

OP-LLCJ150028 1..14


Polysemy and Synonymy in
Syntactic Dependency Networks
............................................................................................................................................................

Radek Čech

Department of Czech Language, University of Ostrava, Ostrava,

Czech Republic

Ján Mačutek

Department of Applied Mathematics and Statistics, Comenius

University, Bratislava, Slovakia

Zdeněk Žabokrtský

Institute of Formal and Applied Linguistics, Charles University in

Prague, Prague, Czech Republic

Aleš Horák

Natural Language Processing Centre, Masaryk University, Brno,

Czech Republic
.......................................................................................................................................

Abstract
The relationship between two important semantic properties (polysemy and syn-
onymy) of language and one of the most fundamental syntactic network proper-
ties (a degree of the node) is observed. Based on the synergetic theory of
language, it is hypothesized that a word which occurs in more syntactic contexts,
i.e. it has a higher degree, should be more polysemous and have more synonyms
than a word which occurs in less syntactic contexts, i.e. it has a lesser degree. Six
languages are used for hypotheses testing and, tentatively, the hypotheses are
corroborated. The analysis of syntactic dependency networks presented in this
study brings a new interpretation of the well-known relationship between fre-
quency and polysemy (or synonymy).

.................................................................................................................................................................................

1 Introduction

A complex network (cf. Newman 2011) is a model of
a system. It contains sets of nodes, representing enti-
ties, and links, representing relations among them.
Syntactic properties of language can be described,
inter alia, as a system which contains both words
and syntactically motivated relations between pairs
of words (cf. Mel’čuk, 1988; Hudson, 2007).
Consequently, it is not surprising that complex net-
works have been used for an analysis of syntax almost

since the complex network theory emerged (Barabási
and Albert, 1999; Barabási, 2002). The analyses of
syntactic complex networks opened new insights
into a language functioning in the last decade.
Specifically, new models of language acquisition
were proposed (cf. Ninio, 2006, 2011; Ke and Yao,
2008; Corominas-Murtra et al., 2010), the relation
between syntax and communication needs was
analysed (cf. Ferrer i Cancho et al., 2005; Ferrer i
Cancho, 2006a), differences of statistical properties
of syntactic networks were used for typological

Correspondence: Radek

Čech, Department of Czech

Language, University of

Ostrava, Reálnı́ 5, Ostrava

701 03, Czech Republic.

E-mail:

cechradek@gmail.com

Digital Scholarship in the Humanities � The Author 2015. Published by Oxford University Press on behalf of EADH.
All rights reserved. For Permissions, please email: journals.permissions@oup.com

1 of 14

doi:10.1093/llc/fqv028

 Digital Scholarship in the Humanities Advance Access published July 25, 2015


studies (cf. Čech and Mačutek, 2009; Liu and Li,
2010; Abramov and Mehler, 2011; Liu and Xu,
2011), the origin of projectivity, i.e. the fact that syn-
tactic dependency crossing occurs but very rarely,
was inquired (cf. Ferrer i Cancho, 2006b, 2008), as
well as the origin of syntax and its role in syntactic
complex networks (cf. Ferrer i Cancho et al., 2005;
Solé, 2005; Liu and Hu, 2008; Liu et al., 2010; Čech et
al., 2011). Despite these achievements, a linguistic
explanation of syntactic network properties is
strongly needed, especially because the majority of
language network analyses were merely descriptive,
and not explanative (cf. Ferrer i Cancho, 2010),
and focused on global network characteristics
(Cong and Liu, 2014a, b). On the other hand, net-
work analyses can bring a finer-grained interpret-
ation of some traditional linguistic problems (Liu,
2011; Gao et al., 2014), as is also presented in this
study in case of polysemy and synonymy functioning,
and the complex network approach have a potential
to be inspiring for a theory of language, according to
Ferrer i Cancho (2014).

One possible way how to explore a syntactic net-
work from a linguistic point of view is to find those
language characteristics which should correlate with
syntactic network properties because of some theor-
etical reasons. Generally, the approach of this kind
was developed by synergetic linguistics in Köhler
(1986, 2005a). It was used to analyse phonology
(Coloma, 2014), lexical properties (Köhler, 2005b;
Köhler and Altmann, 1993; Wang, 2014), syntax
(Köhler, 1999, 2007, 2012; Čech and Mačutek,
2010; Čech et al., 2010, 2011; Liu, 2011; Gao et al.,
2014), and morphology (Prün and Steiner, 2005).
The present study follows this approach, and its aim
is to scrutinize the relationship between syntax
properties modelled by complex network (specific-
ally, a number of syntactically motivated lexical con-
texts of a given word are taken into account; in the
complex network, it is determined by a degree of the
node representing the word; see Section 3 for more
details), on the one hand, and semantic properties
(word polysemy and synonymy), on the other. The
degree of the word is determined by the number of
syntactic dependencies in all syntactic trees in a
corpus; an aggregation of these trees makes the syn-
tactic complex network.

We hypothesize that a word—actually, we con-
sidered a lemma, i.e. a canonical word form1—
which occurs in more syntactically motivated lexical
contexts, i.e. it has a higher degree, should be more
polysemous and have more synonyms2 than a word
which occurs in less contexts, i.e. it has a lesser
degree. Based on this deduction, we put forth
hypotheses as follows:

(1) the higher the out-degree/in-degree of a word,
the more meanings it has;

(2) the higher the out-degree/in-degree of a word,
the more synonyms it has.

