The learner as corpus designer
Guy Aston
Advanced School for Interpreters and
Translators, University of Bologna
1. Introduction
In recent years it has been
suggested that it may be both useful and motivating for teachers and learners
to construct their own corpora to analyse with appropriate interrogation
software. Such suggestions include the construction of:
·
collections
of texts written by learners themselves (Seidlhofer 2000), or which have been
previously read by them during their courses (Gavioli 1997, Willis 1998);
·
collections
of texts which illustrate a particular text-type and/or domain of use (Bowker
1998; Maia 2000a, b; Pearson 1998, 2000), or which illustrate particular
linguistic features (Bertaccini & Aston 2001; Varantola 2000).
Proposals of the first type see
‘home-made’ corpora as a means to cast further light on previously encountered
texts and linguistic features, where analysis is facilitated by virtue of the
learners’ prior familiarity with the texts involved. Proposals of the second
type instead see them as a means to investigate unfamiliar features, domains or
text-types which are inadequately documented by existing resources –
particularly in the area of translator training, where the ability to construct
specialized corpora is increasingly seen as fundamental for the identification
of domain-specific terminology (Pearson 1998, 2000; Maia 2000b).
The main ideas
underlying these proposals appear to be that:
·
‘home-made’
corpora may be more appropriate for learning purposes than pre-compiled ones,
insofar as they can be specifically targeted to the learner’s knowledge and
concerns;
·
such
corpora permit analyses which would not otherwise be readily feasible,
providing a specialized hypertextual environment for the study of particular
texts and instances;
·
by
compiling corpora for themselves, learners may gain insight into how to select
and use corpora appropriately, acquiring skills and knowledge which may be of
value to them in the future.
Against these potential
benefits, however, we must balance the costs of corpus construction. A
considerable amount of work is likely to be involved, and in comparison with
corpora constructed by professional researchers, the quality of the product is
likely to be relatively low. Home-made corpora will typically be more
opportunistic, less carefully designed and edited, and less comprehensively
encoded and annotated than those compiled by experts. Consequently, teachers
and learners may be unconvinced that it is worth building corpora of their own.
In this paper I shall outline an intermediate strategy which, I argue, can
provide some of the same benefits while considerably reducing these costs, by
providing specialized environments for particular areas of study, while at the
same time offering insights into how to design, select and use corpora
appropriately.
2.
Corpora in language learning
In both corpus linguistics and
language pedagogy it is a well-established principle that design must be based
on an analysis of users’ objectives. From this perspective, there are at least
five types of corpus-based activity that appear relevant to language learners
(Aston 2000, 2001b):
·
form-focused
activity, aiming to establish and practice the use of particular linguistic
features (“data-driven learning”: Johns 1991);
·
meaning-focused
activity, aiming either to establish meanings in a particular corpus text or to
understand the concepts referred to and the functions realized in a particular
text-type – what we might term ‘data-driven cultural learning’;
·
skill-focused
activity, aiming to develop particular reading skills and strategies,
particularly of a ‘bottom-up’ variety (Brodine 2001);
·
reference
activity, where corpora are used for support in tasks involving other texts, in
particular as aids to reading, writing and translating (e.g. Owen 1996;
Zanettin 2001 amongst many others);
·
browsing
activity, where learners alternate between the previous types of activity in
serendipitous explorations of the corpus (Bernardini 2000a, b).
Home-made corpora can
lend themselves to all these uses. We may briefly exemplify them in relation to
a learner-designed corpus of astrophysics research articles (Raffa 2000). From
a form-focused perspective, this corpus is an excellent resource for
identifying astrophysical terminology and establishing its uses. From a
meaning-focused one, it provides many opportunities to learn about white
dwarfs, black holes, red giants, etc., as well as about the general methodology
of astrophysical research. From a reading skills perspective, it can provide
focused practice in such areas as the parsing of complex nominal groups, or the
resolution of anaphoric and cataphoric reference in this kind of text. And
obviously, it can serve as a reference tool while reading, writing or
translating astrophysics research articles (the corpus was originally designed
to provide a resource for non-native speakers engaged in astrophysical
research), since it provides an intertextual background against which to
construct and evaluate interpretative or productive hypotheses. Last but not
least, it is a corpus that can be browsed serendipitously, travelling from one
linguistic or cataclysmic variable to another.
