Conversation text types: A multi-dimensional analysis

Douglas Biber
Northern Arizona University, USA
[email protected]

Abstract
Multi-dimensional (MD) analysis is a methodological approach that applies multivariate statistical techniques
(especially factor analysis and cluster analysis) to the investigation of register variation in a language. The
approach was originally developed to analyze the full range of spoken and written registers in a language. Early
studies focused on English register variation (Biber 1985, 1986 and 1988), while later studies have applied the
same approach to Somali, Korean, Tuvaluan, Taiwanese, and Spanish.
Surprisingly, these studies have found some striking similarities in the underlying dimensions that distinguish
among spoken and written registers in these diverse languages. It is even more surprising that MD studies of
restricted discourse domains have also uncovered dimensions that are similar in linguistic form and function to
the more general studies of register variation.
The present study presents an MD analysis of a single register: conversation. Three primary dimensions of
variation are identified, and then cluster analysis is used to distinguish among six conversation text types. The
dimensions and text types are interpreted in linguistic and functional terms.
The authors expectations were that a unique set of dimensions would emerge to characterize the variation
among conversational texts. Instead, the three dimensions identified here turn out to be closely related to
dimensions identified in previous analyses of general register variation. Taken together with previous studies, the
present study of conversation raises the possibility of universal dimensions of variation.

1. Introduction
Multi-dimensional (MD) analysis is a methodological approach that applies multivariate
statistical techniques (especially factor analysis and cluster analysis) to the investigation of
register variation in a language. The approach was originally developed to analyze the range
of spoken and written registers in English (Biber 1985, 1986 and 1988). There are two major
quantitative steps in an MD analysis: (1) identifying the salient linguistic co-occurrence
patterns in a language; and (2) comparing spoken and written registers in the linguistic space
defined by those co-occurrence patterns. In a third step, it is possible to identify groupings of
texts text types that are maximally similar in their multi-dimensional profiles.
Almost any linguistic feature will vary in its distribution across registers, reflecting the
discourse functions of the feature in relation to the situational characteristics of each register
(see, e.g., the grammatical descriptions in the Longman Grammar of Spoken and Written
English; Biber et al., 1999). However, individual features cannot reliably distinguish among
registers: There are simply too many different linguistic characteristics to consider, and
individual features often have idiosyncratic distributions. Instead, analyses based on linguistic
co-occurrence and alternation patterns are required to uncover general register differences.
The theoretical importance of linguistic co-occurrence has been emphasized by linguists such
as Firth, Halliday, Ervin-Tripp, and Hymes. Brown and Fraser (1979: 38-39) observe that it

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
16 DOUGLAS BIBER

can be misleading to concentrate on specific, isolated [linguistic] markers without taking into
account systematic variations which involve the co-occurrence of sets of markers. Ervin-
Tripp (1972) and Hymes (1974) identify speech styles as varieties that are defined by a
shared set of co-occurring linguistic features. Halliday (1988: 162) defines a register as a
cluster of associated features having a greater-than-random...tendency to co-occur.
The MD approach gives formal status to the notion of linguistic co-occurrence, by providing
empirical methods to identify and interpret co-occurrence patterns as underlying dimensions
of variation. The co-occurrence patterns comprising each dimension are identified quantita-
tively through factor analysis. It is not the case, though, that quantitative techniques are
sufficient in themselves for MD analyses of register variation. Rather, qualitative techniques
are required to interpret the functional bases underlying each set of co-occurring linguistic
features. The dimensions of variation have both linguistic and functional content. The
linguistic content of a dimension comprises a group of linguistic features (e.g., nominaliza-
tions, prepositional phrases, attributive adjectives) that co-occur with a high frequency in
texts. Based on the assumption that co-occurrence reflects shared function, these co-occur-
rence patterns are interpreted in terms of the situational, social, and cognitive functions most
widely shared by the linguistic features. That is, linguistic features co-occur in texts because
they reflect shared functions.
Several experiments have been carried out to evaluate the reliability (and to a lesser extent
validity) of the original MD analysis of register variation in English. For example, Biber
(1990) shows that factor analyses carried out on split corpora result in nearly the same
dimensions of variation, as long as the texts in those corpora are sampled to include equiva-
lent ranges of register variation. Biber (1993) shows how these dimensions can be used to
predict the register category of unclassified texts with a high degree of accuracy (using
discriminant analysis). And Biber (1992) uses confirmatory factor analysis to test the good-
ness of fit of several factorial models determined on theoretical grounds, confirming the basic
structure identified using exploratory factor analysis in the 1988 analysis.
While early MD studies focused on register variation in English, subsequent studies have
applied the same approach to Somali, Korean, Tuvaluan, Taiwanese, and Spanish (see, e.g.,
Biber, 1995; Jang, 1998). Although these studies all apply the same methodological approach,
they are carried out independently. In each case, a corpus was designed to represent the range
of spoken and written registers found in the target culture, and a computational tagger was
written to capture the grammatical structure of the target language. The set of linguistic
variables used in each analysis includes the full range of lexical/grammatical distinctions that
are relevant in the target language. Despite this fact, the resulting MD analyses have turned
out to be strikingly similar in some respects. In particular, the analyses of all languages have
uncovered dimensions relating to interactiveness/involvement versus informational focus, the
expression of personal stance, and narrative versus non-narrative discourse (see Biber, 1995,
especially Chapter 7).
The MD methodological framework has also been applied to more restricted discourse
domains.1 These include analyses of elementary school registers (Reppen, 1994 and 2001),

1
There have also been several studies of specific registers that apply the dimensions that were identified and
interpreted in the 1988 MD analysis of spoken and written variation in English (see, for example, the collection
of studies in Conrad and Biber, 2001). It is important to note that these studies do not entail separate MD
analyses. That is, these studies apply the dimensions identified in the 1988 MD analysis of English to some new
discourse domain, but they do not undertake new MD analyses (i.e., involving a new factor analysis).

