ARCTOS

Acta Philologica Fennica

VOL. XLVII

HELSINKI 2013
INDEX

Neci̇p Fi̇kri̇ Ali̇can – Rethinking Plato's Forms 11

Holger Thesleff

Ann Brysbaert Set in Stone? Socio-Economic Reflections on Human and 49

Ani­mal Resources in Monumental Architecture of Late
Bronze Age Tiryns in the Argos Plain, Greece

Gualtiero Calboli A propos de l'ode d'Horace 4,9, en défence de Marcus 97

Lollius. Quelques observations

Ulrike Ehmig Risikobewältigung bei Schwangerschaft und Geburt in 111

der römischen Antike: lateinische dokumentarische und
archäologische Zeugnisse

Rudolf Haensch Von Poppaea zu Pulcheria – Das Bemühen um göttlichen 131

Beistand bei der Geburt eines kaiserlichen Nachfolgers

Kai Juntunen The Arrogant Armenian – Tiridates (Bagratuni) in Cassius 153

Dio and Movses Khorenats'i

Nikolaos Kälviäinen Levels of Style in Byzantine Greek and the Role of Syntactic 173
Complexity: A Quantitative Analysis of the Sentence
Structure of Three Early Byzantine Hagiographic Texts

Antonio Pistellato Gaius Caesar, or the Ideal Non-princeps: A Tiberian Issue 199

Ari Saastamoinen Physical and Visual Characteristics of Latin Building 219

Inscriptions. The Case of North Africa

Elina M. Salminen – Myrrhine's Ball Revisited 243

Mika Kajava

Olli Salomies A Note on the Speeches of the Prosecutors in Cicero's pro 257
Milone 35–6

Heikki Solin Analecta epigraphica CCLXXXV– CCXCI 265

De novis libris iudicia 301

Index librorum in hoc volumine recensorum 393

Libri nobis missi 399

Index scriptorum 403

Arctos 47 (2013) 173–198

LEVELS OF STYLE IN BYZANTINE GREEK AND THE ROLE OF

SYNTACTIC COMPLEXITY: A QUANTITATIVE ANALYSIS OF
THE SENTENCE STRUCTURE OF THREE EARLY BYZANTINE
HAGIOGRAPHIC TEXTS1

Nikolaos Kälviäinen

1. Levels of style in Byzantine Greek

A characteristic of Medieval Greek literature that readers have always been in-
stinctively aware of is its linguistic heterogeneity. Traditional linguistic forms
dating as far back as Homer and Classical Attic continued to be cultivated by
learned Byzantines alongside a range of varieties ultimately derived from the
Hellenistic Koine, the internationalized Attic cemented as the lingua franca of
the Eastern Mediterranean in Hellenistic times, but influenced to varying degrees
by later developments in the history of the spoken language. The metaphor of
verticality is often applied to this kind of variation, which in sociolinguistics is
referred to as diglossia, with two main language systems, H(igh) and L(ow), in
use in the same speech community.2
1 I would here like to express my warmest gratitude to a number of people in both Helsinki and
Rethymno, who in one way or another helped me complete my Master's thesis (University of
Helsinki 20.8.2012), on which this paper is based: my instructors, Prof. Mika Kajava and Dr.
Hilla Halla-aho for their helpful and diligent guidance; Dr. Marina Detoraki for her enthusias-
tic support and feedback; Dr. Marja Vierros for her valuable comments; Prof. Fred Karlsson
for pointing me to a series of studies of syntax which proved most useful in both theory and
practice; and my father, MS Timo Kälviäinen for helping with inferential statistics. I also wish
heartily to thank Otto Nieminen for correcting my English expression. The responsibility for
any remaining errors lurking in the pages of the present paper are of course solely my own.
Finally, to Georgia: επιτέλους είμαι σπίτι!
2 N. Toufexis, "Diglossia and register variation in Medieval Greek" BMGS 32.2 (2008) 208–
09. Toufexis is concerned specifically with the Late Byzantine period, but the diglossic situa-
tion is present during the entire Byzantine millennium, ultimately dating back to antiquity (cf.
the dichotomy of Atticism and Koine).
174 Nikolaos Kälviäinen

There is, however, no question of two distinctly separate linguistic systems

being in use.3 Throughout Byzantine history the conventional terms H and L can
be seen as the two idealized extreme poles of a linguistic continuum stretching
from the elite literature of highly educated Constantinopolitans, with their Atti-
cizing and classicizing pursuits, through the various levels of educated, bureau-
cratic and ecclesiastical Koine all the way down to "vernacular" literature with
relatively few archaisms and comparatively heavy influence from spoken forms
of Greek.4 Of these two poles, H enjoys higher prestige and is associated with
learning and elite literature, whereas L is less prestigious and is associated with
everyday uses of language. L is based on the spoken language of Greek-speaking
Byzantines, whereas H is no-one's native tongue; the ability to use H is acquired
through extensive education and as a consequence it is, in the words of Toufexis,
"far from homogeneous, as it incorporates more than one register or variety with
different linguistic characteristics in correlation both with the genre of each text
and the educational level of each author".5 Of course, the written L is heteroge-
neous as well, being necessarily influenced by the conventions of the more con-
servative written registers, since one could not learn to write without learning at
least some H.6
Apart from genre and the author's own literary background, there is an-
other important factor affecting the choice of register: that of the target audience's
educational level. It has often been observed, in the case of hagiographical works
intended to cater to the spiritual needs of common people, that even educated
authors such as Palladius, John Moschus and Leontius of Neapolis would some-
times deliberately write in a low register in order for less educated audiences to be
able to derive benefit from their edifying tales.7 In fact, one can easily appreciate
the stylistic difference by examining the proems of texts such as Moschus' Pra-
tum Spirituale and Palladius' Historia Lausiaca, the rhetorical character of which

3 Toufexis (above n. 2) 210.

4 See, e.g., G. Horrocks, Greek: A History of the Language and its Speakers, Chichester 20102,
207–30 for a concise account of this variation in the language of Byzantine literature and p. 244
for the coalescing of the middle registers after the Early Byzantine period.
5 Toufexis (n. 2) 211–12.
6 Toufexis (n. 2) 212.
7 Cf., e.g., K. Hult, Syntactic Variation in Greek of the 5th century A.D., Göteborg 1990,
28; Horrocks 2010 (n. 4) 225–26; D. Hesseling, Morceaux choisis du Pré Spirituel de Jean
Moschos, Paris 1931, 49; I. Ševčenko, "Levels of Style in Byzantine Prose", JÖB 31.1 (1981)
295–96.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 175

contrast starkly with the simpler, lower style of the main text.8 Thus, the relation-
ship between education and the use of H in writing is not always as straightfor-
ward as it might seem. It is also worth mentioning that direct speech in Byzantine
narrative texts seems to be rendered quite often by a lower linguistic form than
that used in the actual narrative.9
Now the obvious question facing the student of Byzantine style is how
do we operationalize this continuum of vertical stylistic variation (i.e. replace
it with something countable) in order to study it? This problem was tackled by
I. Ševčenko in an insightful paper published some thirty years ago.10 Ševčenko
adopted the concept of "Three Styles", which have been used in rhetorical and
stylistic theory since ancient times, and posited three main levels of style: high
style, characterized by Atticisms, linguistic classicism and periodic syntax; mid-
dle style, denoting works in a less complex style, influenced by the Scripture and
patristic writings, syntactically less periodic and tending towards parataxis; low
style, that of works with plentiful vernacular vocabulary, influenced by the New
Testament and the Psalter (i.e. the most commonly read parts of the Scripture)
and simple paratactic syntax.11
This tripartite division, simplified as it is in order to be able to capture sig-
nificant generalizations, is criticized by Wahlgren on the grounds that it presents
Byzantine literature as divisible into categories that "are stable across the centu-
ries without being influenced by each other".12 Wahlgren rightly points out that
style should not be oversimplified into "levels" as if they were the only stylistic
divisions to be made in Byzantine literature.13 It is of course to be hoped that a fu-

