Problems impacting the quality of automatically built ontologies Toader GHERASIM1 and Giuseppe BERIO2 and Mounira HARZALLAH3 and Pascale KUNTZ4 Abstract. Building ontologies and debugging them is a time- different researches recently focused on that issue [13, 2, 24]. How- consuming task. Over the recent years, several approaches and tools ever, as far as we know, a generic standardized description of these for the automatic construction of ontologies from textual resources errors does not still exist. It seems however a preliminary step for the have been proposed. But, due to the limitations highlighted by ex- development of assisted construction method. perimentations in real-life applications, different researches focused In this paper, we focus on the most important errors that affect on the identification and classification of the errors that affect the on- the quality of semi-automatically built ontologies. To get closer the tology quality. However, these classifications are incomplete and the operational concerns we propose a detailed typology of the different error description is not yet standardized. In this paper we introduce types of problems that can be identified when evaluating an ontology. a new framework providing standardized definitions which leads to Our typology is inspired from a generic standardized description of a new error classification that removes ambiguities of the previous the notion of quality in conceptual modeling [18]. And, our analysis ones. Then, we focus on the quality of automatically built ontologies is applied on a real-life situation concerning the manufacturing of and we present experimental results of our analysis on an ontology pieces in composite materials for the aerospace industry. automatically built by Text2Onto for the domain of composite mate- The rest of this paper is organized as follows. Section 2 is a state- rials manufacturing. of-the art of the ontology errors. Section 3 describes a framework which provides a standardized description of the errors and draws correspondences between our new classification and the main errors 1 Introduction previously identified in the literature. Section 4 presents our experi- Since the pioneering works of Gruber [15], ontologies play a ma- mental results in the domain of composite materials manufacturing. jor role in knowledge engineering whose importance is growing with More precisely, we analyze errors affecting an ontology produced by the rise of the semantic Web. Today they are an essential component an automatic construction tool (here Text2Onto) from a set of tech- in numerous applications in various fields: e.g. information retrieval nical textual resources. [22, 20], knowledge management [26], analysis of social semantic networks [8] and business intelligence [27]. However, despite the 2 State-of-the art on ontological errors maturity level reached in ontology engineering, important problems In the literature, the notion of ”ontological error” is often used in a remain open and are still widely discussed in the literature. The most broad sense covering a wide variety of problems which affect the on- challenging issues concern the automation of ontology construction tology quality. But, from several studies published this last decade, and their evaluation. we have identified four major denominations associated to comple- The increasing popularity of ontologies and the scaling changes of mentary definitions: (1) ”taxonomic errors” [14, 13, 9, 2], (2) ”design this last decade have motivated the development of ontology learn- anomalies” or ”deficiencies” [2, 3], (3) ”anti-patterns” [7, 25, 23], ing techniques. Promising results have been obtained [6, 5]. And, and (4) ”pitfalls” or ”worst practices [23, 24]. although these techniques have been often experimentally proved to be not sufficient enough for constructing ready-to-use ontology [5], their interest is not questioned in particular in technical domains [17]. 2.1 Taxonomic errors Few recent works recommend an integration between ontology learn- From the pioneering works of Gomez-Perrez [14], the denomination ing techniques and manual intervention [27]. ”taxonomic error” is used to refer to three types of errors that affect Whatever their use, it is essential to assess their quality through- the taxonomic structure of ontologies: inconsistency, incompleteness out their development. Several ontology quality criteria and dif- and redundancy. Recently, extensions have been proposed to non- ferent evaluation methods have been proposed in the literature taxonomic properties [3], but in this synthesis we focus on taxonomic [19, 4, 11, 21, 1]. However, as mentioned by [28], defining ”a good errors. ontology” remains a difficult problem and the different approaches Inconsistencies in the ontology may be logical or semantic. More only permit to ”recognize problematic parts of an ontology”. From precisely, three classes of inconsistencies in the taxonomic structure an operational point of view, error identification is a very important have been detailed: circularity errors (e.g. a concept that is a special- step for the ontology integration in real-life complex systems. And, ization or a generalization of itself), partitioning errors which pro- 1 LINA, UMR 6241 CNRS, e-mail: toader.gherasim@univ-nantes.fr duce logical inconsistencies (e.g. a concept defined as a specializa- 2 LABSTICC, UMR 6285 CNRS, email: giuseppe.berio@univ-ubs.fr tion of two disjoint concepts), and semantic errors (e.g. a taxonomic 3 LINA, UMR 6241 CNRS, e-mail: mounira.harzallah@univ-nantes.fr relationship between two concepts that is not consistent with the se- 4 LINA, UMR 6241 CNRS, e-mail: pascale.kuntz@polytech.univ-nantes.fr mantics of the latter). Incompleteness is met when concepts or relations of specialization ontology). One class corresponds to the functional dimension: ”re- are missing, or when some distributions of the instances of a concept quirement completeness” (RC, when the ontology does not cover its between its sons are not stated as exhaustive and/or disjoint. specifications). And, two classes correspond to the usability dimen- In the opposite way, redundancy errors are met when a taxonomic sion: ”ontology understanding” (OU, information that