Register as a predictor of linguistic variation

DOUGLAS BIBER

Abstract

Over the last two decades, corpus analysis has been used as the basis for
s­everal important reference grammars and dictionaries of English. While these
reference works have made major contributions to our understanding of En­
glish lexis and grammar, most of them share a major limitation: the failure to
consider register differences. Instead, most reference works describe lexico-
grammatical patterns as if they applied generally to English.
The main goal of the present paper is to challenge this practice and the
u­nderlying assumption that the patterns of lexical-grammatical use in English
can be described in general/global terms. Specifically, I argue that d­escriptions
of the average patterns of use in a general corpus do not accurately describe
any register. Rather, the patterns of use in speech are dramatically different
from the patterns in writing (especially academic writing), and so minimally
an adequate description must recognize the two major poles in this continuum
(i.e., conversation versus informational written prose).
The paper begins by comparing two general corpus approaches to the study
of language use: variationist and text-linguistic. Although both a­pproaches can
be used to investigate the use of words, grammatical features, and registers, the
two approaches differ in their bases: the first gives primacy to each linguistic
token, while the second gives primacy to each text. This difference has i­mportant
consequences for the overall research design, the kinds of variables that can be
measured, the statistical techniques that can be applied, and the particular re­
search questions that can be asked. As a result, the importance of register has
been more apparent in text-linguistic studies than in studies of linguistic v­ariation.
The bulk of the paper, then, argues for the importance of register at all lin­
guistic levels: lexical, grammatical, and lexico-grammatical. Analyses com­
paring conversation and academic writing are discussed for each level, show­
ing how a general ‘average’ description includes some characteristics that are
not applicable to one or the other register, while also omitting other important
patterns of use found in particular registers.

Corpus Linguistics and Linguistic Theory 8–1 (2012), 9 – 37 1613-7027/12/0008–0009

DOI 10.1515/cllt-2012-0002 © Walter de Gruyter
10 D. Biber

Keywords: register differences, linguistic variation, conversation, academic

writing, research designs

1. Introduction

One major contribution of corpus research over the past 40 years has been
the increasing awareness that lexis and grammar are intimately intertwined.
Numerous studies have exploited corpus resources to describe the systematic
lexical associations of a target grammatical construction (cf. the survey of
studies in Kennedy 1998: 121–154). For example, some of the most common
verbs in English occur most of the time in the simple present tense (e.g., think,
know, want, mean), while others occur more often in the simple past tense (e.g.,
said, came, took) (Kennedy 1998: 123). The modal would usually occurs with
a simple verb in written academic writing (the expense would fall ), while the
modal can more often occurs with be + past participle (the procedures can be
applied to . . .) (Kennedy 1998: 133). The preposition between most c­ommonly
occurs as a noun modifier in written English (e.g., difference between, relation­
ship between, agreement between), in contrast to the preposition through, which
more often has an adverbial function (e.g., go through, pass through, come
through) (Kennedy 1998: 142–143; cf. Kennedy 1991). A more recent study of
this type is Römer’s (2005) detailed description of the verbs associated with
progressive aspect.
Several major reference grammars have also employed corpus investiga-
tions to identify the words associated with grammatical constructions (e.g., lists
of the verbs and adjectives that can control a that-clause or a to-clause). One of
the earliest grammars to include extensive lexical information of this type is
the Comprehensive Grammar of the English Language (Quirk et al. 1985; see
also Quirk et al. 1972), while the Collins COBUILD English Grammar (1990),
the Longman Grammar of Spoken and Written English (Biber et al. 1999), and
the Cambridge Grammar of English (Carter and McCarthy 2006) are more
recent examples.
These grammars all take a deductive (‘corpus-based’) approach: g­rammatical
constructions are distinguished on the basis of traditional linguistic criteria,
and then the set of words associated with those constructions are identified
through corpus analysis. In contrast, the ‘pattern grammar’ reference books
(Francis et al. 1996, 1998) take the opposite approach, beginning with words
and then identifying the (grammatical) “phraseology frequently associated
with (a sense of ) a word” (Hunston and Francis 2000: 3). These books show
that there are systematic regularities in the associations between sets of words,
grammatical frames, and particular meanings on a much larger scale than it
 Register as a predictor of linguistic variation 11

could have been possible to anticipate before the introduction of large-scale

corpus analysis.
Surprisingly, most of these previous reference books and studies do not re-
port quantitative findings, and it is not clear that specific quantitative analyses
were undertaken as part of the research endeavor. These studies sometimes
suggest that they are based on quantitative analysis, claiming to describe the
words that ‘commonly’, ‘frequently’, or ‘usually’ occur with a given construc-
tion. However, in most cases, no actual quantitative findings are reported.1
There are major advantages to reporting quantitative findings. First, a quan-
titative approach requires explicit operational definitions and accountability in
the linguistic analysis. When specific quantitative findings are not reported, the
reader must simply take it on faith that the identified lexical-grammatical pat-
terns are in fact ‘common’ or ‘frequent’. Similar to observational research in
the natural sciences and social sciences, corpus research is based on ‘samples’
(corpora) which are carefully designed to represent ‘populations’ ( particular
registers and dialects in a particular language). However, it is virtually incon-
ceivable in other disciplines that a researcher would go to the trouble of de-
signing and collecting such samples but then not analyzing the quantitative
distributions of variables in those samples.
But more importantly, quantitative findings provide essential information
about language use by documenting the extent to which a lexical-grammatical
association holds. That is, most linguistic phenomena are not distributed in a
simple binary opposition of ‘frequent’ versus ‘rare.’ Rather, there is normally a
continuous range of variation, and quantitative findings are required to ade-
quately describe those patterns. As a result, all the analyses presented below
employ quantitative analysis.
However, there is a second characteristic of most major reference works that
is less often recognized as a shortcoming: the failure to consider register differ-
ences (cf. Biber and Conrad 2009). That is, most reference works describe
lexico-grammatical patterns as if they applied generally to English. When quan-
titative findings are reported, they are based on an entire corpus (e.g., the Brown
Corpus or the ICE-GB Corpus), which is taken to represent English generally.
The main goal of the present paper is to challenge this practice and the
u­nderlying assumption that the patterns of lexical-grammatical use in English
can be described in general/global terms. Specifically, I argue that descriptions
of the average patterns of use in a general corpus do not accurately describe
any register. Rather, the patterns of use in speech are dramatically different
from the patterns in writing (especially academic writing), and so minimally an
adequate description must recognize the two major poles in this continuum
(i.e., conversation versus informational written prose).
The paper begins by comparing two general corpus approaches to the study
of language use: variationist and text-linguistic. Although both approaches can
12 D. Biber

be used to investigate the use of words, grammatical features, and registers, the
two approaches differ in their bases: the first gives primacy to each linguistic
token, while the second gives primacy to each text. This difference has impor-
tant consequences for the overall research design, the kinds of variables that
can be measured, the statistical techniques that can be applied, and the par­
ticular research questions that can be asked. As a result, the importance of
register has been more apparent in text-linguistic studies than in studies of
linguistic variation.
The bulk of the paper, then, argues for the importance of register at all lin-
guistic levels: lexical, grammatical, and lexico-grammatical. Analyses com-
paring conversation and academic writing are discussed for each level, show-
ing how a general ‘average’ description includes some characteristics that are
not applicable to one or the other register, while also omitting other important
patterns of use found in particular registers. (The data for several of these case
studies are taken from the Longman Grammar of Spoken and Written English,
referred to as LGSWE below.)