8 For Palladius, cf. Hult (n. 7) 28 and the discussion in I. Ševčenko, "Additional remarks to the
report on levels of style", JÖB 32.1 (1982) 214.
9 See, for example, M. Hinterberger, "How should we define vernacular literature?",
http://www.mml.cam.ac.uk/greek/grammarofmedievalgreek/unlocking/Hinterberger.pdf
(quoted 14.10.2011) 2006, 8; Horrocks (n. 4) 254; also N. Kälviäinen, "Ἀναζητώντας τὶς
ἀπαρχὲς τοῦ γραπτοῦ δημώδους λόγου: Στατιστικὰ στοιχεῖα γιὰ τὴν ἔκφραση τοῦ
μελλοντικοῦ χρόνου στὴν πρώιμη βυζαντινὴ ἁγιογραφία", in the proceedings of the confer-
ence Neograeca Medii Aevi VII (forthcoming).
10 Ševčenko (n. 7).
11 Ševčenko (n. 7) 291. Toufexis' subdivision of the higher register H into classicizing Greek
and Schriftkoine is easy to equate to Ševčenko's high and middle levels, while his L would
roughly correspond to Ševčenko's low style: see Toufexis (n. 2) 210 and 212.
12 S. Wahlgren, "Byzantine Literature and the Classical Past", in E. Bakker (ed.) A Companion
to the Ancient Greek Language, Chichester 2010, 528.
13 Wahlgren (n. 12) 529. Ševčenko himself was clearly aware of this, as he contrasted the "ver-
176 Nikolaos Kälviäinen

ture classification of Byzantine styles will be able to locate texts in a multidimen-

sional relation to others, instead of having just a single vertical axis. However,
every investigation has to start from somewhere, and Ševčenko's classification
– in part precisely because of its simplicity – is as good a starting point as any.
Ševčenko's system of levels of style is further criticized by Kazhdan as
being "primarily grammatical" and therefore failing to "make clear the link be-
tween the ideas expressed and the mode of expression"; neither does it reflect
e.g. individual styles.14 What Kazhdan has in mind is that the "style" of a literary
work consists of much more than just linguistic style: it is the result of a com-
plex interplay of language, discourse structure etc. However, as Dover points out,
it is necessary to distinguish between linguistic and non-linguistic style or we
risk being unable to cope with the complexity of the data.15 Similarly, Jeffries &
McIntyre argue that dividing linguistic style into "levels" corresponding to the
traditional categorization of the structure of language (phonology, morphology,
syntax etc.) has the advantage of enabling the researcher to concentrate on one
level at once.16 Surely this applies also to distinguishing linguistic style from
other aspects of style. If we try to bring too many factors into play at once, the
task becomes impossible.
With this in mind, we will here concentrate on a single aspect of the syn-
tactic level of linguistic style, namely the complexity of sentence structure. Ob-
viously this restricts the scope of the generalizations we will be able to make, as
we cannot automatically expect a linguistically archaizing author to write in a
syntactically complex style (or vice versa).17 Even so, syntactic complexity as
a stylistic factor should not be left unexplored, since it is arguably the easiest of
all the structural categories of language (such as morphology, syntax, vocabulary
etc.) to generalize across the board.

tical levels" with "horizontal kinds" of style such the ἰδέαι of Hermogenes' rhetorical theory
(i.e. level-internal variation conditioned by genre, text/discourse type etc.): Ševčenko (n. 7)
290.
14 A. Kazhdan, A History of Byzantine Literature (650–850), Athens 1999, 162.
15 K. Dover, The Evolution of Greek Prose Style, Oxford 1997, 3–4.
16 L. Jeffries – D. McIntyre, Stylistics, Cambridge 2010, 35.
17 Cf. Ševčenko (n. 8) 223.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 177

2. Syntactic complexity

So far the discussion has been theoretical in nature, but in order to test in prac-
tice the scientific validity of levels of style coupled with syntactic complexity as
a conceptual framework for a stylistic classification of Byzantine literature, we
need empirical data. We accordingly need to develop a method for quantitatively
analysing the syntactic complexity of Byzantine texts.18
Syntactic complexity essentially refers to "syntactic structures which ne-
cessitate increased parsing and processing effort".19 In other words, increased
syntactic complexity means structural differences that make a text harder to read.
Complexity in itself, however, is something far too complex to be directly mea-
sured: in order to make use of syntactic complexity, we will need to operational-
ize it as one or more variables that can be credibly linked to the abstract concept
of complexity.
First of all, it has been established that sentence length can be utilized as an
accurate index of syntactic complexity.20 Counting the amount of words per sen-
tence is an economical method of establishing differences in syntactic complexity

18 Although this aspect of linguistic style remains largely unexplored in Classical and Byzan-
tine philology, there are a few studies that must be cited as having inspired the present paper.
T. Webster, "A Study of Greek Sentence Construction", AJPh 62 (1941) 385–415 studies the
evolution of Classical Greek sentence structure by using syntactic variables such as sentence
length and embedding depth. Webster's study can be criticized for overlooking stylistic vari-
ation inside a given work (see Dover [n. 15] 50), but as a pioneering attempt it is noteworthy.
H. Hunger, "Stilstufen in der Geschichtsschreibung des 12. Jahrhunderts: Anna Komnene und
Michael Glykas", BSEB 5 (1978) 137–70 demonstrates in a short qualitative analysis that Anna
Comnene's syntactic structure is more complex than that of Michael Glycas in terms of both
sentence length and the use of subordinate structures. Similarly, W. de Melo, "Zur Sprache
der republikanischen carmina Latina epigraphica: Satzumfang, Satzkomplexität und Diath-
esenwahl", in P. Kruschwitz (ed.), Die metrischen Inschriften der römischen Republik, Berlin
2007, 97–120 used, among others, mean T-unit length (though without calling it a T-unit) and
the ratio of main clauses and subordinate clauses to show that the syntax of Lucretius' De
rerum natura is more complex than that of Terence's Eunuchus and the early Latin carmina
epigraphica. Finally, Dover's book (n. 15) on Classical Greek prose style is a treasure trove
of ideas and observations on the use of syntactic variables such as T-unit (which Dover calls
MCF or "main clause-finite verb unit") and the analysis of Greek sentence structure and genre-
conditioned stylistic variation.
19 B. Szmrecsányi, "On Operationalizing Syntactic Complexity", in G. Purnelle – C. Fairon –
A. Dister (eds.), Actes du colloque JADT 2004 (Journées internationales d'Analyse statistique
des Données Textuelles), Louvain-La-Neuve 2004, 1031.
20 Szmrecsányi (n. 19) 1037–38.
178 Nikolaos Kälviäinen