makes under- relationship can be directly deduced by logical inference from the standability more difficult e.g. concept label polysemy or label syn- other relationships of the ontology, or when concepts with the same onymy for distinct concepts, non explicit declaration of inverse rela- father in the taxonomy do not share any common information (no tions or equivalent properties) and ”ontology clarity” (OC, e.g. vari- instances, no children, no axioms, etc.) and can be only differentiated ations of writing-rule and typography for the labels). by their names. It is easy to deduce from this classification that some pitfalls should belong to different classes associated to different dimensions (e.g. the fact that two inverse relations are not stated as inverse is 2.2 Design anomalies both a ”no inference” (NI) pitfall and an ”ontology understanding” (OU) pitfall). Another attempt [24] proposed a classification of the Roughly speaking, design anomalies mainly focus on ontology un- 24 identified pitfalls in the three error classes (inconsistency, incom- derstanding and maintainability. They are not necessarily errors but pleteness and redundancy) given by Gomez-Perrez et al. [14]. But, undesirable situations. Five classes of design anomalies have been these classes are concerned by the ontology structure and content, described: (1) ”lazy concepts” (leaf concepts in the taxonomy not and consequently four pitfalls associated with the ontology context implied in any axiom and without any instances); (2) ”chains of in- do not fit with this classification. heritance” (long chains composed of intermediate concepts with a In order to highlight the links between the different classifications, single child); (3) ”lonely disjoint” concepts (superfluous disjunction Poveda et al. tried to define a mapping between the classification in 7 axiom between distant concepts in the taxonomy which may disrupt classes deduced from the dimensions defined by Gangemi et al. [11] inference reasoning); (4) ”over-specific property range” (too specific and the 3 error classes proposed by Gomez-Perrez et al. [14]. How- property range which should be replaced by a coarser range which ever, this task turned out to be very complex, and only four pitfall fits the considered domain better); (5) ”property clumps” (duplica- classes exactly fit with one of the error classes. For the other, there is tion of the same properties for a large set of concepts instead of the overlapping or no possible fitting. inheritance of these properties from a more general concept). 3 The framework 2.3 Anti-patterns The state of the art briefly presented in the previous section shows Ontology design patterns (ODP) are formal models of solutions com- that the terminology used for describing the different problems im- monly used by domain experts to solve recurrent modeling problems. pacting on the quality of ontologies is not yet standardized and that Anti-patterns are ODP that are a priori known to produce incon- existing classifications do not cover the whole diversity of problems sistencies or unsuitable behaviors. [23] also called anti-patterns ad- described in the literature. hoc solutions specifically designed for a problem even if well-known In this section we present a framework providing standardized def- ODP are available. Three classes of anti-patterns have been described initions for quality problems of ontologies and leading to a new clas- [7, 25, 23]: (1) ”logical anti-patterns” that can be detected by logi- sification of these problems. The framework comprises two distinct cal reasoning; (2) ”cognitive anti-patterns” (possible modeling errors and orthogonal dimensions: errors vs. unsuitable situations (first di- due to misunderstanding of the logical consequences of the used ex- mension) and logical facet vs. social facet of problems (second di- pression); (3) ”guidelines” (complex expressions valid from a logical mension). and a cognitive point of view but for which simpler or more accurate Unsuitable situations identify problems which do not prevent the alternatives exist). usage of an ontology (within specific targeted domain and applica- tions). On the contrary, errors identify problems preventing the usage of an ontology. 2.4 Pitfalls It is well known that one ontology has two distinct facets: an on- Pitfalls are complementary to ODPs. Their broad definition covers tology can be processed by machines (according to its logical speci- problems affecting the ontology quality for which ODPs are not fication) and can be used by humans (including an implicit reference available. Poveda et al. [24] described 24 types of experimentally to a social sharing). identified pitfalls as, for instance, forgetting the declaration of an in- The remainder of the section is organized alongside the second di- verse relation when this latter exists or of the attribute range. And mension (i.e. logic vs. social facet) and within each facet, errors and they proposed a pitfall classification which follows the three evalu- unsuitable situations are defined. The framework is based on ”nat- able dimensions of an ontology proposed by Gangemi et al. [11]: ural” analogies between respectively social and logical errors and (1) structural dimension (aspects related to syntax and logical prop- social and logical unsuitable situations. erties), (2) functional dimension (how well the ontology fits a pre- defined function), (3) the usability dimension (to which extent the 3.1 Problem classification ontology is easy to be understood and used). Four pitfall classes cor- 3.1.1 Logical ground problems respond to the structural dimension: ”modeling decisions” (MD, sit- uations where OWL primitives are not used properly), ”wrong infer- The logical ground problems can be formally defined by consider- ence” (WI, e.g. relationships or axioms that allow false reasoning), ing notions defined by Guarino et al. [16]: e.g. Interpretation (Ex- ”no inference” (NI, gaps in the ontology which do not allow infer- tensional first order structure), Intended Model, Language, Ontology ences required to produce new desirable knowledge), ”real world and the two usual relations , ` provided in any logical language. modeling” (RWM, when commonsense knowledge is missing in the