2. Perspectives on ‘frequency’: Linguistic versus text-linguistic

variation

Corpus‑based studies generally have one of two primary research goals: 1) to

describe the variants and use of a word or linguistic structure, or 2) to describe
differences among texts and text varieties, such as registers or dialects. The
first goal relates to classic studies of linguistic variation (e.g., the choice be-
tween active and passive voice), while the second is used to investigate text-
linguistic variation, contrasting the words and grammatical structures typically
found in different texts and varieties.
Biber, Conrad and Reppen (1998: 269–274) as well as Biber and Jones
(2009) identify three major types of research design that have been employed
in corpus research. The primary difference among these research design types
is the unit of analysis (or the ‘observations’), which in turn makes each design
type appropriate for one of the above two research goals:
– Type A studies: the unit of analysis is each occurrence of a linguistic fea-
ture. Type A studies are thus designed for Research Goal 1 (describing the
variants of a linguistic structure).
– Type B studies: the unit of analysis is each individual text. Type B studies
are thus designed for Research Goal 2 (describing the differences among
texts and text varieties).
– Type C studies: the unit of analysis is the entire corpus (or different sub-
corpora). Type C studies can be used for either Research Goal 1 or 2, but
they do not permit the use of inferential statistics.
 Register as a predictor of linguistic variation 13

The units of analysis are the ‘observations’ that are described in a study. For the
most part, the observations in Type A studies do not have quantitative charac-
teristics, while the observations in Type B studies are analyzed in terms of
quantitative characteristics. (Type C studies differ from both of the others in
that there are actually very few observations – usually only 2 or 3 observations
– because each sub-corpus is treated as an observation.)
For example, a Type A study of relative clauses might have the goal of
p­redicting the choice of relative pronoun (who, which, that). Each relative
clause would be an observation, coded for restrictive versus non-restrictive
function and for the animacy of the head noun. All three variables (relative
pronoun, clause type, head noun type) in this study would be nominal rather
than numeric.2
In this case, descriptive statistics give the frequencies for each combination
of categories (e.g., how many occurrences of the relative pronoun that are used
with animate head nouns). However, it is difficult to document variation (or
dispersion) across the corpus, and the distribution of linguistic variants across
texts is not considered in this type of analysis (see also Gries 2006). Most Type
A studies obtain frequencies for the corpus as a whole, but give no consider-
ation to variation among the texts in a corpus.
Statistically, this type of design must be analyzed using non-parametric
techniques, such as chi-squared or log-likelihood. As a result, a Type A study
is ideal for studying the proportional preference for one or another variant, or
the proportional extent to which a linguistic variant occurs with particular
c­ontextual factors. However, this design type is not well suited to studying the
overall extent to which a linguistic feature is used in texts.
The important point here is that Type A research designs do not provide the
basis for determining rates of occurrence, so they cannot be used to determine
if a feature or variant occurs more often in one register or another. This is po-
tentially confusing, and even published research studies sometimes make this
mistake in interpreting statistical analyses: describing proportional preference
for one variant over another as if it was the same as a higher rate of occurrence.
Type A studies can tell us what the preferred variant is in a register, and how
registers differ in their reliance on a particular variant. For example, Figure 1
shows the proportional use of that versus 0 complementizer in verb + that-
clause constructions, as in:
The commission agreed that this solution . . .
versus
I thought [0] you did it
Figure 1 compares conversation, newspaper writing, and academic prose, based
on a sample of 1,000 that-clauses taken from each of the registers. (The actual
proportions are given in Table 1; cf. LGSWE 1999: 680).
14 D. Biber

Figure 1. Proportional preference for Verb + that versus Verb + 0 in three registers

Table 1. P
 roportional preference for Verb + that versus Verb + 0, based on a sample of 1,000
that-clauses taken from conversation, newspaper writing, and academic prose

Conversation Newspaper Academic prose

Verb + that 141 (14%) 733 (73%) 940 (94%)

Verb + 0 859 (86%) 267 (27%)    60 (6%)
Total 1,000 (100%) 1,000 (100%) 1,000 (100%)

It would be easy to look at Figure 1 (and Table 1) and incorrectly conclude

that verb + that is most common in academic writing: 6 times more frequent
than in conversation, and also considerably more frequent than in newspaper
writing. But in fact, Table 1 does not provide the basis for such conclusions,
and they are actually incorrect.3
Rather, Figure 1 (and Table 1) shows that when a that-clause is used in
a­cademic prose, it will almost always retain the that complementizer. When a
that-clause is used in conversation, it will usually omit the complementizer.
That-clauses in newspaper writing usually retain the complementizer, but this
preference is less pronounced than in academic prose. These are genuine
r­egister differences. However, it would be incorrect to therefore conclude that
that-retention is more common in academic prose than in newspaper writing or
conversation.
In contrast, Figure 2 presents that the actual rates of occurrence for verb +
that. Figure 2 shows that the rate of that-retention is much higher in newspaper
 Register as a predictor of linguistic variation 15

Figure 2. Rates of occurrence for Verb + that and Verb + 0 in three registers

writing than in academic writing, and the rate in conversation is only some-
what lower than in academic writing. This is because that-clauses overall are
much more frequent in conversation and newspaper writing than in academic
writing. As a result, both variants (with that and 0) occur with much higher
rates in newspaper writing than in academic writing.
It is surprisingly common for researchers to confuse these two perspectives
on variation, or to at least use statements that are misleading to the naïve r­eader.
The main problem here has to do with claims that a linguistic feature is “fre-
quent.” Consider, for example, the following statement from Szmrecsanyi and
Hinrichs (2008: 297): “The s-genitive is, on the whole, more frequent in spo-
ken data than in written . . .” It would be natural to interpret this statement to
mean that a speaker will produce more s-genitives than a writer will. Or put
another way, a listener will encounter more s-genitives in a conversation than
a reader would in a written text. However, that interpretation is not intended by
Szmrecsanyi and Hinrichs, and in fact, it is not accurate.
The pattern being described by Szmrecsanyi and Hinrichs is one of propor-
tional preference, not frequency of occurrence in texts. That is, s-genitives are
proportionally preferred over of-genitives in speech, while of-genitives are
proportionally preferred over s-genitives in writing: “FRED [i.e. a spoken cor-
pus] exhibits the highest percentage of the s-genitive (59.6%), Brown [i.e., a
written corpus] (36.2%) the lowest.” (Szmrecsanyi and Hinrichs 2008: 297).
But, from a text-linguistic perspective, s-genitives are actually much more fre-
quent in writing than in speech. Thus, Figure 4.6 in LGSWE (1999: 302) shows
16 D. Biber

that there are only c. 800 s-genitives per million words of conversation, in
contrast to c. 2,300 s-genitives per million words in academic writing, and
c. 9,000 s-genitives per million words in newspaper writing.
The pattern here is the same as the case study for that-retention presented
above. Thus, when a speaker uses a genitive construction, it is most likely to be
an s-genitive. So, considering only conversation, s-genitives are more frequent
than of-genitives. However, genitives overall are much more frequent in writ-
ing than in speech. As a result, when speech is compared to writing, both of-
genitives and s-genitives occur more frequently in writing. That is, even though
the s-genitive is proportionally preferred in conversation, it still is much less
frequent than the s-genitive in writing.
The point here is not to criticize the Szmrecsanyi and Hinrichs (2008) study,
which is an exemplary study employing a carefully considered research design
and sophisticated statistical methods to analyze this aspect of grammatical
variation. Rather, the point is to emphasize how easy it is to confound the
variationist and the text-linguistic perspectives when reporting frequency re-
sults. This is not an obscure methodological quibble. Rather, the two perspec-
tives are completely different in their practical implications. The variationist
perspective has the goal of comparing linguistic variants: whether one or the
other variant is preferred. These preferences can be compared across registers,
but that analysis cannot tell us the actual extent to which a variant is used in
texts. In contrast, the text-linguistic perspective (a Type B study) has the goal
of providing a linguistic description of texts, by describing the density of gram-
matical features in texts. These studies directly tell us the density of occurrence
of a feature (or variant) in texts from different registers.
There are a few general points worth emphasizing here. The first is the
i­mportance of distinguishing between variationist research designs and text-
linguistic designs: variationist designs investigate proportional preferences,
while text-linguistic designs investigate the rates of occurrence in texts. But a
more general point is a cautionary one: the text-linguistic perspective is often
the more natural interpretation, and thus it is easy for authors (and readers) to
misleadingly use the language of ‘frequency’. When a linguistic feature is
d­escribed as occurring ‘frequently’, we expect to encounter numerous occur-
rences of the feature in texts (the text-linguistic perspective). The proportional
perspective is more difficult to describe: that when a linguistic feature does
occur, it usually has certain characteristics – even if the feature itself occurs
infrequently. Thus, it is essential to be explicit about the nature of the patterns
in variationist studies: that they represent proportional preferences but not nec-
essarily frequent occurrence in texts.
In summary, a Type A research design – studies of linguistic variation where
each linguistic token is an observation – cannot describe the rates of occur-
rence in texts and registers. This design type can identify register influences on
 Register as a predictor of linguistic variation 17

linguistic variation – different proportional preferences in different registers –

but it cannot tell us how frequently we will encounter a linguistic feature in
texts from a register. In contrast, a Type B study – studies of text-linguistic
variation where each text is an observation – is specifically designed for such
research questions. However, as the following sections show, register differ-
ences are important in both design types.4

