Simulation of Language Evolution based on Actual Diachronic Change

Extracted from Legal Terminology

Makoto Nakamura

, Yuya Hayashi

and Ryuichi Matoba

Japan Legal Information Institute, Graduate School of Law, Nagoya University,

1 Furo-cho, Chikusa-ku, Nagoya, Aichi, 464-8601, Japan

National Institute of Technology, Toyama College, 1-2, Ebie-neriya, Imizu City, Toyama, 933-0293, Japan

Keywords:

Language Evolution, Simulation, Iterated Learning Model, Cognitive Bias, Statute, Legal Terminology.

Abstract:

Simulation studies have played an important role in language evolution. Although a variety of methodologies

have been proposed so far, they are typically too abstract to recognize that their learning mechanisms properly

reﬂect actual ones. One reason comes from the lack of empirical data recorded for a long period with explicit

description. Our purpose in this paper is to show simulation models adapt to actual language change. As

empirical diachronic data, we focus on a statutory corpus. In general, statutes deﬁne important legal terms

with explanatory sentences, which are also revised by amendment. We proposed an iterated learning model, in

which an infant agent learns grammar through his/her parent’s utterances about legal terms and their semantic

relations, and the infant becomes a parent in the next generation. The key issue is that the learning situation

about legal terms and their relations can be changed due to amendment. Our experimental result showed that

infant agents succeeded to acquire compositional grammar despite irregular changes in their learning situation.

1 INTRODUCTION

A goal of the study on language evolution, or evo-

lutionary linguistics, is to explain the origins of the

structure found in language (Hurford, 2002). This

study has been increasingly interdisciplinary, involv-

ing collaborations between linguists, philosophers,

biologists, cognitive scientists, robotics, mathemati-

cal and computational modelers (Lyon et al., 2007).

In particular, simulation studies have played an im-

portant role in the ﬁeld of language evolution. A very

important function of simulation is to prove whether

a prediction actually and consistently derives from a

theory (Cangelosi and Parisi, 2002).

Although a variety of methodologies have been

proposed so far (Briscoe, 2002), they are typically too

abstract to recognize that their learning mechanisms

properly reﬂect the actual ones. Natural language,

however, is not such a simple phenomenon. Abstract

models could include only simple learning mecha-

nisms, which would be hard to deal with complicated

phenomena appearing in natural language. The main

challenge in language evolution is a lack of empiri-

cal data, that is, spoken language leaves practically

no traces. Therefore, it would be helpful if there are

language resources recorded in a long period with ex-

plicit description.

To solve this problem, we introduce a Japanese

statutory corpus. In particular, we focus on legal

terms deﬁned in a provision, each of which consists of

a tuple of a legal term and its explanatory sentences.

Legal statutes are not only established but also of-

ten amended by changes in social situations. In some

cases, legal terms are also revised, added, and deleted,

depending on the scale of the amendment. Therefore,

an amendment to provisions for legal terms implies

a drastic change of the entire act. The terminology

for legal terms must deal with such temporal changes

that are dependent on amendment acts. Our purpose

in this paper is to show the simulation models for lan-

guage evolution properly deals with actual language

changes.

We employ simulation models for grammar

acquisition based on the iterated learning model

(ILM) (Kirby, 2002), which shows a process of gram-

matical evolution through generations. This approach

has often been used in simulation models concerning

language evolution (Nakamura et al., 2015). One im-

portant reason for this comes from its robustness for

syntactic learning from input sentences. As long as

it is learning from a single parent, an infant agent re-

ceives sentences derived from a consistent grammar;

Nakamura M., Hayashi Y. and Matoba R.

Simulation of Language Evolution based on Actual Diachronic Change Extracted from Legal Terminology.

DOI: 10.5220/0006291903430350

In Proceedings of the 9th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2017), pages 343-350

ISBN: 978-989-758-219-6

343

it is possible to acquire a concise grammar.

This paper is organized as follows. In Section 2,

we explain how to extract data for actual language

changes from statutory texts. In Section 3, we intro-

duce ILMs and our proposed model, which are exam-

ined in Section 4. Finally, we conclude in Section 5.

