Does corpus linguistics exist? Some old and new issues

Jan Aarts

University of Nijmegen

Abstract

This paper deals with a number of methodological questions in corpus linguistics. It
discusses various kinds of linguistic data and examines the nature and use of corpus data
and the function of corpus annotation. These topics are put within the context of two recent
issues in corpus linguistics: the new evidence that has become available from spoken
corpora such as COLT and the spoken part of the BNC, and the renewed discussion of
corpus linguistic methodology, notably about the distinction between the corpus-based and
the corpus-driven approaches.

1.

Introduction

It is difficult to attach a date to the birth of corpus linguistics, but if we take it to
mean computer corpus linguistics (see Leech 1992) we can locate its origin early
in the 1960s, at the time when the computerization of the the Brown Corpus was
finished. For about a decade and a half the use of corpora belonged very much to
the linguistic underground, and the term ‘corpus linguistics’ did not come into use
until it began to gain some respectability in the late seventies or early
eighties.Today it is taken seriously by the majority of linguists. Not by all of them
– Chomsky’s position, for example, is still as negative as it was in the sixties:
[Corpus linguistics] doesn’t exist. If you have nothing, or if you are stuck, or if
you’re worried about Gothic, then you have no choice. (Chomsky in an interview
with Bas Aarts; Aarts 2000: 5)

Other theoretical linguists take a less diehard view. Fillmore recently

described his position with respect to the use of corpora as follows:

My peace-making position [in Fillmore 1992 – JA] was that one
couldn’t succeed in the language business without using both
resources: any corpus offers riches that introspecting linguists will
never come upon if left to their meditations; and at the same time,
every native speaker has solid knowledge about facts of their
language that no amount of corpus evidence, taken by itself, could
support or contradict. (Fillmore 2001: 1)

That feelings about this issue can still run high became apparent in a recent
discussion on the corpora list, where (after almost 40 years!) we could renew the

2

Jan Aarts

acquaintance with some arguments we hadn’t seen come by for some time. In
many such discussions it is not clear what exactly is at issue, however. Is it a
discussion about the question which data sources are the most appropriate for
linguistic research? Is it a clash between different methodologies? Or is it –
perhaps more basically – a disagreement about what should the object of
linguistic study? More generally, such issues come down to the question: what
kind of linguistic discipline is corpus linguistics? Is it a linguistic discipline in its
own right? Or, to borrow Chomsky’s provocative words: does corpus linguistics
exist?

1

In this paper I want to deal with such questions by looking at two

important things going on in corpus linguistics at the moment. First, there appears
to be a new focus on spoken language – this is beginning to receive the same sort
of attention that the written language has received so far. Notably there is a
widening of the attention given to the spoken language so as to also include its
syntax and to contrast this with the syntax of the written language. This new
attention showed itself, for example, in Miller and Weinert’s book Spontaneous
Spoken Language: Syntax and Discourse (1998). On the basis of corpus material
from English, German and Russian, they analyse the differences between the
syntax and discourse of spontaneous speech and written language, claiming that
the differences are such that they have far-reaching consequences for theory-
oriented linguistic research. In 1999 another book appeared that bears the stamp
of a new focus on the spoken language: the Longman Grammar of Spoken and

Written English (Biber et al.). And finally, an important new resource for the
research into spoken English became available in 2000, namely Anna-Brita
Stenström’s Corpus of London Teenage Language (COLT). Containing the
spontaneous speech of a young generation of native speakers, it is bound to put its
stamp on the research of spoken English in the near future; it has already
generated a considerable list of publications.

2

A second important recent development is a renewed interest in and

discussion of questions of methodology. I have already mentioned the recent
dicussion on the corpora list, which centred on the old question of intuitive data
vs. corpus data. The issue returned at the 2001 ICAME conference, notably in the
keynote lecture by Fillmore, entitled ‘Armchair linguistics vs. corpus linguistics
revisited’ (Fillmore 2001), a title which echoed his earlier paper delivered ten
years previously at the Stockholm Nobel Symposium on corpus linguistics
(Fillmore 1992). At the same conference another contribution on methodology
was made by Tognini-Bonelli. This was only sideways related to the intuitive vs.
corpus data debate, but more concerned with methodological matters that arise
once the linguist has chosen for the use of corpus data. It thus raises an issue that
is internal to corpus linguistics and therefore probably more important for the
discipline. It is concerned with the question how dominant a status should be
assigned to the data in corpus linguistic research. Putting this differently, the
question is not what data the linguist thinks it is appropriate to use, but what he