The relation  is used to express both that one interpretation I is a model of a logical theory T , written as I  T (i.e. all the formulas reasoning according to the targeted ontology applications (for- in T are true in I, written for each formula ϕ ∈ T , I  ϕ), and mally, for some specific formula ϕ, true in the intended models also for expressing the logical consequence (i.e. that any model of a O  ϕ, cannot be derived O 0 ϕ within those suitable reasoning logical theory T is also a model of a formula, written as T  ϕ). The systems); relation ` is used to express the logical calculus i.e. the set of rules used to prove a theorem (i.e. any formula) ϕ starting from a theory The most common logical ground unsuitable situations are T , written as T ` ϕ. listed below. These situations impact negatively on the ”non func- Examples and formalizations hereinafter are provided by using a tional qualities” of ontologies such as reusability, maintainability, ef- typical Description Logics notation (but easily transformable in first ficiency as defined in the ISO 9126 standard for software quality. order or other logics). The usual logical ground errors are listed below. 6. Logical equivalence of distinct artifacts (concepts / relationships / instances) i.e. whenever two distinct artifacts are proved to be 1. Logical inconsistency corresponding to ontologies containing log- logically equivalent; for example, A and B are two concepts in O ical contradictions for which a model does not exist (because the and O  A = B; set of intended models is never empty, an ontology without mod- 7. Symmetrically, logically indistinguishable artifacts i.e. whenever els does not make sense anyway; formally, given an ontology O it is not possible to prove that two distinct artifacts are not equiv- and the logical consequence relation  according to the logical alent from a logical point of view; in other words, if not possi- language L used for building O, there is no interpretation I of ble to prove anyone of the following statements: (O  A = B), O such that I  O). For example, if an ontology contains the (O  A ∩ B ⊆⊥) and (O  c ⊆ AandO  c ⊆ B); this case following axioms B ⊆ A (B is a A), A ∩ B ⊆ > (A and B (7) can be partially covered in the case (3) above whenever in- are disjoint), c ⊆ B (c is instance of B), then c ⊆ A and tended models provide precise information on the equivalence or c ⊆ A ∩ B, so there is a logical contradiction in the definition of the difference between A and B; this ontology; 8. OR artifacts i.e. an artifact A equivalent to a disjunction like C∪S, 2. Unadapted5 ontologies wrt to intended models6 i.e. an ontology A 6= C, S but for which, if applicable, it does not exist at least a for which something that is false in all (some of) the intended common (non optional) role / property for C and S or because C models of L is true in the ontology; formally, there exists a for- and S have common instances; in the first case, a simple formal- mula ϕ such that for each (for some) intended model(s) of L, ϕ ization can be expressed by saying that it does not exist a (non is false and O  ϕ. For example, if we have in the ontology two optional) role R such that O  (C ∪ S) ⊆ ∃R.>; in the second concepts A and B that are declared as disjoint (O  A ∩ B ⊆⊥) case, an even simpler formalization is O  c ⊆ C and O  c ⊆ S, and in each intended model there exists an instance c that is com- being c one constant not part of O; the first case targets potentially mon between A and B (i.e. c ⊆ A ∩ B), then the ontology is heterogeneous artifacts such as Car ∪ P erson, with probably unadapted; no counterpart in the intended models, thus possibly leading to 3. Incomplete ontologies wrt to intended models i.e. an ontology for unadapted ontologies according to case (2) above; the second case which something that is true in all the intended models of L, is targets potential ambiguities as, for instance, one role (property) not necessarily true in all the models of O; formally, there exists R logically equivalent to a disjunction (R1 ∪ R2 ) being (R1 ∩ R2 ) a formula ϕ such that for each intended model of L, ϕ is true and satisfiable; O 2 ϕ. As an example, if in all the intended models C ∪ B = A, 9. AND artifacts i.e. one artifact A equivalent to a conjunction like and the ontology O defines B ⊆ A and C ⊆ A, it is not possible C ∩ S, A 6= C, S but for which, if applicable, it does not exist to prove that C ∪ B = A; at least a common (non optional) role / property for C and S; 4. Incorrect (or unsound) reasoning wrt the logical consequence i.e. this case is relevant to limit as much as possible some potentially when some specific conclusions are derived by using suitable rea- heterogeneous artifacts such as Car ∩P erson, possibly leading soning systems for targeted ontology applications even if these to artifact unsatisfiability; conclusions are not true in the intended models and must not be 10. While some case of unsatisfiability of ontology artifacts (concepts, derived by any reasoning according to the targeted ontology appli- roles, properties etc.) can be covered by (2) because intended mod- cations (formally, when a specific formula ϕ, false in the intended els may not contain void concepts, unsatisfiability tout-court is not models O 2 ϕ, can be derived O ` ϕ within any of those suitable necessarily an error but a situation which is not suitable for ontol- reasoning systems); ogy artifacts (i.e. given an ontology artifact A, O  A ⊆ ⊥); even 5. Incomplete reasoning wrt the logical consequence i.e. when some if in ontologies it might be possible to define what must not be true specific conclusions cannot be derived by using suitable reasoning (instead of what must be true), this practice is not encouraged; systems for targeted ontology applications even if these conclu- 11. High complexity of the reasoning task i.e. whenever something sions are true in intended models and must be derived by some is expressed in a way that complicates the reasoning, while there 5 We use the term ”unadapted” instead of ”incorrect” ontologies because it exist more simple ways to express the same thing; remains unclear if intended models are defined for building the ontology 12. Ontology not minimal i.e. whenever the ontology