All
these types of activity, it should be stressed, lend themselves to being
contextualized in a framework of communication, since they provide numerous
opportunities for report and discussion of linguistic, social, cognitive and
methodological issues. Thereby they allow not only extensive communicative
practice, but can also develop linguistic awareness and encourage learning
autonomy (Aston 2000, 2001b).
2.1 Why
make your own corpora?
Why, however, should learners bother
to construct their own corpora in order to engage in activities of these kinds?
To use an analogy, making your own corpus seems rather like making your own
fruit salad. Why make your own when you can buy a tin off the supermarket
shelf? The reasons (for both corpora and fruit salads) appear similar:
·
Control.
You can devise your
own recipe, choosing your own ingredients, thereby obtaining assortments that
may be unavailable in pre-packaged versions. There was no publicly-available
corpus of astrophysics research articles which could be used to investigate the
particular linguistic and conceptual characteristics of the latter. The BNC,
for instance, contains only 10 written texts which mention astrophysical white
dwarfs – too small and hetergoeneous a sample to warrant generalizations in
this domain.
·
Certainty.
If you make your
own fruit salad, you have a good idea of what went into it, and this makes it
easier to decide what that strange-looking bit was, or why it tastes too bitter
or too sweet. It is much easier to interpret concordances or numerical data if
you know exactly what texts a corpus consists of, since this allows a greater
degree of top-down processing. It takes some time to gain sufficient
familiarity with a pre-packaged corpus to recognize particular texts, or to
interpret results in the light of its particular quirks. With one you have made
yourself it is easier to make adjustments, and to recognize the limits to
inferences.
·
Creativity.
Corpus-making, like
cooking, can be fun, giving scope for individual panache. It is also gratifying
when your fruit salad turns out to be delicious, or your corpus a useful
resource.
·
Critical
awareness. Through
trial and error, and consulting books and experts, you will probably become a
better chef (whether of corpora or fruit salads) as you compare the effects of
different proportions of different ingredients, or discover that mixing popular
science with research articles is not always a good idea. Even if you are
unsatisfied with the results of your efforts, the experience of making your own
seems likely to make you a more critical corpus user, increasing awareness of
how design affects the results – results which are (to quote Sinclair 1991: 13)
“only as good as the corpus”.
·
Communication.
Making your own
corpus or fruit salad can have more social spin-offs than opening a supermarket
tin, providing lots to talk about with co-constructors and with other chefs, as
well as with the consumers of the end product. Making your own opens up a whole
range of opportunities for learners to discuss how best to compile and encode
corpora for particular purposes, and to discuss how good they effectively are
for these purposes and how they might be improved.
2.2 Why
use standard pre-packaged corpora?
Since there is a market for tinned
fruit salad, there must presumably be some arguments in its favour.
Pre-packaged corpora typically offer advantages compared with the home-made
variety in terms of:
·
Reliability.
A pre-packaged
corpus (provided it is well-designed and fits your needs) is likely to give
more reliable results. Just as tinned fruit salads are subject to quality
controls based on market research, it is more likely that a pre-packaged corpus
will be reasonably “representative” of the population it aims to cover (Biber
1993), and carefully “balanced” amongst the different types of text which make
up that population.
·
Documentation.
Pre-packaged
corpora generally provide better documentation than home-made ones. With an
off-the-shelf fruit salad, it is easier to find out the exact sugar content,
and exactly how many calories you are consuming per portion. Pre-packaged
corpora will generally include metatextual information about individual texts
and their sources, and categorizations of their contents. They may also
incorporate details of text structure, annotation of part-of-speech or
syntactic features, etc.
·
Designer
software. Many
off-the-shelf corpora come with specially-designed interrogation software, such
as the SARA interface, which was designed to allow the BNC’s metatextual
documentation to be exploited in interrogating the corpus. All-purpose, plain
text concordancers such as Wordsmith Tools (Scott 1999) or MonoConc
(Barlow 1998) cannot generally interpret such information satisfactorily.
