An ontology is a formal representation of a domain modeling the entities in the domain and their relations. When a domain is represented by multiple ontologies, there is need for creating mappings among these ontologies in order to facilitate the integration of data annotated with and reasoning across these ontologies. The objective of this paper is to present our experience in aligning two large anatomical ontologies and to reflect on some of the issues and challenges encountered along the way. The anatomical ontologies under investigation are the Foundational Model of Anatomy (FMA) and GALEN. Our approach to aligning concepts is automatic, rule-based, and operates at the schema level, generating mostly point-to-point mappings. It uses a combination of domain-specific lexical techniques and structural and semantic techniques (to validate the mappings suggested lexically). It also takes advantage of domainspecific knowledge (lexical knowledge from external resources such as the Unified Medical Language System, as well as knowledge augmentation and inference techniques). Overall, the lexical alignment followed by structural validation identified 3,029 pairs of equivalent concepts in the FMA and GALEN, accounting for about 4% of all FMA concepts and 32% of all GALEN concepts.
Over the past four years, as part of the Medical Ontology Research project at the U.S. National Library of Medicine, we have developed domain knowledge-based techniques for aligning large anatomical ontologies, with the objective of exploring approaches to aligning representations of anatomy differing in formalism, structure, and domain coverage. We started by aligning concepts point-to-point in two large ontologies of human anatomy, using lexical and structural techniques [1]. We later tested these techniques on other pairs of anatomical ontologies, both within and across species [2, 3]. We also investigated the complex alignment of groups of concepts [4] and that of relationships [5]. Finally, we investigated the possibility of deriving the indirect alignment of two ontologies through their direct alignment to a reference ontology [6]. The objective of this paper is to present our experience in aligning two large anatomical ontologies point-to-point and to reflect on some of the issues and challenges encountered along the way. In particular, we want to show the importance of domain-specific knowledge in our alignment strategies. Many features of our system are specific to the domain of anatomy, which is why ‘anatomy’ is the only OAEI data set to which we applied our methods. 1.1
The approach to aligning anatomical ontologies presented here is automatic, rulebased, and operates at the schema level, generating point-to-point mappings. It uses a combination of domain-specific lexical techniques (to map entities at the element, not instance level) and structural and semantic techniques (to validate the mappings suggested lexically). It also takes advantage of domain-specific knowledge (lexical knowledge from external resources such as the Unified Medical Language System® (UMLS®) [7], as well as knowledge augmentation and inference techniques).
The many ontology alignment systems available include PROMPT [8], CUPID [9], FCA-Merge [10], HCONE-Merge [11], and GLUE [12]. With AnchorPrompt [13], we share the notion of “anchor” (i.e., a pair of related terms across ontologies, established by lexical similarity in our case) and the use of shared paths between anchors across ontologies to validate the similarity among related terms. Therefore, AnchorPrompt is undoubtedly the system to which our approach is the most closely related. The major differences between AnchorPrompt and our approach can be summarized as follows. AnchorPrompt creates a sophisticated similarity score based on path length and other features. In contrast, we use a simpler validation scheme based on paths restricted to combinations of taxonomic and partitive relations, suitable for the anatomical domain. Unlike AnchorPrompt, our approach does not rely on path length and is therefore less sensitive to differences in granularity between ontologies. AnchorPrompts requires ontologies to be accessible in the frame-based Open Knowledge Base Connectivity (OKBC) protocol, while our approach is not specific to any particular formalism. 1.2
We identify one-to-one concept mappings between the Foundational Model of Anatomy (FMA) and GALEN using lexical resemblance between concept names and then validate the mappings through shared hierarchical paths among concepts across ontologies.
The lexical alignment identifies shared concepts across ontologies based on lexical similarity between concept names. Both preferred concept names and synonyms, if any, are used in the lexical alignment process. Lexical similarity is assessed through exact match and after normalization. The normalization program distributed with the UMLS provides a linguistically-motivated model for lexical resemblance adapted to the specificity of biomedical terms, abstracting away from minor differences in terms including case, hyphen, inflection and word order variations [14].
Concepts exhibiting similarity at the lexical level across ontologies are called anchors, as they are going to be used as reference concepts in the structural validation and for comparing associative relationship. Additional anchors are identified through synonymy in an external resource: the Unified Medical Language System (UMLS). More specifically, two concepts across ontologies are considered anchors if their names are synonymous in the UMLS Metathesaurus (i.e., if they name the same concept) and if the corresponding concept is in the anatomy domain (i.e., has a semantic type related to Anatomy).