and is therefore suitable for large-scale research in cases where a more detailed
analysis would be impossible, or for fast and easy analysis in general.
Sentence length as a variable, however, possesses one fundamental draw-
back: the lack of an unambiguous definition of the concept "sentence". While a
definition based on punctuation is commonly used, it cannot here be considered
an optimal approach, as in most cases whatever punctuation is found in the manu-
scripts neither reflects a consistent system nor originates from the author's own
pen. Modern editors normally replace it with a more or less intuitive punctuation
(influenced by their own working language) and significant differences can on
occasion be detected in the practice of different editors.21 Thus, we cannot rely
on an author's punctuation as a guide to analysing his sentences, as is the case
with modern languages endowed with reasonably standardized writing systems.
On the other hand, a lexical definition based on identifying coordinating conjunc-
tions does not work either, as most comparable items can be used alternatively
as conjunctions or as discourse particles (cf. the "biblical" sentence-initial καί).
As an alternative Hunt proposes the concept of the minimal terminable unit
or "T-unit", defined as a single main clause plus any subordinate clauses depen-
dent on it.22 In essence, the T-unit corresponds to the sentence in all respects save
that it does away with the coordination of main clauses, which would be difficult
to define. However, despite the usefulness of the T-unit variable, structural length
does not automatically translate to structural complexity.23 That is, measuring
length probably only works as an index of syntactic complexity due to the general
probability that longer syntactic units contain more complex structures. We will
thus take a step further in the hope that more complex variables will enable us to
support the conclusions drawn from measurements of T-unit length.
It is generally accepted that, apart from length, syntactic complexity is
related to the number, type and depth of embedding or subordination in a text.24
Subordination is defined as "the nonsymmetrical relation holding between two

21 Dover (n. 15) 27. Cf. also K. Hunt, Grammatical Structures Written at Three Grade Levels,
Champaign, Illinois 1965, 8 for a similar observation, in his case concerning the revision of
schoolchildren's punctuation by English teachers.
22See Hunt (n. 21) 21. Essentially the same idea is proposed in Dover (n. 15) 28. Cf. also de
Melo (n. 18) 101–02.
23 Szmrecsányi (n. 19) 1032–33.
24 K. Beaman, "Coordination and Subordination Revisited: Syntactic Complexity in Spoken
and Written Narrative Discourse", in D. Tanner – R. Freedle (eds.), Coherence in Spoken and
Written Discourse, Norwood 1984, 45.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 179

clauses such that one is a constituent or part of (i.e. dependent upon) the other".25
In order to understand what this means, we need to consider the fundamental
structural characteristics of language. When looking at language at the phono-
logical level, we see phonological segments strung one after another in a monodi-
mensional chain. At the syntactic level, however, linguistic theory postulates a
two-dimensional model of language structure, with two different ways of com-
bining elements of the same type (such as clauses) to form larger structural units
(such as sentences or T-units): iteration and recursion.
The main difference between iteration and recursion, as formulated by
Karlsson, is that iteration "yields flat output structures, repetitive sequences on
the same depth level as the first instance" while recursion "builds structure by in-
creasing embedding depth".26 In other words, elements combined through itera-
tion are "concatenate",27 syntactically equal, while elements combined through
recursion form a hierarchic structure28 in which some elements are subordinate
to others of the same type (i.e. embedded in them). Such hierarchic structures are
considered more complex than flat iterative structures: to quote one definition of
complexity, "increased complexity is, at its most general level, increased hierar-
chic organization; that is, an increase in the number of hierarchic levels within a
system".29
In sentence structure, where structural units called sentences are formed
from smaller structural units called clauses, iteration manifests itself as coordina-
tion, where syntactically equal clauses are combined, often by means of conjunc-
tions such as καί. The result of recursive combination of clauses, on the other
hand, is known as subordination, where a clause is syntactically dependent on
another; this relationship is in turn marked by conjunctions such as ὅτι.
Now the contrast between iteration/coordination and recursion/subordina-
tion is relevant to the issue of syntactic complexity precisely because hierarchical
structures are considered more complex than flat iterative structures. A funda-
mental difference can be discerned between the two in actual language use as
well: whereas iteration appears practically unconstrained (i.e. there are no theo-

25 Beaman (n. 24) 55.

26F. Karlsson, "Syntactic recursion and iteration", in H. van der Hulst (ed.), Recursion and
Human Language, Berlin 2010, 45.
27 Karlsson (n. 26) 46.
28 Karlsson (n. 26) 46.
29 T. Givón, The Genesis of Syntactic Complexity: Diachrony, ontogeny, neuro-cognition, evo-
lution, Amsterdam 2009, 4.
180 Nikolaos Kälviäinen

retical limits to how many syntactically equal clauses can be coordinated into a
single sentence) and successive application of iterative clause combination can
very well result in unusually long sentences,30 the actually occurring use of re-
cursion to create structural complexity has been found much less impressive in
comparison.31
Thus, according to Karlsson, even extremely long and complex sentences
seem to owe their size and complexity mostly to iteration rather than recursion:
it appears the greatest clausal embedding depth32 (or one of the greatest in any
case) achieved in Karlsson's example sentence from James Joyce (12,931 words)
is 6 levels of depth, while the number of coordinated clauses certainly runs in the
hundreds, if not thousands.33 Similarly, empirical data from the study of several
European languages shows that the distribution of subordinate clauses at differ-
ent levels of embedding depth follows a falling curve, with an added level of
embedding corresponding to a drop in the frequency of clauses occurring at that
level.34 Furthermore, it seems that more complex styles (e.g. legal language) tend
to have more clauses at deeper levels of embedding than syntactically simpler
varieties, an observation directly related to the issue at hand.35
Thus far, then, we have seen that iteration/coordination and recursion/sub-
ordination as strategies of clause combination differ greatly in actual usage, with
recursion heavily limited and iteration lacking similar constraints. What is the
reason behind this discrepancy? The answer is assumed to lie in the relatively
greater cognitive processing difficulty of more complex, "deep" hierarchical re-
cursive structures in comparison with less complex, "flat" iterative structures.
In other words, it is considered that increasing recursive combining of elements
30 Karlsson (n. 26) 46.
31 For an in-depth theoretical discussion, see Karlsson (n. 26) 50–65; cf. also R. Laury – T.
Ono, "Recursion in conversation: What speakers of Finnish and Japanese know how to do", in
H. van der Hulst (ed.), Recursion and Human Language, Berlin 2010, 69–70 and 84–85.
32 The term refers the structural depth achieved through successive application of recursive
cycles. The embedding depth of a subordinate clause is the "vertical" relation of the level of the
hierarchic structure it occupies to that of the main clause it is dependent upon (either directly or
indirectly, through other embedded clauses). The level of embedding occupied by a clause is 1
+ the level of embedding of its superordinate clause, with all main clauses assigned level 0. In
other words, subordinate clauses directly dependent on a main clause are assigned level 1, the
subordinate clauses embedded in these level 2 and so on.
33 Karlsson (n. 26) 46–47.
34 See F. Karlsson, "Multiple final embedding of clauses", International Journal of Corpus
Linguistics 15.1 (2010) 95–99.
35 See Karlsson (n. 34) 96–97.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 181

leads precisely to the kinds of "syntactic structures which necessitate increased

parsing and processing effort" that were mentioned above. There is a large body
of psycholinguistic research that indicates that increasing subordination and es-
pecially increasing embedding depth may strain the ability of the human brain to
process language structure.36
What all this means for us is that clausal subordination is a prime candidate
for a successful index of syntactic complexity in texts, on a par with T-unit length.
Thus, the ratio of main and subordinate clauses in a text can be used as a simple
and easily presentable index of complexity.37 Furthermore, in order to account
for the complex hierarchic structures produced by recursive clause combining,
we can measure the distribution of clauses at different levels of embedding depth.
Having explored the theoretical background of syntactic complexity, we
may now comment on our expectations as to its relevance to the stylistic varia-
tion observed in Byzantine Greek. Our previous experiences with texts of varying
registers as well as common logic suggest that greater complexity will be found
in those texts that belong to higher registers. This would also make sense insofar
as learning to write in a classicizing style was a symbol of status in Byzantine
literary society: the more complex the style and the greater the efforts required to
master it, the greater the prestige associated with it in the elite literary circles.38
Since the liberal use of embedding and other complexity-increasing strategies
can be assumed to contribute heavily towards a style that is progressively more
difficult to understand, we can hypothesize that it will accordingly augment the
social standing of those who do manage to cope with it.

