Automated Compliance Veriﬁcation in ATM using Principles from

Ontology Matching

Audun Vennesla nd

1,2

, Joe Gorman

, Scott Wilson

, Bernd Neumayr

and Christoph G. Schuetz

Department of Software Engineering, Safety and Security, SINTEF, Trondheim, Norway

Norwegian University of Science and Technology, Trondheim, Norway

EUROCONTROL, Brussels, Belgium

Johannes Kepler University Linz, Linz, Austria

Keywords:

Compliance Veriﬁcation, Air Trafﬁc Management, Ontology Matching, Ontology Engineering.

Abstract:

Compliance with standard information models in diverse and complex domains such as Air Trafﬁc Mana-

gement is an important but highly challenging task. The challenges stem from the fact that the information

models are often extensive, the diversity of the domain leads to diverging terminology, and the manual mapping

of information elements necessary to assess compliance is very labor-intensive. This work proposes ways in

which compliance veriﬁcation techniques, currently based on manual t echniques, can be supported and partly

automated by means of a set of basic ontology matching techniques. We have evaluated these techniques in an

experiment involving seven datasets consisting of various ATM ontologies that have been transformed to OWL

from t heir original UML representations. A comparative analysis with two other state-of-the-art matching sy-

stems shows that some of our proposed matching techniques obtain good quality alignments, especially when

they are combined using simple strategies. The evaluation also reveals that identifying equivalence relations

is a far easier task than identifying other types of semantic relations.

1 INTRODUCTION

Within Air Trafﬁc Mana gement (ATM), the ATM In-

formation Reference Model (AIRM) is the standard

informa tion reference model. Information models

targeting information exchange in the ATM network

should be compliant with the AIRM in order to foster

interoperability. However, assuring such compliance

currently requires a signiﬁcant amount of manual ef-

fort both during model development, when modellers

investigate potentially re-usable elements in the refe-

rence mo del, and after com pletion of the model when

its compliance with the AIRM has to be veriﬁed and

maintained for governance purposes.

Several initiatives in the realm of aviation have

investigated the feasibility of introd ucing semantic

technologies as a means for improving informatio n

management. One of the latest contributions to se-

mantic developments in ATM is the NASA ATM On -

tology

. Furthermore , the H orizon 2020

BEST p ro-

ject

looks at ATM inf ormation management an d r e -

https://data.nasa.gov/ontologies/atmonto/index.html

https://ec.europa.eu/programmes/horizon2020/

http://www.project-best.eu/

presents AIRM as well as other information models

(for expressing a eronautical information and weather

informa tion) as OWL (Web Ontology Language) on-

tologies.

Ontology matching, a sub-discipline of ontology

engineer ing, investigates techniques for (semi) auto-

matic identiﬁcation of semantic correspondences bet-

ween onto logies. Our hypothesis is that o ntology ma-

tching techniques lend themselves well to p rovide au-

tomated support for compliance veriﬁcation, and can

reduce much of the human effort that is curre ntly re-

quired for compliance veriﬁcation in ATM. Further-

more, such automated support can motivate re-use of

standardised in formation elements in ATM, preven-

ting interoperability threats and unnecessary use of

development resources.

Based on this, we investigate to what extent onto-

logy matching pr inciples can offer automated support

for identifying semantic correspond e nce between dif-

ferent ATM-related mode ls and the AIRM. While the

quality of ontology matching systems has improved in

recent years, there is no superior system or techniqu e

that performs best in all co ntexts and settin gs. We

therefore suggest an app roach where the input ATM

Vennesland, A., Gorman, J., Wilson, S., Neumayr, B. and Schuetz, C.

Automated Compliance Veriﬁcation in ATM using Principles from Ontology Matching.

DOI: 10.5220/0006924000410052

In Proceedings of the 10th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2018) - Volume 2: KEOD, pages 41-52

ISBN: 978-989-758-330-8

ontologies ar e proﬁled according to a set of ontology

analysis metrics b e fore they are matched.

The performance of ontology ma tching systems is

evaluated in the Ontology Alignment Evaluation Ini-

tiative (OAEI), an ann ual evaluation campaign for on-

tology matching systems. Many systems obtain close

to pe rfect alignment quality in OAEI. However, one

may argue that some of the more popular track s in

OAEI are simple and limited in sco pe. Moreover,

many of the participating ontology matching systems

are very much tuned in on the ontologies represented

in the OAEI tracks as there have bee n few changes

made to the tracks over the last year (E uzenat et al.,

2011). Additional benchmarks, preferably involving

ontologies describing knowledge models new to the

research area are sought. We have developed our own

datasets for evaluating the quality of our tec hniques.

The datasets consists of ATM ontologies and grou nd

truth mappings veriﬁed by human experts in the ATM

ﬁeld.

This research addresses the following research

questions:

• To what extent can ontology matching principles

support inform a tion modeling and compliance ve-

riﬁcation processes within ATM?

• Which ontology matching techniques perform

better this context?

• Are these techniques capable of capturing the full

range of sema ntic correspondences deﬁned by hu-

man experts?

The main contributions fro m this work are:

• We provide datasets consisting of real ontologie s

from the ATM domain that c a n be used to evaluate

other ontology matching systems and technique s.

The reference alignments in these datasets are ve-

riﬁed by experts in the AT M domain.

• We describe matching techniques for identifying

equivalence and other semantic relations as well

as a set of strategies for combining the ir computed

alignments. The performance of the techniques

and combinatio n strategies is compared against

well-acknowledged and state of the art ontolo gy

matching systems.

2 BACKGROUND

2.1 Semantic Interoperability in SWIM

System Wide Information Management (SWIM) co-

vers a complete change in paradigm of how informa-

tion is managed along its full lifecycle and across the

whole European ATM system

. O ne objective is the

establishment of a network-centric informatio n envi-

ronment in Europe, in contrast to toda y’s informa-

tion management which is typically based on point-

to-poin t message transfer, limited use of standards,

and tightly co upled APIs hinder ing interop erability.

It is further recognised that global interoperability and

standardisation are essential.

2.1.1 ATM Informatio n Reference Model

The ATM Information Reference Model (AIRM) is

one of the essential elements of realising SWIM.

AIRM is a reference model that addresses semantic

interoperability through h a rmonised and a greed deﬁ-

nitions of the information being exchanged in ATM

AIRM is formalised in UML (Uniﬁed Modeling Lan-

guage) .

The AIRM model is organised into different sub-

ject ﬁelds, where each subject ﬁeld includ e s elements

for particular areas of ATM, such as the aircraft,

the airpo rt infrastructure, meteorological information,

etc.

Semantic interoperability within ATM is accom-

plished by ensuring that all information being excha n-

ged within ATM is compliant with the intended se-

mantics as deﬁned in AIRM.

2.2 Exchange Models

ATM Exchange models deﬁne the structure and con-

tent of digital inform ation exchanged between ATM

systems. These exchange models have to be compli-

ant with the AIRM.

