Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M.

(eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

The Multidimensional Analysis Tagger

Andrea Nini

University of Manchester

Abstract

This chapter introduces and describes the Multidimensional Analysis Tagger (MAT), a

computer program for the analysis of corpora or single texts using the multi-dimensional

model proposed by Biber (1988). The program uses the Stanford Tagger to generate an initial

tagged version of the input, which is then used to find and count the original linguistic

features used in Biber (1988). The program then plots the text or corpus on to Biber’s (1988)

dimensions and assigns it a text type as proposed by Biber (1989). Finally, MAT offers a tool

to visualize the features of each dimension in the text. The software was tested for reliability

by comparing the dimension scores produced by MAT for the LOB corpus against the ones

obtained by Biber (1988) in his original analysis. This test shows that MAT can largely

replicate Biber’s results. The software was also tested on the Brown corpus and the results not

only confirm the reliability of MAT in calculating the dimension scores, but also suggest that

Biber’s (1988) dimensions and text types can be generalized and applied to other data sets. As

a further example of a MAT analysis, a study of a corpus of threatening and abusive letters is

reported. Although this corpus did not contain the balanced sample of registers required to

perform a new multi-dimensional analysis, MAT allowed a text type analysis of the corpus to

be performed through a comparison with Biber’s (1988; 1989) model of English register

variation.

1. Introduction

About thirty years ago, Biber's (1988)⁠ Variation across Speech and Writing revolutionized

our understanding of registers by introducing factor analysis for the extraction of latent

1
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

dimensions of variation from patterns of co-occurrence of linguistic features, a methodology

later called multi-dimensional analysis. The use of this new methodology also led to a

sounder understanding of the most important linguistic and extra-linguistic factors that

influence register variation in English. Multi-dimensional analysis was adopted in a large

number of other studies on the language used in various registers from academic language (e.

g. Biber, 2003; Gray, 2013) ⁠to the most recent web registers (Grieve, Biber, and Friginal

2011; Titak and Roberson 2013; Biber and Egbert 2016). The flexibility of multi-dimensional

analysis for linguistic research is also demonstrated by its various other applications, such as

the study of author styles (Biber and Finegan 1994)⁠, sociolects (Biber and Burges 2000)⁠,

regional variation (Grieve 2014),⁠ or diachronic register variation (Biber and Finegan 1989).

Beside the value of multi-dimensional analysis itself, Biber (1988) has also uncovered

some very valuable insight on the patterns of variation across the registers of the English

language. By extracting the underlying dimensions of variation for a corpus balanced for

registers, Biber (1988) was able to propose a set of six dimensions that can account for the

linguistic variation in the most important registers of the English language. These six

dimensions represent patterns of co-variation of linguistic features and were functionally

interpreted according to their constituting features and the registers that they characterized.

These original six dimensions are summarized in Table 3.1.

Table 3.1: Short descriptions and summary of the six dimensions of register variation for

English found by Biber (1988).

2
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Description

Dimension 1 Low scores on this Dimension Involved production features:

Involved vs. Informational Discourse indicate informationally dense private verbs, that-deletions,

discourse, as in the case of contractions, present tenses,

academic prose, whereas high second person pronouns, do as

scores indicate that the text is pro-verb, analytic negations,

affective and interactional, as demonstrative pronouns,

for conversations. emphatics, first person pronouns,

pronoun it, be as main verb,

causative subordinations,

discourse particles, indefinite

pronouns, hedges, amplifiers,

sentence relatives, wh- questions,

possibility modals, non-phrasal

coordinations, wh- clauses,

stranded prepositions.

Informational production

features: nouns, average word

length, prepositions, type/token

ratio, attributive adjectives

Dimension 2 The higher the score on this Narrative concerns features: past

Narrative vs. Non-Narrative Concerns Dimension the higher the tenses, third person pronouns,

narrative concern, as in the perfect aspects, public verbs,

case of works of fiction. synthetic negations, present

participial clauses

3
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Dimension 3 Low scores on this Dimension Context-dependent discourse

Context-Independent Discourse vs. indicate dependence on the features: time adverbials, place

Context-Dependent Discourse context as in the case of a sport adverbials, general adverbs.

broadcast, whereas high scores Context-independent discourse

indicate independence from features: wh- relative clauses on

context, as for example in object position, pied-piping

academic prose. relatives, wh- relative clauses on

subject position, phrasal

coordinations, nominalizations

Dimension 4 The higher the score on this Overt expression of persuasion

Overt Expression of Persuasion Dimension indicate the more features: infinitives, prediction

the text explicitly marks the modals, suasive verbs, conditional

author’s point of view as well subordinations, necessity modals,

as their assessment of split auxiliaries

likelihood and/or certainty, as

for example in professional

letters.

Dimension 5 The higher the score on this Abstract information features:

Abstract vs. Non-Abstract Dimension the higher the conjuncts, agentless passives, past

Information degree of technical and abstract participial clauses, by passives,

information, as for example in past participial WHIZ deletion

scientific discourse. relatives, other adverbial

subordinators

Dimension 6 High scores on this Dimension On-line informational

On-Line Informational Elaboration indicate that the information elaboration features: that clauses

expressed is produced under as verb complements,

certain time constraints, as for demonstratives, that relative

example in speeches. clauses on object position, that

clauses as adjective complements

4
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

The value of these dimensions transcends their role in the English language as subsequent

research has also demonstrated a striking cross-linguistic validity for Dimension 1 and

Dimension 2 across several languages from different families (Biber 1995; Biber 2014).

In addition to the discovery of the six dimensions, Biber (1989)⁠ later introduced the

use of cluster analysis to find out the characteristic text types of the English language, that is,

clusters of texts that are linguistically similar in terms of the six dimensions. Using cluster

analysis, this study found out that the same corpus used in Biber (1988) could be divided into

eight text types, a summary of which is presented in Table 3.2.

Table 3.2: Short description and summary of the eight text types for English found by Biber

(1989).

Text type Characterizing registers Dimension profile Description

Intimate interpersonal telephone conversations high score on D1, low Text type that usually

interaction between personal friends score on D3, low score includes interactions that

on D5, unmarked scores have an interpersonal

for the other Dimensions concern between close

acquaintances

Informational interaction face-to-face interactions, high score on D1, low Text type that usually

telephone conversations, score on D3, low score includes personal spoken

spontaneous speeches, on D5, unmarked scores interactions focused on

personal letters for the other Dimensions informational concerns

Scientific exposition academic prose, official low score on D1, high Text type that usually

documents score on D3, high score includes informational

on D5, unmarked scores expositions focused on

for the other Dimensions conveying technical

5
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

information

Learned exposition official documents, press low score on D1, high Text type that usually

reviews, academic prose score on D3, high score includes informational

on D5, unmarked scores expositions focused on

for the other Dimensions conveying information

Imaginative narrative romance fiction, general high score on D2, low Text type that usually

fiction, prepared score on D3, includes texts with an

speeches unmarked scores for the extreme narrative

other Dimensions concern

General narrative press reportage, press low score on D1, high Text type that usually

Exposition editorials, biographies, score on D2, includes texts that use

non-sports broadcasts, unmarked scores for the narration to convey

science fiction other Dimensions information

Situated reportage sport broadcasts low score on D3, low Text type that usually

score on D4, unmarked includes on-line

scores for the other commentaries of events

Dimensions that are in progress

Involved persuasion spontaneous speeches, high score on D4, Text type that usually

professional letters, unmarked scores for the includes persuasive

interviews other Dimensions and/or argumentative

discourse

Besides the pioneering of factor analysis and cluster analysis for the analysis of

registers, another achievement of the findings of the two studies above is the elaboration of a

model of register variation for the English language that is predictive. Using the results of the

multi-dimensional analysis it is possible to determine how a text, corpus, or even register

behaves linguistically in comparison to other registers of English. In essence, the multi-

dimensional model represents a base-rate knowledge of English that allows the description or
6
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

evaluation of other texts or registers.

