POMap: An Effective Pairwise Ontology Matching System

A. Laadhar

, F. Ghozzi

, I. Megdiche

, F. Ravat

, O. Teste

and F. Gargouri

University of Toulouse, IRIT (CNRS/UMR 5505), 118 Route de Narbonne, 31062 Toulouse, France

University of Sfax, MIRACL, Sakiet Ezzit 3021, Tunisia

Keywords:

Semantic Web, Ontology Matching System, Syntactic Matching, Structural Matching.

Abstract:

The identiﬁcation of alignments between heterogeneous ontologies is one of the main research issues in the

semantic web. The manual matching of the ontologies is a complex, time consuming and an error prone task.

Therefore, ontology matching systems aims to automate this process. Usually, these systems perform the

matching process by combining element and structural level matchers. Selecting the optimal string similarity

measure associated with its threshold is an important issue in order to enhance the effectiveness of the element

level matcher, which in turn will improve the whole ontology system results. In this paper, we present POMap,

an ontology matching system based on a syntactic study covering element and structural levels. For the element

level matcher we have adopted the best conﬁguration based on the analysis of the performances of many string

similarity measures associated with their thresholds. For the structural level, we have performed a syntactic

study on both subclasses and siblings in order to infer the structural similarity. Our proposed matching system

is validated and evaluated on the Anatomy, the Conference and the Large Biomedical tracks provided by the

benchmark of OAEI 2016 ontology matching campaign.

1 INTRODUCTION

To ensure the interoperability between different plat-

forms, ontologies has been used as the essence of the

semantic web. An ontology can model a particular

domain as well as the relationships between its enti-

ties in order to allow data to be shared, reused, inte-

grated and queried by different stakeholders. Ontol-

ogy matching is the process of ﬁnding a set of map-

pings between the entities of two or more ontologies

representing a similar domain (Shvaiko and Euzenat,

2013). Therefore, the ontology matching community

has been proposing a variety of strategies in order to

automate the ontology matching process.The ontol-

ogy matching process takes as an input two ontolo-

gies, which contain a set of entities. Usually the map-

pings between two entities is evaluated by a score

of similarity ∈ [0,1]. This score deﬁnes the degree

of conﬁdence of a mapping relationship. This conﬁ-

dence value is conﬁgured automatically by the match-

ing system. A matcher is an algorithm that processes

two ontologies and returns the set of mappings be-

tween them along with a similarity value. Usually,

a matching system combines several distinct match-

ers depending on the input ontologies characteristics.

This combination aims to improve the ﬁnal align-

ment effectiveness. A single matcher is not sufﬁcient

to achieve a high-quality alignments results. Match-

ers are classiﬁed into two main levels: element level

matchers and structure level matchers (Shvaiko and

Euzenat, 2013). Each level can rely on different tech-

niques and methodologies such as syntactic, semantic

and structural based matching.

In this paper, we address the problematic of on-

tology matching while focusing on the syntactic char-

acteristics of both element level and structural level.

As a ﬁrst step, we focus on ﬁnding the optimal sim-

ilarity measure and threshold in order to perform the

element level matching process. The two target on-

tologies are matched using the element matcher while

exploring all the names of an ontology class, such as

labels, local name and synonyms. We performed a

series of tests in order to deﬁne the optimal similar-

ity measure and threshold. Since we are adopting a

sequential composition, the output of element level

matcher is the input of the structural based matcher.

Dealing with the structural matching, Almost the ex-

isting research works focus on the propagation of the

syntactic similarity score in order to derive new struc-

tural mappings. However, this propagation do not

take into consideration the syntactic characteristics of

the target entities. Therefore, we propose two struc-

Laadhar A., Ghozzi F., Megdiche I., Ravat F., Teste O. and Gargouri F.

POMap: An Effective Pairwise Ontology Matching System.