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 17

job interview language (White, 1994), television talk shows (Connor-Linton, 1989), 18th
century written and speech-based registers (Biber, 2001), university spoken and written
registers (Biber, 2003), and academic subregisters (e.g., Grabe, 1987; Kanoksilapatham,
2003). Many of these studies have identified dimensions of variation similar to those found in
the cross-linguistic studies, especially relating to the same functional concerns of interactive-
ness/involvement versus informational focus, the expression of personal stance, and narrative
versus non-narrative discourse.
This result is surprising for two reasons. First, the statistical technique of factor analysis
like all correlational techniques requires variability. Two linguistic variables cannot be
shown to correlate unless the texts included in an analysis represent a wide range of variation
for those variables. Similarly, factor analysis cannot reliably identify sets of co-varying
linguistic features unless the texts included in the analysis represent a wide range of variation
for the full set of features. Thus, factor analysis is most appropriate for general analyses of
spoken and written texts, which represent an extensive range of variation for almost any
linguistic feature (see the detailed analyses in Biber et al., 1999). In contrast, it might be
assumed that factor analysis is less appropriate for analyses of texts from a single, restricted
discourse domain, because that domain will represent a much smaller range of variation.
Second, to the extent that there is linguistic variability among the texts in a restricted dis-
course domain, there is no reason to assume that it would be similar to the patterns of varia-
tion found in a general-purpose corpus. We would rather expect to find different linguistic
features varying in a restricted domain, reflecting the specific functional differences found in
that domain. In the MD analyses, these specific patterns of linguistic variation should result in
dimensions of variation that are unique to each discourse domain.
Previous MD analyses have shown that restricted discourse domains represent sufficient
linguistic variability for the successful application of this methodological approach. More
surprisingly, these analyses show that some of the same basic dimensions of variation seem to
be fundamentally important across restricted and general discourse domains. (In addition,
there are other dimensions that are unique to a particular domain.) This repeated finding
that some dimensions occur across languages and across general and restricted discourse
domains raises the possibility of universal dimensions of register variation.
The present study further explores this possibility by undertaking an MD analysis of linguistic
variation within a single spoken register: conversation. Factor analysis is used to identify the
linguistic dimensions of variation operating in this discourse domain, and then cluster analysis
is used to identify conversation text types that are well-defined in that multi-dimensional
space. The following sections describe these analyses, followed by discussion of the more
general theoretical implications for the study of register variation.

2. Overview of methodology in the Multi-Dimensional approach

A Multi-Dimensional analysis follows eight methodological steps:
1. An appropriate corpus is designed based on previous research and analysis. Texts are
collected, transcribed (in the case of spoken texts), and input into the computer. The situ-
ational characteristics of each spoken and written register are noted (e.g., communicative
purpose, production circumstances, etc.).
2. Research is conducted to identify the linguistic features to be included in the analysis,
together with functional associations of the features.

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
18 DOUGLAS BIBER

3. Computer programs are developed for automated grammatical analysis, to identify or

tag all relevant linguistic features in texts.
4. The entire corpus of texts is tagged automatically by computer, and all texts are edited
interactively to insure that the linguistic features are accurately identified.
5. Additional computer programs compute normed counts of each linguistic feature in each
text of the corpus.
6. The co-occurrence patterns among linguistic features are analyzed, using factor analysis.
7. The factors are interpreted functionally as underlying dimensions of variation.
8. Dimension scores for each text are computed; the mean dimension scores for each register
are then compared to analyze the salient linguistic similarities and differences among the
registers being studied. The functional interpretation of each dimension is refined based on
the distribution among registers.

3. Preliminary steps for the MD analysis of conversation: Corpus and

linguistic features
3.1. The conversation corpus
The corpus used for the present analysis is taken from the Longman Spoken and Written
English Corpus (LSWE Corpus; see Biber et al., 1999: chapter 1). Only the British English
sub-corpus of conversation was analyzed here; this sub-corpus includes 164 texts containing
c. 4 million words. (A large part of this corpus was also included in the BNC sample of
conversation.) Texts were collected by asking participants to carry tape recorders for several
days, recording their daily interactions. The language collected in this way is conversational,
but most text files are very large and actually include many different conversations. Partici-
pants would generally turn the tape recorder off in between conversations, but each text file in
the corpus includes all the conversations that were recorded on a single tape.
The LSWE/BNC corpus of conversation is large enough to provide the basis for a multi-
dimensional analysis. Texts were collected over many days by many different participants,
representing a wide range of social backgrounds. As a result, the corpus should represent the
range of linguistic variability found within conversation. However, much of that variability is
lost when the corpus is analyzed in its current form, with each text file combining multiple
conversations. Individual conversations can vary with respect to situation and purpose, and to
the extent that there is linguistic variability among conversational texts, it will be associated
with those situational/communicative differences. As a result, the first step in the present
analysis was to segment text files into individual conversations (based on the internal headers
included in each text file). Table 1 shows that the 164 text files in the LSWE conversational
corpus were segmented into 2,926 individual conversations. 760 of these conversations were
shorter than 200 words. Because of the difficulties in obtaining reliable rates of occurrence for
linguistic features in shorter texts, these shorter conversations were excluded from subsequent
analysis.
Table 1 shows that the conversations included in the analysis are on average quite long (1,775
words), with the longest conversations being almost 14,000 words.

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 19

# of text files: 164 (conversation transcripts from the LSWE Corpus)

# of words: 3,930,000
Individual conversations longer than 200 words: 2,166
Individual conversations shorter than 200 words: 760 (dropped from subsequent analyses)
Total individual conversations: 2,926
Length of individual conversations included in the analysis:
mean = 1,775 words min = 200 words max = 13,776 words
Table 1. Initial segmentation of the conversation corpus into individual conversations

3.2. Linguistic features used for the analysis

After the conversation corpus was segmented, each conversation was automatically tagged
using the Biber grammatical tagger. The current version of this tagger incorporates the
corpus-based research carried out for the Longman Grammar of Spoken and Written English
(Biber et al., 1999). The tagger identifies a wide range of grammatical features, including
word classes (e.g., nouns, modal verbs, prepositions), syntactic constructions (e.g., WH
relative clauses, conditional adverbial clauses, that-complement clauses controlled by nouns),
semantic classes (e.g., activity verbs, likelihood adverbs), and lexico-grammatical classes
(e.g., that-complement clauses controlled by mental verbs, to-complement clauses controlled
by possibility adjectives). Appendix A lists the full set of features analyzed here.

