Language competition in a population of migrating agents
Dorota Lipowska1 and Adam Lipowski2
arXiv:1702.07888v2 [physics.soc-ph] 4 May 2017
1
Faculty of Modern Languages and Literature, Adam Mickiewicz University, Poznań, Poland
2
Faculty of Physics, Adam Mickiewicz University, Poznań, Poland
Influencing various aspects of human activity, migration is associated also with language formation. To examine the mutual interaction of these processes, we study a Naming Game with migrating
agents. The dynamics of the model leads to formation of low-mobility clusters, which turns out to
break the symmetry of the model: although the Naming Game remains symmetric, low-mobility
languages are favoured. High-mobility languages are gradually eliminated from the system and the
dynamics of language formation considerably slows down. Our model is too simple to explain in
detail language competition of migrating human communities, but it certainly shows that languages
of settlers are favoured over nomadic ones.
I.
INTRODUCTION
The constatation that complex systems can be regarded as composed of many interacting subunits opens
up a possibility of studying them with methods that were
primarily developed in the context of physical manybody systems. Such approach turned-out to be very successful [1], and lead to the emergence of new research
fields like socio- or econophysics [2]. Even certain linguistic problems can be studied using methods with a
strong physical flavour. Language emergence [3] or death
[4], its diversification [5] and diffusion [6], time-evolving
structure [7], appearance of grammar or linguistic categories [8], and language learning [9] are just a few examples of problems, where physicists’ contributions might
prove to be valuable.
Of course, language, as one of our human attributes,
is interrelated with many other forms of our activity. Social interactions, economical status or political situation
influence the way language is acquired and changed, or
sometimes falls into oblivion. Language, as an integral
part of our culture and way of life, is also intricately related to migrations of people [10, 11]. Various tribes, ethnic groups, or even entire nations firmly settled certain
areas, while some others, due to various reasons, almost
constantly migrate. Migration might mix as well as separate human communities and language formation processes should be thus strongly influenced by such a factor. Moreover, some modern trends, especially globalization, most likely increase people’s migrations [12]. Some
researches even suggest that merging multinational and
multicultural migrants creates in some areas a new kind
of super-diverse societies, and to describe their intercommunication, traditionally understood languages do not
seem to be sufficient [13].
It would be desirable to have some general understanding of how migration affects the language formation processes and perhaps vice versa. As for the language formation, an interesting class of models originates from the
so-called Naming Game [14]. In this model a population
of agents negotiates a language (or, more generally, conventional forms). Although the dynamics might depend
on, for example, the structure of the interaction network,
typically the model reaches a consensus on the language.
The process of language formation resembles the ordering dynamics of Ising or Potts models accompanied, due
to the symmetry of the Naming Game, by a spontanous
symmetry breaking. One can even introduce the notion
of an effective surface tension to explain some dynamical
characteristics of the Naming Game [15, 16].
In the present paper, we thus examine the Naming
Game in a population of migrating agents. When mobility of agents is uniform in the entire population, the
model is very similar to the Naming Game of immobile
agents. However, an interesting behaviour appears when
the mobility depends on the language used by an agent.
In such a case, the dynamics turned out to break the symmetry of the Naming Game, favouring low-mobility languages. During the coarsening, agents form low-mobility
clusters that effectively attract and convert high-mobility
neighbours. As a result, the low-mobility agents become
more widespread, which considerably slows down the dynamics. Of course, our model is too simple to explain
the intricacies of language competition in settled and nomadic communities, nevertheless, it shows that the (difference in) mobility has a strong effect on such proccesses.
II.