2 DIACHRONIC CHANGES IN

LEGAL TERMS

In this section, we introduce diachronic changes in

legal terms. Section 2.1 explains how to extract legal

terms from statutory texts. Section 2.2 shows these

evolutionary changes with examples.

2.1 Extraction of Legal Terms

Figure 1 shows an excerpt from the act dealing with

the change of the term “Gas Business.” The amended

act is shown in Figure 2.

What are recognized as legal terms to be collected

depends on the purpose (Winkels and Hoekstra, 2012;

Nakamura et al., 2016). In this paper, we deﬁne le-

gal terms as those explicitly deﬁned prior to use in a

statute, each of which consists of a tuple of a legal

term in the quotations and its explanation.

An article for deﬁnition of legal terms often con-

sists of a number of paragraphs, each of which deﬁnes

a legal term. They are described with boilerplate ex-

pressions including a legal term and its explanatory

sentence, which can be extracted with a set of regular

expression rules. The underlined phrases

in Figure 1

match one of the rules. As a result, the legal term

“Gas Business” and its explanation can be extracted.

A deﬁned term also appears in parentheses fol-

lowing a phrase as its explanation in the main text.

Abbreviations of terms are often deﬁned in parenthe-

ses. An example is shown in Figure 2, where the

term “Speciﬁed Gas Generating Facility” is deﬁned

in parentheses. We extracted the explanation, the un-

derlined part

in Figure 2, from the beginning of the

deﬁnition to just before the beginning of the parenthe-

ses. Note that some explanatory sentences of a term

include other legal terms as its hypernym or hyponym,

which enables us to extract hyponymy relations be-

tween legal terms.

We referred to the Japan Law Translation Database

System (http://www.japaneselawtranslation.go.jp/) for the

English translation of these acts. When there is no trans-

lation for the acts or act titles in the website, we manually

translated them using the database.

The original statute does not include the underlines,

which were added by the author.

Gas Business Act (Act No. 51 of 1954)

✓ ✏

(Deﬁnitions)

Article 2 (1)

The term “Gas Business” as used

in this Act shall mean the business of supplying

gas via pipelines to meet general demand.

✒ ✑

Figure 1: Excerpt from the Gas Business Act (Act No. 51

of 1954).

Gas Business Act (Act No. 51 of 1954)

✓ ✏

[As of October 12, 1970]

(Deﬁnitions)

Article 2 (1) The term “General Gas Utility

Business” as used in this Act shall mean the

business of supplying gas via pipelines to meet

general demand (excluding, however, businesses

generating gas at a gas generating facility pre-

scribed in paragraph (3) and supplying such gas

via pipelines). (*snip*)

(3) The term “Community Gas Utility Busi-

ness” as used in this Act shall mean

the business

of generating gas at a simpliﬁed gas generating

facility speciﬁed by aCabinet Order (hereinafter

referred to as a “Speciﬁed Gas Generating Fa-

cility”) and supplying such gas via pipelines to

meet general demand at not less than 70 gas ser-

vice points within one housing complex.

(*snip*)

(5) The term “Gas Business” as used in this

Act shall mean a General Gas Utility Business

or Community Gas Utility Business.

✒ ✑

Figure 2: Excerpt from the Gas Business Act (Act No. 51

of 1954) as of October 12, 1970.

We successfully extracted legal terms, their ex-

planations, and their relations. We found 27,737

terms and 36,698 relations. The precision of the re-

lations was 88.0%, which might be improved with

additional regular expression rules (Nakamura et al.,

2016). The diachronic terminology shows legal terms

evolves reﬂecting the social change, although legal

documents are the most conservative in terms of lan-

guage change.

2.2 Example of Diachronic Change

We cite the Gas Business Act (Act No. 51 of 1954)

as an example to explain diachronic changes in le-

gal terms. As of 2013, this act has been amended

34 times, at least ﬁve of which include the revision

of terms and deﬁnitions in Articles 2 and 39-2. Fig-

ures 3(a) to 3(c) show the diachronic changes in the

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

344

terms at three time points:

(1) At the new enactment, only two terms, “Gas Busi-

ness” and “Gas Facilities,” were deﬁned in the

Gas Business Act (Act No. 51 of 1954), which

came into effect as of April 1, 1954 (Figure 3(a)).