4. Identifying and interpreting the dimensions of variation in English

conversation
As noted above, the Multi-Dimensional approach to register variation uses factor analysis to
reduce a large number of linguistic variables to a few basic parameters of linguistic variation.
In MD analyses, the distribution of individual linguistic features is analyzed in a corpus of
texts. Factor analysis is then used to identify the systematic co-occurrence patterns among
those linguistic features the dimensions and then texts and registers are compared
along each dimension.
Table 2 gives the full factorial structure for the analysis in this case, while Table 3 summa-
rizes the important linguistic features defining each dimension (i.e., features with factor
loadings over + or .3). Only 27 of the original 120+ linguistic features were retained in the
final factor analysis. Several features were dropped because they were redundant or overlap-
ped to a large extent with other features. For example, the counts for common verbs, nouns,
and adjectives overlapped extensively with the semantic categories for those word classes,
even though the counts were derived independently. In other cases, features were dropped
because they were extremely rare in conversation. Several of these features were combined
into a more general class. For example, the seven phrasal verb types were combined into a
single feature. Similarly, the five specific types of postnominal modifying clause were
combined into a single relative clause feature. Finally, many features were dropped either
because they did not vary across conversational texts, or because they shared little variance
with the overall factorial structure of this analysis (as shown by the communality estimates).
The solution for three factors was selected as optimal. These three factors account for only
36% of the shared variance, but they are readily interpretable, and subsequent factors
accounted for relatively little additional variance. Given that only 27 linguistic variables were
retained in the final factor analysis, the solution with 3 factors was considered optimal.

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
20 DOUGLAS BIBER

Factor1 Factor2 Factor3

Major Factor 1
Features:
wrdlngth 0.75638 0.06122 -0.08615
n_nom 0.52756 0.02817 -0.08707
prep 0.46987 0.04783 0.05038
abstrctn 0.46893 0.13866 -0.11843
rels 0.44690 0.16282 -0.09168
adj_attr 0.35287 -0.17778 -0.08123
allpasv 0.29913 0.08863 0.09965
contrac -0.43589 0.24817 -0.18706
pro1 -0.39020 0.23735 0.08420
pro2 -0.36418 -0.07955 -0.25500
actv -0.31900 -0.15923 0.03750
Major Factor 2
Features:
that_del -0.02783 0.67341 0.38349
mentalv -0.01213 0.66432 -0.07223
fact_vth 0.14676 0.54415 0.08570
lkly_vth 0.01636 0.42516 0.01091
lklyadvl 0.38093 0.40397 -0.07723
sub_all 0.07570 0.35545 -0.00143
gen_hdg 0.33564 0.34281 -0.08784
factadvl 0.28542 0.33556 0.01889
n 0.16827 -0.52004 -0.04842
wh_ques -0.23266 -0.34870 -0.10075
Major Factor 3
Features:
pasttnse -0.04272 -0.04189 0.79494
nonf_vth -0.12568 0.14077 0.60910
commv -0.13631 0.06964 0.58285
pro3 0.00932 0.10243 0.52077
pres -0.45977 0.26021 -0.51128
allmodal -0.23413 0.20474 -0.26521
Inter-Factor Correlations
Factor1 Factor2 Factor3
Factor1 1.00000 -0.24213 0.09272
Factor2 -0.24213 1.00000 0.08185
Factor3 0.09272 0.08185 1.00000
Table 2. Results of the factor analysis: 3 factor solution; Promax rotation.
Each factor comprises a set of linguistic features that tend to co-occur in the conversations
from the conversation corpus. Factors are interpreted as underlying dimensions of variation
based on the assumption that linguistic co-occurrence patterns reflect underlying communica-
tive functions. That is, particular sets of linguistic features co-occur frequently in texts
because they serve related communicative functions. Features with positive and negative
loadings represent two distinct co-occurrence sets. These comprise a single factor because the
two sets tend to occur in complementary distribution: when a conversation has high frequency

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 21

of the positive set of features, that same conversation will tend to have low frequencies of the
negative set of features, and vice versa. In the interpretation of a factor, it is important to
consider the likely reasons for the complementary distribution between positive and negative
feature sets as well as the reasons for the co-occurrence patterns within those sets.

Dimension 1: Information-focused vs. interactive discourse

Features with positive loadings: word length, nominalizations, prepositional phrases, abstract nouns,
relative clauses, attributive adjectives, passive verb phrases, (likelihood adverbs, general hedges)
Features with negative loadings: present tense verbs, contractions, 1st person pronouns, 2nd person
pronouns, activity verbs

Dimension 2: Stance vs. context-focused discourse

Features with positive loadings: that-deletions, mental verbs, factual/mental verb + that-clause,
likelihood/mental verb + that-clause, likelihood adverbs, adverbial clauses, general hedges,
factual adverbs
Features with negative loadings: nouns, WH-questions

Dimension 3: Narrative-focused discourse

Features with positive loadings: past tense verbs, 3rd person pronouns, non-factual/communication
verb + that-clause, communication verbs, that-deletions

Features with negative loadings: present tense verbs

Table 3. Summary of the factorial structure

For example, the positive features on Factor 1 (e.g., long words, nominalizations, preposi-
tional phrases, abstract nouns, relative clauses, etc.) all relate to informational purposes. These
features are mostly associated with elaborated noun phrases and a dense integration of infor-
mation in a text; previous MD studies have shown these features to be typical of written non-
fictional registers intended for specialist audiences (see, e.g., Biber, 1995; Biber and Finegan,
2001).
In contrast, the negative features on Dimension 1 reflect a focus on the immediate interaction
and activities: present tense verbs, contractions, 1st and 2nd person pronouns, and activity
verbs. The overall interpretation of Dimension 1 is thus relatively straightforward, showing
that conversations tend to be either informational or interactive, but not both. The func-
tional label Information-focused versus interactive discourse is proposed for this dimension.
The positive features on Dimension 2 are mostly linguistic features that express stance:
personal attitudes or indications of likelihood. In the 1988 MD study of spoken and written
register variation, several of these features were shown to co-occur typically with interactive
and reduced structure features (on Dimension 1). In contrast, the analysis here shows that
stance-focused discourse is not necessarily highly interactive discourse, and vice versa. (This
dimension also includes several specific features that were not distinguished in the feature set
used for the 1988 analysis, such as likelihood/mental verb + that-clause and factual adverbs).
The negative pole of Dimension 2 shows a surprising co-occurrence of only two features:
nouns and WH-questions. In past analyses, nouns have co-occurred with other stereotypically
literate features (like adjectives, prepositional phrases, etc.), while WH-questions have co-
occurred with stereotypically oral and interactive features. The interpretation here must
consider why these two features would tend to co-occur in conversations, and why they would

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
22 DOUGLAS BIBER

tend to occur in a complementary distribution to stance features. Consideration of texts with a

high frequency of these two features indicates that they are used together to reflect a focus on
the larger context. WH-questions the what, who, where, when and how directly
ask about that context, and nouns are the primary device used to refer to it. Thus, considering
both positive and negative poles, we propose the interpretive label Stance-focused versus
context focused discourse for Dimension 2.
Finally, Dimension 3 is composed of stereotypically narrative features past tense verbs,
3rd person pronouns, and communication verbs controlling that-clauses; the only negative
feature on this dimension is present tense verbs. Given this grouping of features, the interpre-
tation as Narrative-focused discourse is uncontroversial.