DOI: 10.5220/0006492201610168

In Proceedings of the 9th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (KEOD 2017), pages 161-168

ISBN: 978-989-758-272-1

tural matchers, guided by the syntactic treatments on

the subclasses and the sibling entities without mea-

suring any structural similarity score. Each structural

matcher focuses on a part of the ontology hierarchy in

order to obtain a new set of alignments. These match-

ers take proﬁt of the hierarchical relationships of an

ontology siblings and subclasses exploiting by a syn-

tactic treatments of common words.

The remainder of this paper is organized as fol-

lows: Section 2 discusses the existing ontology

matching systems. Section 3 presents the POMap

ontology system including the element and structural

level matchers. Section 4 evaluates and discusses the

obtained results. Conclusion and suggestions for fu-

ture work are shown in section 5.

2 RELATED WORKS

Matching Levels. Usually, a matching system em-

ploys different matchers which in turn follow dis-

tinct matching techniques (Otero-Cerdeira et al.,

2015)(Megdiche et al., 2016).

The element level matchers determine the map-

pings between two entities without taking into con-

sideration the existing relationships between these en-

tities. The state of the art element level matchers

are mostly relying on string based techniques, which

compute the similarity between two entities through

syntactic similarity measures. Language-based tech-

niques can also be employed by the element level

matchers. These techniques rely on natural language

processing (NLP) such as the stop words removal and

lemmatisation in order to derive meaningful terms.

Constraint based techniques can be also employed in

order to analyse the internal structure of an entity like

the domain, range as well as the cardinality of the at-

tributes.

Structure level matchers can use external knowl-

edge while taking proﬁt of the relationships between

the entities (classes, instances and properties). Re-

searchers have been employing several techniques

such as graph-based and instance based. The graph

based techniques compute structural similarities be-

tween neighbouring entities based on their position

in the ontology. Following this intuition, Similarity

Flooding (SF) (Melnik et al., 2002) is a well-known

structural matcher inspired by the idea of structural

similarity propagation. This technique propagates the

structural similarity score among the ontology candi-

date mappings. In another research work, (Cruz and

Sunna, 2008) proposed two structural matchers: De-

scendants Similarity Inheritance (DSI) and Siblings

Similarity Contribution (SSC). For a given set of ini-

tial mappings discovered by an element level matcher,

the main idea of DSI and SSC is to reﬁne the initial

element level mappings with weighted scores accord-

ing to the descendants and the siblings. DSI slightly

outperform the Similarity Flooding method. How-

ever, even after applying these structural matchers,

the value of F-measure still not good enough in terms

of effectiveness. We argue that this inefﬁciency is

caused by the bad quality of the initial alignments re-

sulted our ﬁrst matcher. In this research work, dif-

ferentiated from the state of the art, we do not mea-

sure any structural similarity score. Therefore, we are

guided by the initial similarity score in order to de-

rive new structural mappings while taking the rela-

tionships between the entities as a discriminating fea-

ture. These initial mappings are the input of the struc-

tural level matcher in order to deduce a new set of

mappings.

Ontology Matching Systems. We review some of

the top relevant and mature matching systems, which

had a signiﬁcant success during the last years of the

OAEI campaign. Our results will be compared with

these matching systems on section 4.

Agreement Maker Light (AML) (Faria et al.,

2013) is an automated large-scale ontology matching

system focusing on the biomedical ontologies match-

ing. AML translates the entities of the input ontolo-

gies into the same language. Then, AML employs

several matchers such as a lexical matcher, an external

background knowledge matcher, a structural matcher

and an instance matcher. AML follows a sequential

composition. The ﬁrst matcher implemented by AML

is the Lexicon Matcher, which searches for exact

matching between the entities names of the two input

ontologies. AML employs also a mediating matcher,

which take proﬁt of some external knowledge sources

in order to enrich the input ontologies by new tex-

tual content. In the OAEI 2016 compaign, AML used

the following external knowledge resources: Uber

Anatomy Ontology (Uberon) (Mungall et al., 2012),

the Human Disease Ontology (DOID) (Schriml et al.,

2011), the Medical Subject Headings (MeSH) (Nel-

son et al., 2001) and Wordnet (Miller, 1995). AML

derives its structural matcher from the DSI algorithm.