(2) The term “Gas Business” was changed to “Gen-

eral Gas Utility Business,” which became a hy-

ponym of the newly deﬁned term “Gas Business”

with the newly added term “Community Gas Util-

ity Business,” by the Act on the Partial Amend-

ment of the Gas Business Act (Act No. 18 of

1970), which came into effect as of October 12,

1970. Likewise, the term “Gas Supplier” and its

hyponyms “General Gas Utility” and “Commu-

nity Gas Utility” were deﬁned. So was the term

“Speciﬁed Gas Generating Facility” as an isolated

term. Note that, unlike language changes as a nat-

ural phenomenon, the sense of legal terms was

forced to change on the enforcement date (Fig-

ure 3(b)).

(3) As of the enforcement of the Act for the Partial

Revision of the Electricity Business Act and the

Gas Business Act (Act No. 92 of 2003), the num-

ber of terms deﬁned in the Gas Business Act was

increased to 15 (Figure 3(c)). In the period be-

tween (2) and (3), the terms “Class-I Gas Equip-

ment,” “Class-II Gas Equipment” and “Wholesale

Supply” were deﬁned, but deleted later. In addi-

tion, the term “Intra-Area Wheeling Service” was

replaced with “Wheeling Service.” These were

basically eliminated by social selection.

3 ITERATED LEARNING

MODELS

In this section, we introduce iterated learning mod-

els for learning grammar in the environment of legal

terms. First, we brieﬂy explain Kirby’s ILM (KILM).

Next, we introduce the modiﬁcation for taking cogni-

tive biases into account. Finally, we make some minor

changes to adapt the model to the new environment.

3.1 KILM

Kirby (Kirby, 2002) introduced the notions of com-

positionality and recursion as fundamental features of

grammar, and showed that they make it possible for a

human to acquire compositional language. Figure 4

illustrates KILM. In each generation, an infant can

acquire grammar in his/her mind given sample sen-

tences from his/her mother. After growing up, the

infant becomes the next parent to speak to a new-

born baby with his/her grammar. As a result, infants

can develop more compositionalgrammar through the

generations. Note that the model focuses on the gram-

mar change in multiple generations, not on that in one

generation. Also, Kirby adopted the idea of two dif-

ferent domains of language (Bickerton, 1990; Chom-

sky, 1986), namely, I-language and E-language; I-

language is the internal language corresponding to

a speaker’s intention or meaning, while E-language

is the external language, that is, utterances. In his

model, a parent is a speaker agent and his/her infant

is a listener agent. The speaker agent gives the lis-

tener agent a pair of a string of symbols as an utter-

ance, and a predicate-argument structure (PAS) as its

meaning. A number of utterances would form com-

positional grammar rules in a listener’s mind, through

the learning process. This process is iterated genera-

tion by generation, and converges to a compact, lim-

ited number of grammar rules.

According to KILM, the parent agent gives the in-

fant agent a pair of a string of symbols as an utter-

ance, and PAS as its meaning. The agent’s linguistic

knowledge is a set of a pair of a meaning and a string

of symbols, as follows.

S/love( john,mary) → hjsbs, (1)

where the meaning, i.e., the speaker’s intention, is

represented by a PAS love(john, mary) and the string

of symbols is the utterance “hjsbs”; the symbol ‘S’

stands for the category Sentence. The following rules

can also generate the same utterance.

S/love(x,mary) → h N/x sbs

N/ john → j,

(2)

where the variable x can be substituted for an arbitrary

element of category N.

The infant agent has the ability to generalize

his/her knowledge with learning. This generaliz-

ing process consists of the following three opera-

tions (Kirby, 2002): chunk, merge, and replace.

Chunk. This operation takes pairs of rules and looks

for the most-speciﬁc generalization.