Despite this potential, the majority of the research on the multi-dimensional analysis

of register variation has focused on using factor analysis and cluster analysis on new data sets.

Relatively speaking, few studies have used previous multi-dimensional models or the original

model itself to describe or evaluate new data. Among these studies, the original multi-

dimensional model has been used especially to study the registers of television programs

(Quaglio, 2009; Al-Surmi 2012; Berber Sardinha 2014; Berber Sardinha and Veirano Pinto

2017) and written or spoken academic registers (Conrad 1996; Conrad, 2001; Biber et al.,

2002)⁠.

These studies are evidence that the model can be useful in many applications that

involve the comparison of new data to a base-rate knowledge of English registers. For

example, the evaluation of similarity of a particular academic text written by a learner of

English to the norm for academic registers of English is such an application. Similarly,

register variation researchers can use the same model to compare a register to the other

registers of English considered in the 1988/1989 model. As opposed to finding new

dimensions, which is an endeavor that brings insight in the internal structure of registers,

contrasting a data set to a general model of English can be another way to bring to light its

register identity.

The application of Biber's original dimensions and text types can also be useful for

those interested in looking at variation within small or unstructured corpora. The first multi-

dimensional analysis was successful in producing a model that describes English registers

because the corpus was carefully sampled by registers and large enough to carry out a

statistical analysis. These two pre-requisites are essential to obtain dimensions that can

adequately capture register variation. However, depending on the data that one wants to

analyze, it might turn out to be impossible to collect a large enough corpus or one that is

7
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

internally stratified enough to produce meaningful register dimensions. In such cases, plotting

the input corpus onto Biber's model of English can be a reasonable approximation to running

a new multi-dimensional analysis. Instead of extracting new dimensions for the register, one

can assess how this new register is different or similar to other registers of the English

language and in doing so finding out its register identity.

The application of Biber's model is however dependent firstly on an empirical

validation of its generalization to new texts and secondly on the development of a tool that

can easily allow other researchers to find the location of a new data set in the English multi-

dimensional space. The present chapter presents research that assesses both points. Firstly, the

chapter introduces the Multidimensional Analysis Tagger (or MAT, freely accessible at

https://sites.google.com/site/multidimensionaltagger/), a computer program that facilitates the

process of applying the original 1988 model to a new data set. After describing its architecture

and validation process, an analysis of the Brown corpus using MAT will be reported to

describe to what extent the model can be applied to new texts. Finally, the chapter concludes

with a demonstration of the applications of MAT for register analysis.

2. The MAT

2. 1 The architecture of MAT

MAT is a computer program that replicates Biber's (1988) tagger, calculates the

dimension scores for each of the dimensions, and then plots the input data onto the multi-

dimensional space while also assigning each text to one of the eight text types identified by

Biber (1989). This whole process is achievable due to the detailed descriptions of the tagging

rules for the linguistic features presented in the appendix of Biber (1988).

After the user has provided an input, MAT returns a tagged version of it using the same

67 linguistic features of Biber (1988). However, MAT does not use the original Biber tagger,

8
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

which is not publicly available, and instead uses the Stanford Tagger (Toutanova et al. 2003)⁠

for the preliminary tagging of basic parts of speech, such as nouns, adjectives, verbs, or

adverbs, followed by Biber's (1988) rules for more complex features, such as sentence

relatives, that as a demonstrative as opposed to a complementizer, and so forth. Although the

original Biber tagger prompted the user with ambiguous cases for certain complex features,

MAT does not implement any manual intervention from the user. However, manual

intervention on the tagged texts can be performed by a user, if he/she wishes, before the

statistical analysis takes place.

Although the tagging rules used by MAT are the same as the original Biber tagger,

since the tagging of basic parts of speech is performed by the Stanford Tagger, the tagged

files returned by MAT are bound to contain some inconsistencies with the original tagger.

While some differences are unavoidable, basic parts of speech attribution generally does not

vary greatly across taggers, and thus results should be compatible to the 1988 results. Indeed,

the reliability of MAT has been tested and the results are reported in the next section below.

After the input has been tagged, in order to calculate the dimension scores, firstly the

occurrences of a feature are counted (with the exception of average word length and

type/token ratio), and then their relative frequency per hundred words is calculated. Finally,

the standardized scores, or z-scores, for each feature are calculated using the standard formula

reported below, where x is the relative frequency of a feature in the user’s input, zx is the

resulting z-score of the feature in consideration, µB is the mean frequency for that feature in

Biber’s (1988) corpus, and σB is the standard deviation of that feature in Biber’s (1988)

corpus:
𝑥 − 𝜇'
𝑧" =
𝜎'

MAT applies these formulas and outputs two files, one with the frequencies (per hundred

words) and one with the z-scores. As described in Biber (1988), the final dimension scores for
9
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

each dimension are calculated by summing or subtracting the z-scores of the dimension

features following the features' polarities within a dimension. For example, for Dimension 1,

the final score is calculated with the formula:

𝐷1 = +𝑧,-./012/2-34 + 𝑧160172821.9:4 + 𝑧;9:1-0;1.9:4 + ⋯ = − +𝑧:9>:4 + 𝑧?9-782:@16 + 𝑧,-2,94.1.9:4 + ⋯ =

In the calculation of dimension scores MAT implements a slight alteration as it includes in

each formula only those variables with a mean higher than 1 in Biber (1988: 77). This change

has been implemented as the features with a mean lower than 1 are rare features of English―

the frequency of which can be highly dependent on sample size. For this reason, if by chance

alone one of these rare features is even slightly more common in the user's input than in the

original corpus, then the z-scores of these features would be abnormal and thus inflate the

dimension scores. The loss of this information does not greatly influence the dimension scores

as, given their rarity, these features contribute very little to the dimensions. In addition to this

change, MAT also offers the possibility to apply a z-score correction, that is, a reduction of

all the z-scores of magnitude higher than 5 to 5, in order to avoid unlikely inflated z-scores

and dimension scores.

Finally, MAT plots the input data in the multi-dimensional space and assigns a text

type to each input text. A graph similar to the graphs displayed in Biber (1988: 172) using

means and ranges is produced for the dimensions selected by the user. Using this graph, the

user can compare their data against a selection of registers. The program will also print out

which register is the most similar to the input data. Another plot is also produced for the text

types mirroring Biber's (1989) visualisation. This plot displays the dimensions horizontally

and the location of each text type as well as the input data on each dimension vertically. In

this way, the user can compare their data to the other text types and assess which text type is

the most similar to the input. The most similar text type is assigned using Euclidean distance

10
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

from the centroids of the clusters reported in Biber (1989).

After carrying out the analysis, the user can also use MAT to visualize of one or more

dimension features in the text. MAT can produce a color-coded file with the selected

dimension features, allowing for the qualitative exploration and interpretation of such

features.

Figure X: Screenshot of MAT interface for Windows.

Figure X shows the interface of MAT for Windows. The third button from the left,

Tag and Analyze, will take as input one text or a folder of texts, tag it, and then return the

input’s location in the multi-dimensional model. The two processes can be done in two

separate steps, for example if the user wants to manually check the quality of the tagging, by

using the first button Tag and then the second button Analyze. Finally, the final button is the

Inspect functionality to visualize the dimension features. More information about how to use

MAT can be found in the manual, which is also freely downloadable at

https://sites.google.com/site/multidimensionaltagger/.

Although great differences are not expected between MAT and the 1988 Biber tagger,

the question of whether and to what extent MAT does indeed replicate the original analysis

can only be tested if the same data set is analyzed and similar results are found. The
11
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

description of such analysis is reported in the section below.

2.2 Testing the reliability of MAT

In order to test whether MAT is reliable a MAT analysis of the original data set used by Biber

(1988; 1989) was carried out. The original data set consisted of the LOB corpus (Johansson,

Leech, and Goodluck 1978) for the published written material, the London-Lund corpus

(Svartvik 1990) for the spoken data, and a small corpus of personal and professional letters

collected by Biber.