·
Convenience.
It is clearly less
effort to use a pre-packaged corpus than to make your own. All you have to do,
as it were, is open the tin. Most readers of this paper will have their own
favourite corpora, and many will feel that using them is vastly preferable to
going through the effort of designing and constructing their own. This is
perhaps the main factor to justify the compromise strategy to be outlined in
the next section.
3. The
pick’n’mix compromise
One way to avoid much of the effort
involved in constructing your own corpus (or fruit salad) is to steal the
necessary ingredients from elsewhere. The web is one prolific source of corpus
ingredients (Bertaccini & Aston 2001), but these may require complex
searches and considerable adaptation before they can be used (Pearson 2000). A
more attractive strategy may be to extract a subcorpus from a larger corpus,
whose texts can be treated as pre-prepared ingredients, just waiting to be
selected in the desired proportions. The analogy here might be with a (fruit)
salad bar, where you put together your own mixture from a series of bowls, each
containing one kind of fruit which is ready washed, peeled and chopped into pieces
of the right size. Here you can control the ingredients, selecting those which
appeal to you. You can decide for a preponderance of raspberries, or indeed
select or avoid one particularly bloated raspberry. You can omit the grapefruit
or add another dollop of cream. The following conversational extract
illustrates the process fairly well (however dubious you may be as to the
result):
PS0A2: Right there’s erm (.) it’s a tin of
fruit salad but I’ve put in some er kiwi and grapes (.) so it’s fresh fruit,
it’s in its own juice, so it’s not in a heavy thick juice, there’s Vienetta or
you can have a bit of each
PS09U: Well I’ll have a little bit of each
then please
(BNC Sampler:
KC2)
As well as an increase
in control, the fruit salad bar provides an increase in certainty (you have a
clearer idea what went in), and also in opportunity for creativity and
communication. Repeated attempts may also lead to increased critical awareness.
The fruit salad bar is convenient, requiring little effort bar that of decision-making.
Substantial reliability can be maintained (assuming the components are subject
to quality control), allowing you to draw on the documentation provided for
each component, and to exploit it in designing and consuming your own mixture.
In the same way, if you construct your own subcorpus
from the ingredients provided by a larger corpus, you can (within the limits of
what is on offer) choose your own text-types, and indeed individual texts. Not
only can you thereby increase control over and certainty as to the content, but
you can also indulge your creativity, and exploit opportunities to communicate
about your strategies and their results. In the process you may – through trial
and error – become more critically aware of what are (and are not) useful
subcorpora, and what are (and are not) appropriate design criteria. As we shall
see in the next section, constructing your own subcorpora in this manner can
maintain much of the reliability and conserve the documentation attached to the
original corpus, as well as allowing you to exploit software specifically
designed for use with that corpus. It is also far less work than compiling a
corpus of your own.
4. The
SARA subcorpus option
Recent releases of the BNC Sampler
(1999) and the just-published BNC World edition (2001) come with enhanced
software (SARA98) which allows users to define and then analyse subcorpora
within the corpus in question. Since SARA can be used with any TEI-conformant
corpus, the procedures outlined below are not, in theory, limited to the BNC,
but can also be applied to other corpora which adopt similar encoding
principles.
The
SARA subcorpus option allows the user to define subcorpora consisting of:
·
one or
more specific texts selected from a list of all the texts in the corpus;
·
all
those texts which contain solutions to a particular query, for instance all
those containing the word “Austria”.
Since SARA permits queries
concerning the metatextual information provided with the texts as well as
regarding their linguistic content, this second procedure can also be used to
define subcorpora consisting of all the texts belonging to a particular design
category (spoken/written, monologue/dialogue, imaginative/informative, etc.),
or to a particular descriptive one (e.g. produced by a particular type of
author/for a particular type of audience, or belonging to a particular genre:
cf. 4.4 below and Lee, this volume). The two procedures can also be combined,
with manual editing of the list of texts obtained from a particular query.
Once
defined, a subcorpus can be saved and used as the basis for subsequent queries.
It is also possible to index a saved subcorpus for easy re-use in subsequent
sessions. In the rest of this paper I illustrate some practical examples of
indexed subcorpora extracted from the BNC Sampler, and relate these examples to
the learner uses of corpora described earlier (cf. 2 above).