The out-degree of a word expresses the total
number of words which are connected to it as its
modifiers, i.e. the number of its syntactically
dependent words, while the in-degree of the word
expresses the total number of words which the word
under consideration is connected to, i.e. the number
of its heads, in an observed sample of language
(see Section 3).

Since it is assumed that these hypotheses are not
language specific, six languages (Czech, Dutch,
English, German, Italian, and Spanish) are used
for their testing. Admittedly, they do not cover all
types of languages (all of them belong to the Indo-
European family), but we consider them as a suffi-
cient sample for the first preliminary analysis in this
field.

The article is organized as follows: the status of
polysemy and synonymy in language is presented in
Section 2; a language material together with the
methodology used is introduced in Section 3;
Section 4 is focused on results; and the article is
ended by Conclusion (Section 5).

2 Polysemy and synonymy
of the word

At least since Zipf (1935), it is well known that the
semantic aspect of language (i.e. a meaning of any
language unit) is closely connected to other language
properties (e.g. relative frequency, degree of intensity
of accent, and degree of crystallization of the config-
uration). Probably the most systematical incorpor-
ation of certain semantic characteristics of language

R. Čech et al.

2 of 14 Digital Scholarship in the Humanities, 2015


into a language model was presented by Köhler
(1986, 2005b, 2012). Specifically, in the synergetic
model of language, Köhler focuses on two semantic
properties of language units, viz. polysemy and syn-
onymy. He hypothesizes particular relationships
among them and other language characteristics,
such as frequency, word length, inventory sizes of
language units, and functional loads of language
units, and tests all hypotheses experimentally.
Consequently, both polysemy and synonymy can be
viewed as language properties which are ruled by a
complex mechanism emerging as a result of intricate
interrelations among so-called communication
requirements (cf. Köhler, 2005b). The system of
these requirements represents a novel way of
developing an approach originally proposed by Zipf
(1949). According to Zipf, the properties of language
are determined by a principle which should have a
crucial impact on the human behaviour in general;
he calls it the principle of least effort.

As for the relationship between the principle of
least effort and polysemy, Zipf (1949: 20–1) pointed
out that

[f]rom the viewpoint of speaker (the speaker’s
economy) who has the job selecting not only
the meanings to be conveyed but also the
words that will convey them, there would
doubtless exist an important latent economy
in a vocabulary that consisted exclusively of
one single word—a single word that would
mean whatever the speaker wanted it to
mean. (. . .) But from the viewpoint of the
auditor, who has the job of deciphering the
speaker’s meanings, the important internal
economy of speech would be found rather in
a vocabulary of such size that is possessed a
distinctly different word for each different
meaning to be verbalized.

As a result of these opposite communication
strategies, an equilibrium regarding the number of
meanings to be conveyed by particular words should
emerge; this assumption was corroborated in Zipf
(1949), Tuldava (1998), Ferrer i Cancho and Solé
(2003), Ferrer i Cancho (2005a), and Kelih (2008).
It was shown that the distribution of the number of
meanings follows certain regularities.

Inspired by these findings, in the following para-
graphs we hypothesize some relationships between
both polysemy and synonymy, on the one hand, and
degrees of nodes representing canonical word forms
(lemmas) in a syntactic complex network, on the
other.

First, it is necessary to realize that the meaning of
a word is strongly influenced by words which are in
a syntactic relation with it in an actual language
usage. Obviously, the presence of syntactically
related words makes it possible to express meaning
refinements; thus, syntactically related words take
part in conveying the meaning by the word under
the consideration. Consequently, it is reasonable to
assume that, with an increasing variability of
syntactic contexts of the word, the variability of
subtle meaning differences of the word also in-
creases. If the differences of meaning are ‘striking
enough’, they are detected by a lexicographer and
they are recorded in dictionaries or databases such
as the WordNet (see below) which are usually used
for the determination of the number of meanings of
the word (i.e. its polysemy).

Further, from the speaker’s point of view, the
presence of syntactically related words enables the
speaker to use a unique word for conveying more
meanings because syntactically related words take
part in its meaning expression and its differenti-
ation. This is in a clear agreement with the speaker’s
economy—the speaker tends to use as few words as
possible for as many meanings as possible
(see above); obviously, the ‘cost’ of the increasing
polysemy of the word is the length (or complexity)
of the syntactic constructions needed for the expres-
sion of the meaning (cf. Köhler, 2012). From the
auditor’s point of view, the presence of syntactically
related words makes the determination of the
meaning of the word easier, cf. the priming effect
(Hoey, 2005). So, the auditor’s economy ‘presses’
the speaker to use words with more meanings in a
syntactically more complex environment because it
makes it for him easier to determine the meaning of
the entire expression. If the speaker would ignore
the auditor’s economy, the probability of auditor’s
misunderstanding increases, and, consequently, the
probability that the auditor asks the speaker for a
new, better explanation increases. From the

Polysemy and synonymy in syntactic dependency networks

Digital Scholarship in the Humanities, 2015 3 of 14


speaker’s point of view, the repetition of the expres-
sion conveying the same meaning is in a clear
contradiction to his communication strategy
(i.e. the speaker’s economy).