contains unnec- or may also be defined independently. However, if intended models are essary information: defined for building the ontology, the term ”incorrect” may be more appro- priate. • Unnecessary because it can be derived or built7 . An example of 6 Intended models should have been defined fully and independently as in the such unsuitable situation is the redundancy of taxonomic rela- case of models representing abstract structures or concepts such as num- tions such as whenever A ⊆ B, B ⊆ C, and A ⊆ C are all bers, processes, events, time and other ”upper concepts”, often defined ac- cording to their own properties. If intended models are not available, some ontology axioms, the last axiom can be derived from the first specific entailments can be defined as facts that should necessarily be true two ones; in the targeted domain (or for targeted applications); specific counterexam- ples can also be defined instead of building entire intended models. 7 Built means that the artifact can be defined by using other artifacts. • Unnecessary because it is not part of the intended models. For 11. Lack of adapted and certified versions of the ontology in various instance, a concept A being part of the ontology (language) but languages requires specific efforts by social actors for understand- not defined by intended models. ing and learning the ontology but also to use the ontology in spe- cific standard contexts (limited compliance); there are no natural analogies; 3.1.2 Social ground problems 12. Socially useless artifacts included in the ontology; a natural anal- Social ground problems are related to the perception (interpreta- ogy is with ontology not minimal. tion) and the targeted usage of ontologies by social actors (humans, applications based on social artifacts like WordNet, etc.). Percep- 3.2 Positioning state of the art relevant problem tion (interpretation) and usage may not be formalized at all. In some classes in to the proposed framework sense, a further distinction between social facet and logical facet is as the distinction between respectively tacit and explicit knowledge. The precise definitions of the proposed framework allow us to clas- There are four social ground errors: sify most of the ontology quality problems described in literature. Ta- ble 1 presents our classification of the different problems mentioned 1. Social contradiction i.e. the perception (interpretation) that the so- in Section 2. Some of the problems described in literature may cor- cial actor gives to the ontology or to the ontology artifacts is in respond to more than one class of problems from our framework, as contradiction with the ontology axioms and their consequences; a the definitions of these problems are often very large and sometimes natural analogy is with unadapted ontologies; ambiguous. 2. Perception of design errors i.e. the social actor perception ac- Table 1 reveals, at a first view, that the proposed framework pro- counts for some design errors such as modeling instances as con- vides additional problems that are not directly pointed out, to our cepts; a natural analogy is with unadapted ontologies; knowledge, in the current literature about ontology quality and eval- 3. Socially meaningless i.e. the social actor is unable to give any in- uation (but may be mentioned elsewhere). These problems are No terpretation to the ontology or to ontology artifacts as in the case adapted and certified ontology version, Indistinguishable artifacts, of artificial labels such as ”XYHG45”; a natural analogy is with Socially meaningless, High complexity of the reasoning task and In- unadapted ontologies; correct reasoning. However, while covered, other problems are, in 4. Social incompleteness i.e. the social actor perception is that one or our opinion, too much narrowly defined in existing literature about several artifacts (axioms and/or their consequences) are missing in ontology quality and evaluation. For instance, No standard formal- the ontology; a natural analogy is with incomplete ontologies; ization is specific to very simple situations while we refer to com- plete non standard theories. The social ground unsuitable situations are mostly related to the A deeper analysis of Table 1 reveals that the ”logical anti-patterns” difficulties that a social actor has to overcome for using the ontology presented in [7, 25] belong to the logical ground category and are especially due to limited understandability, learnability and compli- focusing on unadapted ontologies error and unsatisfability unsuit- ance (as defined in ISO 9126). As for the logical ground unsuitable able situation. The ”non-logical anti patterns” presented in [7, 25] situations, it is difficult to dress an exhaustive list; the most common partially cover the logical ground unsuitable situations. The ”guide- and important are listed below. lines” presented in [7, 25] span only over unsuitable situations from both logical and social ground category. 5. Lack of or poor textual explanations i.e. when there are few, no or What is qualified as ”inconsistency” in [14] span over errors and poor annotations; prevents understanding by social actors; there unsuitable situations and also (as in the case of ”semantic incon- are no natural analogies; sistency”) over the two dimensions (logical and social), making, in 6. Potentially equivalent artifacts i.e. the social actors may identify our opinion, the terminology a little bit confusing. According to our as equivalent (similar) distinct artifacts as in the case of artifacts framework, we perceive ”circularity in taxonomies”, as defined in with synonymous or exactly the same labels assigned to distinct [14], as an unsuitable situation (logical equivalence of distinct arti- artifacts; a natural analogy is with logically equivalent artifacts; facts) because, from a logical point of veiw, this only means that ar- 7. Socially indistinguishable artifacts i.e. the social actors would not tifacts are equivalent (not requiring a fixpoint semantics). However, be able to distinguish two distinct artifacts as, for instance, in the ”circularity in taxonomies” can be seen also within a social contra- case of artifacts with polysemic labels assigned to distinct arti- diction if actors