Biomedical computational systems bene ts from the use of ontologies. However, interconnectivity between these systems is a challenge, specially when the ontologies supporting each system are described in di erent natural languages. Ontology alignment plays a key role in data exchange. Existing ontology matching approaches usually provide only equivalent type of relation in the generated mappings. In this article, we propose a re nement technique to enable the update of the semantic type of the mapping beyond equivalence. Our approach relies on information from the ontology evolution. Our evaluation considered LOINC releases in di erent languages. The results demonstrate the usefulness of ontology evolution changes to support the process of mapping re nement.
Advancements in biomedical research require relying upon vast arrays of voluminous, dynamic, heterogeneous and complex datasets, resulting in di culties to use and reuse available data. This generates an ever greater demand for adequate computer-supported methods for automatically locating, accessing, sharing, analyzing and meaningfully integrating data. Biomedical information systems have intensively relied on semantic technologies such as ontologies to turn the semantic of information explicit for machines.
The number of ontologies created in di erent languages grows as their use
increases [
The semantic relations identi ed during the matching process can be
expanded through mapping re nement. We di erentiate relation from relationship,
where the former represents a mapping, and the latter represents concept
connections in an ontology. Re nement can modify or enrich semantic relations. For
instance, during the re nement process, an equivalence ( ) relation (i.e., a
relation de ning that two interrelated concepts are equivalent) can be modi ed to
an is-a (v) (i.e., representing a relation in which one concept is a specialization
of the other) [
Matching approaches are usually approximative and identify mappings based
on relatedness between concepts. The challenges of mapping re nement are due
to the di culties in establishing semantic relations between concepts, beyond the
relatedness identi ed by the traditional matching procedures. In this context,
enriched semantic correspondences in ontology mapping might boost ontology
merging [
Ontologies are constantly evolving, by adding and removing concepts and relationships over time. These changes indicate how concepts and their relationships with each other evolved. In this article, we investigate whether ontology changes are a valuable source of information to enhance the correspondences found between concepts beyond equivalences based on the type of semantic relation. We use the ontology change operation a ecting ontology entities to re ne the semantic relation in already established mappings. We believe that the use of this information might provide an understanding of how concepts were updated over time to support the decision and application of actions required to modify the type of semantic relation in cross-language mappings.
In this investigation, we de ne mapping re nement actions and explore them in re nement procedures based on categories of ontology changes. We propose and formalize mapping re nement for addition changes and for revision changes.
We conduct an evaluation to assess our proposal concerning releases of the Logical Observation Identi ers Names and Codes (LOIN C) and alignments between English and Spanish language. We applied the de ned re nement procedures based on the ontology changes computed from version to the other of LOIN C. Our study reveals a promising approach on the use of ontology evolution changes to enhance semantic relations in mappings.
The remainder of this article is organized as follows: Section 2 presents related work. Afterwards, Section 3 presents a set of formal de nitions including the research problem. Section 4 reports on our proposal for re nement of crosslingual mappings. Section 5 presents the experimental evaluation conducted with biomedical ontologies and their alignments. Section 6 discusses our ndings and lessons learned. Finally, Section 7 wraps up the article and points out future research.
Several approaches were developed to address the ontology mapping problem.