4.1 A
specific text as subcorpus
Scrolling through the list of texts
in the BNC Sampler, I was struck to discover the Monster Raving Loony Party’s
Draft manifesto for the British General Election of 1992 (AP6: Figure 1). Given
the resurgence of extremist political parties in Europe today, I felt that
participants at this conference, like many learners, might share my curiosity
concerning this text:
Figure 1. Selecting texts for inclusion in the
subcorpus
If we select this text and save it
as a subcorpus, we can then begin to pose queries about it. In the first place,
we can simply ask for a list of the most common words in it, shown in Figure 2:
Figure 2. Words whose frequency in the
subcorpus is greater than 20
One striking feature here is the
high frequency of the modals will and shall. This is presumably
because manifestos announce programmes, providing declarations of intent as to
future action. This is confirmed when we look at the results of a query for
these two forms (followed by be + past participle: Figure 3):
Figure 3. Shall/will be VVN in
the Monster Loony Party Manifesto
While the distribution of shall
and will in these citations is not easy to account for, the concordance
clearly demonstrates how a subcorpus consisting of just one text can highlight
its distinctive formal characteristics, and also cast light on its style and
meanings – as well as providing ample opportunity for discussion.
4.2
A bad language subcorpus
Subcorpora need not, of course, be
limited to a single text. If we carry out a query in the spoken texts of the
BNC Sampler for forms beginning with the characters fuck, we find 225
occurrences in 14 texts. In Figure 4, these texts are listed in order of the
number of occurrences found (see the Query 1 column), so we can easily
select all those which contain these forms as a subcorpus:
Figure 4. Texts containing forms of fuck[1]
In a browsing activity,
this subcorpus could be employed to explore various aspects of bad language
use. For example, we can generate a list of the collocates of the forms in
question to cast light on their typical usage within these texts (Figure 5).
What emerges most strikingly here is the collocate oh, which occurs no
less than 45 times in a span of 3 words to the left and to the right. As a
curiosity, it then comes naturally to ask what other words (if any) oh
precedes in this subcorpus (Figure 6):
Figure 5. Collocates of forms of fuck
in the bad language subcorpus, distinguished by part-of-speech and ranked by
frequency
Figure 6. Right collocates of oh in
the bad language subcorpus, ranked by z-score
Ranking the collocates in a span of
2 words to the right of oh in order of significance (Figure 6), we
obtain a list which suggests that we have indeed created a subcorpus of bad
language texts, including a number of other expressions with oh which
learners wishing to improve their abusive competence might explore to their
profit. It might also be of interest to compare the usage of male and female
speakers – is it women or men who say oh
dear? This is another area which can be investigated thanks to the metatextual
information provided in the corpus and the specialized design of the
interrogation software.
Clearly, the number of
texts included in this subcorpus is very small, and we cannot assume that they
constitute a reasonable cross-section of spoken texts involving bad language.
Nonetheless they may still enable the user to generate, if not to definitively
test, hypotheses as to use in this area. The other subcorpora discussed in the
next sections are similarly too small to permit definitive conclusions, but
they can again provide interesting suggestions as to language use in certain
kinds of contexts.
4.3 Subcorpora
of encoded categories of texts
The spoken texts in the BNC fall
into two main classes, demographic and context-governed. These labels, which
refer to the way in which recordings were collected, distinguish free
conversations from talk recorded in settings of an institutional nature –
classrooms, courtrooms, business meetings, and the like. The context-governed
texts may be either monologue or dialogue – a feature which is again indicated
in the BNC text classification. A search for context-governed monologue in the
BNC Sampler finds 17 texts, while a search for context-governed dialogue finds
29. We can use the same procedure as in the last section to list the texts
which match these queries, and to form separate subcorpora of context-governed
monologue and dialogue texts.