assign distinct meanings to the various involved ar- facts; a natural analogy is with logically indistinguishable arti- tifacts. The problems presented as ”incompleteness errors” in [13] facts; belong to the incomplete ontologies class of logical errors. The ”re- 8. Artifacts with polysemic labels may be interpreted as union or in- dundancy errors” fits, in our classification, within the ontology not tersection of their several rather distinct meanings associated to minimal class of logical unsuitable situations. labels; a natural analogy is therefore with OR and AND artifacts. None of the ”design anomalies” presented in [2] is perceived as a 9. Flatness of the ontology (or non modularity), i.e. ontology pre- logical error. Two of them correspond to a logical unsuitable situation sented as a set of artifacts without any additional structure, espe- (logically undistinguishable artifacts), one to a social error (percep- cially if coupled with a important number of artifacts; a natural tion of design errors) and the last one to a social unsuitable situation analogy is with high complexity of the reasoning task but also pre- (no standard formalization). venting effective learning and understanding by social actors; Concerning ”pitfalls” [24], the most remarkable fact concerns 10. Non-standard formalization of the ontology, using a very specific what we call incomplete reasoning. Indeed, introducing ad-hoc re- logics or theory, requires a specific effort by social actors for un- lations such as is a, instance of , etc., replacing the ”standard” re- derstanding and learning the ontology but also to use the ontology lations such as subsumption, member of , etc., should not be con- in standard contexts (reduced compliance); there are no natural sidered as a case of incomplete ontologies but as a case of incomplete analogies; reasoning. This is because accepting a specific ontological commit- Table 1. Positioning state of the art relevant problem classes in to the proposed framework. Framework State of the art problems 1 Logical inconsistency â inconsistency error: ”partition errors - common instances in disjoint decomposition” â inconsistency errors: ”partition errors - common classes in disjoint decomposition”, ”semantic inconsistency” â logical anti-patterns: ”OnlynessIsLoneliness”, ”UniversalExistence”, ”AndIsOR”, ”EquivalenceIsDifference” 2 Unadapted ontologies â pitfalls: P5 (wrong inverse relationship, WI), P14 (misusing ”allValuesFrom”, MD), P15 (misusing ”not some”/”some not”, WI), P18 (specifying too much the domain / range, WI), P19 (swapping ∩ and ∪, WI) Errors â incompleteness errors: ”incomplete concept classification”, ”disjoint / exhaustive knowledge omission” â pitfalls: P3 (”is a” instead of ”subclass-of”, MD), P9 (missing basic information, RC & RWM), P10 (missing 3 Incomplete ontologies disjointness, RWM), P11 (missing domain / range in prop., NI & OU), P12 (missing equiv. prop., NI & OU), P13 (missing inv. rel., NI & OU), P16 (misusing primitive and defined classes, NI) 4 Incorrect reasoning 5 Incomplete reasoning â pitfalls: P3 (using ”is a” instead of ”subclass-of”, MD), P24 - using recursive def., MD) â inconsistency error: ”circularity” 6 Logical equivalence of dis- Logical ground â pitfall: P6 (cycles in the hierarchy, WI) tinct artifacts â non logical anti-pattern: ”SynonymeOfEquivalence” 7 Logically indistinguishable â pitfall: P4 (unconnected ontology elements, RC) artifacts â design anomalies: ”lazy concepts” and ”chains of inheritance” Unsuitable situations 8 OR artifacts â pitfall: P7 (merging concepts to form a class, MD & OU) 9 AND artifacts â pitfall: P7 (merging concepts to form a class, MD & OU) inconsistency error: ”partition errors - common classes in disjoint decomposition” 10 Unsatisfiability â logical anti-patterns: ”OnlynessIsLoneliness”, ”UniversalExistence”, ”AndIsOR”, ”EquivalenceIsDifference” 11 High complexity of the rea- soning task â redundancy error: ”redundancy of taxonomic relations” â pitfalls: P3 (using ”is a” instead of ”subclass-of”, MD), P7 (merging concepts to form a class, MD & OU), P21 12 Ontology not minimal (miscellaneous class, MD) â non logical anti-pattern: ”SomeMeansAtLeastOne” â guidelines: ”Domain&CardinalityConstraints”, ”MinIsZero” â inconsistency error: ”semantic inconsistency” â logical anti-pattern: ”AndIsOR” 1 Social contradiction â pitfalls: P1 (polysemic elements, MD), P5 (wrong inv. rel., WI), P14 (misusing ”allValuesFrom”, MD), P15 (misusing ”not some”/”some not”, WI), P19 (swapping ∩ and ∪, WI) Errors â pitfalls: P17 (specializing too much the hierarchy, MD), P18 (specifying too much the domain / range, WI), P23 (using incorrectly ontology elements, MD) 2 Perception of design errors â non logical anti-pattern: ”SumOfSome” â design anomaly: ”lonely disjoints” 3 Socially meaningless Social ground â pitfalls: P12 (missing equiv. prop., NI & OU), P13 (missing inv. rel., NI & OU), P16 (misusing primitive and 4 Social incompleteness defined classes, NI) 5 Lack/poor textual explana- â pitfalls: P8 (missing annotation, OC & OU) tions 6 Potentially equiv. artifacts â pitfalls: P2 (synonym as classes, MD & OU) Unsuitable situations 7 Indistinguishable artifacts 8 Polysemic labels â pitfalls: P1 (polysemic elements, MD & OU) 9 Flatness of the ontology â pitfalls: P20 (swapping label and comment, OU), P22 (using different naming criteria in the ontology, OC) 10 No standard formalization â guidelines: ”GroupAxioms”, ”DisjointnessOfComplement” and ”Domain&CardinalityConstraints” â design anomaly: ”property clumps” 11 No adapted and certified ontology version 12 Useless artifacts â pitfall: P21 (using a miscellaneous class, MD & OU) ment for building intended models, ad-hoc relations can be defined Table 2. What problems are expected in automatically built ontologies. in the same way as standard relations. However, using standard rea- soning it is expected (and even proved once fixing the logics) that Types of problems Expected (Yes/No) and Why reasoning algorithms are incomplete. However, adding artifacts may N (no axiom is defined ⇒ contradictions are also solve some incompleteness and may also be useful for speeding 1. Logical inconsistency unexpected; but they remain possible in the up reasoning. case of future