For instance, Trojahn et al. [
Aleksovski et al. [
The main approaches available in literature for re nement are based on
external resources or manual pattern de nition. The work conducted by Arnold &
Rahm et al. [
The literature has demonstrated the e ects of ontology evolution in
mappings [
The use of ontology evolution in mapping re nements has not been investigated in the literature. Our proposed approach in this investigation di ers from the above mentioned proposals because we rely only on ontology change operations, an information obtained from the ontology itself, without depending on external tools or resources. We leverage the information obtained from ontology change operations to identify re nement actions applicable to mappings. To the best of our knowledge, this approach has not been investigated in the literature. We demonstrate how the evolution of concepts can be useful to enrich the semantics of correspondences already established. 3
Ontology. An ontology O speci es a conceptualization of a domain in terms of
concepts, attributes and relationships [
Similarity between concepts. Given two particular concepts c1 and c2, the
similarity between them can be de ned as the maximum similarity between each
couple of attributes from c1 and c2. Formally:
sim(c1; c2) = arg max sim(a1x; a2y)
(1)
where sim(a1x; a2y) is the similarity between two attributes a1x and a2y denoting
concepts c1 and c2 respectively. We can compute this similarity at di erent
linguistic levels: character, string, and semantic level [
Mapping. Given two concepts c1 and c2 from two di erent ontologies, a mapping m12 can be de ned as: m12 = (c1; c2; semT ype; conf ) (2) where semT ype is the semantic relation connecting c1 and c2.
The following types of semantic relation are considered: equivalent [ ],
narrowto-broad [ ], broad-to-narrow [ ]. For example, concepts can be equivalent (e.g.,
\cabeca" \head") { \cabeca" in Portuguese language { one concept can be less
or more general than the other (e.g., \thumb" \dedo") {\dedo" in Portuguese
language { or concepts can be somehow semantically related ( ). The similarity
between c1 and c2 indicates the con dence (conf ) of their relation [
Ontology change operations. An ontology change operation (OCO) is dened to represent a change in an attribute, in a set of one or more concepts or in a relationship between concepts. OCO is classi ed into two main categories: atomic and complex changes (cf. Table 1). Each OCO in the atomic category cannot be split into smaller operations, whereas each one in the complex category is composed of more than one atomic operation. For instance, the operation chgA(c; a; v) is composed of two atomic operations, delA(a; c) and addA(a; v).
Change operation Description
A addC(c) Addition of a new concept c 2 OXj
t delC(c) Deletion of an existing concept c 2 OXj 1
o addA(a; c) Addition of a new attribute a to a concept c 2 Oj Xj1 1
m delA(a; c) Deletion of an attribute a from a concept c 2 OX
i addR(r; c1; c2) Addition of a new relationship r between two concepts c1 and c2 which belongs to OXj 1
c delR(r; c1; c2) Deletion of an existing relationship r between two concepts c1 and c2 which belongs to OXj 1
chgA(c; a; v) Change of attribute a in concept c with the new value v
C smuobvsetiCtu(cte;p(c1i;;pc2j)) MReopvliancgemofencotnocfecpotncce(patndciit2s OsuXjbt1rebey) cfroonmcepcotnccje2ptOpXj1 to concept p2
om smpelirtg(ec(i;CCkr;)cj) SFpulsiitonofocfoancseepttocfim2uOltiXjpl1e icnotnoceapstestCokf resOulXjtin1ginctooncceopntcsepCtrcj 2OOXjXj
p toObsolete(c) Sets status of concept c to obsolete (c is no longer available)
l delInnerC(ci; pj) Deletion of concept ci where pj 2 sup(ci) and sub(ci) 6= ; from ontology OXj 1
ex daedldLIenanfeCr C(c(ic;ip;jp)j) ADdeldeittiioonnoofflaeasfucboncocenpctepcti cwihuenredeprjt2hesucopn(ccie)patnpdj s2usbu(cpi()c=i) ;tofrtohmeoonnttoollooggyyOOXjXj 1
addLeafC(ci; pj) Addition of leaf concept ci where pj 2 sup(ci) and sub(ci) = ; to the ontology OXj
revokeObsolete(c) Revokes obsolete status of concept c (i.e., c becomes active)
We denote successive ontology versions derived from evolution by Oj 1 and Oj
to identify ontologies created in time j 1 and j. Changes may occur from one
version to another, and we consider existing tools to automatically detect change
operations [
Given a mapping m12 2 LjXY associated with a concept c1 a ected by changes in the ontology, the challenging issue is to determine an exact and suited action of re nement to apply to m12. To address this challenge, we de ne and formalize a set of mapping re nement actions (cf. Subsection 4.1).