Does corpus linguistics exist? Some old and new issues

3

should do with them. In this connection Tognini-Bonelli distinguishes the corpus-
based approach and the corpus-driven approach. She describes the two as
follows:

... corpus-based work relates corpus data to existing descriptive
categories adding a probabilistic extension to theoretical parameters
which are already received, i.e. established without reference to
corpus evidence. On the other hand, corpus-driven work attempts to
define the categories of description step by step, in the presence of
specific evidence from the corpus. (Tognini-Bonelli 2001b: 99)

Whether a researcher will prefer a corpus-based or a corpus-driven approach
depends on the answers he will give to a number of basic questions: 1. what
kind of evidence does he think corpus data provide? 2. does he want to leave
room for other than corpus data? 3. what should be the role played by
previous linguistic research that was not based on corpus data? As said, these
are basic questions and a discussion of them and of the corpus-based vs.
corpus-driven distinction is called for. In what follows I shall look more
closely at the two issues mentioned; I’ll start with the methodological
question and after that discuss the new research into the spoken language in
relation to it. At the end we return briefly to the question posed in the title.

2.

Methodology

Any discussion of corpus linguistic methodology must start from the data – after
all the discipline bears the name of a linguistic data source and in its present form
arose as a challenge to the preference of the linguistic mainstream for intuitive
data. In this section we first look at the different kinds of data that linguists have
used and are using. After that we discuss the different ways in which linguists use
corpus data in their research. In the last sub-section we examine the role of
annotation in corpus linguistic research.

2.1

Types of linguistic data

A survey of the different kinds of linguistic data is presented in Figure 1
(reproduced from J. Aarts 2000). The major distinction is between intuitive and
non-intuitive data. In the practice of linguistic research it is still more or less
tacitly assumed that intuitive data provide direct access to a speaker’s intuitive
knowledge of the language, that is, to some sort of internalized rule system, but
this has been shown not to be the case; instead, intuitive data appear to reflect
what a speaker believes to be true of language behaviour in concrete situations
(see, for example, Snow and Meyer 1977; Nagata 1989; Shohamy 1996).

4

Jan Aarts

Figure 1. Types of linguistic data

Introspective data were the favourite type of intuitive data during some decades in
the previous century; it was usual for the linguist to consult his own intuitions
about the language. It has been pointed out repeatedly that such data are
unreliable: they are subjective and unverifiable and there is confusion of the two
roles the linguist plays – that of native speaker and that of linguistic researcher.
Some of these drawbacks can be overcome by collecting informant data: the
linguist does not consult his own intuitions, but those of others. The collection of
informant data is a difficult and time-consuming job (see, for example, Quirk and
Svartvik 1966; de Mönnink 2000). As a consequence, informant data have not
been used very frequently nor very extensively.

Non-intuitive data are provided by what people actually say and write. The

material has to be on record in some form or other to serve as workable data for
linguistic research; this is always the case with written and printed material. It is
not surprising, therefore, that until quite recently, if it was claimed that linguistic
studies were based on ‘authentic material’, this was usually written material. The
distinction that is made in Figure 1 between anecdotal and corpus data has not to
do with the type of data involved, but with the manner of collecting them. Briefly
put, anecdotal data are collected unsystematically, while corpus data are compiled
systematically. Most of the standard reference grammars of English of the first
half of the previous century contain anecdotal data. They were collected in a
haphazard way: whenever the grammarian happened to read something that
caught his attention, he made a record of it. Thus, Jespersen (pt II, 1914: vi) talks
about ‘my quotations, which I have collected during many years of both
systematic and desultory reading’ and Poutsma (pt I, 2nd ed., 1928: vii) says:

anecdotal data

corpus data

other native
speakers:
informant data

linguistic
introspection

non-intuitive

intuitive

linguistic
data

Does corpus linguistics exist? Some old and new issues

5

[My] quotations have partly been borrowed from dictionaries and from
other grammars. But by far the majority of them have been collected by
myself from numberless and varied sources: not only literary and scientific
books, but also periodicals and newspapers.

Other twentieth-century grammarians have said similar things (see, for example,
Schibsbye 1965: v).