Analysing
these two subcorpora, we find considerable differences. If we compare the 200
most frequent words in each, we find that could, had, he, know, their, were,
when, who, and your are ranked more than 20 positions higher in the
monologue subcorpus, while ’ll, ’m, any, no, pounds, right, yeah and
yes are more than 20 positions higher in the dialogue subcorpus. The differences
in the frequencies of yeah, yes, no and right suggest that
speakers may be less concerned with explicitly negotiating agreement when they
hold a monopoly of the floor (for instance, we find that there are no
occurrences of all right in the monologue subcorpus). There also appears
to be a difference in the use of pronouns: for instance, we find that we is
relatively more frequent in dialogue, and you in monologue (Figure 7):
|
|
we |
you |
|
Monologue |
2014 |
4253 |
|
Dialogue |
4949 |
6635 |
Figure 7. Occurrences of we and you in the monologue and dialogue
subcorpora
Perhaps this is again due to the
unwillingness of speakers in monologue contexts to claim shared attitudes with
their audiences, given that the latter have little chance to disagree. By
examining sample sets of citations, it may however be possible to advance other
hypotheses to account for these differences, and in any case for learners to
reflect on the linguistic differences between monologue and dialogue settings.
4.4 Subcorpora
based on other categorizations
It is also possible to define
subcorpora which are based on different categorizations from those originally
encoded by the corpus designers. Lee (this volume) provides a personal
categorization of the written texts in the BNC, based on criteria such as
academic vs non-academic, prose vs poetry, fiction vs non-fiction.[2]
We can use Lee’s lists to define subcorpora from the Sampler corresponding to
such categories as:
·
academic
non-fiction (13 texts);
·
non-academic
non-fiction (15 texts);
·
prose
fiction (13 texts).
Looking at wordlists for
these three subcorpora, we discover a number of items which appear to be more
common in one than in the other two. For example, the following adverbs in -ly
all occur more than 15 times in the subcorpus indicated, and less than 10 times
in each of the others:
academic non-fiction: accordingly, essentially, eventually, largely, namely, notably, respectively, surprisingly
;
non-academic non-fiction: effectively, merely, normally, obviously, possibly, specially
;
·
prose fiction: carefully, quietly, slightly,
slowly, softly, surely, truly.
If we take just one of the items from the academic non-fiction list, largely, we can follow the traditional procedures of data-driven inductive learning to explore its uses in this genre (Johns 1991). On the one hand, largely appears to qualify participial predicates with a negative semantic prosody, collocating with
expressions like annihilated by, confined to, denied to,
ignored, invalidated by and limited to. On the other hand, it appears
to qualify linking expressions introducing causes, such as accounted for,
based upon, because, dependent upon and due to (Figure 8):
Figure 8. Largely in academic
non-fiction (Lee’s categorization)
There are two apparent exceptions to
these patterns in this concordance – largely by catching fish and largely
genetic – which I leave to the reader to account for.
A concordance of this
kind casts light on the language of the text-type in question by providing a
limited, relatively homogeneous set of citations, which are easier to
categorize and interpret than ones taken from a broader variety of texts. While
in no way permitting an exhaustive account of the ways in which the word largely
can be used, the concordance clearly illustrates two uses which would seem
to play a significant role in this text-type, and which may therefore be of use
to those who are learning to deal with such texts in their reading or writing.
A subcorpus of this kind
could also be used in other ways: to generate exercises in reading academic
prose (for instance in parsing nominal groups containing particular heads:
Brodine 2001), or as reference tools to assist learners with other features of
academic writing. Taking instead a browsing perspective, investigating largely
may lead the learner to examine near-synonyms (mostly, mainly, to a large
extent, for the most part), or to examine other collocates of such
expressions as confined to and limited to.
Nor
need the learner’s investigations be confined to this subcorpus alone. We have
seen that largely appears particularly frequent in texts classed as
written academic prose, but in analysing its uses in those texts, we have not
posed the question of whether it has the same uses in other text types. The
entry for largely in the Collins Cobuild dictionary (1995) suggests that
this may not be the case, since it cites examples from the Bank of English
which fit into neither the causative or the negative prosody category: The
fund is largely financed through government borrowing … I largely work with
people who already are motivated … Their weapons have been largely stones.