The mapping re nement actions are part of re nement procedures, playing a key role to improve the quality of mappings. The objective is to enrich the mapping set by considering di erent semantic relations between concepts, for instance, equivalence relations can be re ned to is-a or part-of.
In this investigation, we study how LjXY can be re ned (e.g., new mapping relations derived) based on ontology changes related to ontology evolution. The re ned output consists of the LXj Y . In particular, we address the following re0 search questions:
How to exploit ontology change operations for mapping re nement? Is it possible to reach mapping re nement without applying a new matching operation in the whole target ontology? What is the impact of using evolution information on the mapping re nement e ectiveness? 4
Re nement of Biomedical Ontology Mappings across Languages We propose and formalize a set of re nement actions aiming at re ning mapping sets (Section 4.1) and how these actions are applicable in a re nement procedure (Section 4.2). 4.1
We present an approach to re ne ontology mappings based on di erent types of ontology changes (Table 1). The proposal explores OCOs for re ning mappings individually. For this purpose, we de ne actions as pre-de ned behaviours of mapping re nement into algorithms designed to enrich ontology mappings according to ontology evolution (cf. Section 4.2).
The distinct actions representing di erent possibilities for re ning mappings include: mapping movement, mapping derivation, semantic relation modi cation and no action. In the following, we formally describe each action. To this end, let m12 2 LjXY be the mapping between two particular concepts c1 2 OXj and c2 2 OYj . The actions are illustrated in Figure 1
Mapping derivation source. This is an action for which an existing mapping from LjXY derives a new mapping with the same target concept and di erent source concept. This action results in addition of a new mapping mk2 to LXj Y . 0 deriveS(m12; ck) ! m12 2 LjXY ^ mk2 2= LjXY ^ (9ck 2 OXj ; mk2 2 L0Xj Y ^ sim(c1; ck) )^
0j m12 2= LXY (3) where sim(c1; ck) denotes the similarity between c1 and ck 2 neighborhood(c1), and denotes the threshold used to compare the derived mapping.
Mapping derivation target. This is an action for which an existing mapping m12 in LjXY derives a new mapping with the same source and a di erent target. This action results in addition of a new mapping m1v to L0Xj Y . deriveT (m12; cv) ! m12 2 LjXY ^ m1v 2= LjXY ^ (9cv 2 OYj ; m1v 2 L0Xj Y ^ sim(c1; cv) )^
0j m12 2 LXY
Semantic relation modi cation. This is an action in which the type of the semantic relation of a given mapping is modi ed. This action is designed for supporting the re nement of mappings with di erent types of semantic relations rather than only considering the type of equivalence relation ( ). modSemT ype(m12; new semT ype12) new semT ype12 2 f?; ; ; ; ^ semT ype12 6= new semT ype12g 0j ! m12 2 LXY ^ (4) (5) The action for the modi cation of semantic relation can be applied in conjunction with the actions of move of mapping and derivation of mapping. That is when moving a mapping, it is also possible to modify the type of the semantic relation of such mapping. The same applies for derivation of mapping.
No Action. This action does not modify any aspect of a mapping m12.
N oAction(m12)
j 0j ! m12 2 LXY ^ m12 2 LXY (6) 4.2
The mapping re nement phase takes into account concepts from one version of
the source ontology to another (OXj 1 and OX
j ) to re ne a candidate mapping
set (suggests modi cations in the set of mappings). The necessary instances of
OCOs are identi ed from one ontology version at time j 1 to another at time j
with a di computation [
The candidate mapping set LjXY undergoes the mapping re nement procedure. We describe the procedure in two phases: 1. The output of executed ontology change detection tools is used to identify mappings with potential of re nement. The identi cation is based on the type of ontology evolution operations that a ected the concepts. For instance, the addition of a concept to an ontology may indicate a specialization j , the concept \Eagle" was added as child of of another concept (e.g., in OX the concept \Bird ", being the former a specialization of the latter). Therefore, any candidate mapping involving the concepts \Eagle" and \Bird " are identi ed with possibility of re nement. 2. After the selection of mappings for re nement, for each selected mapping from LjXY , an action is executed based on the type of ontology change. The action may include a direct decision to perform modi cation in the semantic relation of the candidate mapping (e.g., a relationship may be replaced with a v), or other appropriate action.