3. Lexical patterns and register

3.1 Lexis from the perspective of linguistic variation: Collocational studies

The importance of register for vocabulary is well-established from corpus

r­esearch (see, e.g., Kennedy 1998: 97–100). To take an easy example, the pro-
nouns I and you are among the most frequent words in the spoken London-
Lund Corpus, but considerably less frequent in the written Brown Corpus. ELT
dictionaries like the Longman Dictionary of Contemporary English (2009)
provide detailed information of this type, explicitly identifying the most fre-
quent words in speech versus the most frequent words in writing.
Despite this high level of awareness concerning the importance of register
differences for word frequencies, most studies of collocation and extended
lexical patterns have disregarded register. The unstated assumption has been
that the lexical collocations of a word will remain constant, regardless of
r­egister.
Many studies have investigated the preferred collocates of specific target
words. For example, Sinclair (1991) describes the uses of phrasal verbs with
set and collocates of the word back. Hunston (2002: Chs. 3 and 4) discusses the
phraseological patterns of several target words, such as recipe, initiative,
c­ondemn, suggestion, point, gaze, leak, and shoulder. Partington (1998), in a
book-length treatment of collocation, provides detailed descriptions of the
phraseological patterns for sheer, pure, complete, absolute, correct, absolutely,
completely, entirely. These studies provide detailed descriptions of the colloca-
tions and preferred uses of a specific target word, and further illustrate how
supposed synonyms are not in fact identical in meaning or use when consid-
ered from this perspective. However, these studies are typical in that they
i­nclude no mention of register or the possibility of different word uses in dif-
ferent registers.
However, corpus investigation of common words shows that this disregard
for register is not justified. For example, contrast the common content-word
collocates for the verbs have, make, and take in conversation versus informa-
tional writing (taken from Conrad and Biber 2009: 13–20):
18 D. Biber

have
Conversation:
dinner, lunch, a drink, fun, a good time, trouble, a hard time, a/no problem
with, kids, children, a baby, a/the chance, an/no idea, a question
Informational writing:
an/no/little effect/impact/influence on, the advantage of, a range of, a wide
variety of, little/no evidence of, no knowledge of, the potential for, little sympa­
thy for, implications for, an interest in, a role in

make
Conversation:
the bed, a phone call, a joke, (a) noise, a sound, an appointment, a deal, plans
to, a living, money, a difference, (a) decision(s), an effort, a mistake, (no)
sense, fun of, time for, sure
Informational writing:
assumptions, comparisons, judgments, choices, decisions, predictions, recom­
mendations, (no) sense, use of, reference to

take
Conversation:
a photo/picture, a bath/shower, a nap, a break, it easy, place, a minute, time,
classes, a test, a message, notes, a car, the bus/train, a ride, a right/left [turn],
a look at, care of, charge of, responsibility for, advantage of, forever, turns
Informational writing:
action, the initiative, the lead in, steps to, the position that, the view that,
a­ccount of, into account, part in, advantage of, precedence over, the form of,
the shape of

From a text-linguistic perspective, the verbs have, make, and take are consider-
ably more frequent in conversation than in informational writing. However, the
perspective of linguistic variation asks different questions: When these verbs
are used in conversation, what are the most common collocates? When these
verbs are used in informational writing, what are the most common collocates?
Are the preferred collocates in conversation the same as those in informational
writing?5
The answers to these questions indicate that register is a fundamentally im-
portant organizing factor for studies of collocation. All three verbs have strong
collocational associations in both conversation and informational writing.
However, those associations are almost entirely non-overlapping.
A collocational analysis of these verbs in a general multi-register corpus (cf.
BNC) might identify many of these combinations. But such an analysis would
miss the point that these are not general collocations. In informational writing,
it is rare to find uses like have lunch, have fun, make a phone call, make a deal,
 Register as a predictor of linguistic variation 19

take a break, take care of, etc. Similarly, in conversation, it is rare to find uses
like have implications for, make assumptions about, take precedence over.
These are all strong collocations, but they are tied to a particular domain of use,
and thus an essential component of their analysis should be documentation of
the register where they typically occur.

3.2 Text-linguistic lexical variation: Frequent lexical sequences

Text-linguistic studies ask a different type of research question from colloca-
tional studies: What are the most common lexical sequences found in texts (or
registers)? That is, in a collocational study of linguistic variation, particular
target words are selected for investigation, and then the analysis identifies the
preferred collocates for each word. There is little consideration given to
w­hether the target word is rare or frequent in texts. Rather, the focus is on the
preferred collocates when the target word occurs, regardless of how often the
word occurs in texts.
In contrast, in a text-linguistic study, no words are pre-selected for investiga-
tion. Rather, the goal is to analyze the texts themselves, to identify the most
frequent lexical sequences in those texts. Given this goal, it is natural that text-
linguistic studies of lexical sequences would consider register differences; and
as a result, these studies have consistently identified fundamentally different
kinds of lexical patterns in different registers.
For example, one series of such studies have been carried out under the
r­ubric of ‘lexical bundles’ (e.g., Biber et al. 1999: Ch. 13; Biber et al. 2004).
Lexical bundles are identified using a frequency-driven approach. In the initial
study of English lexical bundles (Biber et al. 1999), a relatively low frequency
cut-off was used: 10 times per million words. However, a sub-set of these
bundles, occurring more than 40 times per million words, was used for detailed
analyses of structural characteristics and discourse functions. Many of these
bundles are actually much more common, occurring more than 200 times per
million words.
Lexical bundles of any length can be analyzed. For example, the initial
d­escription of English bundles was based on 3-word, 4-word, and 5-word
s­equences, but only 4-word sequences were considered in the more detailed
analyses carried out by Biber et al. (2004). A further defining characteristic is
that a multi-word sequence must be used in multiple texts to be counted as a
lexical bundle (at least five different texts), to guard against idiosyncratic uses
by individual speakers or authors. Most bundles are distributed w­idely across the
texts in a corpus. For example, even the least common lexical bundles in con-
versation or classroom teaching are usually used in at least 20 different texts.
The initial analysis of lexical bundles in English (Biber et al. 1999: Ch. 13)
compared the patterns of use in conversation and academic prose, based on
20 D. Biber