In sum, because both the out-degree and the in-
degree of the word represent the number of syntac-
tically related words in a corpus, we hypothesize
that the higher out-degree (or in-degree) of the
word, the more meanings the word has.

As for the relationship between synonymy and
the degree of the word, it should be viewed as a
consequence of the relationship between polysemy
and synonymy. The more meanings a word obtains,
the more semantic domains of other words it pene-
trates and, consequently, these words become its
synonyms (cf. Köhler, 1986; Wimmer and
Altmann, 2001). Hence, because the relationship
between the degree and polysemy is assumed, also

Fig. 2 The network containing the first fifty lemmas from the English treebank

Fig. 1 The tree graph expressing the structure of the sentence ‘Tom saw my parents yesterday in Boston’ based on the
dependency grammar formalism. Links between words represent the syntactic dependency relations, the arrows express
the direction of the dependency

R. Čech et al.

4 of 14 Digital Scholarship in the Humanities, 2015


the relationship between the degree and synonymy
should be hypothesized, i.e. the higher out-degree/
in-degree of the word, the more synonyms it has.

3 Language material and
methodology

For empirical studies on the relation between vari-
ability of the syntactic context and polysemy of
words, we naturally need two types of data
resources: (1) syntactically annotated data, and (2)
dictionary data with enumerated words’ meanings.
In our experiments, the role of the former type is
played by dependency treebanks, while wordnets are
used as the latter type. We managed to interlink the
two types of information for six languages so far:
Czech, Dutch, English, German, Italian, and
Spanish. In spite of additional treebanks and word-
nets being simultaneously available for several other
languages, we were not able to include them into
our experiments due to various technical obstacles,
such as an insufficient size, a missing lemmatization,
or an incompatible tokenization.

We did not use the original treebank shapes for
building dependency networks, but we used their
‘Prague dependency treebank—harmonized’ forms,
as introduced in Zeman et al. (2012), where many
treebanks were transformed in order to maximize
the compatibility of the resulting dependency trees
with annotation guidelines for so-called analytical
trees (surface-syntactic dependency trees) of the
Prague Dependency Treebank (Hajič et al., 2006).
For instance, subordinating conjunctions are heads

of subordinating clauses according to the Prague
dependency treebank style, analogously to prepos-
itions being heads of prepositional groups. We do
not claim that the Prague Dependency Treebank
style is superior to other treebanks’ conventions
when it comes to building syntactic networks, but
we prefer to treat all the languages as uniformly as
possible, and the Prague Dependency Treebank har-
monization was the only normalization that was
readily available to us.

We used data originating from the treebanks
listed below (cf. Zeman et al., 2012, for more details
on the respective resources and associated normal-
ization procedures):

Czech: Prague dependency treebank 2.0 (Hajič
et al., 2006),
Dutch: Alpino treebank (van der Beek et al.,
2002),
English: Penn treeBank 2 (Surdeanu et al.,
2008),
German: Tiger treebank (Brants et al., 2002),
Italian: Italian syntactic-semantic treebank
(Montemagni et al., 2003),
Spanish: AnCora (Taulé et al., 2012).

The first WordNet database was published by
Miller et al. (1993) at the University of Princeton.
Since then, more than seventy national wordnets
were created following the same principles as the
original Princeton WordNet for English (cf. Horák
et al., 2008).

The data in wordnet databases are organized as
networks of basic entities called ‘synsets’, synonym
sets. Each synset corresponds to one meaning of a
word or a collocation. The synonymy relation in
wordnet is not the standard ‘strict’ synonymy,
where synonymous words are simply identical
(or nearly identical) in meaning (e.g. ‘pretty’ and
‘handsome’). The word meanings in synsets are
‘in a near synonymy’ relation—they are synonymous
in the sense that they can be exchanged in the ‘same
contexts’. For example, the synset ‘exist:1, be:4’ in
wordnet relates the fourth meaning (generally called
a ‘literal’) of ‘to be’ with the first meaning of ‘to exist’
with the definition of ‘to have an existence’.

For the purpose of the current experiment, we
used the data of six languages in the form resulting

Table 1 Ranked distribution of out-degrees in the Czech

treebank

Rank Out-degree

1 7,441

2 3,489

3 2,112

4 1,333

5 914

. . . . . .

11,939 0

11,940 0

Polysemy and synonymy in syntactic dependency networks

Digital Scholarship in the Humanities, 2015 5 of 14


from the EuroWordNet and BalkaNet EU projects.
The studied wordnets have the following statistics:

Czech WordNet: 32,116 words and collo-
cations, 28,448 synsets, 43,958 literals;
Dutch WordNet: 56,329 words and colloca-
tions, 44,128 synsets, 70,356 literals;

English WordNet: 156,588 words and col-
locations, 117,597 synsets, 207,018 literals;
German WordNet: 17,098 words and colloca-
tions, 15,132 synsets, 20,453 literals;
Italian WordNet: 32,964 words and collo-
cations, 40,406 synsets, 48,475 literals;