5. Identifying and interpreting conversation text types

Most MD studies have been undertaken to investigate the patterns of variation among
registers: varieties of language that are defined by their situational (i.e. non-linguistic)
characteristics (see Biber, 1994). Conversation is an example of a register according to this
definition, as is newspaper reportage, classroom lectures, personal letters, and academic
research articles. Registers can be defined at any level of specificity, depending on the extent
to which the situational characteristics are specified. For example, academic prose is a very
general register, while academic research articles, psychology research articles, and method-
ology sections in experimental psychology research articles are registers defined at increasing
levels of specificity. The original MD studies (Biber, 1986 and 1988) analyzed a wide range
of general spoken and written registers in English, while many subsequent analyses have
applied those dimensions to the analysis of other more specialized registers (see, e.g., the
studies in Conrad and Biber 2001).
These analyses have shown that there are important, systematic linguistic differences among
registers. Those linguistic differences exist because of the functional basis of MD analysis:
linguistic co-occurrence patterns reflect underlying communicative functions. Registers differ
in their situational/communicative characteristics, and as a result, the dimensions identify
important linguistic differences among registers. However, it is important to note that the
register categories are defined in situational rather than linguistic terms.
A complementary perspective on textual variation is to identify and interpret the text catego-
ries that are linguistically well defined, referred to as text types. Text type distinctions have
no necessary relation to register distinctions. Rather, text types are defined such that the texts
within each type are maximally similar in their linguistic characteristics, regardless of their
situational/register characteristics. However, because linguistic features have strong func-
tional associations, text types can be interpreted in functional terms.
Text types and registers thus represent complementary ways to dissect the textual space of a
language. Text types and registers are similar in that both can be described in linguistic and in
situational/functional terms. However, the two constructs differ in their primary bases:
registers are defined in terms of their situational characteristics, while text types are defined
linguistically.
In the MD approach, text types are identified quantitatively using Cluster Analysis, with the
dimensions of variation as predictors. Cluster analysis groups texts into clusters on the basis
of shared multi-dimensional/linguistic characteristics: the conversations grouped in a cluster
are maximally similar linguistically, while the different clusters are maximally distinguished.
This approach has been used to identify the general text types in English and Somali (see

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 23

Biber, 1989 and 1995). The present section describes the text types that can be distinguished
linguistically within the single register of conversation.
The dimensions of variation (see Section 4 above) are used as linguistic predictors for the
clustering of conversations. The individual feature counts are first standardized so that each
feature has a comparable scale with a mean of 0.0 and a standard deviation of 1. (The stan-
dardization was based on the overall means and standard deviations for each feature in the
conversation corpus.) Then, dimension scores were computed by summing the standardized
frequencies for the features comprising each of the three dimensions. The cluster analysis is
based on the three dimension scores for each conversation.
The methodology in this analytical step can be illustrated conceptually by the 2-dimensional
plot in Figure 1. Each point on Figure 1 represents a conversation, plotting the scores for that
conversation on Dimensions 1 and 2. The numbers in the figure show the cluster number for
each conversation, based on the results of the cluster analysis. Conversations that are similar
in their dimension scores are grouped together as a cluster, or text type. For example, the
conversations labelled with a 1 on Figure 1 all have large positive scores on Dimension 1
(the vertical axis) and large negative scores on Dimension 2 (the horizontal axis). In contrast,
Cluster 2 has positive scores on both Dimensions 1 and 2.

Dimension 1

30 |
|
|
| 1 1 1 1
| 1 1
|
20 |
|
|
|
|
|
10 | 2 2 2
| 2 2 2
| 2 2 2
| 6 6 6 6666 666666 46 4 4 44 22222 2
| 666 6 6 6666 66666 664664644446 4444 444 4
| 6 666 666666 66466 4466464444444 4 44444
0 | 6 66 66 66 64466466 4 4 444 4
| 5 55 55 55555555555 55
| 555555555555555 555555 555 55555
| 33 555555 5555 5 5555 5555555555
| 3 33 33 3333333 3333333333 5 5 5 555 55 55
| 3 333 3 3333333 333 3333
-10 | 3 3 333 3 3333333333 3 3 3
| 3 3 3 3 3
|
|
|
--|---------|---------|---------|---------|---------|---------|---------|---------|---------|---------|------
-10 -8 -6 -4 -2 0 2 4 6 8 10

Dimension 2

Figure 1. Plot of VBDUs along Dimension 1 vs. Dimension 2 (showing all DUs with a distance < 3
from the cluster centroid. Symbol is value of CLUSTER; NOTE: 194 obs hidden.)

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
24 DOUGLAS BIBER

Cluster analysis performs this grouping statistically, based on the scores for all three dimen-
sions. Figure 1 shows the distribution across only two dimensions (1 and 3); these two dimen-
sions were chosen because they provide a good visual display of how the conversations within
each cluster are grouped based on their dimension scores. However, the actual cluster analysis
uses all three dimension scores to identify the groupings of conversations that are maximally
similar in their linguistic characteristics.
Cluster analysis is an exploratory statistical technique. The FASTCLUS procedure from SAS
was used for the present analysis. Disjoint clusters were analyzed because there was no theo-
retical reason to expect a hierarchical structure. Peaks in the Cubic Clustering Criterion and
the Pseudo-F Statistic (produced by FASTCLUS) were used to determine the number of clus-
ters. These measures are heuristic devices that reflect goodness-of-fit: the extent to which the
texts within a cluster are similar, while the clusters are maximally distinguished. In the pre-
sent case, these measures had peaks for the 3-cluster solution and for the 6-cluster solution.
The latter was chosen for subsequent analyses because it provided greater discrimination
among the specialized clusters, facilitating the interpretation of those clusters as conversation
text types.
Figure 1 shows the distribution of these six clusters in only a 2-dimensional space, whereas
the cluster analysis is actually based on a 3-dimensional space. It turns out that the third
dimension is also important in defining some clusters. For example, Cluster 4 is not sharply
delimited in terms Dimensions 1 and 2, but all conversations in this cluster have large positive
scores on Dimension 3 (narrative).
Tables 4 and 5 provide a descriptive summary of the cluster analysis results. Table 4 shows
the number of conversations grouped into each cluster, while Table 5 gives descriptive
statistics for each dimension across the clusters. The clusters differ notably in their distinct-
iveness: the smaller clusters are more specialized and more sharply distinguished linguisti-
cally. For example, Cluster 1 has only 40 conversations; linguistically, the conversations
grouped in Cluster 1 have extremely large positive scores on Dimension 1 (informational);
large negative scores on Dimension 2 (context-focused); and scores near 0.0 on Dimension
3 (narrative). At the other extreme, Cluster 5 is a general text type: it is large (680 conver-
sations) and relatively unmarked in its dimension scores.