analysis of c. 5-million-word sub-corpora from each register. Not surprisingly,

there are almost ten times as many 3-word bundles as 4-word bundles. It is
perhaps more surprising that there are many more lexical bundles in conversa-
tion than in academic writing, and this pattern is even stronger for the longer
bundles.
The most important finding from these studies for the purposes of the p­resent
paper is that the set of common lexical bundles in conversation is completely
different from the set of common lexical bundles in academic writing. For
e­xample, six 4-word bundles are extremely frequent in conversation, occurring
over 100 times per million words: I don’t know if, I don’t know what, do you
want to, I don’t want to, I was going to, are you going to. In contrast, only two
4-word bundles occur over 100 times per million words in academic writing:
on the other hand, in the case of. Other frequent 4-word bundles in academic
writing include: one of the most, the nature of the, as a result of, on the basis
of, in the form of.
In sum, register is fundamentally important for the description of frequent
lexical sequences, to a much greater extent than previously anticipated. In fact,
the sets of common lexical bundles are nearly disjunctive between conversa-
tion and academic prose. For example, only four of the 133 four-word lexical
bundles that occur over 20 times per million words in conversation were also
found to occur over 20 times per million words in academic writing – an over-
lap of only three percent. All other lexical bundles were found to be distinctive
for conversation versus academic prose.
The lexical bundle framework was extended in Biber (2009b) to allow for
variable patterns, identifying differences between registers at an even more
basic level. In that study, each 4-word lexical bundle was coded for its ‘pattern
type’, depending on the extent to which each slot was variable or fixed. For this
purpose, a simple cut-off of greater or lesser than 50% was used for each slot
in a 4-word sequence. That is, if more than 50% of the associated 3-word-
combination is accounted for by the particular word occurring in a slot, then
that slot is categorized as relatively fixed; otherwise, the slot is categorized as
relatively variable. For example, the sequence    the case of occurred 617
times in the academic writing corpus analyzed for this study. 506 of those
o­ccurrences – or 82% – were preceded by the word in. Thus, in is a ‘fixed’ slot
in this 4-word sequence. In contrast, the sequence in the    of occurred
6,325 times in this corpus, and the most common filler – case – occurred only
506 times (or 8% of the total). Thus, the third word is a variable rather than
fixed slot in this sequence.
The 2009 study shows that spoken discourse relies heavily on 3-word and
4-word fixed sequences of words, like I don’t know *, * you want to, I don’t
want to, and are you going to. In contrast, written academic discourse is com-
posed of lexical patterns that consist of invariable function words with an
 Register as a predictor of linguistic variation 21

i­ntervening variable slot filled by many different possible content words (e.g.,
the * of the, in the * of, to the * of ).
These different patterns also have strong grammatical correlates. For ex-
ample, the continuous fixed sequences in conversation consist of both function
words and content words. In contrast, the fixed slots in the academic writing
patterns are usually function words, while the variable slots are usually content
words. However, there are other differences having to do with the structural
correlates of these lexical patterns: verb phrase and clause fragments in the
case of conversational lexical sequences, but noun phrase and prepositional
phrase fragments in the case of the academic writing patterns. I return to these
associations in Section 5 below.

4. Grammatical variation and register

4.1 Grammar from the perspective of linguistic variation

Section 2 above has already introduced studies of grammatical variation that
treat each token of the grammatical feature as an observation. For example,
Figure 1 (above) showed how registers can differ in their proportional prefer-
ence for a grammatical variant, while a comparison of Figure 1 and Figure 2
showed how the rates of occurrence for variants in texts provide a dramatically
different perspective on register differences.
In recent years, grammatical variation has been investigated using sophis­
ticated multivariate statistical techniques, such as logistic regression (e.g.,
Szmrecsanyi 2005; Hinrichs and Szmrecsanyi 2007). In general, these studies
have not considered register as a predictor. However, when the influence of
register/genre has been investigated, as in Riordan (2007) and Szmrecsanyi
and Hinrichs (2008), it has been found to be an important predictor; for
e­xample, “. . . we have seen that more often than not, individual factors – for
instance, possessor animacy or thematicity of the possessor – have fairly dif-
ferent impacts in spoken and written data.” (Szmrecsanyi and Hinrichs 2008:
307)
The advantage of multivariate techniques like logistic regression is that
they are able to isolate the predictive strength of a factor when numerous
o­ther factors are also considered. However, the importance of register for
s­tudies of grammatical variation can also be illustrated with simple descriptive
statistics.
For example, we can further explore the retention versus omission of that in
verb + that/0 constructions (introduced in Section 2 above). Several gram-
matical factors have been hypothesized to influence this choice, such as the
frequency of the matrix verb, and the presence of an intervening NP between
22 D. Biber

Figure 3. P
 roportional preference for Verb + that versus Verb + 0: The influence of co-­referential
subjects, in conversation vs. newspaper prose

the matrix verb and the that-clause. One factor that has been hypothesized to
favor that-omission is co-referential subjects in the matrix clause and the that-
clause, as in:

I thought [0] I would go

In this example, it is not grammatically possible to interpret the second I as the

direct object as thought. Thus, the use of a co-referential subject in the that-
clause is a relatively explicit signal that a new clause is beginning, even with-
out the complementizer that.
In contrast, when the two subjects are not co-referential, there is greater
need for the complementizer that to signal the start of a that-clause, as in:

The Secretary argued that armed intervention was not the answer.

This factor is interesting for our purposes here because its influence is dra-
matically different in different registers. Thus, Figure 3 above ( based on Biber
et al. 1999: 681) shows that the choice of grammatical subject has little influ-
ence in conversation: the complementizer that is omitted over 80% of the time,
regardless of whether the grammatical subjects are co-referential or not. How-
ever, we see a dramatically different pattern in newspaper writing: when the
construction has co-referential subjects, the complementizer that is omitted
over 70% of the time, while that is omitted only c. 15% of the time when
 Register as a predictor of linguistic variation 23

Figure 4. P
 roportional preference for Verb + that versus Verb + 0: The influence of an interven­
ing NP, in conversation vs. newspaper prose

the subjects are not co-referential. Thus, the strength of this factor is mediated
by register: essentially no influence in conversation ( because that-omission is
already the norm), versus a very strong influence in newspaper writing ( b­ecause
that-retention is the general register norm).
Other grammatical factors favor the retention of that, such as the presence of
a noun phrase intervening between the matrix verb and that-clause (e.g., They
told him that it’s dangerous). These grammatical factors show the opposite
interaction with register, having a very strong influence in conversation but
little influence in informational writing. Thus, Figure 4 above ( based on Biber
et al. 1999: 682) shows that the presence of an intervening noun phrase in con-
versation results in the complementizer that almost always being retained,
while it is retained less than 15% of the time when there is no intervening noun
phrase. In contrast, the presence of an intervening noun phrase has little influ-
ence in newspaper writing, because the norm is already to retain the comple-
mentizer that. Thus, even without an intervening noun phrase, the complemen-
tizer that is retained c. 75% of the time.
In sum, similar to the collocational patterns for individual words, register is
a strong predicting factor for studies of grammatical variation. This factor
i­nteracts with other contextual influences: a contextual factor with a strong
influence in one register might have a minimal influence in another register. As
the following section shows, grammatical differences across registers are even
more notable when approached from a text-linguistic perspective.
24 D. Biber

4.2 Grammar from the perspective of text-linguistic variation

Multi-dimensional studies (e.g., Biber 1988, 1995, 2006) have used a text-­
linguistic approach to describe the grammatical characteristics of different
texts and registers. One of the most important findings from these studies is the
extent to which spoken registers rely on clausal grammar, while informational/
academic written registers rely on phrasal grammar. For example, Dimension
1 in both the 1988 study of general spoken and written registers, as well as the
2006 study of university spoken and written registers, shows this same opposi-
tion: between verbs and finite dependent clauses (e.g., conditional clauses and
WH-clauses) co-occurring frequently with pronouns and adverbials in spoken
registers, versus nouns, attributive adjectives, and prepositional phrases co-
occurring frequently in written academic registers.
Several other studies have documented this fundamental difference between
spoken and informational-written discourse in more detail (Biber and Clark
2002; Biber 2009a). For example, Biber and Gray (2010) use a text-linguistic
design to analyze the rates of occurrence for clausal versus phrasal features
associated with grammatical complexity (i.e., dependent clause types versus
phrasal modifiers in noun phrases). That study contrasts the patterns of use
in conversation versus academic research writing, based on analysis of a
4-­million-word corpus of American English (AmE) conversation (723 texts)
and a 3-million-word corpus of academic research articles (429 texts). Two
structural factors turn out to be important when accounting for the differences
between these two registers:

1. structural type:
a. clauses, especially finite dependent clauses, are preferred in speech
b. (non-verbal) phrases are preferred in academic writing
2. syntactic function:
a. clausal constituents (adverbials and complement clauses) are p­referred
in speech
b. noun phrase constituents (noun modifiers and noun complements) are
preferred in academic writing

Tables 2 and 3 provide selected details of these general trends, comparing the
mean scores for many of the most important complexity features.
In sum, from a text-linguistic perspective, the grammar of conversation is
dramatically different from the grammar of informational writing (cf. Biber,
Gray, and Poonpon 2011). This does not represent an absolute difference be-
tween speech and writing, because written registers like email can employ the
grammatical discourse styles typical of speech (e.g., Biber and Conrad 2009:
Ch. 7). However, informational writing is fundamentally different from con-
versation (and spoken registers generally) in its heavy reliance on phrasal
 Register as a predictor of linguistic variation 25