Table 2 Mean values of out-degrees/in-degrees, the mean number of synonyms, and the mean number of meanings in

the Czech treebank

Out-degree Synonyms Polysemy In-degree Synonyms Polysemy

507.94 8.59 5.61 313.95 3.70 3.71

214.34 6.21 4.27 137.93 4.92 3.74

167.74 5.26 3.47 96.63 4.27 3.40

137.61 4.59 3.42 76.24 3.46 2.95

116.47 4.64 3.38 63.45 5.60 3.78

100.53 5.18 3.32 54.50 3.68 2.89

88.95 4.15 2.92 47.97 3.28 2.74

78.83 3.40 2.54 42.27 2.96 2.54

70.18 3.96 3.31 37.44 3.83 2.84

62.61 3.18 2.56 33.36 2.95 2.63

56.80 3.37 2.52 30.03 3.08 2.48

51.96 3.18 2.32 26.91 3.31 2.44

47.47 3.10 2.50 24.24 2.51 2.18

43.41 3.32 2.59 22.51 3.18 2.43

39.97 3.07 2.39 20.54 2.75 2.21

37.06 2.67 2.23 18.50 2.66 2.10

33.94 3.11 2.33 16.43 2.34 1.97

31.03 2.33 2.11 14.45 2.40 2.16

28.46 2.89 2.11 13.00 2.43 1.97

26.51 2.75 2.11 12.00 2.11 2.02

24.49 2.91 2.28 11.00 1.82 1.74

22.50 2.99 2.10 10.00 2.51 1.97

20.48 2.50 2.16 9.00 2.10 1.85

18.49 2.19 1.85 8.00 2.28 1.79

17.00 2.41 1.79 7.00 1.95 1.69

16.00 2.35 2.07 6.00 2.10 1.74

15.00 2.16 1.83 5.00 1.87 1.63

14.00 2.47 1.90 4.00 1.94 1.59

13.00 2.26 1.94 3.00 1.82 1.54

12.00 2.11 1.84 2.00 1.64 1.43

11.00 2.10 1.65 1.00 1.64 1.35

10.00 2.31 1.75 0.00 1.86 1.30

9.00 2.28 1.79

8.00 2.15 1.67

7.00 1.92 1.76

6.00 1.97 1.62

5.00 1.96 1.54

4.00 1.76 1.54

3.00 1.83 1.52

2.00 1.66 1.41

1.00 1.56 1.42

0.00 1.21 1.34

R. Čech et al.

6 of 14 Digital Scholarship in the Humanities, 2015


Spanish WordNet: 32,606 words and colloca-
tions, 30,350 synsets, 52,362 literals.
Using these wordnet data, the nodes of the
treebank-based syntactic networks were
labelled with two values:
Number of synonyms—the number of other lit-
erals in all synsets for the lemma,
Number of meanings—the number of different
literals/synsets for the lemma.

For example, the verb ‘intend’ has four mean-
ings/synsets in the English WordNet:

(1) intend: 1, mean: 4, think: 7;
(2) intend: 2, destine: 2, designate: 4, specify: 6;
(3) mean: 1, intend: 3;
(4) mean: 3, intend: 4, signify: 1, stand for: 2.

The node ‘intend’ thus received nine synonyms
and four meanings.

In constructing the syntactic dependency net-
works, the methods developed by Ferrer i Cancho
et al. (2004) and Liu (2008) were followed. Each
node of the complex network represents a particular
lemma. Two nodes are linked if there is a depend-
ency relation between the respective lemmas in the
treebank. The links are directed; they go from the
head to the modifier (see Fig. 1).

A single graph is used in our analysis, i.e. only
unique connections between particular lemmas are
counted. Thus, a global syntactic dependency net-
work is constructed by accumulating sentence struc-
tures, and the network should be viewed as an
emergent property of sentence structures (Ferrer i

0 100 200 300 400 500

2
4

6
8

CZ

0 50 100 150

3
4

5
6

DU

0 100 200 300 400 500

5
1

0
1

5

EN

0 100 200 300 400 500

1
.0

1
.5

2
.0

2
.5

GE

0 20 40 60 80

1
3

5
7

IT

0 50 100 150 200 250 300

4
6

8
1

0

SP

Fig. 3 Relation between the mean out-degree (x-axis) and the mean number of synonyms (y-axis)

Polysemy and synonymy in syntactic dependency networks

Digital Scholarship in the Humanities, 2015 7 of 14


Cancho et al., 2004; Ferrer i Cancho, 2005b). The free
software Pajek 2.05 (cf. de Nooy et al., 2005) was used
for network creation and computing. For an illustra-
tion, Fig. 2 shows the syntactic dependency complex
network containing the first fifty lemmas from the
English Treebank by Surdeanu et al. (2008).

4 Statistical Methodology

We follow the procedures described in Čech et al.
(2011), which we remind here briefly. First, words
which have zero meanings (it is determined by the
absence of synsets in the WordNet) were omitted
from our analyses. In the following step, rank-
frequency distributions of out-degrees were

constructed (for each language separately). The
rank-frequency distribution is then exploited in
the process of construction of the figures in
Section 5 (ranks were used in the binning proced-
ure; see below). Values of out-degrees are ordered
from the highest to the lowest, the highest value
receives the rank 1, the second highest one the
rank 2, etc. Thereafter, we consider the ranks to be
values of an auxiliary random variable and the out-
degrees their frequencies. This auxiliary variable
was used because the data are highly skewed—his-
tograms constructed directly from both out- and
in-degree values contain huge numbers of empty
bins.