Cluster Frequency RMS Std Maximum Distance from Nearest Distance Between
Deviation Seed to Observation Cluster Cluster Centroids
1 40 4.6276 19.8319 2 18.2629
2 116 4.3710 16.9770 4 9.5460
3 496 3.2839 18.7622 5 8.6697
4 308 3.4828 18.2692 6 8.2551
5 680 3.2643 16.2853 4 8.3268
6 526 3.2447 17.4048 4 8.2551

Table 4. Summary of the Cluster Analysis

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 25

Cluster Means
Cluster Dim. 1 Dim. 2 Dim. 3
1 22.15 -5.08 -0.99 (Informational context-focused)
2 7.67 5.87 0.93 (Informational stance-focused)
3 -8.04 -5.19 -2.88 (Interactive context-focused)
4 2.12 -0.31 5.61 (Narrative)
5 -4.15 1.74 0.55 (Unmarked interactive)
6 2.63 -4.46 -1.50 (Unmarked context-focused)

Cluster Standard Deviations

Cluster Dim. 1 Dim. 2 Dim. 3
1 5.02 5.19 3.46
2 5.05 4.47 3.41
3 3.75 3.42 2.54
4 3.72 3.28 3.42
5 3.11 3.49 3.17
6 3.75 3.54 2.22

Table 5. Cluster descriptive statistics for each dimension

The clusters can be interpreted as Conversation Text Types, because each cluster represents a
grouping of conversations with similar linguistic profiles. Figure 2 compares the linguistic
characteristics of the four most distinctive of these conversation types, plotting their mean
dimension scores. The general conversation types clusters 5 and 6 are not plotted in
Figure 2.

Figure 2: Multi-Dim ensional profile for Conversational Text Types 1-4

(Note: Dimension 1 has been transformed to a scale of 10 for comparison)

12

10
Informational context focused
8

6
Informational stance focused
4

2
Interactive context focused
0

-2 Narrative focused
-4

-6

-8

Dim. 1: Dim. 2: Stance vs. Dim. 3: Narrative

Informational vs. context focused
interactive
Dim ensions

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
26 DOUGLAS BIBER

Taken together, Table 4 and Figure 2 provide the basis for the interpretation of each conver-
sation type. (These interpretations are refined by consideration of individual conversations
from each type.)
Type 1 is the most specialized, with the fewest number of texts (only 40, or about 2% of the
conversations in the corpus). Linguistically, these conversations are extremely informational
(Dimension 1) and focused on the context (Dimension 2). Text Sample 1 provides an example
of a conversation from this cluster. This text illustrates the dense use of informational
features, such as nominalizations (e.g., conversation, sophistication, agreement, possibility,
information), other long words (e.g., paperwork, computer-wise, computerized, consequently),
attributive adjectives (modern, massive, great, preliminary, certified), passives (be/getting
inundated), prepositional phrases (to you, about the conversation, with Alec, on a piece of
paper), and relative clauses (things that youre liable to get asked). Although texts from this
cluster would be considered interactive and involved in comparison to written expository
texts, they are highly informational in comparison to other conversational texts.

Text Sample 1 Conversation from Cluster 1: Informational, Context-focused

A: I, I want to talk to you about er the conversation I had with Alec <name> yesterday, he seems to be inundated
with having to get details about <unclear> on his er, all his paperwork and so on, and he seems to be inundated
and he sounded a bit low, quite frankly, to me yesterday on the phone that he was getting inundated with all this
B: Mm, mm
A: work. I said Im quite sure there must be something that could be done computer-wise
B: Right
A: but he sort of pooh-poohed it and sort of said well you know, were getting a bit too old for all this modern
sophistication of computers and so on, well I said well quite frankly I am not totally in agreement with you,
because as you probably know Clyde <name> was looking into a program which will could alleviate a lot
B: Yes I know, I know
A: of the work, that I do, but I
B: yes its on the <unclear>
A: would tell you right here and now, er Im still retaining my bible you know the book
B: Yeah, yes, yes
A: that I have downstairs, because its, if it was to be computerized, it would be a massive great bloody great
volume
B: Yes
A: and I would be carrying this around and it just wouldnt be feasible
B: Quite, right
A: so he said that apparently whenever he came back to B S H he was told by Neville roughly about eighteen
hundred acres would be sort of his target
B: Target, right
A: and its, its multiplied by about three or four times that you see
B: Oh right, right, right
A: so consequently hes getting inundated, he really is apparently under pressure
B: Mm, mm, right
A: so this is why I raised the very conversation about it
B: Right, right
A: and er, I said well look Ill have a word with the erm, with <name> and see if he can think of anything that
might
B: Yes alleviate the point
A: all things in mind that are possible on er, on er computer, and he said that he hadnt much time to think about
it and said well look, maybe over Easter
B: Mm
A: put down on a piece of paper what essentials you want done
B: Right
A: and what things that youre liable to get asked
B: Right, mm, mm
A: so hes going to do that, so I said well look, do you mind if I had a wee sort of preliminary talk with him

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 27

B: Right
A: see if, if its a possibility
B: Right
A: What hes looking for is certified numbers, field numbers
B: That sort of
A: all this sort of information

Type 2 is also relatively specialized (with only 116 conversations grouped into this cluster).
Linguistically, this conversation type is relatively informational (Dimension 1) but especially
marked for being highly stance-focused (Dimension 2). (This conversation type should be
contrasted with Type 5: a much larger cluster that is stance-focused and highly interactive
rather than informational.) Text Sample 2 illustrates the typical linguistic characteristics of
Conversation Type 2. Notice especially the frequent mental verbs (e.g., know, think, expect,
want), stance verbs controlling that-clauses, usually with the that omitted (e.g., would have
thought, I think, I suppose), and the frequent hedges and stance adverbs and adverbials
(surely, obviously, really, actually, probably, certainly, to be perfectly frank).2 Texts from this
cluster are informational, in that they are focused on discussion of a particular topic rather
than the immediate interpersonal interaction, but their primary purpose is the expression of
personal stance in relation to that topic.