S/love( john, pete) → ivnre

S/love(mary, pete) → ivnho

⇒







S/love(x, pete) → ivn N/x

N/ john → re

N/mary → ho.

(3)

Merge. If two rules have the same meanings and

strings, replace their non-terminal symbols with

one common symbol.

Simulation of Language Evolution based on Actual Diachronic Change Extracted from Legal Terminology

345

(a) Legal terms and relations in the Gas Business Act (Act No. 51 of 1954)

(b) Legal terms and relations as of enforcement of the Act on the Partial Amendment of the Gas

Business Act (Act No. 18 of 1970)

tricity Business Act and the Gas Business Act (Act No. 92 of 2003)

Figure 3: Dynamics of deﬁnitions in the Gas Business Act.

Figure 4: Iterated learning model.

Replace. If a rule can be embedded in another rule,

replace the terminal substrings with a composi-

tional rule.

In Kirby’s experiment (Kirby, 2002), a constant

number of predicates and object words (ﬁve for exam-

ple) are employed. Also, two identical arguments in

a predicate like love(john, john) are prohibited. Thus,

there are 100 distinct meanings (5 predicates × 5 pos-

sible ﬁrst arguments × 4 possible second arguments)

in a meaning space.

The key issue in ILM is to create a poverty of stim-

ulus, which explains the necessity of universal gram-

mar (Chomsky, 1980). Kirby (Kirby, 2002) modeled

it as learning through bottlenecks, which are rather

necessary for the learning. As long as an infant agent

is given all sentences in the meaning space during

learning, he/she does not need to make a composi-

tional grammar; he/she would just memorize all the

meaning-sentence pairs. Therefore, agents are given

a part of sentences in the whole meaning space. The

total number of utterances the infant agent receives

during learning is parameterized. Since the num-

ber of utterances is limited, the infant agent cannot

learn the whole meaning space; thus, to obtain the

whole meaning space, the infant agent has to gen-

eralize his/her own knowledge by self-learning, i.e.,

chunk, merge, and replace. The parent agent receives

a meaning selected from the meaning space, and ut-

ters it using his/her own grammar rules. When the

parent agent cannot make an utterance because of a

lack of grammar rules, he/she invents a new rule. This

process is called invention. Even if the invention does

not work to complement the parent agent’s grammar

rules, he/she utters a randomly composed sentence.

3.2 Meaning Selection ILM with

Cognitive Biases (MSILMB)

The iterated learning model has been expanded for

examining the relationship between language acqui-

sition and cognitive biases. Several studies have sug-

gested that cognitive biases work effectively in the

ﬁrst language acquisition (Imai and Gentner, 1997;

Markman, 1990). Cognitive bias, which is common to

all human beings, involves systematic errors in judg-

ment and decision-making due to cognitive limita-

tions, motivational factors, and/or adaptations to nat-

ural environments (Wilke and Mata, 2012).

These biases work in a joint attention framework

where two individuals, a parent and an infant, share a

state of an environment. For example, the parent and

infant are looking at a rabbit which are eating car-

rots, and the parent utters ‘Gavagai.’ In this situation,

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

346

Figure 5: Learning Process on Meaning Selection ILM.

the infant cannot infer the meaning of ‘Gavagai’ log-

ically, i.e., there are many possibilities of its mean-

ing, such as a rabbit, a white animal or an action of

eating. This problem is well known as Gavagai prob-

lem (Quine, 1960). The learning environment of in-

fants in ﬁrst language acquisition is very close to this

situation. However, they overcome this problem, and

acquire their ﬁrst language at an overwhelming pace.

For this infants’ phenomenal learning, several studies

have suggested that the infants infer meanings efﬁ-

ciently to limit possibilities in a situation using con-

straints, that is cognitive biases, and identify a mean-

ing of the utterance (Imai and Gentner, 1997; Hansen

and Markman, 2009).