3

Anecdotal data are inferior to corpus data in two respects. In the first

place, the grammarian will only collect examples that ‘catch his attention’ and is
not interested in what he expects to find. Jespersen said that, whenever he could,
he selected sentences that were in some way ‘striking’ (ibid.). For this reason
anecdotal data are not only necessarily skewed, but the subjective element
involved in their collection is not recoverable. Another, more important,
drawback of anecdotal data is that they become isolated from both their micro-
and macro-context; that is to say, they are deprived of their immediate verbal
context and the contextual features that are relevant for their specific form, and
become separated from their larger sociolinguistic context, including genre,
register, style, domain, etc. As said above, anecdotal data provide, next to
intuitive data, the material for most of the twentieth century reference and other
grammars. Sometimes the emphasis is on intuitive data, as is the case in Sweet’s
grammar where we find hardly any authentic data – practically all the examples
he gives to illustrate points of syntax are made-up sentences. Other grammarians
use anecdotal data exclusively (Jespersen, Kruisinga) or predominantly
(Poutsma). In view of the fact that both intuitive data and anecdotal data have a
good deal of subjective colouring, it is perhaps at first sight a bit surprising that
their descriptions of the English language are so homogeneous. This can partly be
explained from the fact that most grammarians relied, of course, to a greater or
lesser extent on previously written grammars. Reading the prefaces of the
grammars one can see for example that Jespersen, Poutsma, Zandvoort and
Strang express their indebtedness to Sweet, while Sweet, Poutsma, Zandvoort,
Strang and Schibsbye say that they are (also) specially indebted to Jespersen –
and this chain of interdependence between grammars can be traced throughout
the century. Another factor is that the variety of English they describe is very
homogeneous: on the whole, not only do they concentrate on British English and
on educated standard language, but they are predominantly concerned with the
written language and with general language about non-specialized subjects.

In most twentieth-century grammars the relation between the description

and the data is not clear. When grammarians use authentic material they usually
regard this as a source to choose their examples from; with the exception of Fries
(1952), they do not treat their material as a data collection that has to be
accounted for. In this connection it is typical that Jespersen and Poutsma speak of
their ‘quotations’ (see above). Even in A Grammar of Contemporary English and
A Comprehensive Grammar of the English Language, which we know to be

6

Jan Aarts

closely related to corpus data, the relation between data and description is not
clear, owing to the fact that the writers decided to ‘refrain from citing actual
utterances with references to sources. Most of our examples are either simplified
versions of actual utterances or invented’ (Greenbaum 1988: 46). I return to the
question of the relationship between description and corpus data below.

2.2

Corpus data: Their nature

The most important difference between corpus data and other kinds of linguistic
data is that in the former case their provenance is known: they are found in the
natural environment of the verbal context, we know in which medium they occur,
to what linguistic or extra-linguistic textual category, what (type of) language
user produced them, and so on. In other words, corpus data reflect the way in
which language is actually used. Now, language use is what most descriptive
grammars profess to want to give an account of. At the same time it must be said
that there is a not inconsiderable number of utterances that one comes across in
corpora but will look in vain for in descriptive grammars of language use. Among
them are broken-off sentences, false starts, repetitions of phonemes, morphemes,
words and (parts of) larger constituents, anacolutha, stretches of text from other
languages or from sub-standard varieties, as well as utterances that the speaker or
writer intended to be ungrammatical; in short, corpora contain among other things
evidence of ‘such grammatically irrelevant conditions as memory limitations,
distractions, shifts of attention and interest and errors ...’ Chomsky 1965: 3). In
other words, what one finds in a corpus is ‘performance’, that is, evidence of
linguistic behaviour. And linguistic behaviour is not only determined by
linguistically relevant, but also by linguistically irrelevant conditions.
Linguistically relevant are all aspects of the verbal context (or ‘micro-context’)
and all the parameters imposed by the sociolinguistic setting (the ‘macro-
context’). It seems reasonable, therefore, to say that a descriptive grammar should
give an account of the system of rules underlying language use together with the
linguistically relevant parameters that determine the form of utterances in their
context. It is of course not at all an easy matter to decide what, in a corpus, is
linguistically relevant and what not. Various terms and notions have been
proposed for the assessment of linguistic relevancy, all of which make an appeal
to some sort of consensus among native speakers, sometimes supported by a
criterion of frequency. Of these, the first was ‘acceptability’ defined by Chomsky
(1965: 11) as the quality of ‘utterances that are perfectly natural and immediately
comprehensible ... and in no way bizarre or outlandish’. Quirk and Svartvik
(1966) tried to operationalize the notion of acceptability, adding a quantitative
dimension. Widdowson (1991) introduces the concept of ‘feasibility’ as a filter
for corpus data; excluded are those data that do not meet a criterion of normality.
Finally, in an earlier article (Aarts 1991) I suggested using the notion of
‘currency’, which is compounded of the frequency of occurrence of a
construction and its ‘normalcy’ which is a qualitative and intuitive notion, closely

Does corpus linguistics exist? Some old and new issues

7

resembling Widdowson’s ‘normality’. All in all, it is likely that there is a
consensus among corpus linguists that not every construct, by virtue of its
appearance in a corpus, has to be taken into account in a description of the corpus
language, but that every corpus will contain constructs that can be excluded on
the basis of their not meeting an intuitive criterion of naturalness or normalcy.