enrichments) Only one of the seven classes of ”pitfalls” [24] perfectly fits in one Y (taxonomic relationships extraction algo- class of our typology: the ”real world modeling” pitfalls belong to the 2. Unadapted ontologies rithms are syntax based 6= from the intended incomplete ontologies logical errors. All the ”ontology clarity” pit- models) falls are social unsuitable situations. All the ”requirement complete- Y (automatically extracted knowledge is ness” pitfalls are logical problems. The ”no inference” pitfalls are 3. Incomplete ontologies limited to concepts and taxonomies 6= from logical or social incomplete ontologies errors. Most (6/9 and 4/5) of the intended models) the ”modeling decisions” and ”wrong inference” pitfalls are consid- 4. Incorrect reasoning N (they might appear for complete formal- ered as errors. The class of ”ontology understanding” pitfalls spans 5. Incomplete reasoning ization of concepts and relationships) over 10 classes of problems, covering logical and social errors and 6. Logical equivalence of Y (automatic tools consider that each term unsuitable situations. distinct artifacts defines a different artifact ⇒ the ontology Most (16/20) of the pitfalls concerning the ”structural dimension” 7. Logically indistin- may contain logically equivalent & logically of the ontology [11] are perceived as errors. All (2/2) the pitfalls guishable artifacts indistinguishable artifacts) concerning the ”functional dimension” of the ontology are logical 8. OR artifacts Y (polysemy of terms directly affects con- problems. cepts / relationships: OR / AND concepts / 9. AND artifacts relationships may appear) 4 Problems that affect the quality of automatically Y (polysemy of terms directly affects con- built ontologies cepts / relationships: these latter may be- 10. Unsatisfiability come unsatisfiable if their polysemic senses Although the proposed framework is general, we are especially con- are combined) cerned by ontologies automatically built from textual resources. We N (few or no axioms are defined ⇒ reason- 11. High complexity of ing remains very basic; but, it can be more therefore aim at pointing the problems that are expected in automat- the reasoning task ically constructed ontologies (i.e. there is evidence of their presence complex if the ontology is further enriched) or they will appear in future enrichments8 of the ontology). We are 12. Ontology not mini- Y (automatic tools introduce redundancies in also interested by the opposite case, i.e. if there are unexpected prob- mal taxonomies) lems in automatically constructed ontologies: it should be noted that unexpected problems are problems that even if the ontology may suf- Y (ontologies are built from limited textual 1. Social contradiction resources which may introduce contradiction fer of them, there is no evidence of their presence/absence for the in taxonomies) ontology as it is (however, these problems may appear in future en- Y (the built ontology may contain concepts richments of the ontology). Our analysis is performed in two steps. In 2. Perception of design that are considered more close to instances the first step (Section 4.1), we point out expected/unexpected prob- errors by the social actor.) lems due to inherent limitations of the tools for automatic ontology Y (several meaningless concepts with ob- construction. In the second step (Section 4.2), we assess the results 3. Social meaningless scure labels are often introduced) obtained in the first step by discussing our experience with the tool 4. Social incompleteness Y (probably due to limited textual corpus) Text2Onto. 5. Lack of or poor textual Y (usually automatic tools do not provide explanations textual explanations) 4.1 Expected and unexpected problems in an Y (automatic tools consider that each term automatically built ontology 6. Potentially equivalent defines a different artifact ⇒ distinct con- artifacts cepts can have synonymous labels ⇒ these In a previous work [12] we have deeply studied four approaches (and latter are perceived as potentially equivalent) associated tools) for the automatic construction of ontologies form Y (the ontology is incomplete ⇒ it contains texts and we compared them with a classical methodology for manual 7. Indistinguishable arti- concepts that can be distinguished only by ontology construction (Methontology). This analysis highlighted that facts their labels; if such concepts have synony- mous labels, they are indistinguishable) none of the automated approaches (and associated tools) covers all the tasks and subtasks associated to each step of the classical manual Y (automatic tools consider that each term 8. Artifacts with poly- defines a different artifact ⇒ it is possible to method. The ignored tasks/subtasks are: semic labels have concepts with polysemic labels) 1. The explicit formation of artifacts (concepts, instances and rela- Y (the ontology is poorly structured and has 9. Flatness of the ontol- no design constraints - e.g. no disjunction ax- tionships) from terms9 ; usually, the automatic tools consider that ogy iom, lazy concepts) each term represents a distinct artifact: they do not group synony- mous terms and do not choose a single sense for polysemic terms 10. No standard formal- N (automatic tools usually can export their ization results in different formalization) 2. The identification of axioms (e.g. the disjunction axioms) 3. The identification of attributes for concepts Y (automatically obtained results closely de- 11. No adapted and cer- pend on the input texts language (often En- 4. The identification of natural language definitions for concepts tified ontology version glish) and certifying them is difficult) 8 Enrichment should be understood as adding artifacts to the existing ones. Y (automatic tools often generate useless ar- 12. Useless artifacts tifacts from additional external resources) 9 A term corresponds to one or several words found in one text. Table 2 provides a complete view of expected and unexpected problems according to our experience and suggest why each prob- lem is expected or not. 4.2 Experience with Text2Onto Table 3. Types of problems identified in the automatically constructed ontology. 