LogMap (Jim

enez-Ruiz and Grau, 2011) is an on-

tology matching system focusing on both the ele-

ment level and the structural level matching of large

biomedical ontologies. The architecture of LogMap

is divided into two main stages: the computation of

candidate mappings and their assessment. In the ﬁrst

stage, LogMap extracts a set of initial candidate map-

pings using a lexical comparison method. These can-

didate mappings are extracted from the inverted index

of each input ontology. An inverted index contains

the labels and the URIs of each ontology. In the sec-

ond stage, LogMap tries to remove the set of incor-

rect candidate mappings by applying lexical, struc-

tural and reasoning techniques.

XMap (Djeddi and Khadir, 2014) deals especially

with the problem of scalability. XMap follows a

sequential composition of three layers: terminolog-

ical Layer, structural layer and alignment layer. In

the terminological layer, XMap employs a transla-

tion matcher in case of multilingual ontologies. Also,

XMap uses a string matcher, which computes the sim-

ilarity between the textual descriptions of two en-

tities by applying some semantic measures. Xmap

employs two background knowledge sources word-

net (Miller, 1995) and UMLS (Bodenreider, 2004)).

During the structural layer, XMap computes the sim-

ilarity between two entities by taking into account

the internal structure of a concept. Finally, in the

alignment Layer, the set similarity scores resulted

by the string matcher, the linguistic matcher and the

structural-based matcher are aggregated using mathe-

matical methods.

CroMatcher (Guli

c et al., 2016) is an ontology

machting system concentrating on the aggregation

of different matchers using a weighted factor. Cro-

Matcher divides its matchers into two main sets:

string based and structure based. The string based

matchers determines the mappings between two en-

tities relying on several techniques such as syntac-

tic similarity measure, instance matching and prop-

erty matching. CroMatcher performs the structural

matching using four distinct matchers: Super En-

tity matcher, sub entity matcher, domain matcher and

range matcher. The matching system aggregates the

results of the string matchers and the structural match-

ers by automatically giving a weight for each sin-

gle matcher. The two obtained matrices of the string

based matcher and structural based matcher are also

combined using an automated weighting factor in or-

der to obtain the ﬁnal set of alignments.

3 THE POMAP ONTOLOGY

MATCHING SYSTEM

3.1 POMap Matching System Overview

The workﬂow of POMap is depicted in Figure 1.

This workﬂow contains several steps beginning by the

OWL ontologies loading until the RDF output align-

ment evaluation. The architecture of POMap contains

three main components: The ontology loading, the

ontology matching and the output alignment evalua-

tion. We will describe in the following sub-sections

the set of activities performed by each component.

Figure 1: The architecture of POMap.

3.2 Ontology Loading Component

We parse a given set of two input ontologies using

OWL API in order to populate our lexical and struc-

tural data structure. The lexical data structure is a

multimap that contains the triplet: a set of classes,

their names as well as the type of each names. A

name has one of these types: URIs, labels or syn-

onym properties. In terms of the structural data struc-

ture, we store all the relationships between classes. A

pre-processing techniques are mandatory to acquire

a cleaner data-set. Since several similarity measures

take into consideration the letter case as a parame-

ter inﬂuencing the similarity score, we set in lower-

case all the input ontology names. Next, we discard

all stopwords as well as the non-alphanumeric char-

acters. We also replace every underscore to a space

character. Finally, we perform an English stemming

process as the one proposed by Porter (Porter, 2001).

3.3 Element Level Matcher

In order to discover mappings, ontology match-

ing systems employ one or more similarity mea-

sures without any reference to the deﬁned thresholds.

Therefore, we tested the different similarity measures

in order to deﬁne the optimal measure and threshold

after performing the pre-processing steps. After the

pre-processing steps, the input ontologies are quite

ready to be processed by our element level matcher.