4

2.3

Corpus data: Their use

I started this paper with two quotations from theoretical linguists to illustrate their
attitude to corpus data. However different in tone they may be, they both illustrate
the same basic attitude towards corpus data: corpus data are supplementary with
respect to intuitive data. As such they may be held in low esteem, as Chomsky
does when he says that they are only relevant ‘if you are stuck’, or highly valued
as Fillmore does, but there can be no doubt about the primacy of intuitive data,
for these rest, as Fillmore (see above) puts it, on the ‘solid knowledge [of every
native speaker] about facts of their language that no amount of corpus evidence,
taken by itself, could support or contradict’. In other words, in the view of both
linguists the corpus is used to jog the linguist’s intuition.

Other linguists assign a more important role to corpus data. They see the

corpus not as a supplementary, but as an indispensable tool for the research into
language use. There are two purposes for which corpus data can be exploited that
could not possibly be achieved by using another kind of data. The first of these is
the examination of language variety – the only way to reveal similarities and
differences between language varieties is to make a comparative study of the
frequency and distribution of linguistic phenomena in corpora representing these
different varieties. So far, this has probably been the most extensive use to which
corpus data have been put; it has been the subject of many individual studies and
this strand of research has come to fruition for the first time in a reference
grammar of English (Biber et al. 1999).

The second research activity for which corpora are indispensable is their

use as a test bed for linguistic hypotheses about the language. What can be tested
is primarily the knowledge accumulated in the literature about one or more
aspects of the language or the ‘complete’ descriptions of the grammar of the
language laid down in reference grammars. In order to test a hypothesis
consistently and exhaustively on corpus data a formal grammar is needed that can
be translated into a computer program which analyses the corpus utterances in the
way specified in the formal grammar. The formal grammar should be a
transparent linguistic tool, the computer program an efficient parser. Testing the
linguistic hypotheses by analysing a corpus will involve of course a test of their
correctness, but what is at least as important is a test of their completeness – does
the formal grammar have sufficient coverage to account for the corpus data? This
is especially important if it is a ‘complete’ grammar that is tested, for in that case
the starting point for the formal grammar is – apart from the linguist’s intuitions –
the knowledge accumulated in reference grammars. And because the description

8

Jan Aarts

found in reference grammars is largely limited to the written language; the
syntactic canon provided by them is therefore bound to contain gaps in the
description.

As said, for the two research activities mentioned above – varieties

differentiation and hypothesis testing – corpora constitute an essential data
source. Both start from the knowledge accumulated in previous research, the
former using it as a basis for the comparative study of language varieties, the
latter testing it for its correctness and completeness. This implies that both
research activities are representative of what Tognini-Bonelli calls the ‘corpus-
based approach’ in the quotation given in the introductory section of this paper.
As early as 1985 Sinclair rejected this approach, saying that:

The categories and methods we use to describe English are not
appropriate to the new material We shall need to overhaul our
descriptive systems. (Sinclair 1985: 252)

In other words, in the face of ‘the new material’ it is no longer possible to
maintain the framework of description built up in the tradition of the English
grammars of the twentieth century. Instead, a ‘corpus-driven approach’ is called
for which emphasizes

... the strict interconnection between an item and its environment ...
[which] leads the corpus-driven linguist to postulate an ‘extended unit
of meaning’ bringing together the lexical, the grammatical, the
semantic and the pragmatic levels. (Tognini-Bonelli 2001a: 11)

This extended notion of meaning is also referred to as a new ‘unit of currency’ for
linguistic description. The corpus-driven approach can therefore be looked upon
as an operationalization of Firth’s ideas about the essential role of context in
linguistic description and his conception of meaning, ideas which also prompted
Quirk’s principle of ‘total accountability’ (Quirk 1968: chapter 7). Needless to
say, in the corpus-driven approach the status of what I have called ‘the
accumulated knowledge about the language’ is called into question.

What is the status of intuitive data in the corpus-based and corpus-driven

approaches? Within the corpus-based approach their status is comparatively clear;
few ‘corpus-based linguists’ will hold that just because something occurs in a
corpus it has to be accounted for in a description of language use. Instead, most
will use a quality of ‘acceptability’ or ‘currency’ or ‘feasibility’ or ‘naturalness’
of utterances as a filter to ‘clean the data’. And the decision whether or not to
assign this quality to an utterance is largely intuition-based. Ultimately, therefore,
a ‘corpus-based linguist’ allows his intuition to overrule his corpus data and
hence gives primacy to the former.