Table 2. Grammatical complexity features typical of conversation

Linguistic feature Conversation Academic WR F significance r2

mean score mean score value

Finite adverbial clauses

total adverbial clauses 7.1 3.6 603.2 <.0001 .35
Verb complement clauses
verb + that-clause 10.8 2.6 2196.7 <.0001 .66
verb + WH-clause 2.7 0.2 1413.9 <.0001 .55
verb + ing-clause 1.3 0.2 842.5 <.0001 .42
verb + to-clause 4.7 3.4 166.6 <.0001 .13

Table 3. Grammatical complexity features typical of academic writing

Linguistic feature Conversation Academic WR F significance r2

mean score mean score value

Noun modifiers – clauses

WH relative clauses 0.9 3.7 858.1 <.0001 .43
nonfinite relative clauses 0.7 4.2 2257.3 <.0001 .66
Noun complement clauses
noun + that-clause 0.1 0.6 474.1 <.0001 .29
noun + to-clause 0.9 2.8 856.8 <.0001 .43
Noun modifiers – phrasal
Attributive adjectives 16.5 57.1 5787.8 <.0001 .84
Nouns as nominal pre-mods 19.0 57.4 1259.2 <.0001 .52
Total prepositional phrases 6.3 51.9 1380.1 <.0001 .94
as nominal post-mods

r­ather than clausal modification. Thus, any description of variation in English

that disregards this opposition will miss out on what is arguably the most im-
portant predictor of grammatical differences.

5. Lexis, grammar, and register

5.1 Variationist and text-linguistic perspectives on the lexis-grammar

interface
The Longman Grammar of Spoken and Written English (LGSWE by Biber
et al. 1999) provides extensive descriptive information on the interaction of
lexis and grammar. Lists of the most common words occurring with many
grammatical constructions are grouped according to their specific frequency
band, including: phrasal verbs ( pp. 409– 412); prepositional verbs ( pp. 416 –
421); perfect aspect verbs ( pp. 463– 464); passive voice verbs ( pp. 477– 480);
26 D. Biber

attributive adjectives ( pp. 511–515); predicative adjectives ( pp. 515–518);

nouns controlling complement clauses ( pp. 648– 653); verbs controlling a
complement clauses ( pp. 667– 670, pp. 688– 693, pp. 710 –714, pp. 741–748);
adjectives controlling complement clauses ( pp. 672– 673, pp. 718–721,
p. 749); stance adverbials ( pp. 867–871); and linking adverbials ( pp. 886 –889).
Other kinds of quantitative information are also provided, such as the extent to
which individual verbs and particular particles/prepositions co-occur in p­hrasal
verbs ( pp. 412– 413) or in prepositional verbs ( pp. 422– 423). Similarly, gram-
matical preferences are identified quantitatively for particular words, such as
lists of verbs that usually occur in present tense or usually in past tense ( p. 459);
verbs that usually occur in progressive aspect or almost never occur in progres-
sive aspect ( pp. 471– 472); and verbs that usually occur in passive voice or
almost never occur in passive voice ( pp. 477– 482).
Both variationist and text-linguistic perspectives on the lexis-grammar
i­nterface are explored in the LGSWE:
1. Linguistic variation: how strongly is a given word associated with a gram-
matical construction?
2. Text-linguistic variation: how often is a lexical-grammatical combination
encountered in natural texts and different registers?
As described above, the first perspective is described in terms of probabilistic
preference, considering the proportional extent to which a word is used with a
particular grammatical construction. For example, although the prepositional
verb base on is not especially frequent in academic writing, it has an e­xtremely
strong association with passive voice, with more than 90% of all occurrences
occurring as passives (LGSWE p. 479). In contrast, the second perspective can
be described by computing rates of occurrence. For example, the verbs be
g­iven, be found, and be seen are the most frequently occurring passive verbs
that a reader will encounter in academic writing (LGSWE p. 478).
Lexico-grammatical associations have been investigated in more detail
through ‘collostructional analysis’, an advanced statistical approach that mea-
sures the strength of association between words and grammatical constructions
(Stefanowitsch and Gries 2003). The methodological goal of collostructional
analysis is to develop a statistical approach that does not assume a normal
d­istribution or homogeneity of variance, and is robust for both rare and high-
frequency collocations. Collostructional analysis achieves this goal by apply-
ing the Fisher exact test, which integrates both proportional preference and
frequency rates of occurrence to produce a single measure that identifies the
words that are most strongly associated with a target grammatical construction
(‘collexemes’). This type of analysis has been applied to numerous case
s­tudies, including an identification of the verbs that are most strongly attracted
to past tense; the verbs that are most strongly attracted to progressive aspect;
 Register as a predictor of linguistic variation 27

the verbs that are most strongly attracted to active versus passive voice and the
verbs that are most strongly attracted to the ditransitive versus to-dative (cf.
Stefanowitsch and Gries 2003; Gries and Stefanowitsch 2004).
Collostructional analysis – similar to previous statistical measures of collo-
cational association – assumes the desirability of a single statistical measure of
linguistic importance (lexico-grammatical association). However, there are
theoretical and practical reasons why it might be preferable to distinguish
b­etween the variationist and text-linguistic perspectives: the words that are
proportionally associated with a grammatical construction (even if they rarely
occur), versus the words that most frequently occur with a g­rammatical construc­
tion (even if those words also frequently have other grammatical associations).
The two types of analysis produce results that are to a large extent comple-
mentary. For example, from the text-linguistic perspective, passive voice con-
structions in academic writing are most frequently found with verbs like be
made, be given, be used, be seen, even though those verbs are also commonly
used in the active voice (compare LGSWE pp. 367–369 and p. 375 with
p. 478). In contrast, the variationist perspective identifies a completely differ-
ent set of verbs have the strongest proportional association with passive voice
in academic writing, occurring as passives over 90% of the time, even though
none of them is especially frequent: be aligned (with), based on, coupled with,
deemed, effected, situated, subjected (to), etc. (LGSWE p. 479). Thus, a reader
will most often encounter passive verbs like be made, be given, be used, and
she/ he will develop the association that when passive voice is employed, it
often incorporates those verbs. In contrast, a reader will less often encounter
passive verbs like be aligned (with) and based on, but because those verbs are
almost never encountered in the active voice, an association in the opposite
direction is developed: when those verbs are used, they are almost always in
the passive voice. The two perspectives are distinct but both are important.

5.2 Register influences on lexis and grammar

Many previous studies of lexico-grammatical associations have disregarded
register differences. For example, probably the most developed exploration of
the lexis-grammar interface is the series of reference books and academic
s­tudies carried out under the rubric of ‘pattern grammar’. Two major reference
books have emerged from this framework (cf. Francis et al. 1996, 1998; Hun-
ston and Francis 2000). These studies investigate the phraseology of i­ndividual
words, showing that there are systematic regularities in the associations be-
tween grammatical frames, sets of words, and particular meanings on a much
larger scale than it could have been possible to anticipate before the introduc-
tion of large-scale corpus analysis. However, there is no indication of register
differences in these studies.
28 D. Biber