One such exchange m odel is the Aeronautical In-

formation Exchange Model (AIXM). AIXM provi-

des a UML data mod e l and associated XML schemas

for representin g the format of digitally commun ic a -

ted aeron a utical inf ormation. AI X M deﬁnes informa-

tion related to, among other things, airports and he-

liports, airspace structures, organisations (including

services they provide ), geographica l points and navi-

gation aids, route information and ﬂying restrictions.

Another exchange model is the ICAO Meteo-

rological Information Exchange Model (IWXXM).

IWXXM provid e s a format for exchanging messages

related to actual and forecasted weather reports at ae-

rodromes, weather conditions along the route, signi-

ﬁcant meteorological information, and advisories re-

lated to volcanic ash even ts and oth er extreme mete-

http://www.eurocontrol.int/swim

http://www.eurocontrol.int/articles/airm-atm-

information-reference-model

KEOD 2018 - 10th International Conference on Knowledge Engineering and Ontology Development

orological conditions (e.g. cyclones). As AIRM a nd

AIXM, IWXXM is originally represented in UML.

2.3 Deﬁning Compliance in ATM

Compliance is deﬁned as (ISO/IEC, 2005):

“



...



the demonstration that speciﬁed require-

ments relating to a product, process, system,

person, or body are fulﬁlled .”

Semantic intero perability in the ATM domain is

supported by having a co mpliance framework deﬁ-

ning a set of artefact-related and procedural require-

ments that nee d to be satisﬁed in order for an inf or-

mation or data construct to cla im compliance with the

AIRM (Wilson et al., 2015).

The result of a compliance veriﬁcation process is a

mapping describing the semantic relation between in-

formation elements of the model claiming compliance

with the AIRM and th e correspondin g AIRM infor-

mation element. The semantic relation is either equi-

valence or a wider sema ntic meaning (Wilson, 2017).

2.4 Representing ATM Information

Models as OWL Ontologies

The aforementioned BEST project transfor med the

referenc e model AIRM an d the exchange mode ls

AIXM and IWMXX to OWL using the mapping ru-

les speciﬁed by the ”Ontology Deﬁnition Metamo-

del Speciﬁcation” developed by Obje ct Manageme nt

Group (OMG) ( O bject Management Gr oup, 2014).

These rules ensu re that the semantics expressed in the

UML models are maintained du ring the transforma-

tion to OWL. Furthermore, the BEST project decom-

posed the ontologie s resulting from this transforma-

tion in to a set of ontology modules following the gui-

delines of d’Aquin (D’Aquin, 2012) and principles of

ontology module extraction from Grau et al. (Cuenca

Grau e t al., 2008).

3 AUTOMATED COMPLIANCE

VERIFICATION

In this section we describe our 3-step process for au-

tomated compliance veriﬁcation as outlined in Figure

3.1 Ontology Proﬁli ng

After the input ontologies have been pre-processed

and parsed to an appropriate re presentation, the in-

1. Matcher Selection 2. Matcher Combination

String-based Matchers

Structure-based Matchers

Lexical Matchers

M1 M2 M3

1. Ontology Proﬁling

Figure 1: Overall Approach for Automated Compliance Ve-

riﬁcation.

put ontologies are analysed according to a set of pro-

ﬁling metrics. These metrics evaluate the terminolo-

gical, structura l and lexical pro ﬁle of th e ontologies

and are computed as an average metric for the ontolo-

gies. Such a process helps to select the most optimal

matchers and reduce p rocessing at run-time caused by

excluding or giving less emphasis to matchers not ca-

pable of contributing to the task at hand. Based on

these metrics, th e set of optimal matchers are identi-

ﬁed given the ontologies to be matc hed.

In the following we describe the proﬁling metrics

we have used in this work.

3.1.1 Compound Ratio (CR)

Compound word s are quite common in ontolog ies. A

compound is a word consisting of one or more indi-

vidual words, such as Aerodr omeProtectionArea. If

the number of occurrences of compounds is high, this

suggests that a m atcher capable of explo iting such

linguistic structures should be emp loye d. Th is also

might suggest that the terminology of the ontology

is quite “uniform”, where existing concept names are

appended (either through preﬁxing or sufﬁxing), for

example when creating sub-classes, instead of using

a riche r and more ﬁne-grained terminology. In such a

case, a string-based matcher could perform well. The

Compound Ratio is compu te d by dividing the num-

ber of classes having a compound name by the total

number of classes in the o ntology.

3.1.2 Annotation Coverage (AC)

Annotation Coverage measures how many concepts

that have a natural language deﬁnition compared to

all concepts in the ontology. If this f raction is high,

then a matcher spec ia lised in ﬁndin g similarity among

annotation properties (comments) should be applied.

This metr ic does not indicate w hether such a ma tc her

will be successful, but rather that if the score is low

such a matcher will probably not contribute much.

3.1.3 Inherita nce Richness (IR)

Inheritance Richness (Tartir et al., 2005 ; Cruz et al.,

2012) measures the structural characteristics of the in-

Automated Compliance Veriﬁcation in ATM using Principles from Ontology Matching

put ontologies as the average number of subcla sses

per class. Hence, if the Inheritance Richness is high,

the concepts in th e ontology have many sub-classes,

something which could be exploited by a structural

matcher. In co ntrast to the other metrics the I R is an

open-e nded positive number rather than a fraction .

3.1.4 Relationship Richness (RR)

Relationship Richness (Tartir et al., 2005; Cruz et al.,

2012) computes the fraction of relations that are dif-

ferent from subClassOf relations and can sugge st to

what extent properties can be exploited to infer con-

cept mappings. If an ontology has a Relationship Ri-

chness close to zero, that would indicate that most

of the relationships are is-a relations, and a structural

matcher could be emphasized. On the other han d, if

the Relationship Richness is high, this indicates that

the ontology has a high fraction of object properties

that could be exploited to infer either class equiva-

lence or other semantic relations.

3.1.5 WordNet Synonym Coverage (WNSYN)

One of the strengths of using WordNet in ontology

matching is to identify (synonymic) relations between

two concepts that other matchers cannot identify, ty-

pically through a shared synset among these concepts.

So, if the degree of synonymy among the input onto-

logies is high, then it is likely that a matcher utilising

WordNet synonyms could co ntribute positively in the

matcher composition. This me tric measures the ex-

tent to which a concept is re presented by synonyms in

WordNet. It is calculated by accumulating the num-

ber of concepts for which there exists a synonym and

then divide this number by the total number of c la s-

ses in each ontology. Whenever the concept name is

a compound word, each compound part of the word

is treated separately. That means that if a compound

concept name (e.g. AerodromeProtectionArea) has a

compound part (e.g . Protection ) for which there is no

set of synonyms in WordNet, it is omitted in the accu-

mulation, a nd the score is reduced.

3.1.6 Proﬁling Scores for the Datasets

Table 1 shows the proﬁling scores according to the

ﬁve intr oduced proﬁling metrics (CR, AC, IR, RR,

WNSYN) for the seven datasets in our experiment

(D1-D7) and the average over all seven datasets

(AVG).