Does corpus linguistics exist? Some old and new issues

9

One would expect an opposite view in the corpus-driven approach, but the

position taken by Tognini-Bonelli is a bit ambiguous. Although she first asserts
that corpus evidence

... casts serious doubts on the role of intuition and introspection,
consistently proven to be unreliable when it comes to representing
exactly how language is really used ... (Tognini-Bonelli 2001a: 91).

she continues in the same passage:

Intuition will still be considered an essential input; it will play a big
part, for instance, in selecting the phenomenon that the linguist will
choose to investigate, and ultimately it will have an important role
when it comes to evaluating the evidence in the corpus. (ibid.)

Two kinds of intuition might be involved here: the intuition the linguist has as a
researcher and his intuition as a native speaker. He will be using the former type
of intuition when he selects the phenomenon to be investigated. It is not quite
clear what is meant by ‘evaluating the evidence in the corpus’. The phrase might
apply to what I called above ‘cleaning the data’, in which case native speaker
intuition seems to be implied, but it may also refer to the researcher’s judgement
as to whether and how new corpus evidence fits in with his previous findings.
Summing up, I think we can say that each of the three attitudes to corpus data
distinguished above leaves room for intuitive data although certainly not to the
same degree.

2.4

The annotation of corpus data

A few words should be said about the annotation of corpus data. I restrict myself
to the annotation for the purposes of linguistic research.
After what has been said above about the corpus-driven approach, it is not
surprising that it rejects annotation. It is argued that the annotation imposes a pre-
existing theory on the corpus data and that therefore it ‘will ensure that the data
will finally fit the theory’ (Tognini-Bonelli 2001a: 73). Another argument that is
put forward against annotation is that it results in a ‘loss’ or a ‘reduction’ of
information (Sinclair 1992: 385 f.; Tognini-Bonelli 2001a: 73). This argument is
a bit difficult to follow. How can the addition of information to items in a corpus
diminish the information contained in the items it is added to, if the two types of
information are of a different order? For the information contained in, say, the
tags added to the words of a text is information about the descriptive framework
that has generated the tags or, more precisely, about the distribution over the text
of the word classes that are part of the descriptive framework; and this
information is entirely different from that contained in the words of the text.

10

Jan Aarts

In the corpus-based approach, which, among other things, wants to collect

quantitative data about the categories of an existing descriptive grammar, or
wants to test an existing grammar on corpus data, (the process of) annotation is an
indispensable tool. For the collection of quantitative data the result of annotation
is important, while the annotation process is especially relevant for the testing of
a grammar. The grammar has a paradigmatic and a syntagmatic dimension: the
former involves assigning (strings of) items to classes, while the latter specifies
what functional relations can hold between the members of these classes in
context. These two linguistic dimensions nicely correspond with the two
computer programs by means of which annotation is usually carried out: a tagger
and a parser. Now, although annotation is very often not carried beyond the
tagging stage, it will be clear that for the testing of a grammar that claims to give
a ‘complete’ description of general language use an integral annotation requiring
both tagging and parsing will be necessary.

Summing up, it can be said that perhaps the most prominent difference

between the corpus-based and corpus-driven approaches is in their attitude to
annotation; whereas it is an indispensable tool for the corpus-based approach, it is
anathema for linguists subscribing to the corpus-driven approach.

3.

Spoken language research and methodological issues

In the introductory section I mentioned some landmarks that point to a heightened
interest in the spoken language. While considerable attention has been paid to
matters relating to intonation and discourse ever since the London-Lund corpus
became available, there is now a shift, or perhaps rather an extension, of interest
to the syntactic structure of spoken language, which may have received its
impetus from the availability of new resources such as COLT and the spoken part
of the BNC. I mentioned two recent books that are representative of the two
strands of corpus-based linguistics I distinguished in Section 2.3: Biber et al.
(1999) and Miller and Weinert (1998). The former extends the scope of a
syntactic description of general English to also include a characterization of
language varieties in terms of the frequency and distribution of structures, and in
doing so leans heavily on the syntactic canon of standard descriptions and on
annotated data reflecting these.

5

(see Biber et al. 1999: 6–7). The second book

raises a more fundamental question with respect to the difference between spoken
and written language: are the differences marginal, so that they can be captured in
a series of ‘footnotes’ to a description of the written language, or are they such
that they have important consequences for linguistic theory, as Miller and
Weinert claim? Both books take existing descriptions as their point of departure,
the one as a means to achieve a result (i.e. varieties differentiation), the other to
test their correctness and completeness. They are therefore firmly ingrained in the
corpus-based approach. This is not the place for an evaluation of the two books
and to assess whether Biber et al. is the last word on varieties differentiation or to

Does corpus linguistics exist? Some old and new issues

11

determine if Miller and Weinert make good their claim; I have mentioned them
because they illustrate the focus on spoken language syntax and because they
represent the two main strands in corpus-based research.