MODEL
In our model, we have a population of agents placed on
a square lattice of linear size L (with periodic boundary
conditions). Initially agents are uniformly distributed
on the lattice with the density (i.e., probability) ρ. Each
agent has its own inventory, which is a dynamically modified list of words. The dynamics of our model combines
the lattice gas diffusion with the so-called minimal version of the Naming Game [17]. More specifically, in an
elementary step, an agent (Speaker) and one of its neighbouring sites are randomly selected. If the selected site
is empty, Speaker moves to this site. If the selected site
is occupied by an agent (Hearer), then the pair SpeakerHearer plays the Naming Game:
• Speaker selects a word randomly from its inventory
and transmits it to Hearer.
2
5
4.5
log10(τ)
4
3.5
ρ=1.0
ρ=0.8, d=1
ρ=0.5, d=1
ρ=0.1. d=1
ρ=0.7, dA=0.3, dB=0.7
ρ=0.7, dA=0.1, dB=1.0
ρ=0.5, dA=0.1, dB=1.0
3
2.5
2
1.5
1
FIG. 1. Illustration of the Naming Game dynamics. In the
case of success—when the word selected by Speaker (dog) is
known to Hearer—they both retain only the transmitted word
in their inventories. Failure occurs when Hearer does not
know the transmitted word, which is then added by Hearer
to its inventory.
• If Hearer has the transmitted word in its inventory,
the interaction is a success and both players maintain only the transmitted word in their inventories.
• If Hearer does not have the transmitted word in its
inventory, the interaction is a failure and Hearer
updates its inventory by adding this word to it.
The unit of time (t = 1) is defined as ρL2 elementary
steps, which corresponds to a single (on average) update
of each agent. In the following, we will refer to words
communicated by agents as languages. Rules of the Naming Game are also illustrated in Fig. 1.
III.
TWO-LANGUAGE VERSION
When ρ = 1, all sites are occupied, thus there is no diffusion and the model is equivalent to an ordinary squarelattice Naming Game. For ρ < 1, a fraction 1 − ρ of sites
is empty and in addition to playing the Naming Game,
agents change their locations from time to time. There
are several characteristics that might be determined for
Naming Game models. To demonstrate some analogies
to Ising-type models, we examined a two-language version of the Naming Game [18]. We measured the average
time τ needed for a system to reach a consensus, i.e.,
the state where every agent has the same language in its
inventory. The initial configuration includes a square of
size M , inside of which all agents have language B in
their inventories, while outside agents have language A.
Both within the square of size M and outside, the agents
are distributed with the uniform density ρ. In Ising-like
models, general arguments, which refer to the notion of a
surface tension and the Laplace law of excesive pressure,
estimate the lifetime of such a bubble as τ ∼ M 2 [19].
Our numerical results (Fig. 2) are in a very good agreement with such estimation both for ρ = 1 and ρ < 1
1.2
1.4
1.6
log10(M)
1.8
2
2.2
FIG. 2. The average lifetime τ of language B whose users are
initially in a square of size M , surrounded by language A users
(L = 300). The straight and dotted lines have slopes corresponding to τ ≈ M 2 and τ ≈ M 1.1 , respectively. The data for
ρ < 1 and language-independent mobility (dA = dB = d = 1)
also seem to obey such scaling. A different behaviour can
be seen for language-specific mobility: the language B is either quickly extinct (dB > dA ) or is relatively persistent
(dB < dA ). Numerical results are averages over 100 independent samples.
(a slight deviation for ρ = 0.1 can be attributed to the
finite size effects). It is thus a strong evidence that for
migrating agents, the domain dynamics in the Naming
Game is also driven by an effective surface tension.
Let us notice that the relation τ ∼ M 2 is expected
to hold when the bubble and its surroundings are thermodynamically equivalent phases. In the Ising models
it means that there is no external magnetic field, which
would favour one of them. In the Naming Game, we
have also such symmetry since the dynamics of the Naming Game does not favour any of the languages used by
agents. We do not present here our additional numerical
results, though we have also measured some other characteristics of the Naming Game with migration (such as
the average time needed to reach a consensus for a system initialized with randomly assigned languages) and
they qualitatively agree with the ordinary ρ = 1 version [15, 16].