So, what do these proﬁling scores tell us? Well,

the Compound Ratio (CR) score tells us that most con-

cept names in these ontologies are compound words

(90 percent of all concept names in all ontologies in-

volved are compounds), suggesting that there could

be a hierarchical structur e where a super concept (e.g.

Wind) h as childre n with concept names that append

their parent (e.g. AerodromeSurfaceWind). This

could be utilised by a subsumption matcher that iden-

tiﬁes for example that Aerodrom eSurfaceWind is a

specialisation (child concept of) of Wind. Further-

more, it suggests that it could be difﬁcult to straight-

forwardly utilise lexical resources su c h as WordNet,

since such resources often hold mostly generic, non-

compounded terms.

The A nnotation Coverage (AC) shows tha t almost

all concep ts are well deﬁned in the sense that they

have a natural language deﬁnition associated with

them. This means that a ma tc her that analyses (si-

milarity) between the concepts’ deﬁnitions should b e

included in the exper imentation.

The Inheritan ce Richness (IR) and the Rela tions-

hip Richness (RR) scores in combination reveal tha t

these ontologies have quite ﬂat structures with few

subclasses per class, but that the representation of r e-

lations (object properties) b etween the classes is rela-

tively high. Based on this, we have not included struc-

tural matchers that infer equivalence rela tions from

the graph-based represen ta tion o f the ontologies. Ho-

wever, we have included matchers that exploit object

properties as means for inferring similarity between

classes should be included.

As earlier mentioned, the fact that most con-

cept names are compounds makes the use of Word-

Net challenging. However, by splitting each con-

cept names into individual compound tokens, e.g.



Aerodrome



Surface



Wind



as in the example used

earlier, we then analyse to w hat extent e ach individual

part has a representation of synonyms in WordNet, re-

sulting in the WordNet Synon y m Coverage (WNSYN).

This metric represents an extension of the WordNet

Coverage used for example in (Tartir et al. , 2005;

Cruz et al., 2 012).

Table 1: Proﬁling scores for all datasets.

Metric D1 D2 D3 D4 D5 D6 D7 AVG

CR .94 .91 .93 .93 .76 .94 .92 .90

AC 1.0 1.0 .99 1.0 1.0 1.0 1.0 1.0

IR 1.13 1.47 1.48 1.46 1.47 1.48 1.46 1.42

RR .59 .56 .56 .57 .5 7 . 57 .57 .57

WNSYN .74 .75 .56 .76 .87 .87 .87 .77

Based on the above proﬁling and discussion,

we implemented the ma tchers presented in the next

section.

KEOD 2018 - 10th International Conference on Knowledge Engineering and Ontology Development

3.2 Matcher Selection

We have implemented the matchers using the

ontology- and ontology matching infrastructure s of

the OWL API (Horridge and Bech hofer, 2011) and

the Alignment API (David et al., 2011).

Some o f the matcher s ide ntify equivalence rela-

tions, w hile others identify other semantic relatio ns

(typically subsumption relations). Table 2 provides a

summary of all implemented matchers.

3.2.1 ISub String Matcher

The ISub String Matcher is a string matching a lgo-

rithm developed by Stoilos et al. (Stoilos, Giorgos

and Stamou, Giorgos and Kollias, 20 05). The ISub

algorithm applies thr ee functions in order to ﬁnd the

similarity b etween two concept names c

and c

. The

ﬁrst function compute s the similarity betwee n the two

strings by iteratively ﬁnding the common substring s.

In the second function it co nsiders the difference as

the length of the remaining characters in the two

strings. I n the third and ﬁnal function, the Wink le r

algorithm (Winkler, 1 999) is used for improving the

results.

3.2.2 Deﬁnitions Matcher

The Deﬁnitions Equivalence Matcher treats deﬁniti-

ons associated with two entities as sets of individual

words. Stopwords that c arry little meaning, such as

’the’, ’a ’, ’is’, etc., are removed before the deﬁnitions

are processed further. As with the other algorithms re-

lying on set-theoretic similarity scores, this algorith m

employs the Jaccard (Jaccard, 1901) set-theoretic si-

milarity measure to compute a similarity score bet-

ween the two deﬁnitions. The Jaccard similarity me-

asure is c omputed by divid ing the intersection of sets

with the union of the sets.

3.2.3 Property Matcher

The Property Matcher bases the equivalence identiﬁ-

cation on similarity of the prope rties associated with

the concepts to be matched. Both object properties

and data properties where the concepts to be matched

are domains are collected into a sin gle set for each

concept and compared with Jacc a rd.

3.2.4 Range Matcher

The Range Matcher measures the similarity of the sets

of range classes of object properties where the con-

cepts c

and c

being matched represent the domain.

If the Jaccard set similarity of the object properties’

range classes is above a certain threshold, this mat-

cher considers that the two c oncepts are equivalent.

3.2.5 WordNet Synonym Matcher

The Word N e t Synonym Matcher (WNSyn) computes

a similarity score based on how many comm on Word-

Net synonyms the two con c epts to be matched are as-

sociated with. Since the ontology proﬁling revealed

that most concept names are compound words, we

split all compounds into a set of compound parts, and

synonyms associated with each part represent indivi-

dual sets o f words taking part in the similarity calcu-

lation. The synonyms associated with the respective

concepts are represented as sets and a similarity score

is computed using Jaccard.

While the previously presented algorithms seek

to identify equivalence relations, the f ollowing algo-

rithms aim to identify other sema ntic relations.

3.2.6 Closest Parent Matcher

The Closest Parent Matcher determines that o ne con-

cept c

is a subclass of concept c

if the superclass of

has a high similarity with c

, as illustrated in Figu re

2. Th is matcher re lies on having a graph representa-

tion of the ontologies. We impleme nt such a graph

representation using the Neo4J

graph database.

AirportHeliportProtectionArea AerodromeProtectionArea

RunwayProtectArea

Equivalent

SubclassOf

Figure 2: The Closest Parent Matcher.

3.2.7 Compound Matcher

The Compound Ma tcher identiﬁes subsumption re-

lations between entities reusin g principles from the

compound strategy from Arnold and Rahm (Arnold

and Rahm, 2014). Compound means that several in-

dividual words are put together to form another word.

Here, parts of compounds in entity name s are identi-

ﬁed and employed as an indicator of a subsumptio n

relation. So, if one or more compo und parts in one

concept name c

are represented as a subset of com-

pounds in another concept name c

, the Compound

Matcher deﬁnes that c

subsumes c

3.2.8 Deﬁnitions Subsumptio n Matcher

The Deﬁnition s Subsumption Matcher considers both

commonality and number of words in the deﬁnitions

https://neo4j.com/

Automated Compliance Veriﬁcation in ATM using Principles from Ontology Matching

Table 2: Overview of Implemented Matchers.