In the example from the Czech language, cf.
Table 1, in the rank-frequency distribution we

0 50 100 150 200 250 300

2
3

4
5

CZ

0 50 100 150 200 250 300

2
3

4
5

DU

0 50 100 150 200 250 300

2
3

4
5

EN

0 50 100 150 200 250 300

2
3

4
5

GE

0 50 100 150 200 250 300

2
3

4
5

IT

0 50 100 150 200 250 300

2
3

4
5

SP

Fig. 4 Relation between the mean out-degree (x-axis) and the mean number of meanings (y-axis)

R. Čech et al.

8 of 14 Digital Scholarship in the Humanities, 2015


assume to have 7,441-times value 1, 3,489-times
value 2, etc; value 11,940 does not occur (it has
frequency 0).

Next, a histogram was created from the rank-
frequency distribution. Its bin width was calculated
first according to Scott (1979). Then, the
bins boundaries were adjusted so that equal values
(i.e. equal out-degrees in the original data) are kept
in one bin (e.g. all ranks corresponding to out-
degree 1 belong to the same bin). The reason for
adjusting the bin widths (which results in a histo-
gram with different bin widths) is that there is no
reasonable hierarchy of word lemmas which share
the same out-degree in a network—they appear in
the same order as they were entered into the tree-
banks (or then found in the process of network

creation). Therefore, we assign all lemmas with the
same out-degree to the same histogram bin.

Then, for each histogram bin we computed the
mean out-degree, the mean number of synonyms,
and the mean number of meanings (e.g. the mean
polysemy) of the lemmas represented by ranks
belonging to the bin. The same procedure was
applied also to in-degrees. We thus obtained six
data sets for out-degrees and six data sets for
in-degrees. This approach (ranked frequencies)
serves as a tool for a better visualization; the tests
(cf. Section 5) were performed on the original data.

The two hypotheses from Section 1 (the higher
out-degree/in-degree of a word, the more meanings
it has; the higher out-degree/in-degree of a word,
the more synonyms it has) were tested using the

0 50 100 150 200 250 300

4
6

8
1

0

CZ

0 50 100 150 200 250 300

4
6

8
1

0

DU

0 50 100 150 200 250 300

4
6

8
1

0

EN

0 50 100 150 200 250 300

4
6

8
1

0

GE

0 50 100 150 200 250 300

4
6

8
1

0

IT

0 50 100 150 200

5
6

7
8

SP

Fig. 5 Relation between the mean in-degree (x-axis) and the mean number of synonyms (y-axis)

Polysemy and synonymy in syntactic dependency networks

Digital Scholarship in the Humanities, 2015 9 of 14


Kendall correlation coefficient (cf. Hollander and
Wolfe, 1999: 382); see Section 5.

5 Results

The mean numbers of synonyms and meanings in
the histogram bins tend to increase quite clearly
with the increasing out-degree/in-degree. The
values for the Czech Treebank3 can be found in
Table 2. Figures 3–6 present the values graphically.
We use the following abbreviations for the lan-
guages: CZ—Czech, DU—Dutch, EN—English,
GE—German, IT—Italian, SP—Spanish.

In addition to the figures, the association
between the variables (out-degree/in-degree and
the number of synonyms, out-degree/in-degree

and the number of meanings) was measured by
the Kendall correlation coefficient4 (cf. Hollander
and Wolfe, 1999). The values of the correlation
coefficient can be found in Table 3. As the original
data without any smoothing were used, the coeffi-
cients are relatively small; nevertheless, the correl-
ation is positive and statistically significant in all
cases (all P-values are below 0.001), which, together
with the trends for binned data in Figs 1–6 corrob-
orates the hypotheses from Section 1 (the higher
out-degree/in-degree of the word, the more syno-
nyms it has; the higher out-degree/in-degree of the
word, the more meanings it has).

Given that sample sizes in Table 3 are quite large,
the (very) small P-values themselves do not say too
much—it is well known that, for large samples,
virtually all null hypotheses are rejected (cf., e.g.

0 50 100 150 200

2
.0

3
.0

4
.0

CZ

0 50 100 150 200

2
.0

3
.0

4
.0

DU

0 50 100 150 200

2
.0

3
.0

4
.0

EN

0 50 100 150 200

2
.0

3
.0

4
.0

GE

0 50 100 150 200

2
.0

3
.0

4
.0

IT

0 50 100 150 200

2
.0

3
.0

4
.0

SP

Fig. 6 Relation between the mean in-degree (x-axis) and the mean number of meanings (y-axis)

R. Čech et al.

10 of 14 Digital Scholarship in the Humanities, 2015


Mačutek and Wimmer, 2013, and references
therein). Therefore, Figs 1–6 provide perhaps a
more reliable evidence.