Examples of anchors, shown in Figure 1, include the concepts Cardiac valve in FMA and Valve in heart in GALEN, identified as anchor concepts because Cardiac valve has Valve of heart as a synonym in FMA and Valve in heart matches Valve of heart after normalization. Additionally, Fibrous ring of mitral valve (with synonym Mitral anulus) in the FMA and Mitral ring in GALEN form an anchor because Mitral anulus and Mitral ring are synonyms, i.e., they are both names for the concept Structure of anulus fibrosus of mitral orifice in the UMLS.
In the structural validation of the lexical alignment, the first step is to acquire the semantic relations explicitly represented in the ontologies. Inter-concept relationships are generally represented by semantic relations <c1, r, c2>, where the relationship r links concepts c1 and c2. Because they form the backbone of anatomical ontologies and are therefore more likely to be represented consistently across ontologies, hierarchical relationships only are considered at this step. These relationships are IS-A and PART-OF, along with their inverses INVERSE-IS-A and HAS-PART, respectively. Having extracted the relations explicitly represented in the ontologies, we then normalize the representation of the relations in each ontology in order to facilitate structural comparisons across ontologies. We first complement the hierarchical relations represented explicitly with their inverses as necessary. Implicit semantic relations are then extracted from concept names (augmentation) and various combinations of hierarchical relations (inference). Augmentation and inference are the two main techniques used to acquire implicit knowledge from the FMA and GALEN. For a detailed analysis of the contribution of each technique, the interested reader is referred to [15].
Complementation. As partial ordering relationships, hierarchical relationships are anti-symmetric. IS-A and PART-OF have inverse relationships, INVERSE-IS-A and HASPART. Except for IS-A, not every relation is represented bidirectionally. For example, <Right breast proper, HAS-PART, Right mammary gland> is explicitly represented in the FMA but its inverse relation is missing. In canonical anatomy, the inverse relations are essentially always valid, although this is not necessarily the case in the real world [16]. For the sole purpose of aligning ontologies, in order to facilitate the comparison of paths between anchors across ontologies, we complement the FMA and GALEN with the inverse relations that are not explicitly represented. For example, we generated the relation <Right mammary gland, PART-OF, Right breast proper>.
Augmentation attempts to represent with relations knowledge that is otherwise embedded in the concept names. Augmentation is based on linguistic phenomena, such as the reification of partitive relations. In this case, a relation <P, PART-OF, W> is created between concepts P (the part) and W (the whole) from a relation <P, IS-A, Part of W>, where the concept Part of W reifies, i.e., embeds in its name, the PART-OF relationships to W. For example, <Neck of femur, PART-OF, Joint> was added from the relation <Neck of femur, IS-A, Component of joint>, where the concept Component of joint reifies a specialized PART-OF relationship. Examples of augmentation based on other linguistic phenomena include <Sweat gland, IS-A, Gland> (from the concept name Sweat gland) and <Extensor muscle of leg, PART-OF, Leg> (from the concept name Extensor muscle of leg). The semantics of nominal modification generally corresponds to subsumption (e.g., the head noun gland modified by sweat is a hypernym of gland). In contrast, the semantics of prepositional clauses introduces by of is not necessarily a partitive relation (e.g., glass of wine is not part of wine). Here, domain knowledge was required to assess what relations can be automatically extracted with high accuracy in the particular context of anatomical terms. We determined that partitive relations could be accurately created from prepositional clauses introduced by of in anatomical terms containing no other prepositions.
Inference generates additional semantic relations by applying inference rules to the existing relations in order to facilitate the comparison of paths between anchors across ontologies. These inference rules, specific to this alignment, represent limited reasoning along the PART-OF hierarchy, generating a partitive relation between a specialized part and the whole or between a part and a more generic whole. For example, <First tarsometatarsal joint, PART-OF, Foot> was inferred from the relations <First tarsometatarsal joint, IS-A, Joint of foot> and <Joint of foot, PART-OF, Foot>. Analogously, <Interphalangeal joint of thumb, PART-OF, Finger> was inferred from the relations <Interphalangeal joint of thumb, PART-OF, Thumb> and < Thumb, IS-A, Finger>.
With these explicit and implicit semantic relations, the structural validation identifies structural similarity and conflicts among anchors across ontologies. Structural similarity, used as positive structural evidence, is defined by the presence of common hierarchical paths among anchors across ontologies, e.g., <c1, PART-OF, c2> in one ontology and <c1’, PART-OF, c2’> in another where {c1, c1’} and {c2, c2’} are anchors across ontologies1. The anchor concepts Cardiac valve in FMA and Valve in heart in GALEN, presented earlier, received positive structural evidence because they share hierarchical paths to some of the other anchors across ontologies. For example, as illustrated in Figure 1, Cardiac valve is related to Heart (PART-OF), to Mitral valve (INVERSE-IS-A) and to Mitral ring (HAS-PART).