Text Sample 2. Conversation from Cluster 2: Informational, Stance-focused

A: No no no one person thats not right.
B: Oh, right.
A: There is no, statutory obligation for the person organizing it
C: Oh, I know.
B: Well not the organizer surely oh I know I would have thought youd have to, <unclear> shoot it
A: Im sure that the social services require psychiatric or
B: Mm, I would of thought so
A: obviously medical <unclear> what youre doing. Mhm but theyre to be qualified people involved. But I
would have expected that the whole thing would have to be operated by, somebody who was qualified.
B: I dont know, because like, you know like the doctors <unclear>
A: <unclear> I think it sort of depends how big that you want to get involved in. If youre just somebody whos
on the outside providing services, to keep the smooth running of it then you dont really have to know anything
about it.
C: Mm.
A: But if youre actually involved in it, and you want to be involved in the people, then I think you have to know
something about it.
B: Well the other evening they were showing something on TV, one of these doctors, doctors practices that are
opting out or whatever. And they got a stockbroker, someone who used to be a stockbroker, actually managing
the whole practice.
A: Yeah.
B: I mean hes obviously not qualified as a doctor.
A: Mhm.
B: So I mean I suppose theyll look at it in the same kind of way, somebody whos got managerial, management
qualities rather than . I suppose people who are interested in the other side of it, the medical side of it,
probably, really be geared up to organizing the money side of it wouldnt they, usually one or the other.
C: So have you done any more calculations on it?
A: Theres nothing really more <unclear> I mean the whole thing is a budget guesstimate. Ive no idea yet, really
what, I mean, you know, for instance I dont know how much ratio staff to patients they need, therefore you
cant really, you know, follow that up because youve no idea what the costs themselves could be.
D: Well you dont know, have you, have you found the statutory requirement for space yet? Per person.

2
Note also the dense use of discourse markers (e.g., I mean, you know) supporting the expression of stance in
this conversation.

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
28 DOUGLAS BIBER

A: I think the thing is going to come unstuck . in the, I think the biggest thing is, I was thinking, is the fact
that youve got to get <unclear> I wouldnt get a commitment from Social Services until they see a property
actually ready for occupation. Now Im not gonna be prepared to go through the whole business and then find
them say oh sorry youre wrong.
C: Property is the biggest bugbear.
A: Yeah. Because I dont think
C: If youre actually sitting on <unclear>
A: I dont think the banks are gonna want to invest. To be perfectly frank. . You see the only way we can get
equity out and put money in ourselves is by selling this place.
C: Yes.
A: Therefore if we dont actually want to live in the same place as the residents, which I certainly wouldnt want
to do, right. Wed have to buy two <unclear> adjoining -

In contrast, Type 3 conversations are much more common (496 conversations grouped into
this cluster, or 23% of all conversations in the LSWE Corpus). These conversations are
extremely interactive and focused on the immediate context, as illustrated by Text Sample 3.
The turns in this conversation are short and highly interactive (notice the dense use of I and
you), and there is a dense use of common nouns together with WH-questions to express
context-dependent information.

Text Sample 3. Conversation from Cluster 3: Interactive, Context-focused

A: Im coming home at lunchtime. Theres milk on the step. Bye-bye.
B: But ... lunchtime
A: Right. Well have to get cracking.
B: what dya mean lunchtime?
A: Well lunchtime Im going to go in and pop back to get things. Ive locked it, it was unlocked. Right.
C: Can I go in the front?
A: Tie your belt up please. Tie your belt up. Okay, speedily . now
B: Oh crash, bang, wallop youre a
A: but both doors were open, you know. Start the car.
C: <screaming>Ah! You happy
A: See the co=
C: now?
A: can you er zip your zips up please? Keira. Can you zip your zip up?
B: I cant.
A: What do you think youll be doing at school today?
B: Recorder concert!
A: Oh! Have you got your recorder? In school?
B: No! Er, yes, yes
A: Yeah.
B: yes.
A: Now, what you gonna be playing?
B: Joe Joe stubbed his toe. Joe Joe stubbed his toe and ... Indian Warrior.
A: Oh!
B: <singing>Big chief, Indian warrior, warrior, warrior. Big chief, Indian warrior. High ... ho! High ...
ho! High ... ho! oh! oh!
A: Right.
B: And erm ... the skateboard ride.
A: <crunches gears> Ooh! That gear. Keeps changing with the
B: Mummy. You know what Ive
A: Skateboard ride?
B: you know what, that I
A: Whats that one?
B: ca= just can play that, I couldnt do recorders that well?
A: Yes.
B: Well now erm, Im really good at it.
A: Can you do all the musical notes?

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 29

B: Yeah.
text omitted
B: Guess what Kirsty was doing when
A: What?
B: we was just practising for recorders?
A: What?
B: She was going like this, and the music was on, she put her feet out and she put the music on her feet.
A: Oh well.

Finally, Type 4 conversations are also relatively common (308 texts). This cluster is relatively
unmarked on Dimensions 1 and 2, but these conversations are extremely narrative in their
Dimension 3 characterization. Text Sample 4 illustrates these characteristics. Note that these
conversations are not necessarily extended stories (although some of them are). Rather, as in
the present case, these conversations can be constructed out of extended discussion of past
events (with frequent past tense verb phrases, 3rd person pronouns, and communication verbs
especially said in this conversation), often coupled with commentary on their immediate
relevance.