3. Methodological considerations

The broad research questions emerging from the preceding discussion – ones that
this study alone cannot attempt to answer – are a) whether the complexity of a

36 See Szmrecsányi (n. 19) 1033–34; Karlsson (n. 34) 101–02; Laury – Ono (n. 31) 79. How-
ever, see also Givón (n. 29) 12–13 for possible difficulties involved in this interpretation. For
the sake of the argument we adopt in this study the view that assumes the existence of a real
psychological correlation between syntactic complexity and cognitive complexity, the latter
being understood here to imply increasing mental processing difficulty.
37 In measuring the performance of different indices of syntactic complexity, Hunt found this
simple subordination ratio less efficient than T-unit length but better than sentence length: Hunt
(n. 21) 23.
38 Ševčenko (n. 7) 304.
182 Nikolaos Kälviäinen

text's sentence structure consistently mirrors its level of style throughout Byz-
antine literature and b) whether this relation can be effectively utilized in order
to create an objective system of classifying Byzantine texts stylistically, i.e. by
using syntactic complexity as an index of the levels of style. These questions can
obviously only be answered when data from the analysis of an extensive corpus
of texts is made available. The contribution of the present study is to demonstrate
empirically that there exists a relation between the stylistic level and syntactic
complexity and that this relation can be quantitatively measured. This aim is
achieved by analysing a small test corpus, the data of which cannot yet be gener-
alized across the board but which serves to indicate why this line of research is,
in the author’s opinion, worth pursuing.
In order to validate theoretically our basic hypothesis for the test corpus
- according to which texts perceived as belonging to higher levels of style mani-
fest correspondingly increasing syntactic complexity - we need to demonstrate
that a systematic correlation exists between two variables: "level of style" and
"syntactic complexity". Both variables, however, are far too abstract to study in
themselves, so they must somehow be operationalized as measurable categorical
variables. In this study, the variable "level of style" is represented by three text
samples (see section 4 below), each assigned to a certain level of style on the ba-
sis of a combination of personal observation as well as judgments and comments
expressed in the scholarly literature.
The variable "syntactic complexity", in turn, is operationalized as the fol-
lowing 5 different syntactic variables involving the length of syntactical units or
the extent of recursive clause combining (subordination/embedding), both con-
cepts theoretically linked to the complexity of syntax, as seen in the discussion
above. Each variable is examined as an independent index of syntactic complex-
ity and the overall image presented by the results yielded by the different vari-
ables is then considered the "final verdict". Variables 1–2 and 4–5 are analysed as
categorical (i.e. not continuous but with a limited number of possible outcomes)
and the categories of each are given here. Due to the fact that the statistical χ2-test
(see below) requires that at least 5 tokens be found in the data for each category
of each variable, the categories have been adjusted after carrying out the analysis
to comply with this requirement.39

39Cf. A. Hakulinen – F. Karlsson – M. Vilkuna, Suomen tekstilauseiden piirteitä: kvantitatiivi­

nen tutkimus, Helsinki 1996, 99.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 183

1. T-unit length (measured in words). T-units are grouped ac-

cording to the amount of words they contain into the fol-
lowing 6 categories: 1–5 words, 6–10 words, 11–15 words,
16–20 words, 21–25 words and 26+ words.
2. T-unit length (measured in clauses). T-units are grouped ac-
cording to the amount of clauses they contain into the fol-
lowing 5 categories: 1 clause, 2 clauses, 3 clauses, 4 clauses
and 5+ clauses.
3. Mean T-unit and clause length. These are simple average val-
ues obtained by dividing the number of words in the sample
by the number of T-units/clauses and the number of clauses
by the number of T-units.
4. Main clauses and subordinate clauses. A simple measure-
ment of the total ratio of main to subordinate clauses in the
sample, with 2 categories.
5. Embedding depth. Measuring the distribution of all subordi-
nate clauses at different levels of embedding, with 5 catego-
ries (levels 1–5) since no clauses at level 6 or deeper were
found in the data (but see n. 73).

Once the raw data has been collected, it is subjected to a statistical χ­­2-test or chi-
square test in order to a) establish the significance of the results and b) evaluate
the extent to which each of the three texts differs from the two others.40 Explain-
ing the basics of inferential statistics is beyond the scope of this article, and the
reader should turn to an exposition of the subject in a basic manual of statistics
for linguists.41
However, in a nutshell, the probability value or p-value obtained by means
of the test is a gauge of the statistical probability that the performance of variable
X (one of our complexity variables) is not conditioned by variable A/B (the "level

40 I have used Preacher's online χ2 calculation tool to compute the values: see K. Preacher, Cal-
culation for the chi-square test: An interactive calculation tool for chi-square tests of goodness
of fit and independence, http://quantpsy.org (quoted 13–18.8.2012), 2001.
41 E.g., E. Levon, "Organizing and Processing Your Data: The Nuts and Bolts of Quantitative
Analyses", in L. Litosseliti (ed.), Research Methods in Linguistics, London 2010, 68–92. See
also Dover (n. 15) 47 for another demonstration of how to use the χ2-test in practice to establish
differences or similarities between Ancient Greek texts composed in different styles.
184 Nikolaos Kälviäinen

of style" variable, i.e. the different texts being compared). For example, a p-value
of p = 0.75 would mean there is a 75% chance that texts A and B are so similar in
their use of the variable X that any observable difference is due purely to chance.
Where to draw the line in these probabilities is a matter of convention; according
to Levon, in the humanities the standard cut-off point is p = 0.05.42 This means
that a p-value of 0.05 or smaller (5% or less chance that A and B do not differ with
respect to X) is conventionally considered proof enough that A and B differ sig-
nificantly from each other with respect to variable X, which is formally expressed
by saying that the difference between A and B with respect to X is statistically
significant at the p < 0.05 level.
Conversely, a p-value greater than 0.05 (over 5% chance) is taken to mean
that the findings are not statistically significant, since pure chance cannot be ruled
out as being responsible for the differences between A and B, and the hypothesis
of A and B belonging to different levels of style has to be rejected with respect
to the use they make of variable X. A and B would thus be considered in our
theoretical model to belong to the exact same style as far as syntactic complexity
is concerned and as far as syntactic complexity is accurately represented by X.
The meaning of statistical significance to the linguist or philologist is that hav-
ing statistically significant findings provides "predictive power to the descriptive
facts"43 and gives a powerful tool for objectively evaluating the differences be-
tween texts.
Apart from the p-value, there are two other values involved in the χ2-test,
the χ2-value and the degrees of freedom (abbreviated df). These are also reported,
although the p-value in itself reveals all that is essential. However, since a pro-
gressively higher χ2-value corresponds to a progressively lower p-value as far as
the df remain the same, the χ2-values can also be compared where the comparison
of extremely small p-values is impractical. (Note that χ2-values, unlike p-values,
cannot be compared if the corresponding df differ!) In short, when comparing
two texts with respect to a single syntactic variable, a high χ2-value and a low
p-value indicate a large difference between the texts, whereas a low χ2-value and
a high p-value indicate a smaller difference (even if that difference may still be
statistically significant if the p-value is low enough).
Having identified the variables to be measured, the question remains how
to analyse a given text in such a manner that the analysis is unconditionally re-
producible and falsifiable. As we have already seen in the case of the concept