The omission of register analyses in the pattern grammar books can be justi-
fied by the magnitude of that project, compiling lists of ‘patterns’ for thousands
of verbs, nouns, and adjectives. Given finite resources, the authors chose to
increase their coverage of words, rather than investigating the possibility of
different patterns occurring in different registers. However, this omission is
also typical of most more restricted lexico-grammatical studies.
One exception to this generalization is Stefanowitsch and Gries (2008), who
investigate the influence of spoken versus written mode on the associations
between lexical items and ‘constructions’. As described above, this study uses
‘collostructional analysis’, a statistical procedure that combines proportional
preference and frequency rates of occurrence to produce a single measure of
association between words and grammatical constructions.
For example, one case study in that paper investigates the verbs that are
most strongly associated with passive voice in speech and writing. Three dif-
ferent statistical comparisons are used:
1) for each verb, contrasting the association with active voice versus passive
voice, carried out separately for the spoken and written modes;
2) for each verb, contrasting the association with the spoken mode versus
written mode, carried out separately for active voice and then p­assive voice.
3) for each verb, checking for cross-over effects: when a verb is associated
with passive voice in the spoken mode but active voice in the written
mode, or vice versa.
The first type of analysis identifies the set of verbs associated with passive
(versus active) voice in speech, and then separately identifies the verbs associ-
ated with passive (versus active) voice in writing. (Stefanowitsch and Gries
2008: 140) Several verbs have associations with passive voice in both modes,
such as: BE + concerned, based, published, associated, confined, designed.
However, other verbs are associated with passive (versus active) voice only in
one of the two modes.
In several cases, the results of this type of analysis are surprising, identifying
verbs that would not normally be associated with the target register. In fact,
many of these findings are difficult to interpret in terms of spoken versus writ-
ten discourse, calling into question the value of this statistical measure for this
application (cf. below). For example, the verbs most strongly associated with
passive voice in speech ( but not associated with passive voice in writing)
i­nclude ‘literate’ verbs like: BE + involved, used, engaged, enclosed, aimed,
distributed, compared, entitled. In contrast, the verbs most strongly associated
with passive voice in writing ( but not associated with passive voice in speech)
include ‘colloquial’ verbs like BE + thought, done, made.
The second type of analysis identifies the set of verbs that are more strongly
associated with passive voice in speech than in writing, and vice versa.
 Register as a predictor of linguistic variation 29

(S­tefanowitsch and Gries 2008: 142). Here again, we find several seemingly
‘literate’ verbs in the ‘spoken-passive’ list, such as BE + concerned, involved,
cross-examined, readmitted, adduced, extruded, while several ‘colloquial’
verbs are in the ‘written-passive’ list, such as BE + had, known, got/gotten,
thought, wanted.
The third type of analysis identified only two verbs with cross-over effects:
find and work, which are both attracted to spoken/active versus written/p­assive.
However, these verbs are claimed to involve different meaning senses in the
two channels, and thus to not represent a genuine register difference. Thus, the
general conclusion that Stefanowitsch and Gries draw is that register differ-
ences are not important for this type of investigation: “the results of channel-
sensitive collostructional analysis are essentially identical to those yielded by
a ‘channel-ignorant’ analysis as far as constructional meaning in the narrow
sense is concerned: we found no interaction between channel and semantics at
all.” (2008: 143)
Before addressing the claimed lack of a register effect, it is useful to discuss
three other important issues that arise in the Stefanowitsch and Gries study.
The first is to emphasize yet again the need to distinguish between the varia-
tionist and text-linguistic perspectives, and specifically to avoid claims about
‘frequency’ in variationist studies that investigate proportional preference. As
noted in Section 2 above, this confusion arises even in some of the most care-
fully designed corpus-based studies. Stefanowitsch and Gries (2008) some-
times seem to fall into this same trap; for example, “the passive construction
occurs relatively frequently with formal vocabulary in both channels.” ( p. 143)
The normal interpretation of this statement is that a speaker would frequently
produce passive constructions with formal vocabulary in speech: a text-­
linguistic perspective. However, Stefanowitsch and Gries (2008) did not actu-
ally consider this perspective and provide no findings to support such a conclu-
sion. In fact, passives are not frequent at all in conversation. For example, the
LGSWE (Figure 6.7; p. 476) shows that passive verbs occur only c. 2,000
times per million words in conversation, contrasted with c. 18,000 times per
million words in academic writing. Further, the verbs that most frequently
o­ccur with passive voice in conversation are not formal vocabulary; rather,
they are mostly everyday, colloquial verbs like BE + made, done, called, put,
told, born, paid (cf. LGSWE pp. 478– 479).
In contrast, the Stefanowitsch and Gries (2008) study apparently reflects the
fact that when a ‘formal’ verb (like engaged, enclosed, adduced, or extruded )
is used in conversation, it is usually used in the passive voice. But this does not
at all mean that such combinations occur frequently in conversation.
The second general issue here is that surprising findings require i­nterpretation
and explanation; it is not sufficient to simply report surprising statistical results
with no discussion. This is especially the case when we are relying on a c­omplex
30 D. Biber

statistical measure, with results that are not necessarily well-understood. Such
an analysis raises two possibilities: 1) that there are completely unanticipated
linguistic patterns identified by the analysis, requiring a radical change in our
understanding of spoken and written discourse; or 2) that the statistical
a­pproach is not measuring what we think it is, and thus the approach itself
r­equires further analysis and interpretation.
For example, Table 6 in the Stefanowitsch and Gries (2008: 142) study lists
the verbs that are most strongly associated with passive voice in speech con-
trasted with writing. The expectation of the reader is that these verbs are some-
how especially typical of speech. The verbs identified by this method, though,
are extremely surprising, including be readmitted, be adduced, be extruded –
and no discussion is provided to explain how these passive verbs are typical of
speech in contrast to writing. Similarly, no discussion is offered to explain
what it means for verbs like BE had, BE got/gotten, BE wanted to be among the
verbs most strongly associated with passive voice in writing (in contrast to
speech).
Finally, these findings point to the need for triangulation of methodological
approaches, and the risks of relying exclusively on any single measure. For
example, multivariate statistics, simple descriptive statistics, and consideration
of linguistic examples in textual contexts should all be considered and recon-
ciled. Similarly, considering quantitative findings from both variationist and
text-linguistic perspectives can provide a more complete description than any
single perspective. Most importantly, analyses based on a single m­ethodological
approach should be presented with explicit discussion of the limitations of that
approach.
For example, descriptive statistics on the use of passive verbs can be used to
illustrate the complementary kinds of information found from the variationist
versus text-linguistic analytical approaches. At the same time, these descrip-
tive statistics identify some completely different patterns from those identified
in the collostructional analysis. For example, an analysis of text-linguistic rate-
of-occurrence identifies five verbs that are especially frequent with the passive
voice in academic writing: BE + found, given, made, seen, used. (LGSWE
p. 478). However, three of those verbs are not identified at all in the first col-
lostructional analysis (verbs associated with passive voice versus active voice
in academic writing): BE + given, found, seen. Similarly, BE + given, found
are not included on the list of passive verbs associated with writing versus
speech (even though they occur c. 100 times more frequently as passive verbs
in academic writing than in speech; cf., LGSWE p. 478).
Other verbs that did not make it to the lists in the Stefanowitsch and Gries
(2008) study have strong proportional use with passive voice in academic
w­riting. For example, the following verbs were not identified in any of the
c­ollostructional analyses, but descriptive statistics show that they all occur as
 Register as a predictor of linguistic variation 31

Table 4. S
 elected verbs showing different patterns for frequency rates of occurrence ( per 1 mil­
lion words) and proportional use of passive voice, comparing conversation (AmE) and
academic writing ( based on analysis of c. 5-million word corpora for each register,
taken from the LSWE Corpus).

Conversation Academic Writing

Rate of Proportional Use Rate of Proportional Use

Occurrence of Passive Voice Occurrence of Passive Voice

Total for all main verbs (finite and non-finite)