Algorithm 1 presents the main procedure to re ne LjXY . The input is the candidate mappings LjXY and the dif f(OXj 1j;OXj ). For each mapping m12 2 LjXY , the algorithm veri es if the concept c1 2 OX was a ected by change operations with the use of the dif f(OXj 1;OXj ). The algorithm then invokes the appropriate procedure for each case by considering addition change operations and revision change operations. If the concept was not a ected by change operations from the dif f(OXj 1;OXj ), then noAction is applied to (m12). The output is the re ned 0j . mapping LXY
We grouped the OCOs into two categories: (i) AdditionOCO adds concepts or information to concepts into the ontology. It consists of OCOs by including: addC(c), addInnerC(cs; ps), addLeaf C(cs; ps), revokeObsolete(c), addA(a; cs) and addR(r; cs1; cs2); (ii) The RevisionOCO group of ontology changes revise existing concepts. It consists of OCOs such as: merge(Ck; cs) and split(ci; Cs). In the following, we explain the procedures involved by Algorithm 1.
AdditionProcedure. This procedure is invoked when c1 was a ected by some OCO in the AdditionOCO group. Algorithm 2 presents the proposed strategy for re ning mappings associated to addition changes. For each mapping m12, the neighborhood of the both c1 and c2 is retrieved to perform a local rematch. The rematch function receives a set of source concepts C1 and a set of target concepts C2 and returns a similarity matrix (simMatrix ). The objective in applying a local rematch is to compare the similarities between the neighborhood of the source and target concepts. The similarity values found then drive modi cations to the semantic relation established in m12.
Require: LjXY ; dif f(OXj 1;OXj ) 1: for all m12 2 LjXY do 2: for c1 2 m12 do 3: if AdditionOCO(c1) 2 dif f(OXj 1;OXj ) then 4: AdditionP rocedure(m12) 5: else if RevisionOCO(c1) 2 dif f(OXj 1;OXj ) then 6: RevisionP rocedure(m12; dif f(OXj 1;OXj )) 7: else 8: noAction(m12); 9: end if 10: end for 11: end for
0j 12: return LXY
Require: m12 1: for c1 2 m12 do 2: neighC1 neighborhood(c1); neighC2 neighborhood(c2); simM atrix(C1; C2) rematch(neighC1; neighC2); 3: for all (c1i; c2i) 2 simM atrix(C1; C2) do 4: if c1i = sup(c1) and (sim(c1i; c2) > sim(c1; c2)) then 5: semT ype relation(c1i; c1); modSemT ypeM (m12; semT ype); deriveS(m12; c1i); 6: end if 7: if (c2i = sup(c2) or c2i = sub(c2)) and
sim(c1; c2i) => sim(c1; c2) then 8: deriveT (m12; c2i); 9: end if 10: end for 11: end for For example, if sim(sup(c1); c2) > sim(c1; c2), the algorithm modi es the semantic relation in m12 to the same semantic relation of sup(c1) and c1 and add a new mapping between sup(c1) and c2. The local rematch also helps establishing a derivation of mapping when the sim(c1; sub(c2)) sim(c1; c2) or sim(c1; sup(c2)) sim(c1; c2).