This matcher is responsible for computing the simi-

larity score between all the class names of the ﬁrst

and the second ontology. To perform that, we fol-

low an all-against-all strategy, which stands for mea-

suring the similarity score between all the the pos-

sible combinations of the entities of the two ontolo-

gies. Then, we choose the maximum similarity score

for every pair of resulted candidate mappings. We

argue that comparing all the names of the two in-

put ontologies is more efﬁcient than comparing only

the labels. For instance, our element level matcher

achieved an F-Measure of 0.830 using only the la-

bels on the anatomy OAEI dataset. However, using

all the names associated with the Anatomy entities,

we succeeded to accomplish an F-Measure of 0.862.

Therefore, we are able to better capture the syntactic

mappings between two ontologies and achieve better

results. We employed similarity measures in order to

compare pairwise names. The efﬁciency of a syntac-

tic matcher is determined by the trade-off between its

resulted F-measure value and its runtime. For exam-

ple, a non-efﬁcient syntactic matcher can produce a

high F-measure in a very long time. The variety of

syntactic similarity measures arises the difﬁculty of

choosing the optimal one associated with its thresh-

old. The optimal choice of the similarity measure and

threshold will immediately improve the alignment re-

sults especially in terms of F-measure. Some re-

searchers (Duan et al., 2010) (Cheatham and Hitzler,

2013) (Ngo et al., 2013) (Sun et al., 2015) analysed

the performance of similarity measures while dealing

with the ontology matching process. However, none

of them analysed the performance of these measures

after performing the pre-processing based on all the

names of an ontology class. Moreover, almost of re-

searchers did not included ISUB (Stoilos et al., 2005)

in their comparison. Consequently, we will reveal the

best similarity measure in terms of F-measure, thresh-

old and runtime. In order to select the optimal F-

Measure, we apply a trimming process using a differ-

ent threshold value for each syntactic similarity mea-

sure. As shown in ﬁgure 2, for every experiment, the

element level matcher has as an input a single mea-

sure associated with a threshold value. The output of

the testing process are the set of evaluation criteria:

the recall, the precision and the F-Measure. During

our series of tests, we vary the threshold value from

the range of 0.1 to 1 with an interval of 0.01.

Figure 2: The input/output schema of the element level test-

ing process.

Then, we compute the F-measure for each thresh-

old interval. The employed tracks in our testing pro-

cess are: the Anatomy track and the Large Biomed-

ical track. For the later, we evaluated the avail-

able similarity measures used by our element level

matcher through the three small fragments: FMA-

NCI , SNOMED-FMA and SNOMED-NCI. Figure 3

draws the variation of F-Measure while adopting dif-

ferent thresholds in the Anatomy track of OAEI. Fig-

ure 4, Figure 5, Figure 6 and Figure 7 show the value

of the optimal F-measure and threshold obtained by

each similarity measure for each OAEI track. The

other tested state of the art syntactic similarity mea-

sures are available at: https://goo.gl/1kUgkH. On all

the three tasks, ISUB with a threshold of 0.9 outper-

form the other similarity measures. We conclude that

ISUB 0.9 is the best measure for ontologies syntactic

matching associated with the pre-processing steps. To

optimise the execution time we took proﬁt of the multi

core architecture of the processor. Therefore, we

adopted an intra-matching parallelism methodology.

Thus, the element level matcher is divided into the

number of available CPU cores and executed in paral-

lel. For a given two input ontologies, the algorithm of

the element level matcher loops classes names of the

source and target indexes and derives for each class

index the set of corresponding names. These names

are compared to each other using a syntactic simi-

larity measure (ISUB). Only the pair of names hav-

ing a threshold above the 0.9 are added to the set of

mappings. Since we are pursuing an alignment multi-

plicity of 1:1, we followed the element level matcher

by a cleaning process of the resulted candidate map-

pings. Therefore, we keep only the maximum similar-

ity score that a single class can have. If two mappings

have the same similarity score we select randomly one

of them.