Working with a subcorpus frequently invites comparison of the results obtained with those from other
subcorpora, or indeed from the whole corpus from which the subcorpus has been
derived. Starting from a limited set of texts of a single type will simplify
this process insofar as it reduces the initial number and variability of
citations – providing, that is, a line of approach to the analysis of samples drawn
from the full corpus (Gavioli 2001).
5. Conclusions
Through these examples I hope to
have illustrated how subcorpora derived from the BNC Sampler can allow learners
to carry out activities of each of the types listed in 2 above, in particular:
·
to study and compare forms in particular texts or
text-types, contrasting these with those in other texts or text-types;
·
to study and compare meanings in particular texts or
text-types, contrasting these with those in other texts or text-types;
·
to carry out focussed reading practice;
·
to adopt appropriate reference tools for particular
tasks;
·
to carry out focussed browsing.
At
the same time, I would argue, subcorpora like these share many of the
characteristics which have motivated proposals to use ‘home-made’ corpora in
language learning (cf. 1 above). To summarize:
·
Subcorpora
can provide small, manageable amounts of data of a more homogeneous nature than
is possible with large mixed corpora, thereby facilitating analysis. It is, of
course, essential for users to recognize that such subcorpora are neither
sufficiently large, nor sufficiently carefully designed, to be considered
“representative” samples of the text-types involved, and that inferences made
from them should not be treated as definitive. However, as I have stressed
elsewhere, language learning appears to be a matter of progressive
approximation on the basis of ever-growing experience (Aston 1997). Thus, while
a learner who sees the two uses of largely presented in 4.4 above cannot
pretend to have fully understood all the potential uses of the word in academic
discourse, s/he arguably has formed an idea of two of the main ways in which it
can be used, and is well placed to refine this knowledge further in the future.
·
Subcorpora can provide a specialized environment for
the study of particular texts and instances. As the subcorpora described in
this paper were all taken from the BNC Sampler (which contains a mere 2% of the
texts in the full BNC), they were extremely small, and also relatively unspecialized.
With the new BNC World edition, however, more highly specialized subcorpora can
be constructed – not just of written academic discourse, but of written
academic discourse in the field of medicine, not just of spoken monologue but
of lectures, and so on.
Increased specialization entails increased homogeneity, and consequently more
precise focussing and reduced dispersiveness in corpus use. Alternatively, much
larger subcorpora can be extracted for categories like those discussed in the
last section, with a corresponding increase in reliability: there are, for
instance, 504 texts classed as written academic prose in the World edition,
which include 2348 occurrences of largely. These larger numbers may
however be difficult for learners to manage, initially requiring analyses of
smaller selections. Consequently there would still seem to be a pedagogic place
for small subcorpora such as those illustrated here, as points of initial focus
for the learner to generate hypotheses which can then be tested against the
larger subcorpora obtainable from the full BNC. Nor should we forget that the complete corpus, with its myriad
paths for the motivated learner to adventure down, is always there to be
consulted (Bernardini 2000).
·
If
learners create and select their own subcorpora for particular tasks, they will
also acquire practice and experience in corpus design which may be of use to
evaluate corpora with which they are unfamiliar, or to create corpora of their
own from less structured sources, such as the Web. These skills would appear
useful not only for would-be readers, writers and translators of specialized
texts, but also for more general-purpose language learners, insofar as the
latter need to develop a sensitivity to genre and register variation. It is
clear that subcorpora extracted from large mixed corpora like the BNC cannot be
expected to satisfy all possible requirements – only a specifically collected corpus of
astrophysics is likely to provide enough information concerning white dwarfs or
the rhetoric of astronomers; only a corpus of learner texts will satisfy the
need to study learner or lingua franca English (Granger 1998, this
volume; Seidlhofer 2000). But, one might argue, it is precisely because
subcorpora have these limits that they can provide valuable ways of learning to
design and use corpora in general.
References
Aston, Guy (1997), “Small and large
corpora in language learning,” in: Lewandowska-Tomaszczyk/Melia (1997), 51-62.
Aston, Guy (2000), “The British
National Corpus as a language learning resource,” in: Battaner/Lopez (2000),
15-40.
Aston, Guy, ed. (2001a), Learning
with corpora, Houston TX: Athelstan.
Aston, Guy (2001b), “Learning with
corpora: an overview”, in: Aston (2001a), 7-45.