Despite some efforts, only the LOB corpus could be retrieved, of which only thirteen

out of its fifteen registers were found: Press Reportage, Press Editorial, Press Reviews,

Religion, Hobbies, Popular Lore, Academic Prose, General Fiction, Mystery Fiction, Science

Fiction, Adventure Fiction, Romantic Fiction, and Humour. The test was therefore carried out

on this data set.

After running MAT on the thirteen available registers of LOB some differences were

observed, but overall Biber's analyses were successfully replicated. The results of the analysis

are displayed in Table 3.3, where the first column identifies a register and the following

columns list the dimension scores obtained by Biber and the ones returned by MAT. The last

column lists and contrasts the distribution of text types for the register using percentages,

from the most common to the least common.

Table 3.3: Comparison of dimension scores and distribution of text types between Biber's

(1988; 1989) analysis of the LOB corpus and a MAT analysis of the same corpus.

Registers D1 D2 D3 D4 D5 D6 Text types

Press reportage -14.02 0.97 2.81 -0.38 0.52 -0.72 59% General
MAT narrative
exposition; 39%
Learned
exposition; 2%
12
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Involved
persuasion;
2% Scientific
exposition
Press reportage -15.01 0.4 -0.3 -0.7 0.6 -0.9 73% General
Biber (1988) narrative
exposition; 25%
Learned
exposition; 2%
Scientific
exposition
Difference 0.99 0.57 3.11 0.32 0.08 0.18

Press editorials -8.4 -0.28 4.38 3.3 1.5 0.33 81% General
MAT narrative
exposition; 7%
Involved
persuasion; 7%
Scientific
exposition;
4% Learned
exposition
Press editorials -10 -0.8 1.9 3.1 0.3 1.5 86% General
Biber (1988) narrative
exposition; 11%
Involved
persuasion; 4%
Learned
exposition
Difference 1.6 0.52 2.48 0.2 1.2 1.17

Press reviews MAT -12.45 -0.74 5.38 -2.32 0.36 -1.01 53% General
narrative
exposition; 47%
Learned
exposition
Press reviews -13.9 -1.6 4.3 -2.8 0.8 -1 47% Learned
Biber (1988) exposition; 47%
General
narrative
exposition; 6%
Scientific
exposition
Difference 1.45 0.86 1.08 0.48 0.44 0.01

Religion -4.26 0.17 4.69 0.85 2.22 1.01 65% General

MAT narrative
exposition; 29%
Involved
persuasion; 6%
Scientific
exposition
Religion -7 -0.7 3.7 0.2 1.4 1 59% General
Biber (1988) narrative
exposition; 18%
Involved
persuasion; 18%
Learned
exposition;
6% Imaginative
narrative
Difference 2.74 0.87 0.99 0.65 0.82 0.01
Hobbies -9.42 -2.1 3.15 1.51 2.54 -0.35 34% General
MAT narrative
exposition; 24%
Learned
exposition; 24%

13
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Involved
persuasion;
18% Scientific
exposition
Hobbies -10.1 -2.9 0.3 1.7 1.2 -0.7 43% General
Biber (1988) narrative
exposition; 21%
Learned
exposition; 21%
Involved
persuasion;
7% Scientific
exposition; 7%
Situated
reportage
Difference 0.68 0.8 2.85 0.19 1.34 0.35

Popular lore -9.58 0.31 3.42 -0.61 1.4 -0.64 36% Learned
MAT exposition; 32%
General
narrative
exposition; 20%
Involved
persuasion;
2% Imaginative
narrative; 9%
Scientific
exposition
Popular lore -9.3 -0.1 2.3 -0.3 0.1 -0.8 36% Learned
Biber (1988) exposition; 36%
Involved
persuasion; 21%
General
narrative
exposition;
7% Imaginative
narrative
Difference 0.28 0.41 1.12 0.31 1.3 0.16

Academic prose -12.16 -2.16 5.38 -0.02 5.14 0.23 56% Scientific
MAT exposition; 24%
Learned
exposition; 14%
General
narrative
exposition;
6% Involved
persuasion
Academic prose -14.09 -2.6 4.2 -0.5 5.5 0.5 44% Scientific
Biber (1988) exposition; 31%
Learned
exposition; 17%
General
narrative
exposition;
9% Involved
persuasion
Difference 1.93 0.44 1.18 0.48 0.36 0.27

General fiction 0.35 6.26 0.03 1.79 -0.45 -0.75 55% Imaginative
MAT narrative; 31%
General
narrative
exposition; 10%
Involved
persuasion;
3% Learned

14
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

exposition
General fiction -0.8 5.9 -3.1 0.9 -2.5 -1.6 51% Imaginative
Biber (1988) narrative; 41%
General
narrative
exposition; 3%
Informational
interaction;
3% Involved
persuasion
Difference 1.15 0.36 3.13 0.89 2.05 0.85

Mystery fiction 0.82 5.76 -0.7 1.55 -0.69 -1.13 67% Imaginative
MAT narrative; 29%
General
narrative
exposition; 4%
Involved
persuasion
Mystery fiction -0.2 6 -3.6 -0.7 -2.8 -1.9 70% Imaginative
Biber (1988) narrative; 23%
General
narrative
exposition; 8%
Situated
reportage
Difference 1.02 0.24 2.9 2.25 2.11 0.77

Science fiction -5.01 6.1 1.08 0.21 -0.54 -0.54 83% General
MAT narrative
exposition; 17%
Imaginative
narrative
Science fiction -6.1 5.9 -1.4 -0.7 -2.5 -1.6 50% General
Biber (1988) narrative
exposition; 33%
Imaginative
narrative; 17%
Situated
reportage
Difference 1.09 0.2 2.48 0.91 1.96 1.06

Adventure fiction -0.85 5.89 -1.29 0.19 -0.97 -1.29 69% Imaginative
MAT narrative; 24%
General
narrative
exposition; 3%
Involved
persuasion;
3% Learned
exposition
Adventure fiction 0 5.5 -3.8 -1.2 -2.5 -1.9 70% Imaginative
Biber (1988) narrative; 31%
General
narrative
exposition
Difference 0.85 0.39 2.51 1.39 1.53 0.61

Romantic fiction 3.55 6.71 -0.88 2.35 -1.26 -1 79% Imaginative

MAT narrative; 17%
General
narrative
exposition; 3%
Involved
persuasion
Romantic fiction 4.3 7.2 -4.1 1.8 -3.1 -1.2 92% Imaginative
Biber (1988) narrative; 8%

15
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

General
narrative
exposition
Difference 0.75 0.49 3.22 0.55 1.84 0.2

Humour -6.19 1.43 1.62 0.43 0.65 -0.56 78% General

MAT narrative
exposition; 11%
Imaginative
narrative; 11%
Involved
persuasion
Humour -7.8 0.9 -0.8 -0.3 -0.4 -1.5 89% General
Biber (1988) narrative
exposition; 11%
Involved
persuasion
Difference 1.61 0.53 2.42 0.73 1.05 0.94

Mean differences 1.24 0.51 2.27 0.72 1.24 0.51

In Dimension 1, Involved versus Informational Discourse, the most important of the

dimensions in terms of variance explained and universality (Biber 1995), score differences

range from 0.28 for Popular Lore to 2.74 for Religion (Mean: 1.24). These differences of the

order or one or two points do not affect the identification of the correct location of a new

input, as Dimension 1 scores in Biber's study range from roughly -20 to 50.

In Dimension 2, Narrative versus Non-Narrative Concerns, again large differences are

not detected, with a range from 0.2 for Science Fiction to 0.87 for Religion (Mean: 0.51). For

this dimension, as for all the other dimensions except for the first one, the range of scores

presented in Biber (1988) roughly ranges from -5 to 5, with a positive score indicating that a

text or register is narrative. Besides the small differences, Table 3 highlights that for all the

registers except two, the sign of the scores is the same as in the original study.