The Meaning Selection ILM (MSILM) (Sudo

et al., 2016) employs the notion of a joint attention

frame in KILM. Figure 5 shows an image of the learn-

ing process on MSILM, in which both parent and in-

fant agents share a situation. The parent agent selects

a part of the situation (M

) which contains multiple

meanings {M

,... ,M

}, and utters U about it to the

infant agent. Once receiving the utterance, he/she

infers its meaning from the presented situation, and

learns a pair of the utterance U and its meaning M

Thus, the infant agent does not always infer the same

pair as the parent’s knowledge, that is, the infant agent

would acquire the grammar rule S/M

→ U, while the

parent’s utterance was derived from his/her knowl-

edge of S/M

→ U.

Some cognitive biases have been employed to

MSILM, and their effectiveness was veriﬁed on sim-

ulation of the ﬁrst language acquisition (Sudo et al.,

2013). Hereafter, we call MSILM with cognitive bi-

ases MSILMB, which employed the following biases:

Symmetry Bias: When the predicate p → q is true,

the symmetry bias allows humans to mislead q →

p being also true. A pair of p and q is put as a

pair of label and object (Imai and Gentner, 1997),

a pair of meaning and utterance (Matoba et al.,

2010) and so on. If an infant agent can gener-

ate the same utterance as the parent agent’s, and

its meaning is found in the presented meaning,

he/she connects the utterance and the meaning.

Otherwise, the infant agent selects one out of the

presented meanings randomly.

Mutual Exclusivity Bias: This is the assumption

that only one label can be applied to each object

in early word learning (Markman, 1990). If an in-

fant agent has already acquired the grammar rule

S/M → U

, he/she dose not connect the meaning

M to any other utterances. In other words, if the

infant agent can generate the utterance of a pre-

sented meaning M

′

in a situation with his/her ac-

quired grammar and the generated utterance is not

the same as the parent agent’s, the infant agent de-

selects M

′

from the candidate of the meaning of

the parent agent’s utterance.

These biases work as “one utterance to one mean-

ing (symmetry bias),” and “one meaning to one utter-

ance (mutual exclusivity bias),” i.e., the effect of these

biases gives a one-to-one relation between a meaning

and an utterance to the infant agent under a multiple

cognition environment like a joint attention frame.

3.3 Our Model

We basically employ MSILMB in infant agents’

learning process. We assume that a parent-infant pair

shares their situation in the environment.

Our model differs from the former one (Sudo

et al., 2013) as follows:

• The situation can be changed about a set of legal

terms and their relations by amendment from gen-

eration to generation. Some terms and relations

remain the same, while the others are deleted,

newly added, or replaced with others.

• The actual meaning space is a subset of the

whole meaning space; speakers can only speak a

scene of the actual situation, while they randomly

choose a meaning from the whole meaning space

in the former model.

• The number of utterances varies depending on the

situation. It is calculated as (the size of the whole

meaning space in generation)/2.

4 EXPERIMENTS

4.1 Experimental Settings

Our experiments aim to examine whether KLIM and

MSILMB properly work in actual situations over the

diachronic change. In order to reproduce the actual

Simulation of Language Evolution based on Actual Diachronic Change Extracted from Legal Terminology

347

diachronic change, we pick up legal terms deﬁned

in the Gas Business Act (Act No. 51 of 1954) from

the Japanese statutory corpus. In the situation, agents

have knowledge about objects corresponding to legal

terms and their relations. One generation in the simu-

lation corresponds to a year in the actual world.

We deﬁned three predicates; one represents a

state of isolation is

isolated, and the others are

for hyponymy relations, that is, is

hypernym of and

is hyponym of. The number of objects n varies along

with amendments. Therefore, the size of the whole

meaning space is calculated as (2 predicates for hy-

ponymy relations × n possible ﬁrst arguments ×

(n− 1) possible second arguments + 1 predicate for

a state of isolation × n possible arguments). Table 1

shows the size of the whole meaning space, the size

of actual meaning space and the number of utterances

in each generation, which corresponds to years after

enforcement of the act.

In KILM, agent’s linguistic knowledge is evalu-

ated by expressivityand the numberof grammar rules.