We present an example to illustrate the AdditionProcedure. Ontology OX evolved over time by generating di erent versions from time j 1 to time j. Figure 2(A) illustrates the changes. A set of candidate mappings LjXY between OXj and OYj , at time j, is given as input for the re nement procedure. Figure 2(B) illustrates the mapping m12 2 LjXY between concepts c1 \Angina" and c2 \Cardiopatia". The re nement procedure requires as input the list of change operations (OCOs) detected from one version of the ontology to another. Similarity values between concept c1 \Angina" and the concepts of the neighborhood of the target concept \Cardiopatia" at time j are calculated via local rematch (cf. Figure 2(C)). If the similarity value between the concepts c1 \Angina" and some neighbor c2i of c2 is higher than the original similarity value given by sim(c1; c2), i.e. sim(c1; c2i) sim(c1; c2), the algorithm derives a mapping between c1 and c2i to re ect this nding (cf. Figure 2(D)).
RevisionProcedure. This procedure is used to re ne mappings when c1 was a ected by some OCO in the RevisionOCO group. Algorithm 3 describes the proposed strategy for the re nement. For each input mapping m12, the algorithm j 1 involved in merge or split ontology change retrieves the concepts from OX operations. In the merge operation, an initial set of concepts Ck OXj 1 gives place to a concept c1 2 OXj . On the other hand, in a split operation, an initial concept c1 2 OXj 1 is split in a set of concepts Cs OXj .
The algorithm extracts the before evolution concepts c1 (in the split) and the set of concepts Ck (in the merge) and computes the similarity between them with c2 2 m12. The algorithm explores the similarity values between c2 and fc1; Ckg to extract information and re ne m12. For example, an useful information for j 1 involved in the re nement is the similarity value between the concept ci 2 OX
split of c1 2 m12 ^ c1 2 Cs and c2. If sim(ci; c2) > sim(c1; c2), we can infer that c1 and c2 do not hold an relation because c1 is the result of the split operation of ci into more speci c concepts, not equivalent concepts. We can use this information to re ne the semantic relation of m12. 5
We evaluate our approach to ontology alignment re nement by applying it to ontologies in the biomedical domain.
We considered the Logical Observation Identi ers Names and Codes (LOIN C) ontology published in English and its linguistic variant in Spanish. LOINC provides a standard for identifying clinical information (laboratory and clinical test results) in electronic reports. LOINC is freely available and widely used in 175 countries4. The English variant of LOINC contains 89,271 entities and the Spanish variant contains 54,599 entities.
LOINC presents a regular update schedule of twice a year, providing an amount of ontology changes in every new version available. Update changes in the ontology entities are provided in every release in a separate document, specifying the change operations undergone by the entities. The version selected for this evaluation was the 2:65, released in December 2018.
Our proposed technique requires an initial mapping set as input. For this purpose, we used the mapping set already established between the two linguistic 4 As of September 12, 2019. variants of LOINC (one ontology in English language and the other in Spanish language). Each entity has a unique permanent identi er named LOINC code (in the sense that it cannot be reused even if the entity is deprecated). This code is invariable across linguistic variants. We use LOINC code to identify equivalent entities between the two selected ontologies. In the release version 2:65, the changes available in the document of updates were AdditionOCO and minor changes in labeling. In particular, we focused our evaluation in AdditionOCO actions. Only the updates performed from version 2:64 to version 2:65 of the ontology in the English variant are considered for this experiment.
Based on this candidate mapping set, we applied our de ned Algorithm 1
to invoke the appropriate re nement actions based on the change operations
undergone by the entities ek 2 LOIN Ce2n:65 participating in the alignment. We
employed the Levenshtein edit-distance [
Our results present real-life examples of the outcome in applying our technique to LOINC entities. Table 2 presents the results as the e ect of applying the re nement actions. The original input mapping set Le2n:65es contains 54599 correspondences. In the last version update, from release version 2:64 to 2:65, over 8000 entities have su ered some OCO, with 1408 entities undergoing AdditionOCO. All of these entities were involved as part of mappings, thus allowing the application of the AdditionProcedure described in Subsection 4.2. The technique has re ned the mapping set and the re nement actions performed generated 1513 new semantically enriched mappings, by increasing the number of mappings to a total of 56113.