Instead, we would like to examine an extension of the
above defined model, in which to each language its own
(thus language-specific) mobility d is assigned. An agent
that changes its language changes thereby also its mobility (which, we hope, might reflect the behaviour in some
human communities). Now, if the chosen neighbouring
site is empty, Speaker migrates to this empty site with
the probability d corresponding to the language it uses.
If Speaker has several languages in its inventory, then one
of them is selected randomly to determine the probability
of migration (though only a very small fraction of agents
have more than one language in its inventories).
Let us notice that such modification affects only the
dynamics of migration while the Naming Game remains
3
200
0.8
180
0.7
160
0.6
140
0.5
ρB
120
100
dA=0.7, dB=0.3
dA=0.3, dB=0.7
dA=0.55, dB=0.45
dA=0.45, dB=0.55
dA=0.51, dB=0.49
dA=0.49, dB=0.51
dA=0.5, dB=0.5
0.4
0.3
80
0.2
60
0.1
40
0
0
20
20
40
60
80
100
120
140
160
180
200
t/1000
0
0
20
40
60
80
100 120 140 160 180 200
FIG. 3. The initial configuration (L = 200, ρ = 0.2) includes
in the center a square of size M = 30 with B-speaking agents
(red/gray), surrounded by A-speaking agents (green/light
gray). B-speakers are less mobile (dB = 0.1) than A-speakers
(dA = 1.0). The figure presents a configuration after t = 5·104
steps. Agents that have both A and B in their inventories are
marked in blue/dark gray.
symmetric. We determined the average time τ in a twolanguage version of this model (Fig. 2). When mobility dB of language B users is larger than mobility dA of
surrounding language A users, τ still increases but considerably slower and perhaps linearly τ ∼ M (the least
square fitting gives τ ∼ M 1.1 but a slight bending of our
data makes the asymtotic τ ∼ M very plausible). We
do not present the estimation of time for dB < dA since
it can be made only for very small M . In turn, for M
above a certain threshold value, the initial bubble instead
of shrinking starts to grow and eventually B-users engulf
the entire system. An example of such growth can be
seen in Fig. 3.
Such behaviour resembles the the behaviour of the
Ising model but in the presence of an external magnetic
field. When the magnetic field does not favour the bubble, a finite velocity of shrinking is expected for large M
and that would explain the growth τ ∼ M [20]. For the
field favouring the bubble, the existence of a threshold
size, above which the bubble will grow indefinitely, is also
a well-known feature. It suggests that in our model the
difference in mobility acts as a magnetic field in the Ising
model and breaks the symmetry of the Naming Game
favouring low-mobility languages.
To confirm that dA − dB is an analogue to the magnetic field in the Ising model, we made simulations of the
system initially divided (say vertically), in which the left
half is filled with agents of mobility dA and the right half
with agents of mobility dB . Depending on the sign of the
difference dA − dB , the interface should move (possibly
at constant speed) either to the left or to the right, and
only for dA = dB it shoud stay more or less in the inital
FIG. 4. The time evolution of the density ρB of language B
users. In the inital state, the lattice (L = 500) is divided
into two halfs containing only A- or B-speakers, respectively
(ρ = 0.7). The vertical interface thus created moves in the
direction which depends on the difference in mobilities of languages. Only for equal mobilities (dA = dB = 0.5), the interface remains immobile. The results are averages over 100
independent runs.
position. Our simulations fully confirmed such scenario
(Fig. 4). The interface always moves in such a way that
a less-mobile language becomes more widespread. Let us
notice that even a very small difference dA − dB is sufficient to favour one language over the other, and only for
precisely the same mobilities dA = dB , the languages are
equivalent.
Let us emphasize that the rules of the Naming Game
do not favour any of the languages, and the bias that
appears for unequal mobilities is generated dynamically.