42 Levon (n. 41) 81.

43 Levon (n. 41) 71.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 185

of "sentence", even basic linguistic terms such as word, clause and sentence are
ambiguous categories and by no means immediately applicable to any text. In the
Master's thesis on which this paper is based, problems related to the definition in
Ancient and Byzantine Greek of concepts such as word, sentence, clause, subor-
dination and coordination were explored in some depth. However, due to limita-
tions of space, what is presented here is only a summary of the most important
(i.e. most likely to impact on the statistics) coding decisions made in the analysis
of our test corpus, leaving out most of the theoretical discussion.

1. Word. In our analysis, word is defined as an orthographic

word (i.e. a sequence of graphemes separated from others by
empty space in the text of the critical edition, including all
clitics, particles, articles etc.), with the exception that words
that have undergone crasis (such as κἀγώ) are counted sepa-
rately.44
2. Clause. We define clause as a syntactic unit consisting of a
single predicate and any elements syntactically dependent on
it, including an initial coordinating or subordinating conjunc-
tion or relative pronoun/adverb, but excepting other clauses.
2.1 Nonverbal predicates (e.g. δεινὸς γὰρ οἶνος [Eur. Cyc. 678])
are treated as normal clauses, i.e. exactly as if the copula
verb εἰμί were overtly present.45
2.2 Elliptic clauses where the predicate has been omitted are
treated as normal clauses, exactly as if the predicate were
present. In cases where coordinate structures make it diffi-
cult to determine exactly how many elliptic clauses are in-
volved, the following rules are applied.46
2.2.1. If the half of the coordinate construction that seems to lack
a predicate verb consists of a single constituent,47 no ellipsis

44 For further discussion cf. Dover (n. 15) 26ff. and de Melo (n. 18) 100–02.
45 Cf. Dover (n. 15) 29.
46Cf. Hakulinen – Karlsson – Vilkuna (n. 39) 10–11 and M. Haspelmath, "Coordination", in T.
Shopen (ed.), Language Typology and Syntactic Description. Vol. II: Complex Constructions,
Cambridge 20072, 4–5 and 37–44.
47 For instructions on how to analyse a clause in terms of its constituent structure, see, e.g., M.
Tallerman, Understanding Syntax, London 1998, 116–25 or R. Van Valin, An Introduction to
186 Nikolaos Kälviäinen

is posited. For example, ἔφαγε τὸν ἄρτον καὶ [τὸ μῆλον]

counts as one clause (the coordination only involves the
noun phrases and not the verb).
2.2.2. If the half of the coordinate construction that seems to lack a
predicate verb consists of two or more separate constituents
with differing syntactic functions, ellipsis of the predicate is
assumed. For example, ὁ Ἰωάννης ἔφαγε τὸν ἄρτον καὶ [ὁ
Κωνσταντῖνος] [τὸ μῆλον] counts as two clauses.
2.3 Monoclausal verb phrases consisting of two or more coordi-
nated verbs are accepted only in cases where the sole constit-
uents being coordinated are the bare, unmodified verbs. For
example, ὁμοθυμαδὸν [καὶ ἔπραττον καὶ ἐστρατεύοντο]
(Xen. Hell. 7,1,22) counts as one clause, but [ὁ πατὴρ αἰεὶ
λέγει] καὶ [σὺ φῂς] καὶ [οἱ ἄλλοι δὲ πάντες ὁμολογοῦσιν]
(Xen. Cyr. 3,3,19) counts as three.48
2.4 Periphrastic constructions such as ἔχω + infinitive or εἰμί +
participle are treated as monoclausal.
2.5 Non-finite clauses (i.e. all infinitival and participial construc-
tions) count as fully fledged clauses. This approach, while
theoretically debatable, is here considered justified on the
grounds that such constructions exhibit verbal syntax on a
par with finite verbs (e.g. nominative or accusative subjects,
accusative direct objects) that is not shared with comparable
non-verbal action nouns (which take at best a genitive object
or subject).49 Note that nouns and adjectives historically de-

Syntax, Cambridge 2001, 110–15 or any other up-to-date introduction to syntax.

48 Cf. Hunt (n. 21) 13–14 and Dover (n. 15) 31 and Hakulinen – Karlsson – Vilkuna (n. 39)
11. Hunt generally dislikes the idea of coordinated verbs counting as separate clauses; Dover is
in theory willing to accept both possibilities; Hakulinen – Karlsson – Vilkuna adopt a position
diametrically opposite that of Hunt and count all overtly expressed verbs as separate clauses.
Our solution is intended as a compromise.
49 For the clausal status of non-finite constructions in general cf., e.g., D. Cheila-Markopoulou,
"Προτασιακότητα και μετοχικές δομές στην Αλεξανδρινή και Μεσαιωνική Ελληνική", in
D. Theofanopoulou-Kontou (ed.), Σύγχρονες τάσεις στην ελληνική γλωσσολογία. Μελέτες
αφιερωμένες στην Ειρήνη Φιλιππάκη-Warburton, Athens 2003, 129; S. Kemmer, "Clause:
Overview", in W. Frawley (ed.) International Encyclopedia of Linguistics I, Oxford 2003,
320; Hakulinen, Karlsson & Vilkuna (n. 39) 12. For the "internal" verbal syntax of non-finite
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 187

rived from participles (e.g. ὁ ἡγούμενος in the sense "hegu-

men") are of course not counted as clauses, as they are (syn-
chronically) not verbal at all.
2.6. Pragmaticalized exclamations such as Κύριε ἐλέησον, ἰδού
or Ποίησον ἀγάπην ("Please"), functioning as discourse
markers rather than true verbal predicates, are not treated as
clauses unless they occur on their own.50
2.7 Grammaticalized lexical items such as ἄφες used as a sub-
ordinating conjunction (cf. Modern Greek ας) in ἄφες ἴδω
τί ἐστιν ὃ ἔχεις (Mir.Artem. 1) are not counted as separate
clauses.
2.8 Parenthetical clauses occurring inside another T-unit are
counted as embedded in it due to the added complexity they
bring to the structure.51
2.9. Parenthetical expressions of saying such as φησίν and ἦ
δ' ὅς, whose function is simply that of markers of reported
speech, are not counted as clauses.52
3. Subordination and coordination. In practice, distinguishing
between subordination and coordination in Greek is not so
difficult as to have an impact on the statistics.53 Some prob-
lematic cases are presented below.
3.1. It is not always clear where to draw the line between rela-
tive clauses and independent sentences introduced by what
is morphologically a relative pronoun or adverb. For the sake
of consistency, in carrying out our analysis ὅθεν, διό etc. as
well as digressive relative clauses54 were always treated as

constructions, cf. M. Haspelmath, Understanding Morphology, London 2002, 230–32 and for
Greek substantivally used (articular) infinitives see A. Rijksbaron, The Syntax and Semantics
of the Verb in Classical Greek. An Introduction, Chicago 20023, 112–14.
50 Cf. De Melo (n. 18) 103 and Laury – Ono (n. 31) 71–74.
51 See Dover (n. 15) 29.
52 See Dover (n. 15) 29.
53 For a detailed account of Classical Greek finite subordinate clauses, see e.g. Rijksbaron (n.
49) 49–94.
54 Dover suggests that digressive relative clauses such as ἐντὸς Ἅλυος ποταμοῦ, ὃς ῥέων ...
ἐξίει should in principle be treated as separate T-units whenever possible: Dover (n. 15) 29. I
188 Nikolaos Kälviäinen

introducing subordinate clauses, though in hindsight this ap-

proach should probably be subjected to criticism.
3.2. In order to avoid unrealistically long T-units, reported speech
is separated from the verb of saying introducing it by posit-
ing an additional T-unit boundary after the verb of saying and
treating the clauses introduced by it as main clauses (with the
exception of the first infinitive in indirect reported speech).55