Contrary to the two dimensions above, the results for Dimension 3, Context-

Independent Discourse versus Context-Dependent Discourse, are not as accurate, with

differences ranging from 0.99 for Religion to 3.22 for Romantic Fiction (Mean: 2.27). Such

differences are much higher in magnitude than the ones previously observed as the range of

16
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Dimension 3 from Biber (1988) spans from -5 to 5. An average difference of more than 2

points can affect the reliable identification of the location of a text in this space.

As opposed to Dimension 3, the scores for the remaining dimensions are again not

largely different from Biber's, ranging, respectively: from 0.19 to 2.25 for Dimension 4, Overt

Expression of Persuasion (Mean: 0.72); from 0.08 to 2.11 for Dimension 5, Abstract versus

Non-Abstract Information (Mean: 1.24), and from 0.01 to 1.06 for Dimension 6, On-Line

Informational Elaboration (Mean: 0.51). Although it could be argued that some differences of

the order of magnitude of 2 could be problematic, these are extreme values, as the more

modest mean differences reveal.

In terms of the distribution of text types, most of the distributions assigned by MAT

are compatible with the ones published in Biber (1989). Despite some differences in the exact

percentages, the order of the text types, from most common to least common, is highly

compatible and the most common error is the shifting of a particular text type of one rank.

Since most text types are unmarked in Dimension 3, the assignation of accurate text types is

unaffected by the discrepancies in Dimension 3 scores noted above.

In conclusion, this analysis has found that MAT can replicate Biber's (1988; 1989)

analyses as well as assign dimension scores and text types that are reliable. The exception

found concerns Dimension 3, were at times moderate differences were observed. Although a

careful analysis of these differences was carried out, the cause of the problem could not be

identified. A possibility is that the difference lies in the procedure used by the Stanford

Tagger to tag basic parts of speech. Qualitative exploration of the z-scores of the Dimension 3

features seems to indicate that the abnormal values are mostly found for the general adverbs

z-scores. An abnormal z-score could either indicate difference in adverbs tagging rules

between the Stanford Tagger and the 1988 Biber tagger or, alternatively, a transcription error

in the mean or standard deviation for adverbs in Biber (1988). Although further investigations

17
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

on the issue will be carried out, it is possible to nonetheless conclude that MAT offers a good

replication of Biber's (1988; 1989) results and that it can be used to plot new data onto its

multi-dimensional space.

3. The reliability of Biber's (1988) original dimensions

After having demonstrated that MAT is successful in replicating Biber's (1988; 1989) results,

a question that can be now answered is the extent to which the model itself is reliable. The

perfect test for such a question is the application of MAT to a data set that includes as many

registers as the ones analyzed in the first study, such as a corpus similar to LOB. Luckily,

such a data set is indeed available as the LOB corpus was created as a British replication of

the Brown corpus (Francis and Kucera 1979)⁠. Since both corpora contain exactly the same

registers, in precisely the same categories, text size, and number, the application of MAT to

the Brown corpus is an excellent way of testing the reliability of Biber's model for a similar

yet new data set. In this section an analysis of MAT on the Brown corpus is reported for the

same thirteen registers that were analyzed in the section above, so that a comparison can be

carried out both with Biber's (1988; 1989) results and with the results of MAT on the LOB

corpus.

MAT results for the Brown corpus show how stable the model is to new data, as well

as to the degree of internal consistency of the analysis with MAT. The results are displayed in

Table 3.4 in a format similar to the previous analysis of the LOB corpus.

Table 3.4: Comparison of dimension scores and distribution of text types between Biber's

(1988; 1989) analysis of the LOB corpus and a MAT analysis of the Brown corpus.

Registers D1 D2 D3 D4 D5 D6 Text types

Press reportage -17.61 0.09 4.51 -1.55 0.85 -1.11 75% Learned
Brown exposition; 20%
18
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

General
narrative
exposition; 4%
Scientific
exposition
Press reportage -15.01 0.4 -0.3 -0.7 0.6 -0.9 73% General
Biber (1988) narrative
exposition; 25%
Learned
exposition; 2%
Scientific
exposition
Difference 2.6 0.31 4.81 0.85 0.25 0.21

Press editorials -10.71 -0.59 4.5 1.39 0.63 -0.28 63% General
Brown narrative
exposition; 7%
Involved
persuasion; 26%
Learned
exposition; 4%
Scientific
exposition
Press editorials -10 -0.8 1.9 3.1 0.3 1.5 86% General
Biber (1988) narrative
exposition; 11%
Involved
persuasion; 4%
Learned
exposition
Difference 0.71 0.21 2.6 1.71 0.33 1.78

Press reviews -13.83 -1.32 5.27 -3.31 0.41 -1.08 59% Learned
Brown exposition; 41%
General
narrative
exposition
Press reviews -13.9 -1.6 4.3 -2.8 0.8 -1 47% Learned
Biber (1988) exposition; 47%
General
narrative
exposition; 6%
Scientific
exposition
Difference 0.07 0.28 0.97 0.51 0.39 0.08

Religion -7.17 -0.11 5.1 0.39 2.11 0.49 35% General

Brown narrative
exposition; 29%
Involved
persuasion; 24%
Learned
exposition; 12%
Scientific
exposition
Religion -7 -0.7 3.7 0.2 1.4 1 59% General
Biber (1988) narrative
exposition; 18%
Involved
persuasion; 18%
Learned
exposition; 6%
Imaginative
narrative
Difference 0.17 0.59 1.4 0.19 0.71 0.51

Hobbies -12.44 -2.66 4.47 -0.86 1.34 -1.15 50% Learned

19
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Brown exposition; 36%

General
narrative
exposition; 6%
Involved
persuasion; 8%
Scientific
exposition
Hobbies -10.1 -2.9 0.3 1.7 1.2 -0.7 43% General
Biber (1988) narrative
exposition; 21%
Learned
exposition; 21%
Involved
persuasion; 7%
Scientific
exposition; 7%
Situated
reportage
Difference 2.34 0.24 4.17 2.56 0.14 0.45

Popular lore -13.3 -0.1 3.9 -1.03 1.38 -0.67 44% Learned
Brown exposition; 42%
General
narrative
exposition; 8%
Involved
persuasion; 6%
Scientific
exposition
Popular lore -9.3 -0.1 2.3 -0.3 0.1 -0.8 36% Learned
Biber (1988) exposition; 36%
Involved
persuasion; 21%
General
narrative
exposition; 7%
Imaginative
narrative
Difference 4 0 1.6 0.73 1.28 0.13

Academic prose -13.58 -2.33 5.93 -0.88 4.48 0.01 38% Scientific
Brown exposition; 38%
Learned
exposition; 23%
General
narrative
exposition; 3%
Involved
persuasion
Academic prose -14.09 -2.6 4.2 -0.5 5.5 0.5 44% Scientific
Biber (1988) exposition; 31%
Learned
exposition; 17%
General
narrative
exposition; 9%
Involved
persuasion
Difference 0.51 0.27 1.73 0.38 1.02 0.49

General fiction -5.83 5.86 0.19 -0.33 -0.44 -1.22 66% General
Brown narrative
exposition; 24%
Imaginative
narrative; 10%
Involved

20
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

persuasion
General fiction -0.8 5.9 -3.1 0.9 -2.5 -1.6 51% Imaginative
Biber (1988) narrative; 41%
General
narrative
exposition; 3%
Informational
interaction; 3%
Involved
persuasion
Difference 5.03 0.04 3.29 1.23 2.06 0.38

Mystery fiction -2.21 5.57 -1.22 0.13 -1.03 -1 46% General

Brown narrative
exposition; 42%
Imaginative
narrative; 13%
Involved
persuasion
Mystery fiction -0.2 6 -3.6 -0.7 -2.8 -1.9 70% Imaginative
Biber (1988) narrative; 23%
General
narrative
exposition; 8%
Situated
reportage
Difference 2.01 0.43 2.38 0.83 1.77 0.9