6 Conclusions

The results presented in the study brought two main
findings. First, we found out that one of the most
fundamental syntactic network properties, the
degree of the node, significantly correlates with
some important semantic properties (polysemy
and synonymy) of language. Moreover, the hypoth-
esis concerning the relationships between degrees
and polysemy (or synonymy) is based on a theoret-
ical linguistic reasoning. Consequently, our findings,
perhaps, advocate the usage of complex network in
linguistic research and they can be viewed as a
constructive response to the call for linguistic
explanation of syntactic network properties
(cf. Cong and Liu, 2014a, b; Ferrer i Cancho,
2014). Second, the empirical corroboration of the
hypothesis concerning the relationship between the
degree and polysemy can be interpreted as a deeper
insight into the well-known relationship between
frequency and polysemy (e.g. Zipf, 1945; Baayen,
and Moscoso del Prado, 2005; Ilgen and
Karaoglan, 2007). In other words, the original
hypothesis concerning the impact of the word
(or lemma) frequency on polysemy presupposes
‘implicitly’ that frequent words occur in more con-
texts and, consequently, this fact leads to an increase
of polysemy. In our study, contexts have been ‘ex-
plicitly’ operationalized—the degrees express the
number of syntactic contexts, in fact. Further,
the use of single graphs diminishes the impact of
the frequency as much as possible in studies of
this kind and it allows observing the impact of a

purely syntactic property (i.e. the degree of node
in a syntactic complex network) of the lemma
on its polysemy. To sum up, our approach experi-
mentally proves the implicit assumption and
reveals more detailed characteristics of the rela-
tionship between these important language
characteristics.

Acknowledgements
This work was supported by the by the Czech
Science Foundation [grant no. P406/11/0268 to
R.Č.]; the Scientific Grant Agency of the Slovak
Republic [grant no.VEGA 2/0047/15 to J.M.]; the
Ministry of Education of the Czech Republic
[LINDAT-Clarin project LM2010013 to A.H.]; the
Ministry of the Interior of the Czech Republic [pro-
ject VF20102014003 to A.H.]; and by the Czech
Science Foundation [project P401/10/0792 to A.H.].

References
Abramov, O. and Mehler, A. (2011). Automatic language

classification by means of syntactic dependency net-
works. Journal of Quantitative Linguistics, 18: 291–336.

Baayen, H. and Moscoso del Prado, M. (2005). Semantic

density and past tense formation in three Germanic
languages. Language, 81: 666–98.

Barabási, A. (2002). The New Science of Networks.
Cambridge, MA: Perseus Publishing.

Barabási, A. and Albert, R. (1999). Emergence of scaling

in random networks, Science, 286: 509–12.

Brants, S., Dipper, S., Hansen, S., Lezius, W. and Smith,
G. (2002). The TIGER Treebank. In Hinrichs, E.
and Simov, K. (eds), Proceedings of the First

Workshop on Treebanks and Linguistic Theories.
Sozopol: Bulgarian Academy of Sciences, pp. 24–41.

Table 3 Kendall correlation coefficients

CZ DU EN GE IT SP

Number of lemmas 11,940 6,582 14,554 6,571 5,261 8,277

Synonyms, out-degree 0.209 0.110 0.229 0.092 0.270 0.180

Meanings, out-degree 0.271 0.238 0.322 0.322 0.286 0.285

Synonyms, in-degree 0.070 0.121 0.188 0.065 0.137 0.093

Meanings, in-degree 0.141 0.212 0.295 0.236 0.179 0.211

Polysemy and synonymy in syntactic dependency networks

Digital Scholarship in the Humanities, 2015 11 of 14


Coloma, G. (2014). Towards a synergetic statistical model
of language phonology. Journal of Quantitative
Linguistics, 21: 100–22.

Cong, J. and Liu, H. (2014a). Approaching human lan-
guage with complex networks. Physics of Life Reviews,
11: 598–618.

Cong, J. and Liu, H. (2014b). Linguistic complex net-
works: Rationale, application, interpretation, and dir-
ections: Reply to comments on ‘‘Approaching human
language with complex networks’’. Physics of Life
Reviews, 11: 644–9.

Corominas-Murtra, B., Valverde, S. and Solé. R. V.
(2010). Emergence of Scale-free Syntax Networks. In
Nolfi, S. and Mirolli, M. (eds), Evolution of
Communication and Language in Embodied Agents.
Berlin: Springer, pp. 83–99.

Čech, R. and Mačutek, J. (2009). Word form and lemma
syntactic dependency networks in Czech: A compara-
tive study. Glottometrics, 19: 85–98.

Čech, R. and Mačutek, J. (2010). On Quantitative
Analysis of Valency in Czech. In Grzybek, P., Kelih,
E. and Mačutek, J. (eds), Text and Language.
Structures, Functions, Interrelations, Quantitative
Perspectives. Wien: Praesens Verlag, pp. 21–9.

Čech, R., Mačutek, J. and Žabokrtský, Z. (2011). The
role of syntax in complex networks: Local and global
importance of verbs in a syntactic dependency network.
Physica A: Statistical Mechanics and Its Applications,
390: 3614–23.

Čech, R., Pajas, P. and Mačutek, J. (2010). Full valency.
Verb valency without distinguishing complements and
adjuncts. Journal of Quantitative Linguistics, 17: 291–
302.

de Nooy, W., Mrvar, A. and Batagelj, V. (2005).
Exploratory Social Network Analysis with Pajek. New
York: Cambridge University Press.

Ferrer i Cancho, R. (2005a). Zipf’s law from a commu-
nicative phase transition. European Physical Journal B,
47: 449–57.

Ferrer i Cancho, R. (2005b). The Structure of
Syntactic Dependency Networks: Insights from
Recent Advances in Network Theory. In Altmann,
G., Levickij, V. and Perebyinis, V. (eds), Problems
of Quantitative Linguistics. Chernivtsi: Ruta, pp.
60–75.