Passive 14,000 2% 110,000 25%
non-passive 645,000 429,000
Passive associated with writing but not speech
high frequency verbs; strong proportional preference
Find passive 26 0.8%    2,900 46%
active 3,030    3,400
high frequency; moderate proportional preference
Show passive 10 0.5%    2,310 28%
active 2,090    6,000
Make passive 180 1.8%    2,500 30%
active 10,200    5,800
moderate frequency; strong proportional preference
Call passive 337 5.5%    1,040 40%
active 5,800    1,553
Passive associated with speech but not writing
low frequency verbs; very strong proportional preference
Allow passive 170 57%     190 7.5%
active 130    2,350
Stick all passive 33 63%       2 21%
(get passive 10)
active 19       8

passives in academic writing over 90% of the time: BE + aligned, coupled

with, deemed, effected, flattened, inclined, obliged, positioned, s­ituated, stained,
subjected to.
When descriptive statistics are considered, and both proportional/v­ariationist
as well as text-linguistic analyses are employed, it is easy to isolate major dif-
ferences between spoken and written discourse. For example, the quantitative
distributions shown in Table 4 for verbs in the passive versus active voice il-
lustrate many of the possible patterns of use, such as very frequent but weak
proportional preference; infrequent but strong proportional preference; etc. In
addition, these case studies illustrate strong differences between conversation
and academic writing, from both analytical perspectives.
The descriptive findings presented in Table 4 show that register is a funda-
mentally important factor for the description of lexico-grammatical a­ssociations,
32 D. Biber

leading to different conclusions from the collostructional analysis of passive

constructions. First of all, the descriptive statistics show that register has a
major influence, while the collostructional analysis concluded that register had
essentially no influence. Beyond that, the lists of specific verbs associated with
the passive in speech or writing are strikingly different in the two a­nalyses.
Finally, the descriptive statistics show that the variationist and text-linguistic
perspectives are clearly independent, indicating that they should be analyzed
separately.
It is not completely clear how to reconcile the quantitative results from the
Stefanowitsch and Gries (2008) collostructional analysis with these d­escriptive
statistics. However, this is an essential task: to be useful, statistical analyses
must be interpretable in linguistic terms. In the case study presented above, it
is difficult to interpret the collostructional results from the perspective of a
normal conversational interlocutor or a typical reader of written texts. For
e­xample, many of the passive verbs with high distinctiveness scores for speech
are clearly not typical of the verbs that a speaker would normally encounter in
conversation (e.g., be readmitted, be adduced, be extruded ). And conversely,
many of the most frequent passive verbs that a listener/reader would encounter
are disregarded by the collostructional analysis (e.g., the verbs be given, be
found, be seen in academic writing). Statistical analyses of corpora often un-
cover patterns of language use that we had not noticed before. But we should
subsequently be able to confirm the results of the statistical analysis in natural
texts: there should not be a disconnect between the language that we observe
in texts and the results of our statistical analysis.
Beyond that, there is reason to question the desirability of a single measure
of lexical association. In particular, the most frequent patterns found in texts
are often not the same as the proportionally preferred associations. It is conve-
nient to have a single score that measures association strength. However, it is
not clear that that approach generally provides the most useful description of
actual language use.
In sum, the main point of this section has been to argue for the importance
of register differences in describing lexico-grammatical patterns of use. The
study by Stefanowitsch and Gries (2008) has been discussed in some detail
because it is one of the few previous studies to investigate register influences
on lexico-grammatical associations. The omnibus measure of collostructional
strength employed in that study suggests that register is not a strong factor
i­nfluencing lexico-grammatical variation, or at least that register “does not
i­nteract substantially with constructional semantics.” ( p. 129) In contrast, I
have argued here that a separate consideration of distributional patterns as­
sociated with the variationist versus text-linguistic perspectives both show
consistent and important differences in lexico-grammatical patterns across
r­egisters.
 Register as a predictor of linguistic variation 33

6. Conclusion

The present paper argues for the importance of register differences at all lin-
guistic levels. However, as background, the paper first distinguishes between
two major approaches to the study of linguistic variation and use: variationist
versus text-linguistic. The variationist approach has been widely used since
the 1960’s for the quantitative study of the ‘sociolinguistic variable’ (see, e.g.,
Labov 1972; Lavandera 1978). More recently, functional linguists from sev-
eral different sub-disciplines have employed similar approaches to study gram-
matical variation, exploring the contextual factors that influence the choice
among related grammatical variants. Variationist studies differ in their linguis-
tic focus ( phonological versus grammatical features) and in the statistical tech-
niques that they employ (e.g., Varbul, logistic regression). However, all varia-
tionist studies share certain characteristics: 1) the research goal is to describe a
linguistic feature, rather than the characteristics of texts; 2) each occurrence of
the target linguistic feature constitutes an observation in the research design;
and 3) the quantitative findings represent proportional preference for one lin-
guistic variant in comparison to other variants.
The text-linguistic approach to linguistic variation differs in all three re-
spects: 1) the research goal is to describe the characteristics of texts, rather than
the characteristics of a linguistic feature; 2) each text constitutes an observa-
tion; and 3) the quantitative findings represent the rates of occurrence of lin-
guistic features in texts rather than the proportional preference for a linguistic
variant in comparison to other variants. Thus, the text-linguistic approach de-
scribes language use from the perspective of a conversational participant or a
normal reader of a text: what features will they encounter most commonly in
spoken interactions or written texts?
The description of passive voice verbs (cf. Section 5.2) illustrates the prac­
tical consequences of this distinction. From a variationist perspective, a­nalyzing
proportional preference, verbs like BE + concerned, involved, cross-examined,
readmitted, adduced, extruded are especially associated with passive voice in
speech. That is, when one of these verbs is used in spoken discourse (or at least
in the particular corpus of speech analyzed for the S&G study), it is likely to
occur as a passive rather than active voice verb. (In this case, these findings are
in contrast to written discourse: these particular verbs are proportionally more
likely to occur as passives in spoken discourse than in written discourse.)
The text-linguistic approach provides a dramatically different perspective,
because most of these verbs are simply not common at all in spoken texts. Thus,
a conversational participant will rarely encounter these verbs in text, whether
they are in the active or passive voice. However, there are other verbs that do
frequently occur with passive voice in conversation, like BE + made, done,
called, put, told, paid. Proportionally, these verbs occur most of the time as
34 D. Biber

active voice verbs, so they are not associated with the passive from a v­ariationist
perspective. But a conversational participant will encounter these passive forms
much more frequently than proportionally-preferred verbs like BE adduced or
BE extruded.
Similar contrasts between the types of information provided by the two
p­erspectives can be given for most linguistic features. The point here is not to
argue that one or the other perspective is correct. However, studies often fail
to distinguish between the two, slipping into descriptions that suggest a text-
linguistic perspective when the data are strictly proportional or variationist.
The secondary goal of the present paper has thus been to emphasize the differ-
ence between these two perspectives, and the need to explicitly characterize
findings as relating to one or the other.
The primary goal, though, has been to argue for the importance of register
differences – in both variationist and text-linguistic studies. That is, the pat-
terns of linguistic variation and use are dramatically different across registers.
The descriptions here have focused mostly on the spoken/written contrast
(e­specially face-to-face conversation versus academic writing), but systematic
differences exist across the full range of registers (see, e.g., Biber and Conrad
2009). These register differences exist across all linguistic levels, including
lexical patterns, grammatical patterns, and lexico-grammatical associations.
Traditionally, most general-purpose corpora were designed to include mul-
tiple registers, and thus many descriptive studies have adopted a text-linguistic
approach and include some information on register differences.6 In recent
years, some variationist studies have also begun to include analysis of register
differences interacting with other factors of the linguistic context (see, e.g.,
Riordan 2007; Szmrecsanyi and Hinrichs 2008). However, it is still the norm
in most studies of collocation and lexico-grammatical associations to disregard
the possible influence of register differences. The main point of the present
paper is that we should instead treat this possible influence as a likelihood: that
the patterns of linguistic variation and use are usually strikingly different in
spoken versus written registers. Thus, the practice advocated here is to begin a
research study with the hypothesis that such register differences exist, and to
include analysis of those differences unless they are empirically shown to be
unimportant.