In the rest of this section I want to concentrate on the second strand of

research: the use of corpus data to test the adequacy and completeness of an
existing descriptive grammar. Unlike varieties differentiation, it does not accept
an existing description as an established fact. It is willing to assume that current
descriptions of English, based as they are mainly on the written medium, ‘are not
appropriate to the new material’, to borrow Sinclair’s words. It is therefore closer
to the corpus-driven approach. However, it does not take Sinclair’s statement as
axiomatic and as a point of departure for research, but merely accepts it as a
possibility.

As I argued above, in order to test an existing description consistently and

exhaustively on corpus data, it needs to be captured in a formal grammar.
However, testing a grammar on corpus data is less easy than it sounds. In the first
place, the test does not simply imply the yes-or-no decision whether a given parse
is successful or fails. The full extent of failure occurs of course when the parser
fails to yield an analysis for a given utterance. But it is also possible that an
analysis is yielded, but is ‘wrong’ in the sense that a structure is assigned which
under no contextual conditions can ever be correct. Finally, it may happen that a
grammatically correct analysis is assigned that is contextually inappropriate.
Whenever an instance of one of these three types of failure occurs, the linguist
must settle the question whether it is due to an error or omission in the description
underlying the formal grammar or due to a mistake in the formalization of the
description – unfortunately there is no one-to-one correspondence between the
type of failure and the location of its cause. The testing of a grammar on corpus
data is therefore a laborious and time-consuming process, requiring the full force
of the linguist’s attention and ingenuity. Every single utterance needs to be
inspected manually, and errors and omissions traced back to their location either
in the formalization of the description or in the description itself; in the latter case
it must be determined whether or not the utterance in question needs to be
accounted for in the description/grammar or not (see Section 2.2).

Now, in the practice of corpus linguistic research it is not usual to go to

the trouble of formalizing a description of general language and converting the
formal grammar into a (tagger and) parser for the sole purpose of testing the
grammar and the description. On the contrary, an effort like that is normally only
made for the purpose of producing a parsed corpus for future research. Such a
project, undertaken by the TOSCA team of the University of Nijmegen, was the
development and implementation of a tagger and parser for the ICE project (see
Greenbaum 1992; Aarts 1992; Aarts et al. 1998) and its subsequent application to
ICE-GB by the Survey team of University College London. Experiences with the
ICE project have taught us that one should continually be on the alert not to
‘stretch’ the grammar too much to make it fit the data (see Aarts and Oostdijk

12

Jan Aarts

1997: 111 ff.). Working in the other direction, a number of problematic utterances
in the spoken material were ‘solved’ by the Survey team by adding to them a
‘normalized’ version which could be dealt with by the parser without any
problems. The possibility that the pressure of having to produce a syntactically
annotated corpus can lead to making ‘the data ... fit the theory’ (Tognini-Bonelli
2001a: 73) in some way or other is therefore not entirely unreal. Ideally, it would
be best to keep the production of a parsed corpus separate from the testing of a
grammar to check the correctness and completeness of its underlying description,
and to carry them out as two distinct research activities.

The confrontation of the TOSCA tagger and parser with spoken material

has led to an enhanced view of the descriptive framework that served as its
starting point, which leads us to believe that the corpus-based approach starting
from a pre-existing description can be effective and rewarding. The enhanced
descriptive framework is shown schematically in Figure 2 (reproduced from Aarts
1999).

Figure 2. Enhanced descriptive framework

The top part of the figure represents what I have called above ‘the

syntactic canon’. It is entirely based on the notion of constituency and the
hierarchical relations among constituents. These relations are nearly always
expressed by syntactic means and are of three kinds: dependency, government
and concatenation. Dependency manifests itself in all subordination/
superordination structures, government in complementation structures, where the
governing element may impose a particular form on the element realizing the
complement, while concatenation involves compounding, coordination or
apposition. The building blocks for the lower part of the figure are also
constituents, but the only ‘structural’ relation with their environment derives from
the one-thing-next-to-the-other principle – there is no attempt whatsoever at
syntactic integration. A distinction can be made between simple parataxis and

Does corpus linguistics exist? Some old and new issues

13

interrupts, which are elements interrupting regular consistency; a subclass of
these can be used to reformulate (a part of) a constituent which immediately
precedes.