Conflicts, on the other hand, are used as negative structural evidence. The first type of conflict is defined by the existence of hierarchical paths between the same anchors across ontologies going in opposite directions, e.g., <c1, PART-OF, c2> in one ontology and <c1’, HAS-PART, c2’> in another. The second type of conflict is based on the disjointness of top-level categories across ontologies (i.e., semantic constraints). 1 The transitive closure of hierarchical relation greatly facilitates paths comparison across ontologies, because complex paths between anchors are represented by a single relation. For example, Point in FMA is a kind of Dimensional entity, while Pointing in GALEN is a Voluntary movement of upper extremity, which is a Process. Dimensional entity and Process being disjoint top-level categories, the two concepts Point in the FMA and Pointing in GALEN are considered semantically distinct, which prevents them from being aligned although they are lexically equivalent (after normalization).
Heart
Heart Fibrous ring of mitral valve
Mitral valve Cardiac valve
FMA Legend
N isa part of
Mitral ring
Mitral valve Valve in heart
GALEN N lexical mapping Normalization UMLS synonymy
involving the FMA and GALEN, we purposely stayed away form any particular formalism in order to avoid distorting the source ontologies during the conversion process. Instead, we simply extracted <concept, relationship, concept> triples from the two ontologies in their original formalism (frames in Protégé for the FMA and the description logic language GRAIL for GALEN) using the Application Programming Interface provided with the ontology. In the OAEI experiment, the FMA and GALEN were converted into OWL Full by the organizers. Since our alignment approach is not designed to specifically take advantage of the OWL formalism, we again extracted concept names and <concept, relationship, concept> triples from the class definitions in OWL.
More specifically, for the FMA, we used rdf:ID to identify concepts, rdf:label and Preferred_name to acquire concept names, and rdfs:subClassOf to acquire taxonomic relations. The various partitive relations represented in the FMA (e.g., part_of, constitutional_part_of, regional_part_of) were acquired using the corresponding properties. An inverse property is specified for 29 properties in the OWL file. The other properties were not used in the alignment.
Because of the deep nesting exhibited by the GALEN file in OWL, we performed the extraction not from the OWL file itself, but from the .pont and .pins files resulting from its conversion into the CLISP format by the OWL plugin in Protégé. The .pins file contains all the information we needed and is the only one we ended up using. (This file is available for download at: http://mor.nlm.nih.gov/pubs/supp/2006-oaeisz/OpenGALEN.pins). We used rdf:ID to both identify concepts and acquire concept names (after tokenization), and rdfs:subClassOf to acquire taxonomic relations. The various partitive relations represented in GALEN (e.g., isPartOf, isComponentOf, IsSurfaceDivisionOf) were acquired using the corresponding properties. No inverse properties are specified in the OWL file. Inverse properties were added manually in 13 cases for alignment purposes. The other properties were not used in the alignment.
study, we consider as only one PART-OF relationship (with HAS-PART as its inverse) the various kinds of partitive relationships present in the FMA (e.g., part_of, constitutional_part_of, regional_part_of) and GALEN (e.g., isPartOf, isComponentOf, IsSurfaceDivisionOf).
semantic mismatches, we added pairwise disjointness axioms between their top-level classes across the two ontologies. For example, as shown earlier, because we define Dimensional entity in the FMA and Process in GALEN as being disjoint top-level categories, Point in FMA – a kind of Dimensional entity – and Pointing in GALEN – a kind of Process – are prevented from being aligned although they are lexically equivalent after normalization. 1.5
The result of our alignment for the ‘anatomy’ data set is available at: http://mor.nlm.nih.gov/pubs/supp/2006-oaei-sz/Zhang&Bodenreider.rdf in the format specified by the OAEI organizers at: http://oaei.ontologymatching.org/2006. 2
The only data set for which we report results is ‘anatomy’. Overall, the lexical alignment followed by structural validation identified 3,029 pairs of equivalent concepts in the FMA and GALEN, accounting for about 4% of all FMA concepts and 32% of all GALEN concepts.
characteristics of the data sets under investigation are listed in Table 1, including the number of classes, concept names, and types of partitive relationships. The number of IS-A and partitive relations extracted from the OWL file and generated by complementation, augmentation and inference is shown in Table 2. Not surprisingly, in both ontologies, a majority of relations come from inference, which performs similarly to a transitive closure of the hierarchical relations. Also listed in Table 2 is the small number of relations removed from the ontologies because they create cycles. FMA 72,560 44,597 0 Lexical alignment. 3,132 matching anchor concepts were identified lexically, accounting for about 4.3% of the FMA concepts and 32.7% of GALEN concepts. Of these, 378 anchors were identified through UMLS synonymy.