Science fiction -4.1 4.79 1.3 0.12 0.79 -0.78 50% General
Brown narrative
exposition; 17%
Imaginative
narrative; 17%
Involved
persuasion; 17%
Learned
exposition
Science fiction -6.1 5.9 -1.4 -0.7 -2.5 -1.6 50% General
Biber (1988) narrative
exposition; 33%
Imaginative
narrative; 17%
Situated
reportage
Difference 2 1.11 2.7 0.82 3.29 0.82

Adventure fiction -6.05 5.88 -0.81 -1.78 -1.05 -1.39 66% General
Brown narrative
exposition; 31%
Imaginative
narrative; 3%
Learned
exposition
Adventure fiction 0 5.5 -3.8 -1.2 -2.5 -1.9 70% Imaginative
Biber (1988) narrative; 31%
General
narrative
exposition
Difference 6.05 0.38 2.99 -0.58 1.45 0.51

Romantic fiction 0.83 6.02 0.41 -0.08 -1.15 -1.08 59% Imaginative
Brown narrative; 31%
General
narrative
exposition; 10%
Involved
persuasion

21
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Romantic fiction 4.3 7.2 -4.1 1.8 -3.1 -1.2 92% Imaginative
Biber (1988) narrative; 8%
General
narrative
exposition
Difference 3.47 1.18 4.51 1.88 1.95 0.12

Humour -6.76 2.96 2.56 -1.16 0.42 -0.46 67% General

Brown narrative
exposition; 22%
Imaginative
narrative; 11%
Learned
exposition
Humour -7.8 0.9 -0.8 -0.3 -0.4 -1.5 89% General
Biber (1988) narrative
exposition; 11%
Involved
persuasion
Difference 1.04 2.06 3.36 0.86 0.82 1.04

Mean difference 2.18 0.53 2.73 0.88 1.18 0.57

With the exception of Press Reportage, MAT assigns text types to the other Press

registers with a very similar distribution compared to the original results. The similar

distribution is of course a reflection of the similarities in the dimension scores, with the

average score differences ranging from 2.6 in Dimension 3 to the impressive 0.07 in

Dimension 1 for Press Review. Press Reportage shows some differences from Biber's studies,

as Learned Exposition becomes the most common text type for this register as opposed to

General Narrative Exposition, which comes second. This difference could be caused by the

lower score in Dimension 1 obtained by this register using MAT. Their narrative character is

nonetheless correctly identified by MAT through the positive score in Dimension 2. In

general, the differences observed could be attributed to differences in the two corpora in

topics or themes covered by the reportages.

For the Religion, Hobbies, and Popular Lore registers a generally strong compatibility

of results is found, despite the fact that these registers are the most likely to contain different

styles and topics compared to the LOB corpus. Religion presents the best results, both in

terms of text type distribution and in terms of dimension score differences. Similarly good
22
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

results are observed for Popular Lore, with the exception of the Dimension 1 difference score

of 4, which is however not too influential given the wide range of Dimension 1. Finally, the

worst results for these categories are the ones for Hobbies, which include a difference as high

as 4.17 for Dimension 3. As Dimension 3 does not have major weight on the attribution of

text types, the impact that this anomaly has on text type distribution is small and leads only to

the swap of first and second positions in the ranks.

The fact that Biber's model is reliable can be better observed in those registers in

which topics and styles are not expected to vary greatly from corpus to corpus, e.g. Academic

Prose. Indeed, for this register an extremely high level of compatibility of results between the

Brown and the LOB corpora was found, as shown by the very small differences in dimension

scores and a rather impressively similar distribution of text types.

Finally, the last registers discussed are the narrative and Humour, most of which

indicate strong compatibility despite the fact that greater variability in styles and topics is

expected in narratives across corpora. The narrative character of these registers is well-

captured by MAT, as all the registers correctly present positive scores in Dimension 2 and

narrative text types in the first ranks. The most common variations in the attribution of text

types concern the alternation of first and second positions between the text type General

Narrative Exposition and Imaginative Narrative. Both text types characterize narrative texts

and the only difference between the two is that Imaginative Narrative tends to be more

involved as it is typical of emotional narratives such as romantic novels. Indeed, for Romantic

Fiction, both Biber's results and the MAT analysis of the Brown corpus show Imaginative

Narrative as the most common text type. For Adventure Fiction and Mystery Fiction,

however, MAT assigns General Narrative Exposition as the first text type, as opposed to

Imaginative Narrative. These differences aside, the results are largely comparable and show

the reliability of the model for narrative texts.

23
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Table 3.5: Comparison of dimension scores and distribution of text types between the analysis

of the LOB corpus and the analysis of the Brown corpus with MAT.

24
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Registers D1 D2 D3 D4 D5 D6 Text types

Press reportage -17.61 0.09 4.51 -1.55 0.85 -1.11 75% Learned
MAT Brown exposition; 20%
General
narrative
exposition; 4%
Scientific
exposition
Press reportage -14.02 0.97 2.81 -0.38 0.52 -0.72 59% General
MAT LOB narrative
exposition; 39%
Learned
exposition; 2%
Involved
persuasion; 2%
Scientific
exposition
Difference 3.59 0.88 1.7 1.17 0.33 0.39

Press editorials -10.71 -0.59 4.5 1.39 0.63 -0.28 63% General
MAT Brown narrative
exposition; 7%
Involved
persuasion; 26%
Learned
exposition; 4%
Scientific
exposition
Press editorials -8.4 -0.28 4.38 3.3 1.5 0.33 81% General
MAT LOB narrative
exposition; 7%
Involved
persuasion; 7%
Scientific
exposition; 4%
Learned
exposition
Difference 2.31 0.31 0.12 1.91 0.87 0.61

Press reviews -13.83 -1.32 5.27 -3.31 0.41 -1.08 59% Learned
MAT Brown exposition; 41%
General
narrative
exposition
Press reviews -12.45 -0.74 5.38 -2.32 0.36 -1.01 53% General
MAT LOB narrative
exposition; 47%
Learned
exposition
Difference 1.38 0.58 0.11 0.99 0.05 0.07

Religion -7.17 -0.11 5.1 0.39 2.11 0.49 35% General

MAT Brown narrative
exposition; 29%
Involved
persuasion; 24%
Learned
exposition; 12%
Scientific
exposition
Religion -4.26 0.17 4.69 0.85 2.22 1.01 65% General
MAT LOB narrative
exposition; 29%
Involved
persuasion; 6%
Scientific
exposition

25
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Difference 2.91 0.28 0.41 0.46 0.11 0.52

Hobbies -12.44 -2.66 4.47 -0.86 1.34 -1.15 50% Learned

MAT Brown exposition; 36%
General
narrative
exposition; 6%
Involved
persuasion; 8%
Scientific
exposition
Hobbies -9.42 -2.1 3.15 1.51 2.54 -0.35 34% General
MAT LOB narrative
exposition; 24%
Learned
exposition; 24%
Involved
persuasion; 18%
Scientific
exposition
Difference 3.02 0.56 1.32 2.37 1.2 0.8

Popular lore -13.3 -0.1 3.9 -1.03 1.38 -0.67 44% Learned
MAT Brown exposition; 42%
General
narrative
exposition; 8%
Involved
persuasion; 6%
Scientific
exposition
Popular lore -9.58 0.31 3.42 -0.61 1.4 -0.64 36% Learned
MAT LOB exposition; 32%
General
narrative
exposition; 20%
Involved
persuasion; 2%
Imaginative
narrative; 9%
Scientific
exposition
Difference 3.72 0.41 0.48 0.42 0.02 0.03

Academic prose -13.58 -2.33 5.93 -0.88 4.48 0.01 38% Scientific
MAT Brown exposition; 38%
Learned
exposition; 23%
General
narrative
exposition; 3%
Involved
persuasion
Academic prose -12.16 -2.16 5.38 -0.02 5.14 0.23 56% Scientific
MAT LOB exposition; 24%
Learned
exposition; 14%
General
narrative
exposition; 6%
Involved
persuasion
Difference 1.42 0.17 0.55 0.86 0.66 0.22