An instance of simple parataxis is found in the italicized part of the

following example:

(1) I have been the radioman for the last couple months or so – last guys got

transferred, I moved up.

An example of summative reformulation is found in (2):

(2) Bartlett was the only member of staff to sport a beard, which bobbed up

and down when he talked, a process Lucy found fascinating to watch.

As the last example shows, parataxis also occurs in the written language,
although it is much more frequent in speech. The utterance in (3) offers a good
example of its frequency – it contains several instances of parataxis of various
kinds, both on phrase level and on clause level:

(3)

and it becomes an accepted wisdom that people in certain areas one of
them is London that have certain income bands the middle classes if you
like the affluent middle classes that they should go private that they will
go private.

Parataxis differs from concatenation in that the latter is usually marked
syntactically and is fully integrated in the constituency of sentences, with the
concatenated constituents always forming a clearly delineated higher-order
constituent. The distinction between constituency and parataxis is not an absolute
one; they can be seen as the two extremes on a scale of syntactisation. Paratactic
and constituent-like variants of constructions can exist side by side: an example is
(paratactic) direct speech by the side of (syntactically integrated) indirect speech
and mixed forms in between these two extremes.

Relations between structural phenomena like constituency and parataxis

and, more in general, extensions and revisions of an accepted descriptive frame-
work will only emerge, of course, if one takes that framework as one’s starting
point, as is usual in the corpus-based approach.

4.

Conclusion

Returning to the question posed in the title of this paper, we can ask ourselves
another: when can we say that a particular linguistic research activity constitutes a
linguistic discipline in its own right? Surely not by virtue of its use of a particular
type of data – corpus linguistics does not distinguish itself from another linguistic

14

Jan Aarts

discipline like theoretical linguistics on account of the fact that it uses corpus
data, nor would theoretical linguistics become a subdivision of corpus linguistics
if all its practitioners should suddenly agree that corpus data are as valuable as
intuitive data. The picture begins to change a little bit if the use of corpus data is
not so much a question of preference but of necessity; neither the corpus-based
nor the corpus-driven approach can do without the use of corpora. And perhaps
the fact itself that these two approaches exist side by side can be regarded as a
sign of an independent discipline – it is a bit difficult to maintain that a discipline
that accommodates two competing methodological approaches does not exist.

However, what really distinguishes one discipline from another is a

differerence in the object of study. And in this respect there really is a difference
between theoretical linguistics and corpus linguistics: whereas the former is
concerned with the phenomenon of language in general (call it the language
faculty, competence, I-language or universal grammar) and its object of study is
therefore language-independent, corpus linguistics is concerned with language
use, that is, with the way in which a given language system manifests itself in a
particular context and cultural environment. It aims therefore not only at an
understanding of the language system, but also of the parameters that determine
the selection from the possibilities offered by the language system and thus shape
the ultimate form that utterances must take in order to be appropriate within their
micro- and macro-context. So yes, corpus linguistics does exist, and it is a
discipline harbouring two mutually exclusive methodologies, the corpus-based
and the corpus-driven approaches, which, in all likelihood, will come up with
different but complementary results that together will give a fuller picture of
language in use.

Notes

1

In what follows I shall go on using the term ‘corpus linguistics’ as if it does.

2

See

www.hit.uib.no/colt

.

A revised version of the corpus is due to appear in the

autumn of 2001.

3

An exception must be made for two grammarians who can be looked upon as

corpus linguists avant la lettre: Fries (1952) and Scheurweghs (1959). Their
grammars are firmly based on corpus data.

4

Sinclair (1984) discusses the ‘naturalness’ of sentences, a term which at first

sight seems to imply the same sort of intuitive judgement. However, he describes
the quality of naturalness as being determined by the degree to which an utterance
obeys the constraints of the context.

5

See Biber et al. 1999: 7: ‘The overriding goal has been to use categories and

terms that are familiar and unobjectionable to the widest range of grammar users.
Since CGEL is terminologically conservative, generally following informed

Does corpus linguistics exist? Some old and new issues

15

tradition in its choice of grammatical terms and categories, we have rarely
departed from its overall framework.’

References

Aarts, B. 2000. Corpus linguistics, Chomsky and fuzzy tree fragments. In C. Mair

and M. Hundt (eds.) Corpus linguistics and linguistic theory. Amsterdam:
Rodopi. 5–13.

Aarts, J. 1991. Intuition-based and observation-based grammars. In K. Aijmer

and B. Altenberg (eds.) English corpus linguistics. Studies in honour of
Jan Svartvik. London: Longman. 44–62.

Aarts, J. 1999. The description of language use. In H. Hasselgård and S. Oksefjell

(eds.) Out of corpora: Studies in honour of Stig Johansson. Amsterdam:
Rodopi. 3–20.