4.2.1 The experimental setup Types of problems Identyfied (Yes/No) and How 1. Logical inconsistency No During the last two years we were implied in a project called ISTA3 that proposed an ontology based solution for problems related to the 2. Unadapted ontologies No integration of heterogeneous sources of information. The application Yes: Some relationships are missing to con- domain was the management of the production of composite compo- 3. Incomplete ontologies nect the 389 lazy concepts; some of them are explicitly indicated in the textual corpus nents for the aerospace industry. In this context, we tried to simplify the process of deploying the interoperability solution in new domains 4. Incorrect reasoning No by using automatic solution for constructing the required ontologies. 5. Incomplete reasoning No The analysis presented in [12] conducted us to choose Text2Onto 6. Logical equivalence of Yes: 3 cycles in the hierarchy; (automatically [6] for the automatic construction of our ontologies. Text2Onto takes distinct artifacts detected by reasoners) as input textual resources from which it extracts different ontologi- Yes: cal artifacts (concepts, instances, taxonomic relationships, etc.) that * 389 lazy concepts (automatically identified by an ad-hoc algorithm) are structured together to construct an ontology. Text2Onto perfor- 7. Logically indistin- *73 groups of ”leaf” concepts; each group is mances for extracting concepts and taxonomical relationships are guishable artifacts composed of concepts that are indistinguish- better than its performances for extracting other types of ontologi- able; (automatically identified by an ad-hoc cal artifacts; consequently, in our tests we used Text2Onto for con- algorithm) structing ontologies containing concepts and taxonomical relation- 8. OR artifacts No ships only. 9. AND artifacts No The textual resource used in the experiment presented in this paper 10. Unsatisfiability No is a technical glossary composed of 376 definitions of the most im- 11. High complexity of portant terms of the domain of composite materials and how are they the reasoning task No used for manufacturing pieces. The glossary contains 9500 words. Yes: one taxonomical relationship can be de- For constructing the ontology we resort to the standard configura- 12. Ontology not mini- duced from two taxonomical relationships tion for the different parameters of Text2Onto: all the proposed al- mal already present in the ontology (automati- gorithms for concepts (and respectively for taxonomic relations) ex- cally identified by an ad-hoc algorithm) tractions have been used and their results have been combined with the default strategy. Yes: 15 taxonomic relationships are jugged 1. Social contradiction semantically inconsistent by the expert The constructed ontology is an automatically built domain ontol- ogy that contains 965 concepts and 408 taxonomic relationships. Yes: 5 concepts that are interpreted as in- 2. Perception of design stances by the expert (units of measure and Some of the central concepts of this ontology are: ”technique”, errors proper names) ”step”, ”compound”, ”fiber”, ”resin”, ”polymerization”, ”laminate”, ”substance”, ”form”. Yes: 21 concepts that have meaningless la- 3. Social meaningless bels, for the expert 4. Social incompleteness Yes 4.2.2 Identified problems 5. Lack of or poor textual Yes: no annotation associated to the ontology explanations or to its artifacts Table 3 summarizes which types of problems have been identified 6. Potentially equivalent Yes: 6 pairs of concepts have synonym la- in the automatically constructed ontology in our experience with artifacts bels, for the expert Text2Onto. It also indicates, when possible, how many problems 7. Indistinguishable arti- have been identified. Most of problems are relatively easy to iden- facts No tify and to quantify (e.g. the number of cycles in the taxonomical 8. Artifacts with poly- Yes: 69 concepts with polysemic labels, for structure), but there are exceptions (e.g. the number of concepts or semic labels the expert taxonomic relationships that are missing from the ontology). 9. Flatness of the ontol- Yes: 389 lazy concepts lead to a poorly struc- ogy tured ontology 4.2.3 Discussion 10. No standard formal- ization No No intended model or use case scenario was available when the ex- 11. No adapted and cer- pert analyzed the automatically constructed ontology. Consequently, tified ontology version No it was able only to make a supposition concerning the logical com- Yes: 28 concepts are not necessary (3 are too 12. Useless artifacts generic, 25 are out of the domain) pleteness of the ontology and no logical error (unadapted ontology, incomplete or incorrect reasoning) was identified. Few logical unsuitable situations are identified, but it is remarkable that they were identified automatically. Unsurprisingly, most of the identified problems are social prob- lems. The analysis in Section 4.1 suggest that most of the problems that [9] M. Fahad and M. Qadir, ‘A framework for ontology evaluation’, in are expected in the automatically constructed ontologies are due to Proc. of the 16th Int. Conf. on Conceptual Struct. (ICCS2008), volume 354, pp. 149–158, (2008). the fact that the automatic tool do not take into account the synonymy [10] M. Fernandez, A. Gomez-Prez, and N. Juristo, ‘Methontology: From and the polysemy of terms when constructing concepts. However, ontological art towards ontological engineering’, in Proc. of the AAAI97 even if Text2Onto, as configured for our test, do not group synonym Spring Symposium Series on Ontological Engineering, pp. 33–40, terms when forming concepts, and allows polysemic terms to be la- (1997). bels for concepts, our test-case reveals that only two types of prob- [11] A. Gangemi, C. Catenacci, M. Ciaramita, and J. Lehmann, ‘Modelling ontology evaluation and validation’, in Proc. of Eur. Sem. Web Conf. lems (socially indistinguishable artifacts and artifacts with polysemic (ESWC2006), number 4011 in LNCS, (2006). labels) may be imputed to this limitation. [12] T. Gherasim, M. Harzallah, G. Berio, and P. Kuntz, ‘Analyse com- Most of the identified problems are related to the fact that the au- parative de methodologies et d’outils de construction automatique tomatically constructed ontology seems to be