Expressivity is deﬁned as how much of the whole

meaning space the agent can utter with his/her gram-

mar rules. In MSILMB, it is important how much

the infant agent acquired language close to the parent

agent’s. Therefore, to evaluate the similarity between

two languages, we employed the language inﬂuence

rate, which is based on similarity between two char-

acter strings. The language inﬂuence rate is calculated

at the end of each generation by comparing between

the enumerations of sentences whom both parent and

infant agents can utter with their grammar rules (See

(Nakamura et al., 2015) for more details of the lan-

guage inﬂuence rate, which is called language dis-

tance).

An utterance is expressed with a string of 10 types

of letters. Agents can invent an utterance by invention

in a range of 2 to 4 letters. A trial of the simulation

stops at the 60th generation, which is the span be-

tween the enforcement of the Gas Business Act and

that of the last amendment act with a 10-year-blank

for learning.

We compare experimental results of KILM and

MSILMB; in the former, since infants receive a pair

of a meaning and a string of symbols from parents,

agents are expected to acquire compositional gram-

mar. Meanwhile in the latter, infants receives only a

string of symbols, which implies they need to infer

what parents talk about. The learning is more difﬁ-

cult, but its learning environment is close to the actual

one.

4.2 Experimental Results

We show experimental results in Figure 6, in which

all the lines are plotted by average of 100 trials.

Figure 6(a) shows the rate of actual meanings

infants can utter by generation. This shows how

much agents can represent actual situations using

their grammar rules. Note that the actual meaning

space is a subset of the whole meaning space. For ex-

ample, in the generations from 1 to 16 corresponding

to Figure 3(a), there are only two actual situations, de-

noted by the meanings: is

isolated(GasBusiness) and

is isolated(GasFacilities). If the parent agent utters

both of them in the limited number of utterances, that

is 3, his/her infant can learn both utterances. In addi-

tion, if those utterances share a common substring in

different meanings, the infant can extract a composi-

tional rule denoting is

isolated(X) by chunking.

Overall, since this process is common in KILM

and MSILMB, the result shows almost the same.

Since the number of utterances is much more than that

of the actual meanings, infant agents are likely to lis-

ten to all kinds of utterances for the actual meanings

in a generation.

Figure 6(b) shows expressivity in the whole mean-

ing space by generation. The thin solid line de-

notes the border of compositionality, which is calcu-

lated as (Size of actual meaning space)/(Size of the

whole meaning space). If the expressivity exceeds

it, the grammar is regarded as compositional. In

the early generations, expressivity in the both models

shows around 0.3, because there are 2 actual mean-

ings against 6 possible patterns. After the ﬁrst amend-

ment at the 17th generation, the expressivity suddenly

dropped down. This is because the number of possi-

ble meaning patterns rose from 6 to 120 due to the

growing number of legal terms from 2 to 8.

Figure 6(c) shows the normalization by the num-

ber of actual meanings for Figure 6(b). Therefore, the

value less than 1 implies agents do not learn composi-

tional grammar, which suggests that agents can utter

little other than what they heard until the 31st genera-

tion. This is because agents receive 60 utterances for

10 actual meaning patterns. In other words, they are

not exposed to poverty of stimulus, which is supposed

to promote learning compositional grammar. As a re-

sult, they just seem to memorize all the utterances

from their parents during the 17th to 31st generation.

We can see if agents learn compositional gram-

mar by checking the number of grammar rules they

acquired. Figures 6(d) and 6(e) show the number of

rules by generation in KILM and MSILMB, respec-

tively. Note that decreasing sentence rules and in-

creasing word rules imply that acquired grammar is

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

348

Table 1: Number of utterances based on meaning space in each generation.