3.4 Structural Level Matching

For a given set of alignments, which are discovered

by the element level matcher, we can enrich these

mappings by a set of new ones by applying the struc-

tural matching. Our structural level matcher is com-

posed of two structural sub-matchers: siblings and

subclasses. During the ﬁrst sub-matcher, we follow

the intuition that states: if two classes are similar, then

their siblings should be somehow similar (Shvaiko

and Euzenat, 2013). In the second sub-matcher, we

remove the common words between the super class

and its subclasses. Moreover, this matcher admit the

idea that: if two classes are similar, then their sub-

classes should be also similar. In that way, the struc-

tural matching improves the initial matching results

of the element level matcher.

3.4.1 Structural Matcher based on Siblings

This structural matcher detects new mapping based

on the siblings of the already discovered alignments

by the element level matcher. These mappings have

a similarity score between ISUB 0.9 and ISUB 0.8.

Figure 3: The variation of F-Measure over the Anatomy track while adapting different similarity measure and thresholds.

Figure 4: Top F-Measure in Anatomy task.

Figure 5: Top F-Measure in FMA-NCI small fragment.

Figure 6: Top F-Measure in FMA-SNOMED small frag-

ment.

We argue that if two entities are matching, then their

siblings will probably match. Therefore, we decrease

the threshold from 0.9 to 0.8. For a given ontology O

and an entity e

, the j

sibling of an entity e is de-

noted as Sibj

, where

∑

Sibj

∈ [0,m] and m is the

Figure 7: Top F-Measure in SNOMED-NCI small frag-

ment.

number of siblings. We consider a single mapping re-

sulted by the element level matcher as: a

= (e

,s)

with Sib1

is the only sibling of e

. If the similar-

ity score between the two siblings is high (still below

the syntactic threshold), then we consider that the two

siblings will certainly match. However, if the two en-

tities e

and e

have many siblings, then we should

compare all the siblings of e

with the siblings of e

Therefore, we perform an all-against-all comparison

between all the siblings of e

and e

. Next, like the

element level matcher, we pursue an alignment mul-

tiplicity of 1:1. The input of the sibling matching al-

gorithm is the discovered mappings by the syntactic

matcher as well as the two input ontologies. The out-

put of this matcher is the new set of structural map-

pings, which are added to the initial mappings dis-

covered by the element level matcher. Since we used

a threshold of 0.8 ISUB for the structural matcher, we

removed all the mappings, which are already discov-

ered by the syntactic matcher.

3.4.2 Structural Matcher based on Subclasses

For the subclasses structural matcher, we are follow-

ing the intuition that if two classes are similar, then

their sub classes are similar (Shvaiko and Euzenat,

2013). Given an ontology O and an entity e of O,

the j

descendant of e

is denoted as Subj

, where

∑

Sub

∈ [0,n] and n is the number of descendants. To

illustrate, if a single mapping is denoted as a

= ( e

, s), Subj

is the descendent of e

and Subj

the descendent of e

. Therefore, Subj

and Subj

are probably similar. However, if there are many sub-

classes of e

and e

, it will be complex to perform

the mapping process between the descendant, espe-

cially when they have close similarity scores. There-

fore, it is necessary to ﬁnd some other discriminating

features (Shvaiko and Euzenat, 2013). To illustrate,

we consider an example of mapping resulted from the

element level matcher: a

= ( e

, e

, s), with s >

threshold and the name of e

= e

= ”heart ventrice”.

For simplicity, as shown in the ﬁgure 8, each entity of

this mapping has only two subclasses:

• Sub1

, Sub2

where the name of Sub1

= heart

left ventricle and the name of Sub2

= ”heart

right ventricle”.

• Sub1

, Sub2

, where the name of Sub1

”left ventricle” and the name of Sub1

= ”right

ventricle”

Then, we remove the common tokens between the

entity e

and its descendants (Sub1

, Sub2

We also remove the common tokens between e

and

its descendants (Sub1

, Sub2

) . Consequently,

the descendants (sub-classes) became:

• Sub1

= ”left” and Sub2

= ”right”

• Sub1

= ”left” and Sub2

= ”right”

Figure 8: Structural matching based on subclasses illustrat-

ing example.