26
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

General fiction -5.83 5.86 0.19 -0.33 -0.44 -1.22 66% General
MAT Brown narrative
exposition; 24%
Imaginative
narrative; 10%
Involved
persuasion
General fiction 0.35 6.26 0.03 1.79 -0.45 -0.75 55% Imaginative
MAT LOB narrative; 31%
General
narrative
exposition; 10%
Involved
persuasion; 3%
Learned
exposition
Difference 6.18 0.4 0.16 2.12 0.01 0.47

Mystery fiction -2.21 5.57 -1.22 0.13 -1.03 -1 46% General

MAT Brown narrative
exposition; 42%
Imaginative
narrative; 13%
Involved
persuasion
Mystery fiction 0.82 5.76 -0.7 1.55 -0.69 -1.13 67% Imaginative
MAT LOB narrative; 29%
General
narrative
exposition; 4%
Involved
persuasion
Difference 3.03 0.19 0.52 1.42 0.34 0.13

Science fiction -4.1 4.79 1.3 0.12 0.79 -0.78 50% General
MAT Brown narrative
exposition; 17%
Imaginative
narrative; 17%
Involved
persuasion; 17%
Learned
exposition
Science fiction -5.01 6.1 1.08 0.21 -0.54 -0.54 83% General
MAT LOB narrative
exposition; 17%
Imaginative
narrative
Difference 0.91 1.31 0.22 0.09 1.33 0.24

Adventure fiction -6.05 5.88 -0.81 -1.78 -1.05 -1.39 66% General
MAT Brown narrative
exposition; 31%
Imaginative
narrative; 3%
Learned
exposition
Adventure fiction -0.85 5.89 -1.29 0.19 -0.97 -1.29 69% Imaginative
MAT LOB narrative; 24%
General
narrative
exposition; 3%
Involved
persuasion; 3%
Learned
exposition
Difference 5.2 0.01 0.48 1.97 0.08 0.1

27
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Romantic fiction 0.83 6.02 0.41 -0.08 -1.15 -1.08 59% Imaginative
MAT Brown narrative; 31%
General
narrative
exposition; 10%
Involved
persuasion
Romantic fiction 3.55 6.71 -0.88 2.35 -1.26 -1 79% Imaginative
MAT LOB narrative; 17%
General
narrative
exposition; 3%
Involved
persuasion
Difference 2.72 0.69 1.29 2.43 0.11 0.08

Humour -6.76 2.96 2.56 -1.16 0.42 -0.46 67% General

MAT Brown narrative
exposition; 22%
Imaginative
narrative; 11%
Learned
exposition
Humour -6.19 1.43 1.62 0.43 0.65 -0.56 78% General
MAT LOB narrative
exposition; 11%
Imaginative
narrative; 11%
Involved
persuasion
Difference 0.57 1.53 0.94 1.59 0.23 0.1

Mean difference 2.84 0.54 0.63 1.33 0.43 0.28

Table 3.5, shows the internal consistency of MAT through a comparison of the

analyses of the LOB and the Brown corpora. The mean differences of the six dimensions

displayed in Table 5 are small in magnitude, ranging from the relatively small 2.84 in

Dimension 1 to 0.28 in Dimension 6. Text type assignation is also very consistent; the only

differences being an inversion of the first and second text types in the ranks.

In conclusion, the results of these analyses suggest that the model proposed by Biber

is valid and that MAT is consistent in applying it. Similar scores and text types are returned if

the model is applied to a new data set with the same registers tested in Biber's original data set

and this is despite the fact that the tagging of basic parts of speech is done with a different

tagger. This is an important result, which shows that Biber’s model contains valid and stable

information about general patterns of register variation in English. That the multi-dimensional
28
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

model of English can be used with data sets other than the original one implies that its

contribution to linguistics and register analysis goes beyond the introduction of a method of

analysis: the model constitutes a base-rate knowledge of linguistic characteristics of registers

in the English language that can be used in new applied and theoretical research. As a

demonstration of such applications, the next section details the analysis of a register that was

not investigated by Biber in 1988 using MAT.

4. Applying Biber's multi-dimensional model of English to a new data set

The last step of this chapter is to demonstrate the usefulness of MAT for register analysis,

both for comparing and contrasting new data to other registers of the English language and for

performing a multi-dimensional analysis of a corpus that does not meet the its requirements.

If a researcher's goal is to study the language of a certain register using the multi-dimensional

method, then he/she should have access to a corpus that is large enough and with a sufficient

internal stratification of situational parameters to allow for dimensions of register variation to

be captured using factor analysis (Biber and Conrad 2009)⁠. Unfortunately, depending on the

type of data this is not always possible and some data sets might not therefore be analyzed in

this way. The example reported in this section concerns a register for which data collection is

highly problematic, i.e. the register of malicious forensic texts.

A malicious forensic text, or MFT, is defined as a written piece of communication that

is abusive, threatening or defaming and that is used as evidence in a forensic case (Nini 2017).

For example, a ransom demand, an abusive or threatening letter, or a slanderous piece of

writing that has been part of an investigation or a court trial would all qualify as MFTs. The

situational characteristic that MFTs have in common and that links them to each other is the

presence of a malicious purpose or speech act, such as a threat. However, texts classified as

MFTs can differ in other situational characteristics and thus can be letters, text messages,

29
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

notes, and so forth. In the corpus considered, MFTs do tend to share many linguistic and

situational characteristics with each other as they tend to have the same situational

characteristics of written communication, such as being written with the possibility of being

edited, absence of shared time and space between the participants involved, presence of only

one recipient with no audience, etc. A full analysis of the situational parameters of the corpus

of MFTs considered can be found in Nini (2017).

Although this register has been looked at before, especially at the pragmatic level and

in particular for threatening texts (Fraser 1998; Napier and Mardigian 2003; Solan and

Tiersma 2005)⁠, there seems to be no study on the register identity of these kinds of texts. This

gap in the literature could be due to the difficulty of accessing such confidential data.

Although forensic linguists working with law enforcement units frequently work with these

kinds of texts, even in this field paucity of this type of data is an issue. Another problem of

these texts is their relative shortness, as often such texts only contain one hundred or less

tokens. Given the data collection problems listed above, finding enough data stratified by

situational parameters so that a multi-dimensional register analysis can be carried out is

difficult. However, thanks to the application of Biber's model through MAT some of these

limitations can be overcome and the register identity of these texts can be investigated.

The MFT corpus here adopted was collected by Nini (2015) and consists of 104 texts,

for a total of 39,188 word tokens and an average text length of 357 tokens (min: 103, max:

1610). Almost all the data set was collected using publicly available sources, such as forensic

linguistics textbooks (e. g. Olsson, 2003)⁠, the FBI Vault⁠ repository of texts

(https://vault.fbi.gov/), or the web through search engine queries. A smaller section of the

corpus was made up of non-public texts made available by forensic linguists who frequently

work on real-life cases in the UK and in the US. The corpus was limited to texts that

contained at least 100 word tokens, as below this value it is not possible to calculate the

30
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

frequency of features reliably (Biber 1993; Biber and Jones 2005)⁠.

Although this corpus contains largely non-standard, unedited texts with misspellings,

slang, non-standard punctuation or capitalization and so forth, MAT performed well in

tagging the corpus. The reliability of MAT for the MFT corpus was tested by hand-checking a

random 20% of the MAT tagged files while counting the number of tagging errors and

calculating the percentage of correct tags for each text. This test revealed that on average 96%

of the tags were correct (min: 87%, max: 100%). After this test, the corpus was used as input

for a MAT analysis as described above.

With the application of Biber's model using MAT, it was possible to add MFTs to the

register analysis of English presented by Biber (1988; 1989). The present section concerns the

dimension scores and text types obtained for the MFTs compared to the remaining dimension

scores and text types from Biber (1988; 1989). Table 3.6 gives the mean dimension scores for

the MFT register in comparison with two registers that are similar in terms of situational

parameters, Personal and Professional Letters.