Structural validation. For the 3,132 anchors, the presence of paths to other anchors is searched in both ontologies, as well as the existence of conflicts is assessed. This information is used for the structural validation of the alignment, the results of which are summarized in Table 3.
Anchors with positive structural evidence. 96.7% of the anchors receive positive evidence, most of them sharing hierarchical paths of the same type (e.g., Cardiac valve 2056 973 9 in FMA and Valve in heart in GALEN, presented earlier). An example of shared “compatible” hierarchical relations is the anchor Pelvic fascia. In both ontologies, this concept is linked to Visceral pelvic fascia, but, although going in the same direction, the relationship is INVERSE-ISA in GALEN and HAS-PART in FMA. For alignment purposes, sharing compatible hierarchical relations is deemed a sufficient condition. Anchors with positive structural evidence are presented in our result file with “=” as value for “relation” and 1.0 as value for “measure”.
Anchors with negative structural evidence. About 0.7% of the anchors represent conflicts between the two ontologies. For example, the relationship between the anchors Apex of bladder and Urinary bladder is PART-OF in GALEN but HAS-PART in FMA. Another type of conflict is represented by the semantic incompatibility between Point (the dimensional entity) in the FMA and Pointing (the process) in GALEN presented earlier. Anchors with negative structural evidence are presented in our result file with “!=” as value for “relation” and 1.0 as value for “measure”.
Anchors with no structural evidence. 2.6% of the anchors do not receive any structural evidence. For example, although linked to Body Cavity (is-a) and Peritoneal Cavity (inverse-isa) in GALEN, Serous Cavity has no connections to other anchors in FMA. The absence of any paths to other anchors represents about two thirds of the cases. The remaining cases correspond to the absence of shared paths to other anchors across ontologies. For example, although Eyebrow is linked to two anchors in FMA (e.g., Set of hairs), and two in GALEN (e.g., Face), none of these paths are shared across ontologies. Anchors with no structural evidence are presented in our result file with “=” as value for “relation” and 0.5 as value for “measure”. 3
3.1
Overall, the results we obtained on this ‘anatomy’ data set are essentially similar to the results obtained in previous experiments from slightly different versions of these two large anatomical ontologies [1]. The main difference is that, in the past, we mapped all of GALEN – not just its anatomy subset – to the FMA, leading to increased ambiguity between anatomical and non anatomical concept names. In contrast to what most teams reported at OAEI 2005, for us, this ‘anatomy’ data set was actually slightly less complex than what we are used to. 3.2
The strengths and weaknesses of our system have been analyzed in previous papers [17]. The major difference with other systems is that we take advantage of domain knowledge throughout the mapping process. For example, we use specific tools and resources, including normalization techniques developed for biomedical terms and synonyms from the Unified Medical Language Systems. We also developed techniques specific to the anatomical ontologies under investigation in order to represent explicitly relations implicitly present in these ontologies. These additional synonyms and relations increase the chances of identifying matches both at the lexical and structural level. Conversely, because most of these techniques are specific to anatomical ontologies, our system is unlikely to perform well on other types of ontologies. Similar techniques would have to be developed for other domains in order to obtain similar levels of performance. Finally, specifying disjointness between toplevel categories was more useful in previous experiments when all of GALEN was aligned with the FMA. For example, because we define Anatomical structure in the FMA and Inert solid structure in GALEN as being disjoint top-level categories, Nail in FMA – a kind of Anatomical structure – and Nail in GALEN – a kind of Surgical fixation device – were prevented from being aligned although they are lexically identical. 3.3
In our opinion, one of the most controversial aspects of the OAEI procedure is the decision of the organizers to convert the anatomical data sets under investigation – the FMA and GALEN – from their original formalism to OWL Full. On the one hand, the intuition is that having to deal with only one formalism will make it easier for participants to align these two massive ontologies. On the other, the result of the conversion is dependent on the original modeling. The FMA’s native environment is frame-based and makes heavy use of metaclasses [18], while GALEN’s original formalism is based on the description logic language GRAIL – the GALEN Representation and Integration Language [19]. The mechanical conversion of these two ontologies to OWL Full is therefore likely to result in differing representations, e.g., because of the difference in the use of metaclasses and instances.