Matcher Target Relation Typ e

ISub Entity - Name Equivalenc e

Deﬁnitions M a tc her Entity - Deﬁnition Equivalenc e

Range Ma tc her Entity - Structure Equivalenc e

Property Matcher Entity - Structure Equivalenc e

WordNet Syn onym Matc her Entity - Lexical Prope rties Equivalence

Closest Parent Matcher Entity - Structure Other

Compound Ma tcher Entity - Name Other

Deﬁnitions Subsumption Matcher Entity - Deﬁnition Other

in order to compute if two entities are in a subsump-

tion relation. If the commonality of the deﬁnitions is

above a certain threshold, we consid e r the size (num-

ber of words) of the deﬁnitions as a qualiﬁer for sub-

sumption, whe re an entity with a smaller deﬁnition

subsumes an entity with a larger de ﬁnition. The ratio-

nale behind this is that the more speciﬁc and detailed

the entity is, the more text is re quired to sufﬁciently

describe it.

3.3 Matcher Combination

The next step of the overall process is to combine the

alignments produced by the matchers in th e previous

step.

3.3.1 Weighted Sequential Combination (WSC)

In the Weighted Sequential Com bination the initial

alignment produced by th e ﬁrst matcher is reﬁned by

each following ma tc her in the sequence. Weight is

added to correspondences that are identiﬁed by two

consecutive matchers. If the correspondence is new,

that is, identiﬁed only by the current ma tc her, o r if the

correspo ndence is only identiﬁed b y the previous ma-

tcher(s) and not the current one, the correspondence

is added to the reﬁned alignment with equally redu -

ced weight. As an example, consider that a matcher

has produced an alignment that is transferred to

, the next matcher in the sequence. If the same cor-

respond ence (the same two entities and the same rela-

tion type) is identiﬁed as correct b y both m

and m

the conﬁdence value associated with this correspon-

dence is increased. On the othe r hand, if a correspon-

dence received b y m

from m

is only identiﬁed as a

correct correspondence by m

and not by m

, this cor-

respond ence is reduced before the alignment is sent

further to m

. The weighting schem e applied in this

study is to add (or reduce) 12 percent to the conﬁ-

dence of the corr esponden ce. Maxim um conﬁdence

is 100 p e rcent (1.0).

3.3.2 Simple Vote

Simple Vote is a parallel combination strategy. Here,

all ma tc hers are run in parallel. The alignments they

produce are initially treated as equally important, but

only th ose correspondences identiﬁed by a predeﬁ-

ned ratio of matchers (for example using the majority

vote, such as three out of ﬁve matchers) are eligible

for the ﬁnal alignment.

3.3.3 Autoweight++

The third co mbination strategy is an implementation

of the Autoweight++ algorithm described in (Guli´c

et al., 2016). As with Simple Vote, this is also a pa-

rallel combination strategy, but a more sophisticated

one, since it includes both matcher conﬁguration and

combination. The concept of highest corresponden-

ces is central in this approach. A correspondence be-

tween two concepts c

and c

is considered a highest

correspo ndence if it has a higher conﬁdence value

than any o ther correspondence that includes either c

or c

. The highest correspondences are used both for

automatically conﬁguring the matching algorithms’

weight and for combining the individual alignments

into an o ptimal ﬁnal alignment. When producing the

ﬁnal alignment (from the aggregated set of cor respon-

dences in the intermediate common alignment), Auto-

weight++ takes an iterative approach. It starts by ta -

king the high est correspondences from an intermedi-

ate common alignmen t. Then in the following iterati-

ons, the correspondences that do not include concepts

taking part in the already established highest corre-

spondences are processed. The algorith m stops when

there are no more corre spondenc es above a given con-

ﬁdence threshold.

4 EXPERIMENTAL EVALUATION

In this section we start by describing the datasets in-

cluded in the experimental evaluation. Th en we pro-

KEOD 2018 - 10th International Conference on Knowledge Engineering and Ontology Development

ceed to explain how the evaluation was performed,

before we present our ﬁndings from the evaluation.

4.1 Datasets

Table 3 shows some statistics associated with the da-

tasets used in the experiments. Each dataset consists

of two ontologies that are m atched pairwise. The on-

tologies used in the datasets include both monolithic

ontologies a nd ontology modules derived from the

AIRM, AIXM and IWXXM ontologies (see section

2.4).

A re ference alignment represent the correct set

of relations between entities in the datasets descri-

bed in the previous chapter, and act as our comp a ri-

son base fo r evaluating the quality of our techniques.

The sour ce material for the reference alig nments are

mapping ﬁles in Excel from real compliance veriﬁ-

cation processes of the exchange models AI XM and

IWXXM. These mapping ﬁles contain manually de-

veloped relation s based on human expert veriﬁcation

between the exchange mo dels AIXM/IWXXM and

AIRM and have been transformed to reference alig-

nments in th e RDF Alignment Format u sing the Java

library Apache POI

. The two rightmost columns des-

cribes the number of relations in each reference align-

ment. If a dataset ha s both equ ivalence relations (EQ)

and other semantic relations (OTH) there are two re-

ference alignments for the dataset, otherwise there is

only one.

The d atasets, that is, the ontology pairs along

with the reference alignments, are available from:

https://github.com/sju-best-project/KEOD18

4.2 Evaluation

Typically, evaluation of onto logy matchin g techni-

ques is performed using precision and rec all against

referenc e alignments holding the true positive relati-

ons based on expert judgment (Euzenat and Shvaiko,

2013). Prec ision is co mputed as the ratio of cor-

rectly found correspondences (according to th e refe-

rence alignment) over the to tal num ber of identiﬁed

correspo ndences. Recall is computed as the ratio of

correctly found correspondences over the total num-

ber of expected correspondences (as expressed in the

referenc e alignment). An evaluation measure that ba-

lances precision and re c all is the F-measure. Th is is

often used as an overall m e asure for representing the

quality of an ontology matching technique or com-

plete system.

https://poi.apache.org/

4.2.1 Baseline Ma tchers

In order to perform a compara tive analysis we have

used two ontology matching systems th a t often rank

as top contenders of th e OAEI campaign as baseline

matchers. LogMap (Ru iz-Jimenez and Grau, 2011) is

a system that of te n competes in several of the OAEI

tracks. In our experiments we ran the standalone ver-

sion (version 2.4) of LogMap. AgreementMakerLight

(AML) (Faria et al., 2013) is, like LogMap , a r e cur-

ring pa rticipant at the OAEI, with several top positi-

ons in the tracks it compe tes in. We used the graphi-

cal user in terface version from April 2016

. As with

our own matcher imple mentations we let LogMap and

AML compute alignments with conﬁdence thresholds

0.5, 0. 7 and 0.9, and only class relations.

4.2.2 Combination Strategies

When using the combination strategies described in

section 3.3, w e combine the alignments produced by

the three best-p erforming individual matchers across

all datasets. For the equivalence relations these are

the Deﬁnitions Matcher (90 percent conﬁdence), the

ISub Matcher (90 percent conﬁdence) and the Word-

Net Synonym Matcher at 90 perc ent conﬁdence.