Text Sample 4. Conversation from Cluster 4: Narrative

A: Ive just explained that to him. And he said he didnt know that, that he would get hold of Sen and ring me
first thing, thing in the morning . er, to tell me why Sen hasnt paid. Hes got the invoice and everything. I
said well youve sent us twenty thou= . I said there is no VAT on it which it should be! Deary me! He says.
Has he got the invoice? I said yes. And I said, weve been having, having the invoice outstanding since October
at two and half thousand pound! I said, you actually owe me six thousand, one hundred and odd! And I said, you
must realize Ive a small company, and thats, in one respect that Ive had to send those conditions because
youre failing to meet the agreed thirty days payment!
B: Yeah.
A: And I said its not on! I said we couldnt survive like that. And he said, well would you like to carry on with
the contract? I said were too far committed now to, I says to back out. I said, you know, we cant back out at
this stage. And I said, but I said if there isnt the payments of the invoice when they are sent . then . you
know, weve go= youve gotta look at it. So that invoice wants -
B: Doing. Yeah.
A: it wants doing and sending, and put in i= put twenty eight days on.
B: Yeah.
A: Had to be paid, it cant be paid by the twenty eighth its er . you know well if I couldve got hold of
David or er, Andrew <name>, I was gonna give Andrew <name> a right bollocking for just pushing it in and he
shouldve sent it to er, Michael <name>, Michael <name>s just got it shoved in front of his nose a= in
Edinburgh. Hes just gone in to see if everythings alright at Edinburgh . and of course, thats why hes had to
report for that. Which was fair play to him,
but bloody Andrew should have told him! Its agreed, the system of stage payments, its all written to him.
B: And youve just spoken to him have you?
A: Ive just spoken to Michael <name>. Michaels great!
B: And you, you . so he understands after hes sent you this?
A: What?
B: Whats going on.
A: Yeah. Because yo= di= I said I had to send that agreement because youre failing to meet the standard
agreement, youre not paying within the twenty eight days or the thirty days!
B: Mm.
A: I said Ive got an invoice outstanding for October, and I said I cant afford to do that! He said, I realize that.
Then he said, we would want you to do that work he said, because youve got a good reputation. It makes,
you know, if we were not gonna go bust just to get twelve months bloody work out of him on a service
contract! You know, but i= if we couldnt, if they . we= . as they said, if they wanted that money back
tomorrow we could only give them half that money back because of what weve got.

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
30 DOUGLAS BIBER

6. Conclusion
The three dimensions identified by this factor analysis of a conversational corpus are surpris-
ingly similar to the dimensions of variation found in the earlier MD analysis of general
spoken and written registers (Biber, 1988). Both analyses have a dimension that reflects the
distinction between involved/interactive versus informational discourse; both analyses have a
narrative dimension; and both have dimensions related to the expression of stance. The large-
scale MD analyses of spoken and written registers in Somali and Korean similarly identified
dimensions associated with these functions; composed of similar kinds of linguistic features.
Even more surprisingly, several MD analyses of restricted discourse domains have identified
dimensions with similar formal and functional correlates (compare, for example, Reppens
(1994, 2001) analysis of elementary school registers with Whites (1994) analysis of job
interview registers). The fact that similar dimensions are found to be basic even in a corpus
restricted to conversation suggests that these might be candidates for universal parameters of
variation.
Comparing the present analysis to previous MD studies provides two complementary per-
spectives on the characteristics of conversation. In comparison to the full range of spoken and
written registers, conversation is distinctive in being extremely interactive, involved, focused
on the immediate context and personal stance, and constrained by real-time production
circumstances. However, when conversation is considered on its own terms, we discover
systematic patterns of variation among conversational texts (see also Carter and McCarthy,
1997; McCarthy, 1998; Quaglio, 2004; Quaglio and Biber, to appear). Interestingly, the
present analysis indicates that the major parameters of variation internal to conversation are a
mirror image to the dimensions of variation that distinguish among spoken and written
registers.

References
Biber D. (1985). Investigating macroscopic textual variation through multi-feature/multi-dimensional
analyses. Linguistics, vol. (23): 337-60.
Biber D. (1986). Spoken and written textual dimensions in English: Resolving the contradictory
findings. Language, vol. (62): 384-414.
Biber D. (1988). Variation across speech and writing. Cambridge University Press.
Biber D. (1989). A typology of English texts. Linguistics, vol. (27): 3-43.
Biber D. (1990). Methodological issues regarding corpus-based analyses of linguistic variation.
Literary and Linguistic Computing, vol. (5): 257-269.
Biber D. (1992). On the complexity of discourse complexity: A multidimensional analysis. Discourse
Processes, vol. (15) : 133-163. (Reprinted in Conrad and Biber (Eds) (2001): 215-240.)
Biber D. (1993). Using register-diversified corpora for general language studies. Computational
Linguistics, vol. (19): 219-241.
Biber D. (1994). An analytical framework for register studies. In Biber D. and Finegan E. (Eds),
Sociolinguistic perspectives on register. Oxford University Press: 31-56.
Biber D. (1995). Dimensions of register variation: A cross-linguistic comparison. Cambridge
University Press.
Biber D. (2001). Dimensions of variation among 18th century speech-based and written registers. In
Conrad S. and Biber D. (Eds): 200-214.

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 31

Biber D. (2003), Variation among university spoken and written registers: A new multi-dimensional
analysis. In Meyer C. and Leistyna P. (Eds), Corpus analysis: Language structure and language
use. Rodopi.
Biber D. and Finegan E. (2001). Diachronic relations among speech-based and written registers in
English. In Conrad S. and Biber D. (Eds): 66-83.
Biber D., Johansson S., Leech G., Conrad S. and Finegan E. (1999). The Longman grammar of spoken
and written English. Longman.
Brown P. and Fraser C. (1979). Speech as a marker of situation. In Scherer K.R. and Giles H. (Eds),
Social markers in speech. Cambridge University Press: 33-62.
Carter R. and McCarthy M. (1997). Exploring Spoken English. Cambridge University Press.
Connor-Linton J. (1989). Crosstalk: A multi-feature analysis of Soviet-American spacebridges. Ph.D.
Dissertation: University of Southern California.
Conrad S. and Biber D. (Eds) (2001). Variation in English: Multi-Dimensional Studies. Longman.
Ervin-Tripp S. (1972). On sociolinguistic rules: Alternation and co-occurrence. In Gumperz J.J. and
Hymes D. (Eds), Directions in sociolinguistics. Holt: 213-250
Halliday M.A.K. (1988). On the language of physical science. In Ghadessy M. (Ed.), Registers of
written English: Situational factors and linguistic features. Pinter: 162-178
Hymes D. (1974). Foundations in sociolinguistics: An ethnographic approach. University of Pennsyl-
vania Press.
Jang S.-Ch. (1998). Dimensions of spoken and written Taiwanese: A corpus-based register study.
Ph.D. Dissertation. University of Hawaii.
Grabe W. (1987). Contrastive rhetoric and text-type research. In Connor U. and Kaplan R.B. (Eds.),
Writing across languages: Analysis of L2 text. Addison-Wesley: 115-138.
Kanoksilapatham B. (2003). A Corpus-based Investigation of Biochemistry Research Articles: Linking
Move Analysis with Multidimensional Analysis. Ph.D. Dissertation. Georgetown University.
McCarthy M. (1998). Spoken Language and Applied Linguistics. Cambridge University Press.
Quaglio P. (in preparation). Conversation and TV Dialogue: A Corpus-based Study of NBCs Friends.
Ph.D. Dissertation. Northern Arizona University.
Quaglio P. and Biber D. (to appear). The grammar of conversation. In McMahon A. and Aarts B.
(Eds), The Handbook of English Linguistics. Blackwell.
Reppen R. (1994). Variation in elementary student writing. Ph.D. Dissertation. Northern Arizona
University.
Reppen R. (2001). Register variation in student and adult speech and writing. In Conrad S. and Biber
D. (Eds): 187-199.
White M. (1994). Language in job interviews: Differences relating to success and socioeconomic
variables. Ph.D. Dissertation. Northern Arizona University.