Table 3.6: Comparison of mean dimension scores and standard deviations for MFTs and

Personal and Professional Letters from Biber (1988).

31
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Personal MFTs Professional

Letters Letters
Mean SD Mean SD Mean SD
Dimension 1 19.5 5.4 1.5 11.5 -3.9 13.7
Involved vs. Informational
Discourse
Dimension 2 0.3 1.0 -0.7 3.9 -2.2 3.5
Narrative vs. Non-Narrative
Concerns
Dimension 3 -3.6 1.8 2.6 3.9 6.5 4.2
Context-Independent
Discourse vs. Context-
Dependent Discourse
Dimension 4 1.5 2.6 3.9 5.7 3.5 4.7
Overt Expression of Persuasion
Dimension 5 -2.8 1.9 0.3 4.0 0.4 2.4
Abstract vs. Non-Abstract
Information
Dimension 6 -1.4 1.6 0.1 2.6 1.5 3.6
On-Line Informational
Elaboration

Table 3.6 shows that the mean dimension scores for MFTs are located between the

two types of letters, with the exception of Dimension 4, and on average closer to Professional

than to Personal Letters.

32
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Figure 3.1: Mean scores and ranges for Dimension 1, Involved vs. Informational Discourse,

for a selection of Biber's (1988) registers compared to the mean and range for MFTs.

In Dimension 1, Involved versus Informational Discourse, (Figure 3.1), MFTs are

rather unmarked texts. With a mean score close to one, the average MFT text contains

language that is not characterized by either dimension poles, similarly to that of General

Fiction, and rather distant from more involved or informational registers, such as

Conversation and Academic Prose, respectively.

33
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Figure 3.2: Mean scores and ranges for Dimension 2, Narrative vs Non-Narrative Concerns,

for a selection of Biber's (1988) registers compared to the mean and range for MFTs.

In Dimension 2, Narrative versus Non-Narrative Concerns, (Figure 3.2), the mean

score below zero suggests that MFTs are on average non-narrative texts situated together with

other non-narrative registers such as Academic Prose and distant from Fiction registers.

34
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Figure 3.3: Mean scores and ranges for Dimension 3, Context-Independent vs. Context-

Dependent Discourse, for a selection of Biber's (1988) registers compared to the mean and

range for MFTs.

The analysis of Dimension 3, Context-Independent versus Context-Dependent

Discourse (Figure 3.3) reveals that MFTs on average have a tendency towards explicitness of

information, just as other written registers, such as Academic Prose and as opposed to the

context-dependency of certain spoken registers, such as Conversations or Broadcasts.

35
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Figure 3.4: Mean scores and ranges for Dimension 4, Overt Expression of Persuasion, for a

selection of Biber's (1988) registers compared to the mean and range for MFTs.

Dimension 4, Overt Expression of Persuasion, (Figure 3.4) presents interesting results,

as the mean score achieved by the average MFT is far higher than any other register analyzed

in Biber's works. In his original study, the highest mean score for this dimension was that of

Professional Letters, as this register is the one that most frequently adopts persuasive

linguistic means. As this analysis reveals, though, MFTs are far superior in the use of

persuasive means than average-scored Professional Letters and the high mean score of MFTs

in Dimension 4 could be regarded as the register ‘signature’. As Nini (2017) reveals by

comparing and contrasting texts with different situational parameters in this same MFT

corpus, this high score in Dimension 4 is due to the presence of threatening texts, which are

characterized by high levels of persuasion features.

36
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Figure 3.5: Mean scores and ranges for Dimension 5, Abstract vs Non-Abstract Information,

for a selection of Biber's (1988) registers compared to the mean and range for MFTs.

In Dimension 5, Abstract versus Non-Abstract Information (Figure 3.5), MFTs are on

average unmarked. They are distant from other more abstract written registers, such as

Academic Prose and also distant from other registers, such as Conversations, in which

abstract discourse is rarely found. As such, for this dimension, MFTs again appear as a kind

of less formal and less abstract written register.

37
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Figure 3.6: Mean scores and ranges for Dimension 6, On-Line Informational Elaboration, for

a selection of Biber's (1988) registers compared to the mean and range for MFTs.

Finally, in terms of Dimension 6, On-Line Informational Elaboration (Figure 3.6),

MFTs are also unmarked and do not show signs of high degrees of on-line information

elaboration.

In terms of text types, MAT reveals that the average text type for MFTs is Involved

Persuasion, with 47% of the texts being classified in this category. The high frequency of

Involved Persuasion text types is not surprising as this text type is marked by high scores in

Dimension 4 and it is the prototypical text type for Professional Letters. The fact that almost

50% of MFTs fall into this text type is therefore a confirmation of their peculiar Dimension 4

scores due to the presence of several texts with threatening content, which often employ

modal verbs and other modality features to convey their stance (Gales 2011; Gales 2015; Nini

2017)⁠.

38
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Although this analysis is not entirely equivalent to finding out the dimensions of

variation for MFTs and/or the text types within this register, there is much value in plotting

the register of MFTs against other registers of the English language. The results of this study

reveal interesting facts about this relatively unexplored register so that it is possible to get a

grasp of its linguistic characteristics even without reading any of the texts and simply by

examining the text types these texts are classified as and the scores that these texts have on

Biber’s dimensions. This type of analysis can also approximate the results of a full multi-

dimensional register analysis, as the evaluation of Biber's dimensions can be used to

understand that, for example, Dimension 4 is a key dimension for MFTs, or that Dimension 5

and 6 are relatively unimportant. This knowledge is useful by itself, but can also inform future

multi-dimensional studies, which in order to find the internal structure of the register, should

perhaps focus on modality and other persuasive linguistic features.

In summary, besides concluding that MFTs are similar to Professional Letters, the

insight given by this analysis is the individuation of the space of variation for these texts and

their role within the ecosystem of the English language. An interested analyst can now predict

which features to expect in MFT texts and with which frequency. Such knowledge can

empower, for example, forensic linguists who are interested in base-rate knowledge of

forensic registers. Similarly, for more theoretical purposes, analyses as the one presented

constitute another piece of the puzzle in search for a comprehensive descriptive and predictive

framework of the registers of English and for the understanding of the nature of the linguistic

features, their extra-linguistic predictive factors, and their history and evolution.

5. Conclusions

The aims of the present chapter were to demonstrate the usefulness of Biber's model of

English register variation as well as to show how this can be applied for new data using MAT.

39
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

The purpose of the chapter was not only to present a method to do so, but also to encourage

new research that exploits the model as a different way to perform multi-dimensional

analysis.

Firstly, the application of MAT to the LOB corpus revealed that despite some

differences in the tagging of basic parts of speech, MAT can obtain largely the same

dimension scores as Biber (1988) and assign text types similarly as Biber (1989). The only

questionable results are found in the Dimension 3 scores, although these do not substantially

affect the attribution of text types. Besides the reliability of MAT, this test also demonstrates

the stability of Biber's original model, which can return the same results even if slightly

different taggers are used.

Secondly, the validity of the model was again demonstrated by applying MAT to the

Brown corpus, a corpus similar to LOB, for which therefore similar attribution of text types

and dimension scores were expected. The large majority of the results prove that Biber's

model is not only valid for LOB, but that it is also applicable to new data sets.

Finally, the last step of this chapter was to demonstrate how the model could be

applied to new data sets using MAT. Although the corpus of malicious forensic texts used

was not large and stratified enough for a full multi-dimensional analysis, the application of

Biber's model of English shed light on the register identity of the corpus, the location of

malicious forensic texts within the English language, and their linguistic peculiarities.

Through this analysis it was revealed that malicious forensic texts are linguistically similar to

professional letters and are often characterized by large scores in Dimension 4, the dimension

of expression of persuasion.