4.3 Results and Findings

In this section we summarise the results and observa-

tions fr om the experiments for each dataset.

4.3.1 Equivalence Relations

Table 4 shows the F-measure score s for all individual

matchers, the combination strategies and the baseline

matchers for each dataset isolated on equ ivalence re-

lations.

Dataset 1 is the largest dataset in terms of relati-

ons in the referenc e alignment, with 12 6 equivalence

relations. Many of the relations consists of c oncepts

where the names are identical, but th ere are also some

“traps”, w here the semantic meaning deviates despite

of name equality. Conversely, there are also relations

where the semantic meaning of the two concepts is

equal, while their names are differ ent.

The best performing ind ividual matcher is the

WordNet Synonym Matcher which obtains an F-

measure of 88 percent. The best combination stra-

tegy is SimpleVote. Here, the true positive relations

identiﬁed by the WordNet Synonym Matcher are sup-

plemented with additional true positive relations from

the other matchers, resulting in an F-measure of close

This was the latest release on github as of April 2018

Automated Compliance Veriﬁcation in ATM using Principles from Ontology Matching

Table 3: Dataset Statistics.

Dataset Ontologies Classes Object Prop. Data Prop. EQ OTH

AIXM-AirportHeliport

347 571 162 126 1

AIRM-Aerod romeInfrastructure

IWXXM-Common

923 1762 494 0 9

AIRM-Monolithic

IWXXM-Metar

961 1807 530 11 7

AIRM-Monolithic

AIXM-Shared

938 1785 518 21 0

AIRM-Monolithic

AIXM-Geom etry

922 1764 506 5 2

AIRM-Monolithic

AIXM-Obstacle

930 1788 501 11 2

AIRM-Monolithic

AIXM-Organisation

925 1776 499 10 0

AIRM-Monolithic

to 92 percent. In compa rison, the best baseline sy-

stem, AML a t conﬁdence 0.9, obtains an F-measure

of 94.6 percent.

Dataset 3 is quite challenging as it includes se-

veral instances of complex (1..n) mappings among the

11 relations in the referenc e alignment. The best score

is achieved by the ISub string matcher which obtains

an F-measure of around 19 percent. ISub identiﬁes

two true positive correspondences. None of the com-

bination strategies are able to improve the score, and

of the baseline systems only AML is able to identify

any true positives (one).

Dataset 4 contains 21 relations in the reference

alignment and they include mostly generic, domain-

indepen dent concept names. The best perform ing in-

dividual matcher is the Deﬁnitions Matcher (c onﬁ-

dence 0.9) obtaining an F-measure of around 83 per-

cent. T he ISub string matcher manages to identify

most relations in the reference alignment at low con-

ﬁdence threshold (hence a high recall), but includes

too many false positives, so the resulting F-measure

becomes quite low at the lower conﬁdence levels. The

overall best alignment quality is achieved by the com-

bination strategy SimpleVote, with an F-measure of

90 percent. The baseline systems AML and LogMap

obtains a maximum F-measure of 95 and 92 percent

respectively.

In Dataset 5 there are 5 relations in the reference

alignment. This is a challenging dataset for matchers

basing their equivalence identiﬁcation o n string simi-

larity of concept names as in only on e of the refe-

rence alignment relations such similarity is noticea-

ble. Here, the Deﬁnitions Matcher at conﬁdence 0.7

obtains the highest F-measure score of 40 percent,

while most of the other matchers are only capable of

identifyin g the aforementioned concept n ame equa-

lity. The Deﬁnitions M atcher iden tiﬁes 3 of the cor-

rect relations from its similarity com putation of natu-

ral language deﬁnitions associated with the concepts.

Dataset 6 contains 11 relations. The best F-

measure score is obtaine d by the WordNet Syn onym

Matcher at conﬁdence level 0.9, which identiﬁes all

true positive correspondences, and manages to dis-

regard more false positives than the ISub Matcher

that also identiﬁes all correct relations. When com-

bining the alignments with the SimpleVote strategy,

we obtain an F-measure of 83 percent.

The reference alignment for Da taset 7 contains 10

relations. Most of the relations in the reference alig-

nment consists of identical concept names (8 out of

10). Here, the best performance of the individual ma-

tchers is achieved by the ISub matcher at conﬁdence

0.9, which identiﬁes all true positive relations but one ,

resulting in an F-measure of 87 percen t. The baseline

systems have better precision (less false positives),

and AML at conﬁdence 0.5 obtains an F-measure of

close to 95 percent.

When we average the F-measure scores a c ross all

datasets, the best individual matcher is the WordNet

Synonym matcher with an F-measure of 55.48 per-

cent. The best combination strategy is Simple Vote

with an average F-measure of 67 .9 percent. N ot unex-

pectedly, the baseline matchers obtain overall h igher

F-measure scores. The best F-measure is achieved

by AML a t conﬁden ce 0.5 (69.74 percen t), followed

by AML a t conﬁde nce 0 .7. Both these conﬁgurations

obtain higher F-measure scores than the b est combi-

nation strategy SimpleVote on third.

KEOD 2018 - 10th International Conference on Knowledge Engineering and Ontology Development

Table 4: F- measure scores for all datasets - Equivalence re-

lations. The best performing individual matcher is highlig-

hted in yellow.