Barlow, Michael (1998), MonoConc,
Houston TX: Athelstan.
Battaner, M. Paz/Carmen López, eds.
(2000), VI jornada de corpus linguistics, Barcelona: Institut
universitari de lingüística aplicada, Universitat Pompeu Fabra.
Bernardini, Silvia (2000a), Competence,
capacity, corpora, Bologna: CLUEB.
Bernardini, Silvia (2000b),
“Systematizing serendipity: proposals for concordancing large corpora with
learners,” in: Burnard/McEnery (2000), 225-234.
Bernardini, Silvia/Federico
Zanettin, eds. (2000), I corpora nella didattica delle lingue. Bologna:
CLUEB.
Bertaccini, Franco/Guy Aston (2001),
“Going to the Clochemerle: exploring cultural connotations through ad hoc
corpora,” in: Aston (2001a), 198-219.
Biber, Doug (1993),
“Representativeness in corpus design,” Literary and linguistic computing
8, 243-257.
Bowker, Lynne (1998), “Using
specialized monolingual native-language corpora as a translation resource: a
pilot study,” Meta 43, 631-651.
The British National Corpus World
Edition (2001),
Oxford: Oxford University Computing Services.
The BNC Sampler (1998), Oxford: Oxford University
Computing Services.
Brodine, Ruey (2001), “Integrating
corpus work into an academic reading course,” in: Aston (2001a), 138-176.
Burnard, Lou/Tony McEnery, eds
(2000), Rethinking language pedagogy from a corpus perspective, Frankfurt
am Main: Peter Lang.
Collins Cobuild Dictionary (1995, 2nd edition), London:
HarperCollins.
Gavioli, Laura (1997), “Exploring
texts through the concordancer: guiding the learner,” in:
Wichmann/Fligelstone/McEnery/Knowles (1997), 83-99.
Gavioli, Laura (2001), “The learner
as researcher: introducing corpus concordancing in the classroom,” in: Aston
(2001a), 108-137.
Granger, Sylviane, ed. (1998), Learner
English on computer, London: Longman.
Johns, Tim (1991), “Should you be
persuaded: two samples of data-driven learning materials,” English language
research journal 4, 1-16.
Lewandowska-Tomaszczyk,
Barbara/James Melia, eds. (1997), Palc ’97: Practical applications in
language corpora, Łódz: Łódz University Press.
Maia, Belinda (2000a). “Making
corpora: a learning process,” in: Bernardini/Zanettin (2000), 47-60.
Maia, Belinda (2000b), “Comparable
and parallel corpora – and their relationship to terminology work and
training”, Paper presented at 2nd Corpus Use and Learning to Translate
conference, Bertinoro.
Owen, Charles (1996), “Do corpora
require to be consulted?” ELT journal 50, 219-224.
Pearson, Jennifer (1998), Terms
in context, Amsterdam: Benjamins.
Pearson, Jennifer (2000), “Surfing
the internet: teaching students to choose their texts wisely,” in:
Burnard/McEnery (2000), 235-239.
Raffa, Giuliana (2000), The
astrophysics research article: a corpus-based analysis. Unpublished
dissertation. Forlì: SSLMIT.
Scott, Mike (1999), Wordsmith
tools ver. 3.0, Oxford: Oxford University Press.
Seidlhofer, Barbara (2000),
“Operationalizing intertextuality: using learner corpora for learning,” in:
Burnard/McEnery (2000), 207-223.
Sinclair, John (1991), Corpus
concordance collocation, Oxford: Oxford University Press.
Varantola, Krista (2000),
“Translators and disposable corpora”, Paper presented at 2nd Corpus Use and
Learning to Translate conference, Bertinoro.
Wichmann, Ann/Steve Fligelstone/Tony
McEnery/Gerry Knowles, eds. (1997), Teaching and language corpora,
London: Longman.
Willis, Dave (1998), “Learners as
researchers,” Paper presented at IATEFL 32nd annual conference, UMIST,
Manchester.
Zanettin, Federico (2001), “Swimming
in words: corpora, translation, and language learning,” in: Aston (2001a),
177-198.