Year From 1954 1970 1985 1994 1999 2003

To 1969 1984 1993 1998 2002 2014

Generation 1-16 17-31 32-40 41-45 46-49 50-60

Number of objects 2 8 10 13 14 15

Number of isolated objects 2 2 1 1 3 2

Number of hyponymy relations 0 8 12 18 16 20

Size of the whole meaning space 6 120 190 325 378 435

Size of the actual meaning space 2 10 13 19 19 22

Number of utterances 3 60 95 163 189 218

0.2

0.4

0.6

0.8

0 10 20 30 40 50 60

Expressivity

in the actual meaning space

Generation

KILM

MSILMB

(a) Expressivity in the actual meaning

space

0.2

0.4

0.6

0.8

0 10 20 30 40 50 60

Expressivity

in the whole meaning space

Generation

KILM

MSILMB

Actual Space

(b) Expressivity in the whole meaning

space

0.5

1.5

2.5

0 10 20 30 40 50 60

Normalized expressivity

Generation

KILM

MSILMB

0 10 20 30 40 50 60

Number of rules

Generation

S+W

(d) Number of grammar rules in KILM

(S: sentence, W: word)

100

0 10 20 30 40 50 60

Number of rules

Generation

S+W

(e) Number of grammar rules in

MSILMB (S: sentence, W: word)

0.2

0.4

0.6

0.8

0 10 20 30 40 50 60

Language influence rate

Generation

KILM

MSILMB

(f) Language inﬂuence rate

Figure 6: Experimental Results.

compositional. During the 17th to 31st generation,

agents in KILM keep around 10 sentence rules for

10 actual meanings, while those in MSILMB seem

to have a decreasing number of sentence rules. The

same can be seen during the 32nd to 40th genera-

tion. Since the symmetry bias works well, agents

make similar utterances regard to have similar mean-

ings. Furthermore, the mutual exclusivity bias pre-

vent agents from making inconsistent rules, which fa-

cilitates acquisition of compositional grammar. As a

result, the normalized expressivity of MSILMB ex-

ceeds that of KILM from the 32nd generation due to

high compositionality.

Learning compositional grammar enables agents

to represent even inexperienced situations. From this

viewpoint, as far as seeing Figure 6(b), neither KILM

nor MSILMB is enough for representing the whole

meaning space. This is because the learning period in

60 generations is too short.

The problem of MSILMB is that grammatical

learning is far from matching actual meanings. In

fact, it is not always true that similar utterances have

similar meanings. Figure 6(f) shows the language in-

ﬂuence rate by generation. KILM keeps higher than

MSILMB in the inﬂuence rate, which denotes agents

in KILM are more likely to speak similar language

to their parents. This is characteristics of MSILMB,

in which learning compositionality takes priority over

interpretation of utterances. For example in the sit-

uation of Figure 3(a), when a parent agent utters

“

bus

” and “

fac

” for is isolated(GasBusiness) and

isolated(GasFacilities), his/her infant agent may

misunderstand the former utterance corresponds to

the latter meaning, and vice versa. This phenomenon

could cause the decrease of the language inﬂuence

rate.

Through the experiments, we showed agents in

MSILMB properly acquired compositional grammar

in the actual situation of language change. An excerpt

of grammar rules that the infant agent acquired at the

60th generation in a trial is shown in Equation (4).

Simulation of Language Evolution based on Actual Diachronic Change Extracted from Legal Terminology

349











S/x

) → C

/is

hypernym of → ei

/is

hyponym of → a

/CommunityGasUtilityBusiness → e

/LargeVolumeGasBusiness → e

/GasBusiness → febja

S/isolate(x

) → e C

/GasFacilities → hbfj

/WheelingService → hieb,

(4)

where S, C

, C

and C

are non-terminal sym-

bols as a category name, and x

, x

and x

are vari-

ables.

5 CONCLUSION

In this paper, we introduced a diachronic legal ter-

minology to simulation models to conﬁrm proposed

models properly deal with natural language phenom-

ena. Legal terms are deﬁned in statutes, in which they

are added, deleted, or replaced with others reﬂecting

social change. Therefore, the change of legal terms

and their relations is unstable and not coherent.

We used KILM and MSILMB for learning com-

positional grammar under the environment. As a re-

sult, KILM showed agents acquired less composi-

tional grammar due to a difﬁcult learning environ-

ment. Meanwhile in MSILMB, agents succeeded to

deal with the environment, although their grammar is

less inﬂuenced on their parents’ one.