!"#$%&' () (* (+ (, (- (. /01

!"#$%&' ()(*$+ %),-$+ .)(/$+ .%)-.$+ .)%($+ /)./$+ -)0($+

!"#$%&0 /')10$+ 1)%1$+ /0)/0$+ *%)%%$+ .%)-*$+ ..)0($+ /%)0*$+

!"#$%&1 (*)*.$+ .')-,$+ ,-)--$+ .()(0$+ 0/)0-$+ '0).*$+ '.)(.$+

2345$%&' -)01$+ %),*$+ %)0-$+ %),'$+ %)(/$+ %)1'$+ .)/1$+

2345$%&0 ..)1-$+ /)-*$+ /)1*$+ /)**$+ -)--$+ /)0/$+ *)/,$+

2345$%&1 (,)*0$+ .1)%'$+ (0),($+ --)--$+ '.)/,$+ ,()1($+ '*)*1$+

6786$%&' *0)./$+ .*)/1$+ *()(0$+ %)%%$+ /')%%$+ 1)%1$+ /-)(1$+

6786$%&0 -().*$+ .()(0$+ *.)-,$+ %)%%$+ .-)--$+ ./)'%$+ /%)%%$+

6786$%&1 /1)-%$+ %)%%$+ *.)-,$+ %)%%$+ %)%%$+ ./)'%$+ .-),($+

79:;"$%&' *%)%%$+ %)%%$+ --)--$+ %)%%$+ ..)0($+ /()(0$+ .,)(-$+

79:;"$%&0 /1)*1$+ %)%%$+ /,)'0$+ %)%%$+ %)%%$+ .*)/1$+ ./)%($+

79:;"$%&1 /().*$+ %)%%$+ /,)'0$+ %)%%$+ %)%%$+ .*)/1$+ ..)'%$+

<:3=>$%&' -.)*($+ %)%%$+ 1),($+ -)*'$+ .%)',$+ /*)-1$+ .-)/1$+

<:3=>$%&0 (/)*%$+ %)%%$+ -.)%-$+ .*)/1$+ /.)%'$+ -,)*($+ /0),0$+

<:3=>$%&1 ,,)-%$+ %)%%$+ (1)-1$+ /%)%%$+ 0()1/$+ 0,)/($+ '')*,$+

<3? (*)*.$+ .1)%'$+ (1)-1$+ --)--$+ 0/)0-$+ 0,)/($+ '().1$+

94@ABCDEF@GG ,%)'0$+ %)%%$+ ,%)%%$+ --)--$+ ,%)%%$+ '0).*$+ '').0$+

3DHIJCKA@C 1.)0($+ .*)/1$+ 1%)%%$+ --)--$+ ,-)--$+ 1*)0*$+ (0)1.$+

9LM$%&' 1*)**$+ 1)11$+ 1')%%$+ --)--$+ 1%)1.$+ 1*)0*$+ (1)0*$+

9LM$%&0 1*)(.$+ ./)'%$+ 1/)-.$+ --)--$+ 1%)1.$+ ,,),1$+ (,)0($+

9LM$%&1 1/),-$+ %)%%$+ 1/)-.$+ --)--$+ 1%)1.$+ ,,),1$+ (()-,$+

MAELNI$%&' 1-)/,$+ %)%%$+ 1/)-.$+ --)--$+ 1%)1.$+ ,,),1$+ (()*'$+

MAELNI$%&0 1/)-0$+ %)%%$+ ,-)--$+ --)--$+ 1%)1.$+ ,,),1$+ (*),.$+

MAELNI$%&1 /1)0-$+ %)%%$+ %)%%$+ %)%%$+ %)%%$+ %)%%$+ *)1'$+

234567"#63789#'"#&:6&;

<";&=67&8!"#$%67:89>;#&4;

4.3.2 Other Semantic Relations

Five of the datasets include other semantic relations

than equivalence. Table 5 shows the F- measure sco-

res for all individual matchers isolated on other re la ti-

ons than equivalence. We also experimented with the

same comb ination strategies for the other correspon-

dences experiments also, using the overall 3 best indi-

vidual matchers and their produced alignmen ts. Ho-

wever, sin c e none o f the combination strategies were

able to produce better quality alignments than the best

individual matcher, we exclude the scores from the

combination strategies for these alignm ents.

In Dataset 1 there is 1 relation in the r efe-

rence alignment holding “oth e r semantic relations”.

The only relation in the reference alignment is

RunwayElement-RunwayElement, which intuitively

suggests an equivalence relation. The reason why this

seemingly equivalent relation is co nsidered a different

semantic relation, is that the natural language deﬁni-

tion in ontology 1 is more speciﬁc than the natural

languag e deﬁnition in ontology 2.

The Closest Parent Matcher with c onﬁdence thres-

holds 0.5 an d 0.7 ide ntiﬁed the true p ositive relation,

but since these alignments also included a very large

number of false positive relations, the pr e cision be co-

mes very poor (6 percent and 9 percent respectively) ,

resulting in very low F-measure score s (0.19 percent).

The Deﬁnitions Subsumption matcher at conﬁ-

dence threshold 0.5 identiﬁed the true positive rela-

tion, but as with the Closest Parent Matcher align-

ments, the p recision and consequently the F-measur e

were very low d ue to a very large number of false po-

sitives.

The reference alignment in Dataset 2 contains 9

relations. The best performing matcher is the Deﬁ-

nitions Subsumption Matcher at conﬁdence thre shold

0.9 which identiﬁed 1 true positive correspondence

and no false positive ones, resulting in a 100 percent

precision. However, since it missed the o ther 8 cor-

respond ences in the reference alignment, the recall

was quite low, resultin g in an F-measure of 20 per-

cent. The Compound Matcher computed also 1 true

positive re la tion, but 1 false positive one. The Closest

Parent Matcher computed 1 true positive relation, and

27 false po sitives. All the true positive relations id en-

tiﬁed by these 3 matchers are different relations, so

in total 3 out of 9 relations in the reference alignment

were identiﬁed.

The relations from the reference alignment not

identiﬁed by any matche r were:

AerodromeForecastWeather < CodeSignificantWeatherQualifierType

AerodromeForecastWeather < CodePrecipitationType

AerodromeForecastWeather < CodeWeatherPhenomenonType

AerodromeForecastWeather < CodeObscurationType

AerodromeSurfaceWindTrendForecast < Wind

AerodromeSurfaceWindTrendForecast < TREND

where X < Y means that X is a spe cialisation of Y.

All these correspondences repre sent complex

mappings, and the manual inspection of the ontolo-

gies suggests that neither structure nor natural lan-

guage deﬁnitions can help infer any type of semantic

relations between these concepts.

Dataset 3 includes 7 relations a nd the only ma t-

cher able to iden tify any true positives is the Com-

pound Matcher (at all conﬁdence thresholds), which

identiﬁes the following two:

Aerod romeSurfaceWind < Wind

Aerod romeRunwayVisualRange < RunwayVisualRange

Here, the use of endocentric compound s c ontributes

to the identiﬁcation of a subsumption relationship.

Endocentric compounds consist of a compound head,

which represent the base meaning of the compound,

and one or more m odiﬁers that serves to narrow the

meaning of the compound as a whole (Arnold and

Rahm, 2014).

The reference alignment in Dataset 5 c ontains two

relations. Both re la tions were id e ntiﬁed by the D eﬁ-

nition Sub sumption Matcher at all conﬁdence thres-

holds, hence a perfect recall. However, the number of

false positives increased with lowering the thresholds.

The best alignment was thus obtained at 90 percent

conﬁdence, with an F- measure of 44.4 percent.

Dataset 6 included also two relations, and the o nly

matcher able to identify a true positive correspon-

dence in this dataset is the Deﬁnitions Subsumption

matcher at conﬁdence thresholds 0.5 and 0.7. Both

alignments identify one true positive correspo ndence,

but contain a large number of false p ositives, resulting

Automated Compliance Veriﬁcation in ATM using Principles from Ontology Matching

in a very low F-measure score of just over 1 perce nt

for the best performing matcher (at conﬁdence level

0.7).

Lookin g at the avera ge F-measure across all data-

sets, the Deﬁnitions Subsumption Matcher performs

best, with an F- measure of 12.9 percent.

Table 5: F-measure scores for all datasets - Other semantic

relations. The best performing individual matcher is high-

lighted in yellow.