Ferrer i Cancho, R. (2006a). When language breaks
into pieces. A conflict between communication
through isolated signals and language. Biosystems, 84:
242–53.

Ferrer i Cancho, R. (2006b). Why do syntactic links not

cross? Europhysics Letters, 76: 1228–35.

Ferrer i Cancho, R. (2008). Some word order biases from

limited brain resources. A mathematical approach.

Advances in Complex Systems, 11: 394–14.

Ferrer i Cancho, R. (2010). Network Theory. In Hogan,

P. C. (ed.), The Cambridge Encyclopedia of the Language

Sciences. Cambridge: Cambridge University Press, pp

555–7.

Ferrer i Cancho, R. (2014). Beyond description.

Comment on ‘‘Approaching human language with

complex networks’’ by Cong & Liu. Physics of Life

Reviews, 11(4): 621–3.

Ferrer i Cancho, R. and Solé, R. (2003). Least effort and

the origins of scaling in human language. Proceedings of

the National Academy of Sciences USA, 100: 788–91.

Ferrer i Cancho, R., Solé, R. V. and Köhler, R. (2004).

Patterns in syntactic dependency networks. Physical

Review E, 69: article no. 051915.

Ferrer i Cancho, R., Riordan, O. and Bollobás, B.

(2005). The consequences of Zipf’s law for syntax and

symbolic reference. Proceedings of the Royal Society of

London Series B, 272: 561–5.

Gao, S., Zhang, H., Liu, H. (2014). Synergetic properties

of Chinese verb valency. Journal of Quantitative

Linguistics, 21: 1–21.

Hajič, J., Panevová, J., Hajičová, E., Sgall, P., Pajas P.,

Štěpánek, J., Havelka, J., Mikulová, M., Žabokrtský,

Z. and Ševčı́ková-Razı́mová, M. (2006). Prague

Dependency Treebank 2.0 (CD-ROM). Philadelphia:

Linguistic Data Consortium.

Hoey, M. (2005). Lexical Priming: A New Theory of Words

and Language. New York: Routledge.

Hollander, M. and Wolfe, D. A. (1999). Nonparametric

Statistical Methods. Hoboken: Wiley.

Horák, A., Pala, K. and Rambousek, A. (2008). The

Global WordNet grid Software Design. In Tanács,

A., Csendes, D., Vincze, V., Fellbaum, C. and

Vossen, P. (eds), Fourth Global WordNet

Conference Proceedings. Szeged: University of

Szeged, pp. 194–9.

Hudson, R. (2007). Language Networks. The New Word

Grammar. Oxford: Oxford University Press.

Ilgen, B. and Karaoglan, B. (2007). Investigation of

Zipf’s law-of-meaning on Turkish corpora. 22nd

International Symposium on Computer and

Information Sciences (ISCIS 2007), Ankara: IEEE, pp.

1–6.

R. Čech et al.

12 of 14 Digital Scholarship in the Humanities, 2015


Ke, J. and Yao, Y. (2008). Analysing language develop-
ment from a network approach. Journal of Quantitative
Linguistics, 15: 70–99.

Kelih, E. (2008). Modelling polysemy in different lan-
guages: A continuous approach. Glottometrics, 16: 46–
56.

Köhler, R. (1986). Zur linguistischen Synergetik: Struktur
und Dynamik der Lexik. Bochum: Brockmeyer.

Köhler, R. (1999). Syntactic structures. Properties and
interrelations. Journal of Quantitative Linguistics, 6:
46–57.

Köhler, R. (2005a). Synergetic Linguistics. In Köhler, R.,
Altmann, G. and Piotrowski, R. G. (eds), Quantitative
Linguistics. An International Handbook. Berlin; New
York: de Gruyter, pp. 760–74.

Köhler, R. (2005b). Properties of Lexical Units and
Systems. In Köhler, R., Altmann, G. and Piotrowski,
R. G. (eds), Quantitative Linguistics. An International
Handbook. Berlin; New York: de Gruyter, pp. 305–13.

Köhler, R. (2007). Quantitative Analysis of Syntactic
Structures in the Framework of Synergetic Linguistics.
In Mehler, A. and Köhler, R. (eds), Aspects of Automatic
Text Analysis. Berlin; Heidelberg; New York: Springer,
pp. 191–210.

Köhler, R. 2012. Quantitative Syntax Analysis. Berlin;
Boston: de Gruyter.

Köhler, R. and Altmann, G. (1993). Begriffsdynamik und
Lexikonstruktur. In Beckmann, F. and Heyer, G. (eds),
Theorie und Praxis des Lexikons. Berlin: de Gruyter, pp.
173–90.

Liu, H. (2008). The complexity of Chinese syntactic de-
pendency networks. Physica A: Statistical Mechanics and
Its Applications, 387: 3048–58.

Liu, H. (2011). Quantitative properties of English verb
valency. Journal of Quantitative Linguistics, 18: 207–33.

Liu, H. and Hu, F. (2008). What role does syntax play in a
language network? EPL, 83: article no. 18002.

Liu, H. and Li, W. W. (2010). Language clusters based on
linguistic complex networks. Chinese Science Bulletin,
55: 3458–65.

Liu, H. and Xu, C. (2011). Can syntactic networks indi-
cate morphological complexity of a language? EPL, 93:
article no. 28005.