As a result, we can match the descendants of

( Sub1

, Sub2

) and the descendants of

(Sub1

, Sub2

) by computing the similarity

score between them (all-against-all strategy). Since

we are following a 1:1 multiplicity, then we keep

only the maximum score for each class of the candi-

date structural mappings, which has a similarity score

above the threshold (Monge Elkan 0.85). Unlike

ISUB, Monge Elkan (Monge et al., 1996) is able to

capture the mappings between two textual sequences,

which are containing numerical values. Furthermore,

Monge Elkan is not recommended for large matching

tasks, as a result we are using it only for the structural

matching process. The new discovered mappings are

added to the list of initial alignments which are identi-

ﬁed by the element level matcher. The mappings data

structure contains the list of the already discovered

mappings from the two input ontologies. We notice

that we pursued a cleaning process of the discovered

mappings by the structural matcher in order to dis-

card the list of mappings already found by the element

level matcher as well as structural matcher based on

siblings.

4 EXPERIMENTS AND

DISCUSSIONS

4.1 Experimental Setup

We developed the POMap system using JAVA and

Eclipse as the integrated development environment.

We choose to perform the experiment process analy-

sis through two biomedical tracks of the OAEI bench-

mark as well as the conference track in order to show

the efﬁciency of POMap in various matching tasks

while dealing with OWL ontologies. We employed

for our series of tests a hardware conﬁguration of:

15Gb of RAM coupled with an Intel Core i5-4570

CPU @3.40 Ghz x 4. This conﬁguration is similar

to the one used by the OAEI benchmark. We com-

pare the performance of POMap with the results of

ontology matching system enrolled in the OAEI 2016

campaign obtained using the Anatomy, Conference as

well as the LargeBio track.

4.2 Experimental Results

Anatomy. The anatomy track is a matching task

between the Adult Mouse Anatomy (2744 classes)

and the NCI Thesaurus (3304 classes) describing the

human anatomy. In terms of F-Measure, POMap

is ranked in the fourth position among 14 match-

ing systems while producing a competitive runtime

of 43 seconds. We achieved an F-Measure of 0.893

without using any external knowledge source unlike

the other top performing matching systems (AML,

LogMap and XMap). Table 1 illustrates the position-

ing of POMap (grey color) compared to the partici-

pated matching systems of OAEI 2016.

Conference. This track consists of matching seven

ontologies describing the conference organisation do-

main. Dealing with the conference dataset, we suc-

ceeded to achieve an F-Measure of 64%, which is an

average score compared to the other matching sys-

tems. Since this dataset includes the matching of

Table 1: POMap results in the anatomy track compared to

the OAEI 2016 systems.

System F-Measure Precision Recall Runtime

AML .943 .95 .936 47

CroMatcher .925 .949 .902 573

XMap .896 .929 .865 45

POMap .893 .931 .859 43

LogMapBio .892 .888 .896 11

FCAMAP .882 .932 .837 117

LogMap .88 .918 .846 24

LYAM .869 .863 .876 799

Lilly .83 .97 .794 272

LogMapLite .828 .962 .728 20

LPHOM .718 .709 .727 1601

ALIN .501 .996 .335 306

DkpAomLite .238 .99 .135 372

DkpAom .238 .99 .135 379

properties, we plan to add the data property and ob-

ject property matchers in our next version of POMap.

Large Biomedical Ontologies. The large Biomedical

Ontologies track aim to ﬁnd the alignments between

three large ontologies: The Foundational Model of

Anatomy (FMA), SNOMED Clinical Terms and the

National Cancer Institute Thesaurus (NCI). Table 2

represents the obtained results for the FMA-NCI

small fragment. We succeed to achieve an F-measure

of 90%. Table 3 illustrates the results of POMap for

the FMA-SNOMED small fragment in which we ob-

tained an F-measure of 84% better than two matures

matching systems. We draws in the table 4 the results

of POMap in the SNOMED-NCI small fragment task

in which we achieved an F-measure of 73%. All the

resulted alignments by our system can be downloaded

at https://goo.gl/gRYtQj.