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
32 DOUGLAS BIBER

Appendix A.
List of grammatical, syntactic, lexico-grammatical, and semantic features
identified by the Biber Tagger
1. Pronouns and pro-verbs
first person pronouns
second person pronouns
third person pronouns (excluding it)
pronoun it
demonstrative pronouns (this, that, these, those as pronouns)
indefinite pronouns (e.g., anybody, nothing, someone)
pro-verb do

2. Reduced forms and dispreferred structures

contractions
subordinator that deletion (e.g., I think [that/0] he went)
stranded prepositions (e.g., the candidate that I was thinking of)
split auxiliaries (e.g., they were apparently shown to )

3. Prepositional phrases

4. Coordination
phrasal coordination (NOUN and NOUN; ADJ and ADJ; VERB and VERB; ADV and ADV)
independent clause coordination (clause initial and)

5. WH-Questions

6. Lexical specifity
type/token ratio
word length

7. Nouns
nominalizations (ending in tion, -ment, -ness, -ity)
nouns
7a. Semantic categories of nouns
animate noun (e.g., teacher, doctor, employee )
cognitive noun (e.g., fact, knowledge, understanding )
concrete noun (e.g., rain, sediment, modem )
technical/concrete noun
quantity noun (e.g., date, energy, minute )
place noun (e.g., habitat, room, ocean )
group/institution noun (e.g., committee, bank, congress )
abstract/process nouns (e.g., application, meeting, balance )

8. Verbs
8a. Tense and aspect markers
past tense
perfect aspect verbs
non-past tense

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
 CONVERSATION TEXT TYPES: A MULTI-DIMENSIONAL ANALYSIS 33

8b. Passives
agentless passives
by passives
8c. Modals
possibility modals (can, may, might, could)
necessity modals (ought, must, should)
predictive modals (will, would, shall)
8d. Semantic categories of verbs
be as main verb
activity verb (e.g., smile, bring, open)
communication verb (e.g., suggest, declare, tell)
mental verb (e.g., know, think, believe)
causative verb (e.g., let, assist, permit)
occurrence verb (e.g., increase, grow, become)
existence verb (e.g., possess, reveal, include)
aspectual verb (e.g., keep, begin, continue)
8e. Phrasal verbs
intransitive activity phrasal verb (e.g., come on, sit down)
transitive activity phrasal verb (e.g., carry out, set up)
transitive mental phrasal verb (e.g., find out, give up)
transitive communication phrasal verb (e.g., point out)
intransitive occurrence phrasal verb (e.g., come off, run out)
copular phrasal verb (e.g., turn out)
aspectual phrasal verb (e.g., go on)

9. Adjectives
attributive adjectives
predicative adjectives
9a. Semantic categories of adjectives
size attributive adjectives (e.g., big, high, long)
time attributive adjectives (e.g., new, young, old)
color attributive adjectives (e.g., white, red, dark)
evaluative attributive adjectives (e.g., important, best, simple)
relational attributive adjectives (e.g., general, total, various)
topical attributive adjectives (e.g., political, economic, physical)

10. Adverbs and adverbials

place adverbials
time adverbials
10a. Adverb classes
conjuncts (e.g., consequently, furthermore, however)
downtoners (e.g., barely, nearly, slightly)
hedges (e.g., at about, something like, almost)
amplifiers (e.g., absolutely, extremely, perfectly)
emphatics (e.g., a lot, for sure, really)
discourse particles (e.g., sentence initial well, now, anyway)
other adverbs
10b. Semantic categories of stance adverbs
non-factual/manner-of-speaking adverbs (e.g., frankly, mainly, truthfully)

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles
34 DOUGLAS BIBER

attitudinal adverbs (e.g., surprisingly, hopefully, wisely)

factual adverbs (e.g., undoubtedly, obviously, certainly)
likelihood adverbs (e.g., evidently, predictably, roughly)

11. Adverbial subordination

causative adverbial subordinator (because)
conditional adverbial subordinator (if, unless)
other adverbial subordinator (e.g., since, while, whereas)

12. Nominal post-modifying clauses

that relatives (e.g., the dog that bit me, the dog that I saw)
WH relatives on object position (e.g., the man who Sally likes)
WH relatives on subject position (e.g., the man who likes popcorn)
WH relatives with fronted preposition (e.g., the manner in which he was told)
past participial postnominal (reduced relative) clauses (e.g., the solution produced by this process)

13. That complement clauses

13a. That clauses controlled by a verb (e.g., we predict that the water is here)
non-factual/communication verb (e.g., imply, report, say, suggest)
attitudinal verb (e.g., anticipate, expect, prefer)
factual/mental verb (e.g., demonstrate, know, realize, show)
likelihood/mental verb (e.g., appear, hypothesize, predict, think)
13b. That clauses controlled by an adjective (e.g., it is strange that he went there)
attitudinal adjectives (e.g., good, advisable, paradoxical)
likelihood adjectives (e.g., possible, likely, unlikely)
other adjectives
13c. That clauses controlled by a noun (e.g., the proposal that he put forward was accepted)
non-factive noun (e.g., comment, proposal, remark)
attitudinal noun (e.g., hope, reason, view)
factive noun (e.g., assertion, observation, statement)
likelihood noun (e.g., assumption, implication, opinion)

14. WH-clauses

15. To-clauses
15a. To-clauses controlled by a verb (e.g., He offered to stay)
speech act verb (e.g., urge, report, convince)
cognition verb (e.g., believe, learn, pretend)
desire/intent/decision verb (e.g., aim, hope, like, prefer, want)
modality/cause/effort verb (e.g., allow, leave, order)
probability/simple fact verb (e.g., appear, happen, seem)
15b. To-clauses controlled by an adjective
certainty adjectives (e.g., prone, due, apt)
ability/will adjectives (e.g., competent, hesitant)
personal affect adjectives (e.g., annoyed, nervous)
ease/difficulty adjectives (e.g., easy, impossible)
evaluative adjectives (e.g., convenient, smart)
15c. To-clauses controlled by a noun (e.g., agreement, authority, intention)

JADT 2004 : 7es Journes internationales dAnalyse statistique des Donnes Textuelles