common with some full verbs (e.g. MEAN, SAY, THINK) through the eighteenth century, so if subject inversion did involve verb-raising this would suggest a much more gradual elimination than Kroch's estimate. Even in some varieties of Modern British English we still have the possessive auxiliary HAVE as in Have you a car?, which seems
to call for a raising analysis although it is also a transitive verb. The problem is well known (Pollock 1989:407), but still lacks a convincing solution (as argued by Kim and Sag 1996). ; Kroch explains the drive to eliminate verb-raising by claiming that syntactic variation is always resolved (by `Grammar Competition') definitively in favour of one pattern. This general claim does not seem to be borne out by some well-known cases. For example, he quotes Taylor's
(1994) work on changes in Greek word-order between Homer and the New Testament, which does indeed show an increasing tendency to prefer SVO rather than the earlier SOV; but the outcome was only a trend, and quite unlike the relatively rigid SOV order found in (say) German subordinate clauses. This flexibility continues into modern Greek, two millenia later, where SVO is still dominant but all five other orders of S, V and O are still found (Tzanidaki 1996; Lascaratou 1989).
In conclusion, Kroch's explanation leaves enough important ends dangling to suggest that we should look for an alternative explanation. A common thread runs through all these criticisms, namely the need to locate changes within the details of the grammar, in relation to individual constructions or lexical items, rather than to aim at a single global variable. The next section will develop this theme.
4. An explanation in terms of word-classes
iiiAn alternative way to explain Kroch's data is in terms of the word-classes `auxiliary verb' and
`full verb'. In this explanation I shall be largely following Warner (1993), supplemented by relevant sections of Denison 1993. Like Warner and Denison, I believe the major change that affected English in the early modern period involved the distribution of grammatical characteristics between these two subclasses of verbs.
Warner's account of developments in this part of English grammar is very different from Kroch's. The unifying factor, which may explain the Constant Rate Hypothesis, is provided by the two word-classes concerned (auxiliary and full verbs), whose membership stayed roughly the same. What changed so dramatically was the range of characteristics which distinguished the classes, but the increase in the fifteenth and sixteenth centuries was part of a much longer development which extended far beyond the period covered by Kroch. In a nutshell, we have moved from the Middle English list of differences in Table 3 to the modern list in Table 4. (The numbers after the examples refer to the pages where Warner quotes them; the sign < means
[Table 3 about here]
[Table 4 about here]
The most striking difference between these two tables is in their length. Since Old English, the class of auxiliary verbs has grown up. As a `baby' category it had very few distinctive characteristics, and a somewhat fluid membership at the points where these characteristics disagreed; but in its modern form it is a perfect example of a well-developed prototype, distinguished by a large number of characteristics which, by and large, all apply to the same range of members but with a penumbra of exceptional forms.
The fourth column of Table 4 shows roughly when the new form or restriction entered
the grammar. The dates show that the development of the auxiliary class is spread over at least three centuries, with Kroch's changes as just one segment of the development. It is true that the sixteenth century seems to have been a particularly important period in this history, but as I pointed out earlier, not all the changes happening at this time can be explained by the loss of verb-raising. Warner (192) explains this `apparently coherent long-term development of an auxiliary group' in terms of cognitive principles, namely Rosch's (1978:28) `principle of cognitive economy':
The task of category systems is to provide maximum information with the least cognitive effort. This principle favours categorization systems in which distinct characteristics are highly correlated (as in the case of our modern auxiliary-verb class). Warner suggests two causal links between this principle and the observed change, but the one which I should like to highlight is that the way we classify human behaviour such as language, unlike our classification of everything else, is not mere classification, but affects the material classified. In the case of language, Rosch's principle provides a feed-back loop: the more closely linked we think two features are in other people's usage, the more we link them in our own usage, which in turn encourages our hearers to link them even more closely in their usage, providing us with even better evidence for their linkage, and so on.
The last column of Table 4 shows whether it was the auxiliary class or the full-verb class that had acquired the new characteristic. For example, the `< adverb' characteristic of auxiliaries is simply a continuation of the ability to occur before an adverb such as never (e.g. We have
never eaten pears) which was previously shared by all verbs (We ate never pears), but which
became distinctive for auxiliaries when full verbs lost it; this is shown as a change to full verbs. In contrast, the `reducible to clitic' characteristic was an innovation which only affected auxiliary verbs. It is therefore misleading to refer to this collection of historical changes as the development of the auxiliary class; it could just as accurately be described as the development of the full-verb class. Indeed, this would be a more accurate description because it would reflect an interesting historical generalisation that emerges from Table 4, namely that full verbs changed before auxiliaries did. All three changes covered by Kroch's discussion involved the relations between full verbs and some other word in the sentence (an adverb, the subject and the negative marker); but all the subsequent changes involved the characteristics of auxiliary verbs. As far as the earlier changes are concerned, it is auxiliaries rather than full verbs that are in direct line of descent from OE verbs; but the reverse is true for the later changes. Consequently, rather than seeing the change as the rise of the auxiliary class, it would be better to see it as simply the separation of two classes, neither of which has any particular priority. Some of the characteristics of OE verbs have been inherited by modern full verbs, and others by modern auxiliaries; but both are still verbs.