!"#$%&' () (* (+ (, (- ./0

!"#$%& $'((#) &'*(#) $'$$#) $'$$#) $'$$#) ('($#)

!"#$%+ $'(,#) $'$$#) $'$$#) $'$$#) $'$$#) $'$*#)

!"#$%, $'$$#) $'$$#) $'$$#) $'$$#) $'$$#) $'$$#)

!-."#$%& $'$$#) (/'(/#) 01'1+#) $'$$#) $'$$#) /',+#)

!-."#$%+ $'$$#) (/'(/#) 01'1+#) $'$$#) $'$$#) /',+#)

!-."#$%, $'$$#) (/'(/#) 01'1+#) $'$$#) $'$$#) /',+#)

234567#$%& $'(&#) ('/+#) $'$$#) *'01#) $'((#) ('0/#)

234567#$%+ $'$$#) (/'(/#) $'$$#) 08'&8#) ('0(#) /'&/#)

234567#$%, $'$$#) 0$'$$#) $'$$#) **'**#) $'$$#) (0'/,#)

4.4 Conclusions from Experimental

Evaluation

The general conclusions are th at the identiﬁcation

of equivalence relations is far easier than identiﬁca-

tion of other semantic relations, such as sub sumption.

There are several contributing factors to this. First,

the reference alignments conta in a variety of diffe-

rent semantic relation types (subsumption, part-whole

relations, less/more general based on narrower/wider

natural language deﬁnitio ns). One such example is

the already mentioned relation betwee n AIXM Run-

wayElement an d AIRM RunwayElem ent which in the

referenc e alignment is of type “less general”. The re-

ason why the ﬁrst is more restrictive than the latter, is

that RunwayElement is describe d in a generic way in

AIRM, where it is deﬁned as “A portion of a runway”,

and there was thus a need for makin g the deﬁnition

more accurate in AIXM. In AIXM RunwayElement

is deﬁned as follows: ’Run way element may consist

of one mor e polygons not deﬁned as oth e r portions of

the runway class’.

Secondly, and especially the case when trying to

identify such relations between IWXXM an d AIRM,

it is very difﬁcult to ﬁnd usable patterns of sp e ciali-

sation in most of the mapping s. Most specialisation

relations in the mapping ﬁles are not identiﬁed even if

we have implemented matchers that utilise terminolo-

gical, structural and lexical patterns.

The quality of the equivalence matching is far bet-

ter. When comparing against two of the top per-

forming ontology matching systems, AgreementMa-

kerLight and LogMap, the pe rformance of our some

of our basic matcher s is fairly good. In two of the

datasets, our matchers obtain higher F-measure than

the baseline systems, and when combining the align-

ments using simple aggregation strategies, we obtain

an F-measure on par with the baseline systems.

Finding a complementary set of matcher s is es-

sential. The three best performing individual mat-

chers are the WordNet Synonym Matcher, the ISub

Matcher and the Deﬁnitions Matcher. These mat-

chers infer equivalence relations based on the con-

cept’s name and natural language deﬁn ition, sugges-

ting that the termin ology expressed by these ontolo-

gies is quite standardised. H owever, while th e se three

matchers often identify many of the same relations,

the Property Matcher and the Range Matcher supple-

ment with relations where such terminological simi-

larity does not exist through the use of property simi-

larity.

In general, we see that combining the alignments

often improves the alignment quality involving equi-

valence c orrespondences. Here, the combination stra-

tegies extract complementary true positive correspon-

dences from each individual alignment, and also helps

reduce the number of false positive corresponden-

ces. For othe r types of correspondences, co mbining

the alignments hurts the quality, resulting in lower F-

measure scores than for the best perfo rming indivi-

dual matchers.

Another observation is that the m atchers utili-

sing properties as me a ns for inferring class similarity

(Property Matcher and Range Matcher) perform bet-

ter when using low conﬁdence thresholds, while the

other matchers perform better at higher conﬁdence

thresholds. The precision is fairly stable, but the r e-

call drops when conﬁdence is increased. The expla-

nation is that at lower conﬁdence, the other matchers

produce too many false positives, reducing the preci-

sion. Finding a goo d compro mise between precision

(i.e. reducing false positives) and recall (i.e. retrie-

ving a s many true positives as possible) is the key to

good perform ance of a matching system.

5 RELATED WORK

To the best of our knowledge, automating complian c e

veriﬁcation with standard information models using

ontology ma tc hing is not investigated elsewhere. Ho-

wever, several studies have investigated the use of

ontology m atching techniques to automate compli-

ance assessment between business processes in actual

use and th ose prescribed by standards (Ternai et al.,

2013; Ternai, 2015; G´abor et al., 201 3; Szab´o and

Varga, 2014), regulations (Sapkota et al., 2013) and

structured end-u ser r equirements (Bakhshandeh et al.,

2015). Wong et al. (Wong et al., 2008) investiga -

ted how ontology matching could reconcile semantic

models extracted from different IT governance stan-

KEOD 2018 - 10th International Conference on Knowledge Engineering and Ontology Development

dards, with th e aim to make it easier for companies to

comply with their rules. In their a pproach they ﬁrst

translated the governance stan dards into an o ntologi-

cal representation using natur a l lang uage processing

(NLP) techniqu e s and then identiﬁed common seman-

tics u sin g ontology m a tc hing techniques.

6 CONCLUSIONS AND FU RTHER

WORK

6.1 Conclusions

Compliance veriﬁcation in ATM is currently per for-

med by manually inspecting large-sized information

models and identifying the semantic relations th a t ex-

ist between them. This is a very laborious process that

could be supported by matching tec hniques. This pa-

per has presented an approa c h for automating compli-

ance veriﬁcation which will reduce the manual effort

related to th e standards compliance p rocess in ATM

and motivate reuse of standardised information ele-

ments. The a pproach is based o n applying quite basic

ontology matching techniques for automa tically id en-

tifying different types of semantic relations among

ontologies describing concepts from the ATM dom-

ain. The matching techniques are selected based on

the results from an ontology proﬁling step that reveal

terminolo gical, structural and lexical characteristics

of the ontologies.

From an experimental evaluation involv ing seven

different datasets we h ave learne d that the proposed

matchers identif y equivalence r elations quite well. A

compara tive evaluation with two state-of-the-art mat-

ching systems show that our in dividual m atchers pe r-

form on par with these more so phisticated systems in

some of the datasets and even perform better in da-

tasets where concept name similarity is not the de-

ciding factor for equivalence. When the alignments

produced by the individual matchers are combined,

the alignment quality nor mally becomes even better.

The overall best combination strategy is Simple Vote,

a strategy based on aggregating alignment rela tions

based on majority vote.

The evaluation also shows that the identiﬁcation

of other semantic relations than equivalence is more

challengin g. Due to diversity in semantic relation

type, relation cardinality (e.g. 1..n r e la tions), large va-

riations in ter minology, structure and lexical c harac-

teristics, the implemented matchers struggled to iden-

tify such relations.