In addition to the results above, the analyses reported also encourage reflections

regarding the importance of using previously generated multi-dimensional models. Since the

introduction of multi-dimensional analysis research by Biber (1988), a lot of attention has

40
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

been given to the method itself and to its potential for new research. A vast array of studies

have applied the multi-dimensional method to registers other than those used in his seminal

study and produced exciting results. However, not many studies have exploited the

dimensions of variation in English or text types produced by other multi-dimensional

analyses. If it is believed that the potential of a multi-dimensional analysis carried out on a

representative corpus of a register is to produce a model of said register, which is descriptive

and predictive, then it is also advisable that these models become the bricks upon which new

research is built. This is particularly the case for those multi-dimensional models that aim at

being comprehensive, such as Biber's model of the English language. Research that builds on

the foundations of previous multi-dimensional studies is very much welcomed, especially for

applied problems. An example of such applications is Crosthwaite's (2016) analysis of a

longitudinal corpus of English for Academic Purposes students' writings, in which MAT is

used to plot student texts onto Biber's model in order to assess their progress. Such

applications reveal the possibilities that past multi-dimensional studies offer and should

encourage researchers using this methodology to make their models available and applicable

for other researchers.

References

Al-Surmi, Mansoor. 2012. “Authenticity and TV Shows: A Multidimensional Analysis

Perspective.” TESOL Quarterly 46 (4): 671–694. doi:10.1002/tesq.33.

Berber Sardinha, Tony. 2014. “25 Years Later: Comparing Internet and Pre-Internet

Registers.” In Multi-Dimensional Analysis, 25 Years on: A Tribute to Douglas Biber,

edited by Tony Berber Sardinha and Marcia Veirano Pinto, 81–105. Amsterdam;

Philadelphia: John Benjamins Publishing Company.

Berber Sardinha, Tony, and Marcia Veirano Pinto. 2017. “American Television and off-

41
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

Screen Registers: A Corpus-Based Comparison.” Corpora 12 (1): 85–114.

doi:10.3366/cor.2017.0110.

Biber, Douglas. 1988. Variation across Speech and Writing. Cambridge: Cambridge

University Press.

———. 1989. “A Typology of English Texts.” Linguistics 27 (1): 3–43.

———. 1993. “Representativeness in Corpus Design.” Literary and Linguistic Computing 8

(4): 243–57. doi:10.1093/llc/8.4.243.

———. 1995. Dimensions of Register Variation: A Cross-Linguistic Comparison.

Cambridge; New York: Cambridge University Press.

———. 2003. “Variation among University Spoken and Written Registers: A New Multi-

Dimensional Analysis.” Language and Computers 46 (1): 47–70.

———. 2014. “Using Multi-Dimensional Analysis to Explore Cross-Linguistic Universals of

Register Variation.” Languages in Contrast 14 (1): 7–34.

Biber, Douglas, and Jena Burges. 2000. “Historical Change in the Language Use of Women

and Men: Gender Differences in Dramatic Dialogue.” Journal of English Linguistics 28

(1): 21–37.

Biber, Douglas, and Susan Conrad. 2009. Register, Genre, and Style. Cambridge; New York:

Cambridge University Press.

Biber, Douglas, Susan Conrad, Randi Reppen, Pat Byrd, and Marie Helt. 2002. “Speaking and

Writing in the University: A Multidimensional Comparison.” TESOL Quarterly 36 (1):

9–48. doi:10.2307/3588359.

Biber, Douglas, and Jesse Egbert. 2016. “Register Variation on the Searchable Web: A Multi-

Dimensional Analysis.” Journal of English Linguistics 44 (2): 95–137.

doi:10.1177/0075424216628955.

Biber, Douglas, and Edward Finegan. 1989. “Drift and the Evolution of English Style: A

42
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

History of Three Genres.” Language 65: 487–517.

———. 1994. “Multi-Dimensional Analyses of Authors’ Styles: Some Case Studies from the

Eighteenth Century.” In Research in Humanities Computing 3, edited by D Ross and D

Brink, 3–17. Oxford: Oxford University Press.

Biber, Douglas, and James Jones. 2005. “Merging Corpus Linguistic and Discourse Analytic

Research Goals: Discourse Units in Biology Research Articles.” Corpus Linguistics and

Linguistic Theory 1 (2): 151–82. doi:10.1515/cllt.2005.1.2.151.

Conrad, Susan. 1996. “Investigating Academic Texts with Corpus-Based Techniques: An

Example from Biology.” Linguistics and Education 8: 299–326.

———. 2001. “Variation among Disciplinary Texts: A Comparison of Textbooks and Journal

Articles in Biology and History.” In Variation in English: Multi-Dimensional Studies,

edited by Susan Conrad and Douglas Biber, 94. Harlow: Longman.

Crosthwaite, Peter. 2016. “A Longitudinal Multidimensional Analysis of EAP Writing:

Determining EAP Course Effectiveness.” Journal of English for Academic Purposes 22.

Francis, Winthrop Nelson, and Henry Kucera. 1979. Manual of Information to Accompany a

Standard Corpus of Present-Day Edited American English, for Use with Digital

Computers. Brown University.

Fraser, Bruce. 1998. “Threatening Revisited.” Forensic Linguistics 5 (2): 159–73.

Gales, Tammy. 2011. “Identifying Interpersonal Stance in Threatening Discourse: An

Appraisal Analysis.” Discourse Studies 13: 27–46. doi:10.1177/1461445610387735.

———. 2015. “Threatening Stances: A Corpus Analysis of Realized vs Non-Realized

Threats.” Language and Law/Linguagem E Direito 2 (2): 1–25.

Gray, Bethany. 2013. “More than Discipline: Uncovering Multi-Dimensional Patterns of

Variation in Academic Research Articles.” Corpora 8 (2): 153–81.

Grieve, Jack. 2014. “A Multi-Dimensional Analysis of Regional Variation in American

43
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

English.” In Multi-Dimensional Analysis, 25 Years on: A Tribute to Douglas Biber,

edited by Tony Berber Sardinha and Marcia Veirano Pinto, 3–35. Amsterdam: John

Benjamins.

Grieve, Jack, Douglas Biber, and Eric Friginal. 2011. “Variation among Blogs: A Multi-

Dimensional Analysis.” Genres on the Web 42: 303–22.

Johansson, Stig, Geoffrey Leech, and Helen Goodluck. 1978. Manual of Information to

Accompany the Lancaster-Oslo/Bergen Corpus of British English, for Use with Digital

Computers. University of Oslo.

Napier, M, and S Mardigian. 2003. “Threatening Messages: The Essence of Analyzing

Communicated Threats.” Public Venue Security.

Nini, Andrea. 2015. “Authorship Profiling in a Forensic Context.” Aston University, UK.

———. 2017. “Register Variation in Malicious Forensic Texts.” International Journal of

Speech, Language and the Law 24 (1). doi:10.1558/ijsll.30173.

Olsson, John. 2003. Forensic Linguistics: An Introduction to Language, Crime and the Law.

London: Continuum.

Quaglio, Paulo. 2009. Television Dialogue: The Sitcom Friends vs Natural Conversation.

Amsterdam; Philadelphia: John Benjamins Publishing.

Solan, Lawrence M., and Peter M. Tiersma. 2005. Speaking of Crime: The Language of

Criminal Justice. Chicago: University of Chicago Press.

Svartvik, Jan. 1990. The London-Lund Corpus of Spoken English: Description and Research.

Lund, Sweden: Lund University Press.

Titak, Ashley, and Audrey Roberson. 2013. “Dimensions of Web Registers: An Exploratory

Multi-Dimensional Comparison.” Corpora 8 (2): 235–60.

Toutanova, Kristina, Dan Klein, Christopher Manning, and Yoram Singer. 2003. “Feature-

Rich Part-of-Speech Tagging with a Cyclic Dependency Network.” In Proceedings of

44
Nini, A. (2019). The Multi-Dimensional Analysis Tagger. In Berber Sardinha, T. & Veirano Pinto M. (eds), Multi-
Dimensional Analysis: Research Methods and Current Issues, 67-94, London; New York: Bloomsbury Academic

HLT-NAACL 2003, 252–59.

45