In our opinion, the asymmetry appears due to a tendency of low-mobility languages to form clusters. Such
low-mobility clusters are relatively resistant upon interactions with high-mobility agents (Fig. 5). That lowmobility languages have a tendency to form clusters can
be seen also in Fig. 3. Indeed, the central area with the
less mobile language seems to be more densely filled than
its surroundings (and initially the entire lattice was filled
with the same density ρ = 0.2). Moreover, the interface
between the languages has a considerably lower density
than the interior of the area with the more mobile language. Apparently, high-mobility agents that are close to
the interface get intercepted by the low-mobility center
(and converted into low-mobility agents).
To support the above arguments, we made simulations where we measured the probability probA that a
small system (L = 6) starting with randomly distributed
agents will reach a consensus with language A. Initially,
languages A and B (and migrations dA and dB ) are
also randomly assigned to agents. Numerical simulations
show (Fig. 6) that only for the number of agents n = 2
and n = 36, we have probA = 0.5. For n = 2, only binary interactions of agents might take place (left panel
of Fig. 5) and the symmetry of the Naming Game im-
4
0.5
0.49
probA
0.48
0.47
0.46
0.45
0.44
dA=1, dB=0.1
dA=1, dB=0.5
0.43
0
5
10
15
20
25
30
35
n
FIG. 5. (Left) Due to the symmetry of the Naming Game, the
A-B pair with equal probabilities ends up in the A-A or B-B
state. When language A is more mobile, the pair is likely
to drift apart, while the B-B pair is more stable. (Right)
Apparently, the A-B symmetry is broken for more complex
interactions. Even if A happens to graft his language onto
a cluster, it is likely to drift appart. Then a single A user,
surrounded by three B neighbours, is likely to be converted
back to B. Thus less mobile clusters are relatively immune to
encounters with more mobile agents and the dynamics (effectively) favours less mobile languages.
plies that probA = 0.5. Similarly, for n = 36 migration
is suppressed and probA = 0.5 is the expected ordinary
Naming Game result. Simulations show, however, that
for any other value of n the symmetry of the model is
broken and the less mobile language (B) is effectively
favoured. It would be certainly desirable to have more
general understanding of the mobility-induced symmetry
breaking that takes place in our model. For example, one
might hope to develop some kind of a coarse-grained description of our model in terms of the Ginzburg-Landau
potential, an approach that turned out to be quite effective in some other agreement-dynamics models [21, 22].
IV.
MULTI-LANGUAGE VERSION
In the present section we examine the multi-language
version of our model. In such a case Fig. 7 shows that
the cluster-formation mechanism is also at work (simulations start from a random distribution of languages and
their mobilities). Initially mobilities were set randomly
from the range 0 < d < 1 but by the time t = 3 · 103 and
especially t = 104 , languages with the largest mobilities
(close to 1) were eliminated. One can clearly see the formation of low-mobility clusters, which grow by depleting
their surroundings from more mobile agents.
We also examined the time dependence of the average
FIG. 6. Probability that consensus will be reached on language A (dA = 1) as a function of the number of agents n.
Simulations were made for a small lattice with L = 6. For
each n we made 106 runs with a random distribution of agents
(initially, languages A and B were also assigned randomly).
Only for n = 2 (pair interactions) and n = 36 (no migration), the dynamics of the model remains symmetric and the
probability to end up with language A is the same as with
language B.
languages
mobility
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.3
0.25
0.2
0.15
0.1
0.05
0
FIG. 7. The configuration obtained after t = 5 · 102 (top),
3 · 103 (middle) and t = 104 (bottom) MC steps (L = 200,
ρ = 0.1). Low-mobility languages form more dense clusters,
surrounded by depleted zones, and gradually grow at the expense of more mobile neighbours. The color bars apply only
to the mobility panels.
5
agents and thus the remaining small-mobility agents are
primarily responsible for a considerably slower dynamics
(Fig. 9).