6.2 Further Work

The “other relations” category includes a variety of

semantic relation types. Analysing the other types

of semantic relations involved and ﬁn ding tec hniques

for their identiﬁcation could lead to more precise ma-

tching results for this category. One example is part-

whole (meronymy) r elations. Investigating pattern s in

names, deﬁnitions and structure that could help reveal

part-whole relations (and other possible relations) and

distinguish them from equivalence and specialisation

relations could lead to better and mo re accurate alig-

nments.

Ontology matching systems often use external re-

sources to facilitate identiﬁcation of semantic rela-

tions. In this work we have employed WordNet as

an external resource, but other more domain-sp eciﬁc

sources could possibly enhance the matc hing results.

One such resource for the aviation domain is Sky-

brary

, a wiki that contains loads of domain know-

ledge related to aviation and ATM. Investigating met-

hods on how a resource such as Skybrar y could be

utilised to suppo rt identiﬁcation of semantic relations

is an interesting further work item.

Scalability is not considered in this work, but is an

important quality to look at, especially when ontolo-

gies are as large as the AIRM ontology (counting over

3000 entities altogether). Some of the matchers re-

quired signiﬁca nt run-time, which probably could be

substantially reduce d by performing a thorough sca-

lability analysis.

ACKNOWLED GEM EN TS

This research has received funding from the SESAR

Joint Und ertaking under grant agreement No 699298

(the BEST project) under the European Union’s Ho-

rizon 2020 research and innovation program. The

views expressed in this paper a re those of the a uthors.

REFERENCES

Arnold, P. and Rahm, E. (2014). Enriching ontology map-

pings with semantic relations. Data and Knowledge

Engineering, 93:1–18.

Bakhshandeh, M., Pesquita, C., and Borbinha, J. L. (2015).

Ontology matching techniques for enterprise architec-

ture models. In OM, pages 236–237.

https://www. skybrary.aero/index.php/Main

Page

Automated Compliance Veriﬁcation in ATM using Principles from Ontology Matching

Cruz, I. F., Fabiani, A., Caimi, F., Stroe, C., and Palmonari,

M. (2012). Automatic conﬁguration selection using

ontology matching task proﬁling. Lecture Notes in

Computer Science (including subseries Lecture Notes

in Artiﬁcial Intelli gence and Lecture Notes in Bioin-

formatics), 7295 L N CS:179–194.

Cuenca Grau, B., Horrocks, I., Kazakov, Y., and Sattler,

U. (2008). Modular reuse of ontologies: Theory and

practice. Journal of Artiﬁcial Intelligence Research,

31:273–318.

D’Aquin, M. (2012). Modularizing Ontologies. In O nto-

logy Engineering in a Networked World, pages 213–

233. Springer Berlin Heidelberg, Berlin, Heidelberg.

David, J., Euzenat, J., Scharffe, F., and Dos Santos, C. T.

(2011). The alignment API 4.0. Semantic Web,

2(1):3–10.

Euzenat, J., Meilicke, C., Stuckenschmidt, H., Shvaiko, P.,

and Trojahn, C. (2011). Ontology Alignment Evalu-

ation Initiative: six years of experience. Journal on

Data Semantics, pages 158–192.

Euzenat, J. and Shvaiko, P. (2013). Ontology Matching.

Springer Science & Business Media.

Faria, D., Pesquita, C., Santos, E., Palmonari, M., Cruz,

I. F., and Couto, F. M. (2013). The agreementmaker-

light ontology matching system. In OTM Confedera-

ted International Conferences” On the Move to Mea-

ningful Internet Systems”, pages 527–541. Springer.

G´abor, A., K˝o, A., Szab´o, I., Ternai, K., and Varga, K.

(2013). Compliance check in semantic business pro-

cess management. In OTM Confederated Internatio-

nal Conferences” On the Move to Meaningful Internet

Systems”, pages 353–362. Springer.

Guli´c, M., Vrdoljak, B., and Banek, M. (2016). CroMat-

cher: An ontology matching system based on automa-

ted weighted aggregation and iterative ﬁ nal alignment.

Web Semantics: Science, Services and Agents on the

World Wide Web, (41).

Horridge, M. and Bechhofer, S. (2011). T he OWL API: A

Java API for OWL Ontologies. Semantic Web Journal,

2(1):11–21.

ISO/IEC (2005). Iso/iec 17000 conformity assessment–

vocabulary and general principles.

Jaccard, P. (1901). Distribution de la ﬂore alpine dans la

Bassin de Dranses et dans quelques regions voisines.

Bulletine del la Soci´et´e Vaudoisedes Sciences N aturel-

les.

Object Management Group (2014). Ontology Deﬁnition

Metamodel (ODM) v1.1. Technical report, Object

Management Group Inc., Needham, OSA.

Ruiz-Jimenez, E. and Grau, B. C. (2011). LogMap : Logic-

Based and Scalable Ontology Matching. In ISWC

2011, pages 273–288.

Sapkota, K., Aldea, A., Younas, M., Duce, D. A., and

Ba˜nares-Alc´antara, R. (2013). Rp-match: a frame-

work for automatic mapping of regulations with or-

ganizational processes. In e-Business Engineering

(ICEBE), 2013 IEE E 10th International Conference

on, pages 257–264. IEE E.

Stoilos, Giorgos and Stamou, Giorgos and Kollias, S.

(2005). A string metric for ontology alignment. In

Proceeding of the International Semantic Web Confe-

rence 2005, pages 624–637. Springer.

Szab´o, I. and Varga, K. (2014). Knowledge-based compli-

ance checking of business processes. In OTM Con-

federated International Conferences” On the Move to

Meaningful Internet Systems”, pages 597–611. Sprin-

ger.

Tartir, S., Ar pinar, I., Moore, M., Sheth, a., and Al eman-

Meza, B. (2005). OntoQA: Metric-Based Ontology

Quality Analysis. IEE E Workshop on Knowledge

Acquisition from Distributed, Autonomous, Semanti-

cally Heterogeneous Data and Knowledge Sources,

pages 45–53.

Ternai, K. (2015). Semi-automatic methodology for com-

pliance checking on business processes. In Internatio-

nal Conference on Electronic Government and the In-

formation Systems Perspective, pages 243–256. Sprin-

ger.

Ternai, K. , Szab´o, I., and Varga, K. (2013). Ontology-based

compliance checking on higher education processes.

In International Conference on Electronic Govern-

ment and the Information Systems Perspective, pages

58–71. S pringer.

Wilson, S. (2017). EUROCONTROL Speciﬁcation for

SWIM Information Deﬁnition version 1.0. Technical

report, Eurocontrol, Brussels, Belgium.

Wilson, S ., Keller, S., Marrazzo, G ., and Suzic, R. (2015).

AIRM Compliance Framework. Technical report.

Winkler, W. E. (1999). The state of record linkage and cur-

rent research problems. I n Statistical Research Divi-

sion, US Census Bureau.

Wong, A., Yip, F., Ray, P., and Paramesh, N. (2008). To-

wards semantic interoperability for it governance: An

ontological approach.

KEOD 2018 - 10th International Conference on Knowledge Engineering and Ontology Development