Inputsmizaepping ATdodtaitlioofncOhaCnOgestyopfe Total ofmchaapnpginesgsa ecting aMppalpypi niangcgtrioseinzsneeamfteenrt 54,599 1,408 1,408 56,113
The re nement of the candidate mapping set relies on the similarity values computed between the concept \Zika virus Ab.IgG " and the concepts of the neighborhood of the target concept \Virus Zika IgG ". To this end, the algorithm 2 performed a cross-lingual local rematch de ned in its step 2. As a result of this operation, the algorithm applies re nement action deriveT and derives a mapping between \Zika virus Ab.IgG " and \Virus zika" (cf. Figure 3(B)). In this investigation, we assumed that ontology evolution is useful to decide on the application of mapping re nement actions and improve the mapping quality outcome. To the best of our knowledge, the use of ontology change operations for mapping re nement has never been proposed in literature. This aspect refers to the key originality of this article. We demonstrated the usefulness of ontology changes to aid the process of ontology mapping re nement in a case of aligned biomedical ontologies.
The actions performed during the re nement procedure enrich the candidate mapping set with semantic context, which is bene cial for ontology merging and system integration. Our proposal de ned algorithms that reach mapping re nement without applying a new matching operation with the whole target ontology. In addition, our technique enables the update of the semantic type of mappings. The current approach focuses only on is-a and part-of relationships. Other relationship types will be addressed in future work.
The main advantage of using evolution information is the possibility of renement without the need of an external resource. The information required to re ne is available in the ontology itself or can be computed based on a early version of the ontology, by using ontology di computation tools to calculate the change history. This is particularly useful when external resources are unavailable to aid in the re nement task. Due to the lack of experimental results concerning mapping re nement in literature, we were unable to compare our method to others approaches.
The use of OCOs for re nement purposes is limited to mappings with at least one participant ontology with multiple versions available to calculate history changes; or a list of updates between versions must be available. Our proposed procedure depends on the set of ontology changes, thus only mappings with entities associated with ontology change(s) are eligible for the procedure. This limits the amount of mappings that can be re ned with this technique. For example, in the conducted evaluation we were able to re ne mappings where one of the participants has undergo AdditionOCO action, because that was the only change action available for the entities participating in mappings.
Our main goal with this evaluation was to assess the usefulness of evolution change information in mapping re nement, by verifying if the semantic relationships in mappings are expanded beyond equivalence in a meaninful way. The correctness of generated output will be evaluated in future work.
Biomedical ontologies usually are syntactically regular and this condition might not be found in other domains. Further investigations are required to verify the applicability of our proposed technique to di erent domains.
The procedure uses similarity measures for local rematch. The selection of the applicable similarity measure depends on the addressed problem, because there are similarity measures which depends on background knowledge, such as, semantic similarity measures relying on semantic networks of a speci c domain. In our experimental evaluation, we chosen a simple, widely used and domain neutral similarity measure. Nevertheless, any similarity measure appropriated for the problem can be used. 7
Ontology mapping re nement remains an open research problem. The result of mapping re nement increases the usefulness of mapping sets, bene ting the semantic data integration of systems. This article proposed an original approach with the use of ontology change operations detected during ontology evolution to leverage mapping re nement. We contributed with the formalization of re nement actions and de ned algorithms to apply them based on ontology evolution operations. We demonstrated the evaluation of the technique using biomedical real-world ontologies across di erent languages. Future work involves to include domain specialists to evaluate the correctness proposed concept mappings and their speci c type of semantic relation. We also plan to investigate this approach in monolingual mappings and evaluate the impact of other similarity measures in the quality of the re nement.
This work is supported by CNPq (grant #307560/2016-3), the S~ao Paulo Research Foundation (FAPESP) (Grant #2017/02325-5). This study was nanced in part by the Coordenac~ao de Aperfeicoamento de Pessoal de N vel Superior Brasil (CAPES) - Finance Code 001.5 5 The opinions expressed in this work do not necessarily re ect those of the funding agencies.