incomplete. d’ontologies a partir de ressources textuelles’, in Proc. of EGC’2011, pp. 377–388, (2011). [13] A. Gomez-Perez, ‘Ontology evaluation’, in Handbook on Ontologies, eds., S. Staab and R. Studer, Int. Handbook on Inf. Syst., pp. 251–274, 5 Conclusion Springer, 1 edn., (2004). [14] A. Gomez-Perez, M.F. Lopez, and O.C. Garcia, Ontological Engineer- In this paper, we have introduced a framework providing standard- ing: With Examples from the Areas of Knowledge Management, E- ized definitions for different errors that have some impact on the Commerce and the Semantic Web, chapter Chap 3.8.2 Taxonomy eval- quality of the ontologies. This framework aims at both unifying var- uation, 180–184, Advanced Information and Knowledge Processing, ious error descriptions presented in the recent literature and com- Springer, 2001. [15] T. R. Gruber, ‘A translation approach to portable ontology specifica- pleting them. It also leads to a new error classification that removes tions’, Knowl. Acquisition, 5(2), 199–220, (1993). ambiguities of the previous ones. During ontology evaluation this [16] N. Guarino, D. Oberle, and S. Staab, ‘What is an ontology?’, in Hand- framework may be used as a support for verifying in a systematic book on Ontologies, 1–17, Springer, 2 edn., (2009). way if the ontology contains errors or unsuitable situations. [17] G. Hirst, ‘Ontology and the lexicon’, in Handbook on Ontologies, eds., R. Studer and S. Staab, Int. Handbook on Inf. Syst., 269–292, Springer, In the second part of the paper we focused on the quality of au- 2 edn., (2009). tomatically built ontologies and we present experimental results of [18] J. Krogstie, O.I. Lindland, and G. Sindre, ‘Defining quality aspects our analysis on an ontology automatically built by Text2Onto. The for conceptual models’, in Proc. of the IFIP8.1 Working Conference results show that a large part of the identified errors are linked to on Information Systems Concepts: Towards a Consolidation of Views the ontology incompleteness. Moreover, it confirms that the identi- (ISCO3), (1995). [19] A. Lozano-Tello and A. Gomez-Perez, ‘Ontometric: A method to fication of logical errors other than inconsistency requires intended choose the appropriate ontology’, Journal of Database Management, models (or at least a set of positive and negative examples) and use 15(2), 1–18, (2004). case scenarii. [20] N. Ben Mustapha, H. Baazaoui Zghal, M.A. Aufaure, and H. Ben Due to the increasing complexity of the software, the identification Ghezala, ‘Enhancing semantic search using case-based modular ontol- ogy’, in Proc. of the 2010 ACM Symposium on Applied Computing, pp. of the origin of each error in the ontology building process remains an 1438–1439, (2010). open question. And a further works consists in associating the iden- [21] L. Obrst, B. Ashpole, W. Ceusters, I. Mani, S. Ray, and B. Smith, ‘The tified errors with the different tasks of an ontology construction (e.g. evaluation of ontologies: toward improved semantic interoperability’, the Methontology tasks [10]). This work could help to improve the in SemanticWeb: Revolutionizing Knowledge Discovery in the Life Sci- quality results of the software by a retro-engineering process and/or ences, ed., K.-H. Cheung C. J. O. Baker, 139–158, Springer, (2007). [22] J.D. Osborne, J. Flatow, M. Holko, S.M. Lin, W.A. Kibbe, L. Zhu, M.I. to design assistant to detect and to solve major errors. Danila, G. Feng, and R. L. Chisholm, ‘Annotating the human genome with disease ontology’, BMC Genomics, 10, 63–68, (2009). [23] M. Poveda, M. C. Suarez-Figueroa, and A. Gomez-Perez, ‘Common REFERENCES pitfalls in ontology development’, in Proc. of the Current topics in ar- tificial intelligence (CAEPIA09), and 13th conference on Spanish asso- [1] M. Almeida, ‘A proposal to evaluate ontology content’, Journal of Ap- ciation for artificial intelligence, (2009). plied Ontology, 4(3-4), 245–265, (2009). [24] M. Poveda, M. C. Suarez-Figueroa, and A. Gomez-Perez, ‘A double [2] J. Baumeister and D. Seipel, ‘Smelly owls design anomalies in on- classification of common pitfalls in ontologies’, in Proc. of Workshop tologies’, in Proc. of 18th Int. Florida Artificial Intelligence Research on Ontology Quality (OntoQual 2010), Co-located with EKAW 2010, Society Conf. (FLAIRS), pp. 215–220, (2005). (2010). [3] J. Baumeister and D. Seipel, ‘Anomalies in ontologies with rules’, Web [25] C. Roussey, O. Corcho, and L. M. V. Blzquez, ‘A catalogue of owl Semantics: Science, Services and Agents on the World Wide Web, 8(1), ontology antipatterns’, in Proc. of the Fifth Int. Conf. on Know. Capture 55–68, (2010). KCAP, pp. 205–206, (2009). [4] A. Burton-Jones, V. Storey, and V. Sugumaran, ‘A semiotic metrics [26] N. H. Shah and M. A. Musen, ‘Ontologies for formal representation suite for assessing the quality of ontologies’, Data Knowl. Eng., 55(1), of biological systems’, in Handbook on Ontologies, eds., R. Studer and 84–102, (2005). S. Staab, Int. Handbook on Inf. Syst., 445–462, Springer, 2 edn., (2009). [5] P. Cimiano, A. Madche, S. Staab, and J. Volker, ‘Ontology learning’, in [27] E. Simperl and Tempich C., ‘Exploring the economical aspects of on- Handbook on Ontologies, eds., R. Studer and S. Staab, Int. Handbook tology engineering’, in Handbook on Ontologies, eds., R. Studer and on Inf. Syst., 245–267, Springer, 2 edn., (2009). S. Staab, Int. Handbook on Inf. Syst., 445–462, Springer, 2 edn., (2009). [6] P. Cimiano and J. Volker, ‘Text2onto - a framework for ontology learn- [28] D. Vrandecic, ‘Ontology evaluation’, in Handbook on Ontologies, eds., ing and data-driven change discovery’, in 2nd Eur. Semantic Web Conf., R. Studer and S. Staab, Int. Handbook on Inf. Syst., 293–314, Springer, eds., A. Montoyo, R. Munoz, and E. Metais, volume 3513, pp. 227–238, 2 edn., (2009). (2005). [7] O. Corcho, C. Roussey, and L. M. V. Blazquez, ‘Catalogue of anti- patterns for formal ontology debugging’, in Atelier Construction d’ontologies: vers un guide des bonnes pratiques, AFIA 2009, (2009). [8] G. Ereteo, M. Buffa, O. Corby, and F. Gandon, ‘Semantic social net- work analysis: A concrete case’, in Handbook of Research on Methods and Techniques for Studying Virtual Communities: Paradigms and Phe- nomena, 122–156, IGI Global, (2010).