Our achievement could contribute to not only lan-

guage evolution, but also some novel ﬁeld of language

processing, because it is a part of huge and complex

problem of creation of systems with learning (self-

learning) abilities to the reactions in previously un-

known situations. For example, it would be useful for

creation of constantly expending library of actions for

robots working on another planets.

Integration of language evolutionand legal knowl-

edge is a challenging theme. Our analysis of the statu-

tory corpus revealed that statutes are excellent data

for pursuing actual language change. Although we,

in this paper, chose the Gas Business Act by chance,

further analysis would lead to synthetic characteris-

tics of legal terms.

ACKNOWLEDGEMENT

This work was partly supported by JSPS KAKENHI

Grant Numbers JP15K00201, JP15K16013.

REFERENCES

Bickerton, D. (1990). Language and Species. University of

Chicago Press.

Briscoe, E. J., editor (2002). Linguistic Evolution through

Language Acquisition: Formal and Computational

Models. Cambridge University Press.

Cangelosi, A. and Parisi, D., editors (2002). Simulating the

Evolution of Language. Springer, London.

Chomsky, N. (1980). Rules and Representations. Basil

Blackwell, Oxford.

Chomsky, N. (1986). Knowledge of Language:Its Nature,

Origin, and Use. Praeger, New York.

Hansen, M. B. and Markman, E. M. (2009). Children’s use

of mutual exclusivity to learn labels for parts of ob-

jects. Developmental Psychology, 45(2):592–596.

Hurford, J. R. (2002). The Roles of Expression and Repre-

sentation in Language Evolution. In The Transition to

Language, pages 311–334. Oxford University Press,

Cambridge.

Imai, M. and Gentner, D. (1997). A cross-linguistic study

of early word meaning: Universal ontology and lin-

guistic inﬂuence. Cognition, 62(2):169–200.

Kirby, S. (2002). Learning, bottlenecks and the evolution

of recursive syntax. In Briscoe, T., editor, Linguistic

Evolution through Language Acquisition: Formal and

Computational Models, chapter 6. Cambridge Univer-

sity Press.

Lyon, C., Nehaniv, C., and Cangelosi, A., editors

(2007). Emergence of Communication and Language.

Springer.

Markman, E. M. (1990). Constraints children place on word

meanings. Cognitive Science, 14(1):57–77.

Matoba, R., Nakamura, M., and Tojo, S. (2010). Efﬁciency

of the symmetry bias in grammar acquisition. Infor-

mation and Computation, 209(3):536–547.

Nakamura, M., Matoba, R., and Tojo, S. (2015). Simulation

of Emergence of Local Common Languages Using It-

erated Learning Model on Social Networks. Inter-

national Journal on Advances in Intelligent Systems,

8(3&4):374–384.

Nakamura, M., Ogawa, Y., and Toyama, K. (2016). De-

velopment of Diachronic Terminology from Japanese

Statutory Corpora. Journal of Open Access to Law,

4(1):16 pages.

Quine, W. V. O. (1960). Word and Object. MIT Press.

Sudo, H., Matoba, R., Cooper, T., and Tsukada, A. (2016).

Effect of the Symmetry Bias on Linguistic Evolution.

Artiﬁcial Life and Robotics, 21(2):207–214.

Sudo, H., Matoba, R., Hagiwara, S., Nakamura, M., and

Tojo, S. (2013). Knowledge Revision based on Ef-

ﬁcacy of Cognitive Biases in First Language Acquisi-

tion. In Proceedings of the 30th Annual Meeting of the

Japanese Cognitive Science Society, pages 343–349.

Wilke, A. and Mata, R. (2012). Cognitive Bias. Encyclope-

dia of Human Behaviour, 1:531–535.

Winkels, R. and Hoekstra, R. (2012). Automatic Extraction

of Legal Concepts and Deﬁnitions. In Legal Knowl-

edge and Information Systems - JURIX 2012: The

Twenty-Fifth Annual Conference, pages 157–166.

ICAART 2017 - 9th International Conference on Agents and Artiﬁcial Intelligence

350