4. The test corpus

The material analysed in this study consists of three samples, each taken from
an Early Byzantine hagiographical work: the Miracles of Ss. Cyrus and John of
Sophronius of Jerusalem (Mir. Cyr. & Jo.), the Pratum Spirituale of John Moschus
(Prat.) and the Miracles of St. Artemius (Mir. Artem.) of unknown authorship.
The texts have been chosen with a view to minimizing the impact of fac-
tors other than register or level of style, i.e. Ševčenko's "kinds of style" such as
genre or text/discourse type (e.g. the distinction of narrative and descriptive com-
positional elements56). Each work is dated to the same period (the late 6th to 7th
centuries) and belongs to one of two hagiographical subgenres with a reasonably
similar internal structure. Mir.Cyr.&Jo. and Mir.Artem. represent the genre of Mi-
racula/Θαύματα, independent collections of short episodes detailing the saintly
hero's miraculous exploits, while Prat. is a collection of short edifying stories,
anecdotes and sayings that illustrate the spiritual prowess of the desert-dwelling
ascetics of the Near East.
In order to use this corpus as an index for the variable "levels of style",
each of the three texts in the following description is tentatively assigned to a
certain level of style on the basis of lexical and content-related criteria as well as
comments and judgments made by other scholars: Mir.Cyr.&Jo. to high style and
Prat. and Mir.Artem. to low style.
Sophronius, patriarch of Jerusalem (ca. 560–638) is known to have been
a teacher of rhetoric before assuming monastic garb, and went on to author both

do not see why they should be treated differently from other relative clauses, however.
55 See Dover (n. 15) 29–30.
56 Cf. Kazhdan (n. 14) 159–60.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 189

rhetorical homilies and hymns in the classical Anacreontic metre.57 It is hardly

surprising, then, that the Miracles of Ss. Cyrus and John is the work that repre-
sents high style in our corpus. Contrasted with the other two texts, its rhetorical
nature is thrown into sharp relief: Mir.Cyr.&Jo. uses a classicizing vocabulary
and morphology, abounds in learned, epideictic digressions, quotes Homer (30,2)
as well as the Scripture, mentions various pagan classics (e.g. 30,4; 64,3; 70,8–9)
and seemingly prefers to report speech indirectly rather than by means of direct
quotations from its protagonists, although these too make the occasional appear-
ance.
Τhe monk John Moschus (ca. 540–619/634), mentor and travelling com-
panion of our Sophronius,58 was certainly a learned man.59 Nevertheless, his ex-
tant work, the Pratum Spirituale, is written in a simple, unaffected language.
It has justly been characterized an important source for the development of the
spoken Greek language of the period.60 The heavily rhetorical style of Sophro-
nius is absent and a distinct vernacular element is present in both morphology and
vocabulary, often in the form of recent loans from Latin. Horrocks characterizes
Moschus' language as "the simple narrative style of educated speakers of the peri-
od, modelled on that of the only vernacular-based literary tradition and including
features that were almost certainly in decline in the popular speech of Móschos'
own time".61 Although by no means consistently vernacular in the manner of
much of the popular literature of the Late Byzantine period, Prat. is doubtless a
good candidate for Ševčenko's low style.
Our third text, the anonymous Miracles of St. Artemius, can be dated ap-
proximately to the years 656–668.62 There can hardly be any question that the
57 See ODB s.v. "Sophronios", 1928; Th. Detorakis, Εισαγωγή στη σπουδή των αγιολογικών
κειμένων, Rethymno 1985, 38; H. Chadwick, "John Moschus and his Friend Sophronius the
Sophist", JThS 25 (1974) 53.
58 ODB s.v. "Moschos, John", 1415; Chadwick (n. 57) 49–56 and 59.
59 Cf. Ševčenko (n. 7) 295–296; Chadwick (n. 57) 50–51; R. Maisano, "Tradizione orale e
sviluppi narrativi nel 'Prato' di Giovanni Mosco", in C. Giuffrida – M. Mazza (eds.), Le tra-
sformazioni della cultura nella tarda antichità. Atti del convegno tenuto a Catania, Università
degli studi, 27 sett.–2 ott. 1982, Rome 1985, 665–70. Also, as is the case in each of our three
texts, the proem is heavily rhetorical in comparison to the rest of the work.
60 Detorakis (n. 57) 38.
61 Horrocks (n. 4) 255.
62 A. Papadopoulos-Kerameus, Varia Graeca Sacra. Сборникъ греческихъ неизданныхъ
­ огословскихъ текстовъ IV–XV вековъ (Записки Историко-Филологическаго Факультета
б
Императорскаго С.-Петербургскаго Университета 95), St. Petersburg 1909, ii; Kazhdan (n.
190 Nikolaos Kälviäinen

Miracles of St. Artemius belong to the lowest stylistic level in Early Byzantine
hagiography. Kazhdan and Papadopoulos-Kerameus describe the use of vernacu-
lar vocabulary as conspicuous, the expression as largely brief, simple and inartis-
tic (with little use of rhetorical figures) and the syntax as clear.63 Papadopoulos-
Kerameus even goes as far as to suggest that, on occasion, the text is so natural
and inartistic that it may approach the speech of contemporary Greeks.64 The
numerous cases of direct speech are indeed rendered in what might be thought of
as a credible approximation of a "written spoken language" (the question of the
fundamental differences between writing and actual speech is, of course, another
matter).
As far as "vertical" levels of style are concerned we expect Mir.Cyr.&Jo.
to differ greatly from the two others in terms of syntactic complexity. At the
same time we expect Prat. and Mir.Artem. to display a very similar (if probably
not identical) use of syntax, since according to our hypothesis they belong to
the same level of style. In terms of "horizontal" kinds of style, we can probably
assume that the difference in content between the hagiographical subgenres of
Miracula and Apophthegmata Patrum does not in itself correspond to differences
in language use, since the two are structurally very similar.
Nevertheless, even if genre, in the broad sense of "the genre of a literary
work as a whole", has been accounted for, any lengthy piece of writing may be
assumed to contain sections that differ from each other in terms of style.65 For
the purposes of this work it has not been possible to subject the selected samples
to such rigorous control as to eliminate this factor entirely. It is simply hoped
that, even if some compositionally heterogeneous material is inevitably included
in the samples, the overall similarity in structure of the three texts will dampen
their impact. Ultimately, it is difficult to ascertain whether possible differences
between the texts are due to genre, personal stylistic choices or other factors (for
example, it may be that Mir.Cyr.&Jo. contains a disproportionate amount of de-
scriptive – as opposed to narrative – elements in the form of learned digressions
such as rhetorical ἐκφράσεις.66