In practice, the OAEI participants cannot reliably take advantage of such features of the underlying representation for identifying equivalent entities across ontologies. This issue was illustrated by the example given by Kalfoglou & Hu in their attempt to align the FMA and GALEN in the 2005 OAEI campaign [20]. They specifically question the mapping established by [17] between Pancreas in FMA and Pancreas in GALEN on the basis that it “is defined as a class in FMA … whereas in GALEN (OpenGALEN) [it is defined] as an instance of [the] class "Body Cavity Anatomy"”. Looking beyond formalism, it is quite clear to domain experts that the entities referred to by Pancreas in the FMA and GALEN are equivalent. For example, in both ontologies, they are defined as a kind of organ and have the same parts, including exocrine and endocrine pancreas, and head, neck, body and tail of pancreas. The containment relations are slightly different in the FMA (retroperitoneal space) and GALEN (abdominal cavity), but compatible for alignment purposes.
As shown recently, converting the entire FMA to OWL is a difficult exercise because it requires some understanding of both the modeling strategy and the domain, which is beyond what can be expected from the mechanical translation provided by the “export to OWL” function in Protégé [21]. However, the mechanical conversion was sufficient for us to reliably extract the information we needed (namely, concept names and taxonomic and partitive relations). Our point here is converting these two ontologies to OWL Full is an overkill and gives the false impression that the formalism can be relied upon. 3.4
In 2005, only two teams participated in the anatomy challenge [20, 22]. Their reports essentially outline the difficulties encountered along the way, including the large size of the anatomical ontologies and the transformation of both ontologies from their native format into OWL Full. Hopefully, more participants will have aligned these two large ontologies in 2006. While anatomical ontologies represent good data sets for evaluating alignment strategies, important issues – beyond size and formalism – remain the absence of a gold standard alignment and the need for domain knowledge. 3.5
While the format required for reporting results is adequate for point-to-point matches across ontologies, it is not clear how other kinds of mappings could be reported. For example, although not reported here, our system can identify group-to-group matches [4] and so do other systems such as AnchorPrompt [13]. Along the same lines, one important feature of our system is to identify non-matches, i.e., concepts exhibiting resemblance at the lexical level, but semantic differences. Such anchors with negative structural evidence are presented in our result file with “!=” as value for “relation” and 1.0 as value for “measure”. However, an alternative representation for such nonmatches would be to use “=” as value for “relation” and a negative value (e.g., -1.0) for “measure”. In this case, the meaning of the value parameter would be somewhat equivalent to that of a correlation coefficient.
In the absence of a reference mapping, it is difficult to evaluate alignment systems on this data set. As a biomedical terminology integration resource, the Unified Medical Language System (UMLS) could provide some elements for evaluating the mappings. However, unfortunately, neither the FMA nor GALEN is part of the UMLS. Moreover, the framework for representing anatomical entities in SNOMED CT – one of the source vocabularies in the UMLS – leads to multiple mappings for many anatomical terms (ambiguity) [3]. In previous work, we analyzed the mapping results produced automatically by several systems [17]. Most mappings were identified by the two systems under investigation, which was reassuring. Analyzing the differences, i.e., mappings identified by one system but not by the other, provided interesting insights about the strengths and weaknesses of each system. Crossvalidation among the various systems in competition in the OAEI campaign, would not provide a definitive evaluation metric, but could help system developers analyze the mappings specific to their systems and those they missed. The idea of evaluating the mappings in the context of an application suggested by [20] is very interesting, but difficult to realize. Since ontology alignment is key to the Semantic Web, a collaboration with the W3C Semantic Web Health Care and Life Sciences Interest Group could provide the context for such an evaluation.
Aligning large anatomical ontologies has generated a lot of interest in the biomedical and computer science communities and successful methods will contribute to create the biomedical component of the Semantic Web. Despite promising advances, aligning anatomical ontologies remains challenging, because of the sheer size of these resources and the need for domain knowledge. Differences in modeling and representation formalism are also an issue, as well as the lack of a reference mapping for evaluation purposes. Competitive evaluation campaigns such as OAEI represent for ontology alignment the same kind of driving force BioCreAtIvE has been for information extraction in biology [23].
Acknowledgments. This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH), National Library of Medicine (NLM), and by the Natural Science Foundation of China (No.60496324), the National Key Research and Development Program of China (Grant No. 2002CB312004), the Knowledge Innovation Program of the Chinese Academy of Sciences, MADIS of the Chinese Academy of Sciences, and Key Laboratory of Multimedia and Intelligent Software at Beijing University of Technology.