Bionotes

Douglas Biber is Regents’ Professor of English (Applied Linguistics) at

N­orthern Arizona University. His research efforts have focused on corpus
l­inguistics, English grammar, and register variation (in English and cross-­
linguistic; synchronic and diachronic). He has written 170 research articles
 Register as a predictor of linguistic variation 35

and 13 books and monographs, including academic books published with

Cambridge University Press (1988, 1995, 1998, 2009), John Benjamins (2006,
2007), the co-authored Longman Grammar of Spoken and Written English
(1999), and three grammar textbooks published by Longman. Email: Douglas.
Biber@nau.edu

Notes

1. In a few cases, corpus linguists have actively argued against the need for quantitative analysis.
One of the best known linguists to take this position is Sinclair, stating that: “some numbers
are more important than others. Certainly the distinction between 0 and 1 is f­undamental, b­eing
the occurrence or non-occurrence of a phenomenon. The distinction between 1 and more than
one is also of great importance . . .” [because even two unconnected tokens constitute] the
r­ecurrence of a linguistic event . . . , [which] permits the reasonable assumption that the event
can be systematically related to a unit of meaning. In the study of meaning it is not usually
necessary to go much beyond the recognition of recurrence [i.e. two independent tokens]. . . .
(Sinclair 2001: 343–344).
2. It is also possible to include quantitative variables in a Type A study. For example, each occur-
rence of a relative clause could be coded for the number of words separating the relative pro-
noun from the gap position. However, studies of linguistic variation generally do not include
such variables.
3. If the three sub-corpora had been exactly the same size, and if the analysis had been based on
a complete sample of all that-clauses, then conclusions of this type would be appropriate.
However, most Type A studies do not meet these two requirements.
4. Type C studies are also designed for text-linguistic research questions, but they do not permit
the use of inferential statistics. The results reported in the LGSWE are actually based on Type
C designs rather than Type B designs.
5. The operational definition of ‘collocate’ can also be approached from both variationist versus
text-linguistic perspectives. The variationist perspective uses statistics like mutual informa-
tion and log-likelihood, which are based on the proportion of both words that co-occur. The
text-linguistic perspective uses simple rate of occurrence, measuring how often the combina-
tion of words is found in texts.
6. Most of these studies use a ‘Type C’ design, treating each sub-corpus as an observation rather
than analyzing each text as an observation. While this still results in a text-linguistic perspec-
tive, it does not permit analysis of dispersion or the extent to which register differences hold
for individual texts.

References

Biber, Douglas. 1988. Variation across speech and writing. Cambridge: Cambridge University
Press.
Biber, Douglas. 1995. Dimensions of register variation: A cross-linguistic comparison. Cam-
bridge: Cambridge University Press.
Biber, Douglas. 2006. University language: A corpus-based study of spoken and written registers.
Amsterdam: John Benjamins.
36 D. Biber

Biber, Douglas. 2009a. Are there linguistic consequences of literacy? Comparing the potentials of
language use in speech and writing. In David R. Olson & Nancy Torrance (eds.), Cambridge
Handbook of Literacy, 75–91. Cambridge: Cambridge University Press.
Biber, Douglas. 2009b. A corpus-driven approach to formulaic language: Multi-word patterns in
speech and writing. International Journal of Corpus Linguistics 14. 381– 417.
Biber, Douglas & Victoria Clark. 2002. Historical shifts in modification patterns with complex
noun phrase structures: How long can you go without a verb? In Teresa Fanego, Maria Jose
Lopez-Couso & Javier Perez-Guerra (eds.), English historical syntax and morphology, 43– 66.
Amsterdam: John Benjamins.
Biber, Douglas & Susan Conrad. 2009. Register, genre, and style. Cambridge University Press.
Biber, Douglas, Susan Conrad & Viviana Cortes. 2004. If you look at . . . : Lexical bundles in
university teaching and textbooks. Applied Linguistics 25. 371– 405.
Biber, Douglas, Susan Conrad & Randi Reppen. 1998. Corpus linguistics: Investigating language
structure and use. Cambridge: Cambridge University Press.
Biber, Douglas & Bethany Gray. 2010. Challenging stereotypes about academic writing: Com-
plexity, elaboration, explicitness. Journal of English for Academic Purposes 9. 2–20.
Biber, Douglas, Stig Johansson, Geoffrey Leech, Susan Conrad & Edward Finegan. 1999. The
Longman grammar of spoken and written English. London: Longman.
Biber, Douglas & James K. Jones. 2009. Quantitative methods in corpus linguistics. In Anke
Lüdeling and Merja Kytö (eds.), Corpus linguistics: An international handbook, 1286 –1304.
Berlin: Walter de Gruyter.
Biber, Douglas, Bethany Gray, and Kornwipa Poonpon. 2011. Should we use characteristics of
conversation to measure grammatical complexity in L2 writing development? TESOL Quar­
terly 45. 5–35.
Carter, Ronald & Michael McCarthy. 2006. Cambridge grammar of English. Cambridge: CUP.
Collins COBUILD English Grammar. 1990. London: Collins.
Conrad, Susan & Douglas Biber. 2009. Real grammar: A corpus-based approach to English. Pear-
son Longman.
Francis, Gill, Susan Hunston & Elizabeth Manning (eds.). 1996. Collins COBUILD grammar pat­
terns 1: Verbs. London: HarperCollins.
Francis, Gill, Susan Hunston & Elizabeth Manning (eds.). 1998. Collins COBUILD grammar pat­
terns 2: Nouns and adjectives. London: HarperCollins.
Gries, Stefan Th. 2006. Exploring variability within and between corpora: some methodological
considerations. Corpora 1. 109–151.
Gries, Stefan Th. & Anatol Stefanowitsch. 2004. Extending collostructional analysis: a corpus-
based perspective on ‘alternations’. International Journal of Corpus Linguistics 9. 97–129.
Hinrichs, Lars & Benedikt Szmrecsanyi. 2007. Recent changes in the function and frequency of
standard English genitive constructions: a multivariate analysis of tagged corpora. English Lan­
guage and Linguistics 11. 335–378.
Hunston, Susan. 2002. Corpora in applied linguistics. Cambridge: Cambridge University Press.
Hunston, Susan & Gill Francis. 2000. Pattern grammar: A corpus-driven approach to the lexical
grammar of English. Amsterdam: John Benjamins.
Kennedy, Greame. 1991. Between and through: The company they keep and the functions they
serve. In Karin Aijmer & Bengt Altenberg (eds.), English Corpus Linguistics, 95–110. London:
Longman.
Kennedy, Graeme. 1998. An introduction to corpus linguistics. London: Longman.
Labov, William. 1972. Sociolinguistic patterns. Philadelphia: University of Pennsylvania Press.
Lavandera, Beatriz R. 1978. Where Does the Sociolinguistic Variable Stop? Language in Society
7. 171–82.
Longman Dictionary of Contemporary English. 2009. London: Longman.
 Register as a predictor of linguistic variation 37

Partington, Alan. 1998. Patterns and meanings. Amsterdam: John Benjamins.

Quirk, Randolf, Sidney Greenbaum, Geoffrey Leech & Jan Svartvik. 1972. A grammar of contem­
porary English. London: Longman.
Quirk, Randolf, Sidney Greenbaum, Geoffrey Leech & Jan Svartvik. 1985. A comprehensive
grammar of the English language. London: Longman.
Riordan, Brian. 2007. There’s two ways to say it: Modeling nonprestige there’s. Corpus Linguis­
tics and Linguistic Theory 3. 233–279.
Römer, Ute. 2005. Progressives, patterns, pedagogy: A corpus-driven approach to English pro­
gressive forms, functions, contexts and didactics. Amsterdam/ Philadelphia: John Benjamins.
Sinclair, John. 1991. Corpus, concordance, collocation. Oxford: Oxford University Press.
Stefanowitsch, Anatol & Stefan Th. Gries. 2003. Collostructions: investigating the interaction
b­etween words and constructions. International Journal of Corpus Linguistics 8. 209–243.
Stefanowitsch, Anatol & Stefan Th. Gries. 2008. Channel and constructional meaning: A collos-
tructional case study. In Gitte Kristiansen & René Dirven (eds.), Cognitive Sociolinguistics,
129–152. Berlin & New York: Mouton de Gruyter.
Szmrecsanyi, Benedikt. 2005. Language users as creatures of habit: a corpus-linguistic analysis of
persistence in spoken English. Corpus Linguistics and Linguistic Theory 1. 113–150.
Szmrecsanyi, Benedikt & Lars Hinrichs. 2008. Probabilistic determinants of genitive variation
in spoken and written English: a multivariate comparison across time, space, and genres. In
Nevalainen, Terttu, Irma Taavitsainen, Päivi Pahta & Minna Korhonen (eds.), The dynamics
of linguistic variation: Corpus evidence on English past and present, 291–309. Amsterdam:
Benjamins.
Copyright of Corpus Linguistics & Linguistic Theory is the property of De Gruyter and its content may not be
copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written
permission. However, users may print, download, or email articles for individual use.