14) 27.
63 Kazhdan (n. 14) 34; Papadopoulos-Kerameus (n. 62) i–ii.
64 Papadopoulos-Kerameus (n. 62) ii.
65 Cf. Dover (n. 15) 46–56.
66 In many places Sophronius adopts the tone of rhetorical ἐγκώμια, as has also been pointed
out in N. Fernandez Marcos, Los Thaumata de Sofronio: Contribución al estudio de la incuba-
ción cristiana (Manuales y anejos de Emérita 31), Madrid 1975, 154.
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 191

However, a result which pairs Prat. and Mir.Artem. as relatively similar

and clearly less complex than Mir.Cyr.&Jo. will corroborate our hypothesis and
suggest this line of research should be pursued further. For the purposes of this
pilot study, the criteria outlined in this section will suffice to enable our results to
demonstrate the fundamental validity of our method.
The next questions concern the size of the samples to be collected and the
method of their collection. In their study on Finnish sentence structure, Karlsson,
Hakulinen & Vilkuna compared samples of different sizes with a view to dis-
covering how much data is required for projects investigating syntactical struc-
ture. They came to the conclusion that for syntactic variables with less than 10
categories a sample containing a few hundred clauses should suffice to provide
a reliable image of the text as a whole, provided of course that the text is inter-
nally reasonably homogeneous; even the differences between their two smallest
test samples (100 and 300 clauses) were not statistically significant for most of
the variables examined.67 In order to be on the safe side, we have aimed at 500+
clauses, selecting samples of approximately 2500 words (the first 2500 words
plus the rest of the T-unit which would otherwise be cut in half). Furthermore,
as an added security measure we have performed a "split-half test" on each of
the samples, analysing the first half of each sample and performing χ2-tests to
compare each of these halves with its parent text. No statistically significant dif-
ferences were detected, with the lowest p-value being 0.45 for embedding depth
in Mir.Cyr.&Jo. Our samples are thus clearly of sufficient size.
One final question concerns the method of collecting the samples. In this
study random sampling of T-units has not been used: the samples are continuous
chunks of text. This is generally not considered a statistically valid approach,
especially if the text is not homogeneous, but Hakulinen, Karlsson & Vilkuna
argue that in studies involving syntax it is perhaps better to have a sample that
works as a coherent text by itself than a random collection of unrelated syntactic
elements.68
The samples have been selected as follows, leaving out the short proem of
each work, since these are composed in a higher style than the main text:69
67 Hakulinen – Karlsson – Vilkuna (n. 39) 99–100 and 102–104.
68 Hakulinen – Karlsson – Vilkuna (n. 39) 101.
69 The samples have been obtained in digital form from TLG (http://www.tlg.uci.edu, quoted
29.7.2012). The texts are those published in Fernandez Marcos (n. 66) 241–402 (Mir. Cyr. &
Jo.), Papadopoulos-Kerameus (n. 62) 1–75 (Mir. Artem.) and J.-P. Migne, "Beati Ioannis Eu-
cratae liber qui inscribitur Pratum", in J.-P. Migne (ed.) Patrologiae cursus completus, Series
Graeca 87c, Paris 1864, 2852–3112 (Prat.).
192 Nikolaos Kälviäinen

Mir.Cyr.&Jo. = Miracula 1,2–14 + miracula 2–7 + the first T-unit of

miraculum 8 (2500 words)

Prat. = Capita 1–19 + the first 3 T-units of caput 20 (2505 words)

Mir.Artem. = Miracula 1–11 + the first 9 T-units of miraculum 12

(2504 words)

5. Presentation and analysis of the data

5.1. T-unit length in words

Mir.Cyr.&Jo. Prat. Mir.Artem.

T-unit length/words
(114 T-units) (270 T-units) (223 T-units)
1–5 8 (7.0%) 84 (31.1%) 64 (28.7%)
6–10 21 (18.4%) 105 (38.9%) 72 (32.3%)
11–15 21 (18.4%) 46 (17.0%) 35 (15.7%)
16–20 17 (14.9%) 16 (5.9%) 22 (9.9%)
21–25 12 (10.5%) 11 (4.1%) 16 (7.2%)
26+ 35 (30.7%) 8 (3.0%) 14 (6.3%)

Fig. 5.1: T-unit length in words (% of total T-units / words per T-unit)
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 193

Mir.Cyr.&Jo. Mir.Cyr.&Jo. Prat.

χ2-test(df = 5)
vs. Prat. vs. Mir.Artem. vs. Mir.Artem.
χ2 = 97.925 p < 0.01 χ2 = 55.82 p < 0.01 χ2 = 9.464 p = 0.09

Table 5.1.2.

Our first variable, T-unit length in words, offers tangible support to our hypothe-
sis. Prat. and Mir.Artem. show closely parallel distributions, while the high-level
Mir.Cyr.&Jo. differs from them dramatically on two separate counts. First, while
the low-level texts show a peak at T-units of 6–10 words, followed by a falling
curve, in Mir.Cyr.&Jo. the peak is shared with T-units of 11–15 words and the fall
that comes after is much less steep.
Secondly, unlike the low-level texts, Mir.Cyr.&Jo. shows a second peak at
T-units of 26 or more words, betraying a significant difference in the way T-units
of different lengths are handled by the two groups of writers: Prat. and Mir.Ar-
tem. prefer small T-units of around 1–10 words and much more rarely use longer
ones, whereas Mir.Cyr.&Jo. seems to strive for a balanced variation of long and
short T-units, with its short ones longer than the short ones of Prat. and Mir.Ar-
tem. This kind of balanced variation in Mir.Cyr.&Jo. is exactly what we would
expect from a sophist trained in the conventions of classical literature, while the
preponderance of short T-units in Prat. and Mir. Artem. reflects their paratactic
"καί-style".
Thirdly, there do seem to be some minor differences between Prat. and Mir.
Artem. in the form of slightly greater T-unit length in Mir.Artem. (note that the χ2
value of Mir.Cyr.&Jo. vs. Mir.Artem. is much lower than that of Mir.Cyr.&Jo. vs.
Prat., although the p-value is in both cases too small to conveniently reveal this
difference). The χ2-test reveals these differences are not statistically significant
at the 0.05 level, but it is a close call, with a p-value of only 0.09. In contrast,
the p-values obtained in comparing either low-level text with Mir.Cyr.&Jo. are
vanishingly small.

5.2. T-unit length in clauses

T-unit length/ Mir.Cyr.&Jo. Prat. Mir.Artem.

clauses (114 T-units) (270 T-units) (223 T-units)
1 20 (17.5%) 113 (41.9%) 86 (38.6%)
2 23 (20.2%) 94 (34.8%) 58 (26.0%)
194 Nikolaos Kälviäinen

3 13 (11.4%) 34 (12.6%) 30 (13.5%)

4 14 (12.3%) 20 (7.4%) 19 (8.5%)
5+ 44 (38.6%) 9 (3.3%) 30 (13.5%)

Table 5.2.1.

Fig. 5.2: T-unit length in clauses (% of total T-units / clauses per T-unit).

χ2-test Mir.Cyr.&Jo. Mir.Cyr.&Jo. Prat.

(df = 4) vs. Prat. vs. Mir.Artem. vs. Mir.Artem.
χ2 = 93.772 p < 0.01 χ2 = 34.722 p < 0.01 χ2 = 19.469 p < 0.01

Table 5.2.2.