After performing the experimental process, we

can notice that some matching systems (AML,

LogMap, XMAP and Cromatcher) are performing

better than POMap. We argue that these systems are

outperforming POMap because of their implementa-

tion of a variety of matchers. However, POMap im-

plements only three matchers: one syntactic matcher

and two structural matchers. For instance, XMap em-

ploys different matchers such as a translation matcher

and a semantic matcher, which uses external back-

ground knowledge sources. However, POMap do not

perform any semantic matching process. Compared

to the runtime of other matching systems, we mention

that our element level matcher is the responsible of an

exponential growth of the execution time. However,

even with these drawbacks, we are able to accom-

plish acceptable results due to the efﬁciency of our

syntactic matcher resulted by a study of the different

similarity measures. Compared to the top perform-

Table 2: POMap results in the FMA-NCI small fragment

compared to the OAEI 2016 matching systems .

System F-Measure Precision Recall Runtime

XMAP .937 .977 .901 17

FCA MAP .935 .954 .917 236

AML .931 .963 .902 35

LogMap .924 .949 .901 10

LogMapBio .923 .935 .910 1712

POMap .900 .953 .852 241

LogMapLite .887 .967 .819 1

LYAM .796 .721 .889 1043

Lilly .657 .603 .721 699

ALIN .625 .995 .455 5811

DkpAom-Lite .615 .652 .577 1698

DkpAom .616 .652 .577 1547

Table 3: POMap results in the FMA-SNOMED small frag-

ment compared to the OAEI 2016 matching systems.

System F-Measure Precision Recall Runtime

XMAP .912 .989 .846 54

FCAMAP .865 .936 .803 1865

POMap .846 .928 .776 1223

AML .825 .953 .727 98

LogMapBio .801 .944 .696 1180

LOGMAP .799 .948 .690 60

LogMap .343 .968 .209 2

Table 4: POMap results in the SNOMED-NCI small frag-

ment compared to the OAEI 2016 matching systems .

System F-Measure Precision Recall Runtime

AML .797 .904 .713 537

LogMap .771 .922 .663 117

LogMapBio .770 .896 .675 3757

POMap .736 .813 .674 6920

XMap .697 .911 .564 .697

LOGMAPLite .693 .892 .567 693

ing systems of OAEI 2016, which are proposing sev-

eral matchers, POMap produces competitive results

using efﬁcient matchers without relying on any exter-

nal knowledge resources. In order to have a better in-

sight of the efﬁciency of the structural matching step,

the following table 5 draws the results before and after

performing the structural matching process. The grey

color highlights the use of the structural matching.

5 CONCLUSION

In this paper, we described POMap a novel matching

system dedicated for the ontologies matching process.

POMap is employing syntactic techniques in order to

accomplish the element and structural level matching.

Table 5: POMap results before and after (grey color) using

the structural matching.

Track F-Measure Precision Recall

Anatomy .862 .942 .795

Anatomy .893 .931 .859

FMA-NCI (small) .897 .957 .844

FMA-NCI (small) .900 .953 .852

FMA-SNOMED (small) .836 .940 .752

FMA-SNOMED (small) .836 .928 .776

SNOMED-NCI (small) .736 .832 .651

SNOMED-NCI (small) .736 .813 .674

After performing a series of experiments, we provided

the top performing similarity measures and thresholds

that can be employed by an element level matcher

in the case of pre-processed ontologies. We have

proposed two structural matchers, which perform the

syntactic treatment over the siblings as well as the sub

classes. These two structural matchers explore the ini-

tial mappings resulted by the element level matcher in

order to improve the matching system results.

As a future work, we plan to adapt our matching

system in order to achieve a better run time over the

larger datasets of the OAEI campaign. While dealing

with other ontology matching ﬁelds rather than the

biomedical domain, other syntactic similarity mea-

sure can outperform the ones that we recommended.