0
-0.5
5.5
-1
ρ=1.0
ρ=0.8
ρ=0.6
ρ=0.3
ρ=0.2
ρ=0.1
-1.5
4.5
4
-2
-2.5
log10(lM)
log10(d)
5
ρ=0.8
ρ=0.6
ρ=0.3
ρ=0.2
ρ=0.1
-3
-3.5
0
3.5
3
2.5
2
1
2
3
4
5
log10(t)
FIG. 8. The time dependence of the average mobility hdi;
simulation for L = 103 with random initial conditions and
averaged over 100 independent runs. The dash-dotted line
corresponds to the decay hdi ∼ t−0.9 .
mobility hdi in the system. Indeed, the numerical results in Fig. 8 confirm that hdi systematically decreases.
For a low density (ρ = 0.1), one can notice relatively
long initial plateaux, related to the fact that the system needs some time to build low-mobility clusters, and
only then the process that favours low-mobility languages
starts. Moreover, in the time decay of the average mobility, one can distinguish two power-law regimes. In the
high-density regime (ρ = 0.6 and 0.8), one has hdi ∼ t−0.9
while in the low-density regime (ρ = 0.1, 0.2, and 0.3),
the exponent is smaller than 0.9 and perhaps even varies
with density.
Certain Naming Game characteristics exhibit a similar power-law behaviour. In Fig. 9 we present the time
dependence of the number of users of the largest language lM . While in the high-density regime (ρ = 1.0, 0.8,
0.6), the increase seems to be universal and lM ∼ t0.8 ,
in the low-density regime (ρ = 0.3, 0.2, 0.1), the powerlaw behaviour has a density dependent exponent. The
behaviour of hdi and lM shows that the dynamics in the
high-density regime is much faster and the Naming Game
characteristics are very similar to those of an ordinary
Naming Game (with ρ = 1). The low-density regime has
a much slower dynamics and we relate such behaviour
to the formation of low-mobility clusters (Fig. 7). It is
likely that the model undergoes a phase transition around
ρ = 0.5, but its more detailed analysis is left for the future.
The two-language version described in the previous
section exhibits a symmetry breaking that speeds up the
dynamics. In the multi-language version, however, we
have initially the entire spectrum of languages and mobilities. The dynamics gradually eliminates large-mobility
[1] C. Castellano, S. Fortunato, and V. Loreto, Statistical
physics of social dynamics, Rev. Mod. Phys. 81, 591
1.5
1
0.5
0
1
2
3
log10(t)
4
5
6
FIG. 9. The time dependence of the number of users of the
most common language lM ; simulation for L = 103 with
random initial conditions and averaged over 100 independent
runs. The dashed line corresponds to the increase lM ∼ t0.8 .
V.
CONCLUSIONS
In summary, motivated by a possible mutual influence
of language formation and migration of human communities, we examined the Naming Game model with mobile
agents. As our main result, we have shown that even
a small difference in a language-specific mobility favours
a low-mobility language. Of course, taking into account
an extreme complexity of human interactions, we are not
even tempted to suggest that our model proves that languages of settlers should outperform nomadic ones, nevertheless, it certainly shows a strong relation between
language formation and migration. In our model, lowmobility languages form clusters and in a low-density
regime this process slows down the dynamics of the Naming Game. Let us also notice that the dynamics of a typical Naming Game (with ρ = 1) rather quickly leads to
the consensus, which not necessarily corresponds with a
relatively stable multi-language structure of the human
population [23]. With this respect, a slower dynamics
and a longer lifetime of the multi-language state (as suggested in Fig. 9) of the proposed model might be more
suitable. Finally, it should be also noted that the Naming Game is one of the models with the so-called agreement dynamics. The Voter or Ising models are yet other
well-known examples of this kind of models and some aspects of mobility in such systems were already examined
[24]. It would be, in our opinion, interesting to examine
their generalizations that taking into account the statedependent mobility.
(2009).
6
[2] D. Stauffer, Social applications of two-dimensional Ising
models, Am. J. Phys. 76, 470 (2008).