Aarts, J. 2000. Towards a new generation of corpus-based English grammars. In

B. Lewandowska-Tomaszcyk and P. Melia (eds.) PALC ’99: Practical
applications in language corpora. Papers from the international
conference at the university of ód . Frankfurt am Main: Peter Lang. 17–
36.

Biber, D., S. Johansson, G. Leech, S. Conrad and E. Finegan. 1999. Longman

grammar of spoken and written English. Harlow, Essex: Longman.

Chomsky, N. 1965. Aspects of the theory of syntax. Cambridge (Mass.): M.I.T.

Press.

De Mönnink, I. 2000. On the move. The mobility of constituents in the English

noun phrase: A multimethod approach. Amsterdam: Rodopi.

Fillmore, C. 1992. ‘Corpus linguistics’ or ‘Computer-aided armchair linguistics’.

In J. Svartvik (ed.) Directions in corpus linguistics. Proceedings of Nobel
Symposium 82, Stockholm, 4–8 August 1991. Berlin: Mouton de Gruyter.
35–60.

Fillmore, C. 2001. Armchair linguistics vs. corpus linguistics revisited. In S. De

Cock, G. Gilquin, S. Granger and S. Petch-Tyson (eds.) ICAME 2001.
Future challenges for corpus linguistics. Proceedings of the 22nd ICAME
conference. Louvain-la-Neuve: Université Catholique de Louvain. 1–2.

Fries, C. C. 1952. The structure of English. New York: Harcourt, Brace and

Company.

Greenbaum, S. 1988. Good English and the grammarian. London: Longman.
Jespersen, O. 1909–1949. A Modern English grammar on historical principles.

London: Allen and Unwin.

16

Jan Aarts

Leech, G. 1992. Corpora and theories of linguistic performance. In J. Svartvik

(ed.) Directions in corpus linguistics. Proceedings of Nobel Symposium
82, Stockholm, 4–8 August 1991. Berlin: Mouton de Gruyter. 105–122.

Miller, J. and R. Weinert. 1998. Spontaneous spoken language: Syntax and

discourse. Oxford: Clarendon Press.

Nagata, H. 1989. Judgments of sentence grammaticality and field-dependence of

subjects. In Perceptual and Motor Skills 69: 739–747.

Poutsma, H. 1928–1929. A grammar of late Modern English. Part I and Part II.

Groningen: Noordhoff.

Quirk, R. 1968. Essays on the English language: medieval and modern. London:

Longman.

Quirk, R. and J. Svartvik. 1966. Investigating linguistic acceptability. The Hague:

Mouton.

Schibsbye, K. 1965. A Modern English grammar. London: Oxford University

Press.

Shohamy, E. 1996. Competence and performance in grammar testing. In G.

Brown, K. Malmkjaer and J. Williams (eds.) Performance and competence
in second language acquisition. Cambridge: Cambridge University Press.
138–151.

Sinclair, J. 1984. Naturalness in language. In J. Aarts and W. Meijs (eds.) Corpus

linguistics: Recent developments in the use of computer corpora in
English language research. Amsterdam: Rodopi. 203–210.

Sinclair, J. 1985. Selected issues. In R. Quirk and H. Widdowson (eds.) English

in the world. Cambridge: Cambridge University Press.

Snow, C. and G. Meijer. 1977. On the secondary nature of syntactic intuitions. In

S. Greenbaum (ed.) Acceptability in Language. The Hague: Mouton. 163–
177.

Strang, B. M. H. 1962. Modern English structure. London: Edward Arnold.
Sweet, H. 1900. A new English grammar, logical and historical. Part I. Oxford:

Clarendon Press.

Tognini-Bonelli, E. 2001a. Corpus linguistics at work. Amsterdam: John

Benjamins.

Tognini-Bonelli, E. 2001b. Towards a corpus-driven approach. In S. D. Cock, G.

Gilquin, S. Granger and S. Petch-Tyson (eds.) ICAME 2001. Future
challenges for corpus linguistics. Proceedings of the 22nd ICAME
conference. Louvain-la-Neuve: Université Catholique de Louvain. 99–
101.

Widdowson, H. G. 1991. The description and prescription of language. In J.

Alatis (ed.) Georgetown University Round Table on Languages and
Linguistics 1991. Linguistics and language pedagogy: The state of the art.
Washington D.C.: Georgetown University Press. 11–24.

Does corpus linguistics exist? Some old and new issues

17

Zandvoort, R. W. 1945. A handbook of English grammar. Groningen: J. B.

Wolters.