Liu, H., Zhao, Y. and Huang, W. (2010). How do local
syntactic structures influence global properties in lan-
guage networks? Glottometrics, 20: 35–9.

Mačutek, J. and Wimmer, G. (2013). Evaluating good-
ness-of-fit of discrete distribution models in

quantitative linguistics. Journal of Quantitative

Linguistics, 20: 227–40.

Mel’čuk, I. (1988). Dependency Syntax: Theory and

Practice. Albany: State University of New York

Press

Miller, G. A., Beckwith, R., Fellbaum, C., Gross, D. and

Miller, K. (1993). Five Papers on WordNet (CSL Report

43). Princeton: Princeton University.

Montemagni, S., Barsotti, F., Battista, M., Calzolari, N.,

Corazzari, O., Lenci, A. and Zampolli, A. (2003).

Building the Italian Syntactic-semantic Treebank. In

Abeillé, A. (ed), Treebanks: Building and Using Parsed

Corpora. Dordrecht: Kluwer, pp. 189–210.

Newman, M. E. J. (2011). Networks. An Introduction.

Oxford: Oxford University Press.

Ninio, A. (2006). Language and the Learning Curve: A

New Theory of Syntactic Development. Oxford: Oxford

University Press.

Ninio, A. (2011). Syntactic Development, Its Input and

Output. Oxford: Oxford University Press.

Prün, C. and Steiner, P. (2005). Quantitative

Morphologie. Eigenschaften der Morphologischen

Einheiten und Systeme. In Köhler, R., Altmann, G.

and Piotrowski, R. G. (eds), Quantitative Linguistics.

An International Handbook. Berlin; New York: de

Gruyter, pp. 227–42.

Scott, D. W. (1979). On optimal and data-based histo-

grams. Biometrika, 66: 605–10.

Solé, R. V. (2005). Syntax for free? Nature, 434: 289.

Surdeanu, M., Johansson, R., Meyers, A., Màrquez, L.

and Nivre, J. (2008). The CoNLL-2008 shared task on

joint parsing of syntactic and semantic dependencies. In

Proceedings of the Twelfth Conference on Computational

Natural Language Learning. Manchester: ACL, pp. 159–

77

Taulé, M., Martı́, M. A. and Recasens, M. (2008).

AnCora: Multilevel annotated corpora for Catalan

and Spanish. In Proceedings of the Sixth International

Conference on Language Resources and Evaluation

(LREC’08). Marrakech, pp. 96–101.

Tuldava, J. (1998). Probleme und Methoden der

Quantitativ-systemischen Lexikologie. Trier: WVT.

van der Beek, L., Bouma, G., Daciuk, J., Gaustad,

T., Malouf, R., van Noord, G., Prins, R. and

Villada, B. (2002). Algorithms for Linguistic

Processing. NWO PIONIER Progress Report.

Groningen: Graduate School for Behavioral and

Cognitive Neurosciences.

Polysemy and synonymy in syntactic dependency networks

Digital Scholarship in the Humanities, 2015 13 of 14


Wang, L. (2014). Synergetic studies on some properties of
lexical structures in Chinese. Journal of Quantitative
Linguistics, 21: 177–97.

Wimmer, G. and Altmann, G. (2001). Two Hypotheses
on Synonymy. In Ondrejovič, S. and Považaj, M. (eds),
Lexicographica �99. Bratislava: Veda, pp. 218–25.

Zeman, D., Mareček, D., Popel, M., Ramasamy, L.,
Štěpánek, J., Žabokrtský, Z. and Hajič, J. (2012).
HamleDT: To Parse or Not to Parse? In Proceedings
of the Eighth International Conference on Language
Resources and Evaluation (LREC’12). Istanbul, pp.
2735–41.

Zipf, G. K. (1935). The Psycho-biology of Language. An
Introduction to Dynamic Philology. Boston: Houghton-
Mifflin.

Zipf, G. K. (1945). The meaning frequency relationship of
words. Journal of General Psychology, 33: 251–66.

Zipf, G. K. (1949). Human Behavior and the Principle of
Least Effort. Cambridge: Addison-Wesley.

Notes
1 Canonical word form is the infinitive form of a verb

and the nominative form of a substantive, adjective,

pronoun, and so on (for example, the lemma of word

forms ‘go’, ‘goes’, ‘went’, ‘gone’ is GO). In this analysis,
simple lemmatization is used; differences among par-

ticular part-of-speeches are not taken into account, e.g.

all forms of a word ‘mean’ fall to one lemma MEAN,
without differentiation of verb, noun, or adjective.

2 The terms polysemy, the number of meanings, and syn-
onymy are used in accordance with the approach pre-

sented by the WordNet (see Section 3).
3 On the webpage http://www.cechradek.cz/data/net-

work_polysemy_synonymy_tables.pdf one can find

values for all languages analysed in the article.
4 The Kendall correlation is a measure of monotonous

relation between two variables. It achieves its maximum

value 1 if the relation ‘the greater x, the greater y’ holds

for all pairs x,y; similarly, its minimum value �1 cor-
responds to the relation ‘the greater x, the smaller y’.

R. Čech et al.

14 of 14 Digital Scholarship in the Humanities, 2015

http://www.cechradek.cz/data/network_polysemy_synonymy_tables.pdf
http://www.cechradek.cz/data/network_polysemy_synonymy_tables.pdf