[3] N. C. Ellis and D. Larsen-Freeman, Language emergence:
Implications for applied linguistics—Introduction to the
special issue, Appl. Linguist. 27, 558-589 (2006).
[4] D. M. Abrams and S. H. Strogatz, Linguistics: Modelling
the dynamics of language death, Nature 424(6951), 900
(2003).
[5] M. Serva and F. Petroni, Indo-European languages tree
by Levenshtein distance, EPL-Europhys. Lett. 81, 68005
(2008).
[6] M. Patriarca and T. Leppänen, Modeling language competition, Phys. A 338, 296-299 (2004).
[7] A. M. Petersen et al., Languages cool as they expand:
Allometric scaling and the decreasing need for new words,
Sci. Rep. 2, 943 (2012).
[8] A. Puglisi, A. Baronchelli, and V. Loreto, Cultural route
to the emergence of linguistic categories, PNAS USA
105(23), 7936-7940 (2008).
[9] R. Reisenauer, K. Smith, and R. A. Blythe, Stochastic
dynamics of lexicon learning in an uncertain and nonuniform world, Phys. Rev. Lett. 110, 258701 (2013).
[10] S. Canagarajah (ed.), The Routledge Handbook of Migration and Language (Routledge, 2017).
[11] C. H. Williams (ed.), Language in Geographic Context
(Multilingual Matters Ltd, 1988).
[12] M. Czaika and H. de Haas, The Globalization of Migration: Has the World Become More Migratory?, Int.
Migration Rev. 48, 283-323 (2014).
[13] J. N. Jörgensen et al., Polylanguaging in superdiversity,
Diversities 13(2), 22-37 (2011).
[14] L. Steels, A self-organizing spatial vocabulary, Artif. Life
2, 319-332 (1995).
[15] A. Baronchelli, L. Dall‘Asta, A. Barrat, and V. Loreto,
Topology-induced coarsening in language games, Phys.
Rev. E 73, 015102 (2006).
[16] D. Lipowska and A. Lipowski, Phase transition and
fast agreement in the Naming Game with preference for
multi-word agents, J. Stat. Mech. Theor. Exp. 2014(08),
P08001 (2014).
[17] A. Baronchelli, M. Felici, V. Loreto, E. Caglioti, and
L. Steels, Sharp transition towards shared vocabularies in multi-agent systems, J. Stat. Mech. Theor. Exp.
2006(06), P06014 (2006).
[18] A. Baronchelli, L. Dall’Asta, A. Barrat and V. Loreto,
Non-equilibrium phase transition in negotiation dynamics, Phys. Rev. E 76, 051102 (2007).
[19] D. A. Huse and D. S. Fisher, Dynamics of droplet fluctuations in pure and random Ising systems, Phys. Rev. B
35, 6841 (1987); I. M. Lifshitz, Kinetics of ordering during second-order phase transitions, Zh. Eksp. Teor. Fiz.
42, 1354 (1962) [Sov. Phys. JETP 15, 939 (1962)].
[20] P. A. Rikvold, H. Tomita, S. Miyashita, and S. W. Sides,
Metastable lifetimes in a kinetic Ising model: Dependence on field and system size, Phys. Rev. E 49, 5080
(1994).
[21] L. Dall’Asta and T. Galla, Algebraic coarsening in voter
models with intermediate states, J. Phys. A: Math.
Theor. 41, 435003 (2008).
[22] F. Vazquez and C. López, Systems with two symmetric absorbing states: Relating the microscopic dynamics
with the macroscopic behavior, Phys. Rev. E 78, 061127
(2008).
[23] D. Lipowska and A. Lipowski, Naming game on adaptive
weighted networks, Artif. Life 18(3), 311-323 (2012).
[24] A. Baronchelli and R. Pastor-Satorras, Effects of mobility on ordering dynamics, J. Stat. Mech. Theor. Exp.
2009(11), L11001 (2009).