Our second variable yields results that closely parallel those of T-unit length in
words as far as differentiating Mir.Cyr.&Jo. from the other two texts is concerned.
The basic pattern is the same. However, what is interesting is that Prat. and Mir.
Artem. appear slightly less similar this time – although the overall shape of their
curves is again parallel, this time Mir.Artem. has a slight peak at 5+ clauses, while
Prat. clearly relies slightly more on very short T-units of 1–2 clauses than Mir.
Artem. does. Furthermore, this time the χ2-test reveals statistically significant dif-
Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 195

ferences between all three authors, and although the difference between Prat. and
Mir.Artem. is the smallest one, it is far from negligible.

5.3. Mean T-unit and clause length

Mean T-unit & clause length Mir.Cyr.&Jo. Prat. Mir.Artem.

words 2500 2505 2504
T-units 114 270 223
clauses 504 540 553
mean T-unit length in words 21.9 9.3 11.2
mean T-unit length in clauses 4.4 2.0 2.5
mean clause length in words 5.0 4.6 4.5

Table 5.3.

Average T-unit length in Mir.Cyr.&Jo. is nearly twice that of Mir.Artem. and over
twice that of Prat., the difference between which is a small one of 1.9 words
only. Mean T-unit length in clauses continues the same story, as does mean clause
length in words, although here there is only a relatively small difference of around
0.5 words between Mir.Cyr.&Jo. on the one hand and Prat. and Mir.Artem. on the
other (these being practically identical). In order to interpret this result, further
study is required to explore to what extent clause length generally varies in Byz-
antine Greek prose.

5.4. Ratio of main to subordinate clauses

Subordination ratio Mir.Cyr.&Jo. Prat. Mir.Artem.

clauses 504 540 553
main clauses 114 270 223
subordinate clauses 390 270 330
ratio of main
22.6% : 77.4% 50.0% : 50.0% 40.3% : 59.7%
to subordinate clauses

Table 5.4.1.
196 Nikolaos Kälviäinen

χ2-test Mir.Cyr.&Jo. Mir.Cyr.&Jo. Prat.

(df = 1) vs. Prat. vs. Mir.Artem. vs. Mir.Artem.
2 2 2
χ = 84.052 p < 0.01 χ = 38.065 p < 0.01 χ = 10.328 p < 0.01

Table 5.4.2.

The simple subordination ratio reveals that Mir.Cyr.&Jo., the high style text, has
approximately three quarters of all his clauses embedded, whereas a full half of
the clauses in Prat. are main clauses. Mir.Artem. is now situated in the middle
even more clearly than was the case with T-unit length in clauses; clear differ-
ences seem to be emerging between the two low-style texts.

5.5. Embedding depth

Embedding Mir.Cyr.&Jo. Prat. Mir.Artem.

depth (390 sub.clauses) (270 sub.clauses) (330 sub.clauses)
1 202 (51.8%) 233 (86.3%) 245 (74.2%)
2 121 (31.0%) 34 (12.6%) 67 (20.3%)
3 54 (13.8%) 2 (0.7%) 16 (4.8%)
4 12 (3.1%) 1 (0.4%) 2 (0.6%)
5 1 (0.3%)

Table 5.5.1.

Fig. 5.5: Embedding depth (% of subordinate clauses / level of depth)

Levels of Style in Byzantine Greek and the Role of Syntactic Complexity 197

χ2-test70 Mir.Cyr.&Jo. Mir.Cyr.&Jo. Prat.

(df = 1) vs. Prat. vs. Mir.Artem. vs. Mir.Artem.
χ2 = 84.523 p < 0.01 χ2 = 38.263 p < 0.01 χ2 = 13.32 p < 0.01

Table 5.5.2.

Embedding depth presents our authors in no different light than the previous vari-
ables. Mir. Cyr. & Jo. is clearly our most complex text, having significantly more
clauses embedded at levels 2 and 3 than either of the others. As with T-unit length
in clauses and the basic subordination ratio, these again differ among themselves
as well, with Mir.Artem. continuing to display greater complexity than Prat. The
difference between Mir.Artem. and Prat. is again statistically significant.
Our data for embedding depth are difficult to compare directly with those
of previous studies involving modern languages since in many cases only finite
clauses are counted. Nevertheless, generally speaking our texts seem to conform
to the picture emerging from previous research, with embedding beyond level 4
being extremely rare and level 4 itself relatively uncommon.71 For our purposes,
then, the critical zone consists primarily of the levels 1–3: it is here that embed-
ding depth bears out the general impression derived from studying our previous
variables, as we can appreciate both the yawning chasm between Mir.Cyr.&Jo.
and Prat. as well as the smaller but nevertheless distinct difference between Prat.
and Mir.Artem. At level 4 the difference between Prat. and Mir.Artem. vanishes
while Mir.Cyr.&Jo. remains distinct, and by the time level 5 is reached all distinc-
tions have by and large been ironed out.

6. Conclusions

These results demonstrate that there indeed exists a relation between the two
basic variables investigated, syntactic complexity and level of style. The sample
identified as belonging to the higher level of style (Mir.Cyr.&Jo.) consistently
manifests greater syntactic complexity than the samples identified as belonging
to a lower level of style (Prat. and Mir.Artem.), with these differences being sta-

70 In the case of embedding depth, due to the low incidence of tokens from level 3 onwards
(especially in Prat.) the χ2-test has been carried out with levels 2–5 collapsed into a single cat-
egory in order to fulfil the requirement of 5+ tokens per table cell.
71 Cf. Karlsson (n. 34) 96–98 and Laury – Ono (n. 31) 77–78.
198 Nikolaos Kälviäinen

tistically significant at the p < 0.05 level. Thus, the basic research hypothesis of
the correlation of level of style and syntactic complexity is corroborated by these
results.
On the other hand, there are two caveats. Firstly, although the vast differ-
ence between high-level text Mir.Cyr.&Jo. and the low-level text Prat. is consis-
tently borne out in the behaviour of all 5 variables, the second text identified as
belonging to a low stylistic level, Mir.Artem. exhibits somewhat greater complex-
ity than Prat.; this difference is statistically significant in 4 out of 5 variables. On
the other hand, it must be stressed that the overall image of syntactic complexity
presented by Mir.Artem. is still relatively close to that of Prat. and that the differ-
ences between the two, while undeniable, should not be exaggerated either.
That statistical significance or lack thereof alone cannot neatly divide our
texts into clearly separate levels of style is of course obvious, since in reality
the "levels of style" are not levels but a continuum with plenty of room for sta-
tistically significant differences between personal styles even within groups of
relatively similar styles. The question remains whether the continuum is an even,
unbroken line across the whole corpus of Byzantine texts or whether such groups
of relatively similar styles, ones that might then be named levels of style, can be
identified. In such a future classification of Byzantine styles on the basis of their
syntax, we might conceivably see Mir.Artem. labelled "lower middle style" or al-
ternatively "low style" with Prat. as "super-low". Before more data are available,
however, all such characterizations remain speculative.
Secondly, the corpus of texts examined in this study is extremely limited
and as such the results cannot be generalized across the board. In order to estab-
lish the applicability of our method to Byzantine literature as a whole, a much
larger corpus is required, with texts from different periods, genres and levels of
style. It goes without saying that the method of analysis needs also to be further
refined. Future research should prepare the ground for a more comprehensive ac-
count of complexity-related syntactic variation in Byzantine literature. Such an
account could then be related to findings made in other areas pertaining to style,
such as the lexicon, morphology, rhetorical devices and so on.

University of Crete