Thus, we will concentrate on automating the process

of ﬁnding the best similarity measure and threshold

depending on the ontology matching context. There-

fore, we target the prediction of the optimal similar-

ity measure associated with its threshold by extracting

various features related to an ontology matching task.

REFERENCES

Bodenreider, O. (2004). The uniﬁed medical language sys-

tem (umls): integrating biomedical terminology. Nu-

cleic acids research, 32(suppl 1).

Cheatham, M. and Hitzler, P. (2013). String similarity met-

rics for ontology alignment. In International Semantic

Web Conference. Springer.

Cruz, I. F. and Sunna, W. (2008). Structural alignment

methods with applications to geospatial ontologies.

Transactions in GIS, 12(6).

Djeddi, W. E. and Khadir, M. T. (2014). A novel approach

using context-based measure for matching large scale

ontologies. In International Conference on Data

Warehousing and Knowledge Discovery. Springer.

Duan, S., Fokoue, A., and Srinivas, K. (2010). One size

does not ﬁt all: Customizing ontology alignment using

user feedback. The Semantic Web–ISWC 2010.

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

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

erlight ontology matching system. In OTM Confed-

erated International Conferences” On the Move to

Meaningful Internet Systems”. Springer.

Guli

c, M., Vrdoljak, B., and Banek, M. (2016). Cro-

matcher: An ontology matching system based on

automated weighted aggregation and iterative ﬁnal

alignment. Web Semantics: Science, Services and

Agents on the World Wide Web, 41.

Jim

enez-Ruiz, E. and Grau, B. C. (2011). Logmap: Logic-

based and scalable ontology matching. In Interna-

tional Semantic Web Conference. Springer.

Megdiche, I., Teste, O., and Trojahn, C. (2016). An extensi-

ble linear approach for holistic ontology matching. In

International Semantic Web Conference. Springer.

Melnik, S., Garcia-Molina, H., and Rahm, E. (2002). Sim-

ilarity ﬂooding: A versatile graph matching algorithm

and its application to schema matching. In Data Engi-

neering, 2002. Proceedings. 18th International Con-

ference on. IEEE.

Miller, G. A. (1995). Wordnet: a lexical database for en-

glish. Communications of the ACM, 38(11).

Monge, A. E., Elkan, C., et al. (1996). The ﬁeld matching

problem: Algorithms and applications. In KDD.

Mungall, C. J., Torniai, C., Gkoutos, G. V., Lewis, S. E., and

Haendel, M. A. (2012). Uberon, an integrative multi-

species anatomy ontology. Genome biology, 13(1).

Nelson, S. J., Johnston, W. D., and Humphreys, B. L.

(2001). Relationships in medical subject headings

(mesh). In Relationships in the Organization of

Knowledge. Springer.

Ngo, D., Bellahsene, Z., and Todorov, K. (2013). Open-

ing the black box of ontology matching. In Extended

Semantic Web Conference. Springer.

Otero-Cerdeira, L., Rodr

ıguez-Mart

ınez, F. J., and G

omez-

Rodr

ıguez, A. (2015). Ontology matching: A litera-

ture review. Expert Systems with Applications, 42(2).

Porter, M. F. (2001). Snowball: A language for stemming

algorithms.

Schriml, L. M., Arze, C., Nadendla, S., Chang, Y.-W. W.,

Mazaitis, M., Felix, V., Feng, G., and Kibbe, W. A.

(2011). Disease ontology: a backbone for disease se-

mantic integration. Nucleic acids research, 40(D1).

Shvaiko, P. and Euzenat, J. (2013). Ontology matching:

state of the art and future challenges. IEEE Transac-

tions on knowledge and data engineering, 25(1).

Stoilos, G., Stamou, G., and Kollias, S. (2005). A string

metric for ontology alignment. The Semantic Web–

ISWC 2005.

Sun, Y., Ma, L., and Wang, S. (2015). A comparative eval-

uation of string similarity metrics for ontology align-

ment. Journal of Information &Computational Sci-

ence, 12(3).