Towards ontology composition from cognitive libraries Stefano Borgo Laboratory for Applied Ontology ISTC CNR, Trento, Italy stefano.borgo@cnr.it Abstract • any change in the formalisation of a notion can lead to logical problems (Ghilardi, Lutz, and Wolter 2006) or can Sometimes knowledge engineers have to come up with force unexpected changes in other notions. new ontologies for some specific domain or applica- tion even though there are already ontological works in These problems arise because ontology artefacts tend to that area. Although there are techniques for ontology be large logical theories about general as well as specific de/composition, a common problem is that the existing notions1 and this makes it impossible to modify the system systems often characterise some notion in a way that relying on just intuition. Logical tools help to verify the con- is not quite what the knowledge engineer needs. The sistency of a modified system and to build specific models, change of a few notions in a given ontology can be chal- especially when we limit ourselves to decidable languages, lenging: it is not easy to understand the impact of these but do not help to implement changes in a system that are changes. aimed to preserve the modeller’s intuitions, nor to model in- In this paper we investigate another route. Assuming tertwined notions like those of action and agent. Further- that the knowledge engineer has to deal with notions at the mesoscopic level that are cognitively clear, we pro- more, to be fully exploited ontology artefacts must cover pose to independently characterised them (in a sense to (most of) the domain of application leading to build large be discussed), and look at how they can be used to build systems that have to do justice of possible competing per- an ontology such that it comprises only the needed no- spectives and ad hoc distinctions. tions and with the right meaning. Here we just explore some important steps of this approach, and list other This paper investigates the construction of ontologies problems that need to be faced. from cognitively motivated modules each focussing on a sin- gle notion. In particular, we rely on intuition, understood as a mix of cognition, domain expertise and common sense, to Introduction identify the needed notions and to select their characterisa- tion. Assuming a library of such modules is available, we Ontology artefacts (Guarino 1998; Guarino, Oberle, and show how a Knowledge Engineer (KEer) can build her/his Staab 2009) are complex formal systems used to ensure own ontology to meet intuition and needs. The paper is ex- readability and transparency of information models and to ploratory in nature and aims to only sketch the approach. enhance interoperability across information systems. How- This is done by looking at how the three notions of agent, ever, the development and tuning of ontologies, no matter action and artefact are understood in literature; how these whether focussed on foundational, reference or domain con- should be formalised to form a library of modules; and how cepts, is still largely left to the personal skills and prefer- these modules can be merged into a single ontology. It also ences of the ontologist (developing an ontology is compara- discusses several problematic issues that need to be resolved ble to a craft) and is one of the major bottlenecks towards to turn this approach into a methodology for ontology con- their wide application. struction. Among the problems reported by ontology users three are Structure of the paper. After a section with preliminary particularly interesting for this paper: observations, we propose to divide notions in two (contextu- • the construction of ontology artefacts is time-consuming alised) types which form the basis of our approach. Section even when trying to reuse existing systems (Fernández- “Cognitive categories” exemplifies key steps of the approach López, Gómez-Pérez, and Suárez-Figueroa 2013; Simperl by discussing the notions of agent, action and artefact. The and Tempich 2006); next section explains how to understand the definitions just introduced and the section after it, “The role of underspec- • the proper characterisation of the notions to include is ified definitions”, clarifies how they are used and why that complex (Guarino 1999) and 1 In this paper we use the generic term ‘notion’ as a rough syn- onym of concept or category. A notion in this sense is formally modeled as a class. form of underspecification is important. Next we explain guishing product and device, and moving the latter category how to deal with the other class of categories (called “struc- into a role taxonomy); or by adding/removing single cate- ture categories”) by relaying on a top-level framework. The gories or articulated branches. In alternative, the KEer can final section highlights the remaining issues and some criti- use ontology modularity and matching approaches roughly cal aspects. consisting in isolating relevant parts of possibly different on- tologies and merging them into a new system, see (Euzenat Preliminary Observations and Shvaiko 2007). For whatever reason one may use modu- There can be a variety of motivations to build an ontology larity (Borgo 2011), the reuse of existing ontological work is and these determine the key categories one needs to include non trivial and, when specific constraints must be satisfied, in the system. When one has practical reasons like a specific it can be demanding and error-prone. application, as opposed to a research plan for systematising a topic or for a foundational ontology inspired by a partic- Another possibility is presented in the following sections ular worldview, he/she usually starts by listing a set of key where we concentrate on the modelling of predicates that terms, their intended meaning and related requirements. An identify classes (see our use of the term ‘notion’ in foot- ontology construction methodology aims to ensure, among note 1) but the method, mutatis mutandis, should be appli- other things, that the categories for these entities are cor- cable to properties and relations as well. rectly modelled and find a natural position in the ontology Starting from the application/domain-driven mandatory framework. Further categories are introduced for different notions, at first the KEer defines (both in natural language reasons, e.g., (a) some categories are added for complete- and formally) each notion she needs to cover. This is done ness (in an ontology for transportation one could include assuming that the other notions used in any of these defini- unicycles although they are not generally considered trans- tions, are already correctly modelled (see the next Section portation means) and (b) some categories are added to struc- for examples). This initial work leads to isolate a list of for- ture the system’s taxonomy (in the previous example, the mally characterised categories and a list of uncharacterised category of physical objects and the subcategory of phys- categories. The latter set collects the notions that were used ical devices with the latter including that of transportation in some formal definition but were not themselves defined. devices). In other words, the KEer associates a term in the formal lan- In practice, in many cases the categories one starts with guage to each notion she considers but formally axiomatises are comparable in the sense that they carve up reality at in the language only some of these terms. There are different some homogeneous level within the so called mesoscopic ways to arrive at these formal definitions but typically one view, i.e., they characterise entities human beings typically follows the understanding of the notion in the given domain perceive, manipulate and find relevant in their everyday (from which the request of a description in natural language) life. Topics related to everyday life (services to the person, and, with the help of competency questions (Grüninger and production of mechanical and mechatronic devices, organ- Fox 1995), fixes some constraints to capture that informal isational and institutional structure, logistics, transportation meaning. Some accompanying axioms may be added espe- etc.) are almost exclusively at these perceptual/cognitive lev- cially to bind the arguments of the occurring relations. When els: people, animals, buildings, public places, appliances, so- this step is completed, the KEer has produced a fairly com- cial roles etc. plete list of categories deemed relevant to occur in the ontol- ogy at stake. Among these categories, some are associated Since most of the ontologies that are built today naturally with a formal (and typically partial) axiomatisation, these deal with categories at these perceptual/cognitive levels, it categories are called domain categories, the other are so far makes sense to leverage on this regularity to improve exist- uncharacterised categories. ing techniques for ontology construction. This paper makes Recall that we target an ontology focusing on notions the first steps to turn this observation into a suitable (and at some homogeneous mesoscopic/cognitive level. General possibly promising) methodology. categories like entity, object, happening and role, should not occur among the domain categories. It is however very well Approach Sketched possible that they occur in the list of uncharacterised cate- Let us assume a KEer wants to construct an ontology for gories. The basic idea is that since these more general no- an application or for a specific company. She starts by em- tions are hard to characterise correctly, it is better to borrow bracing that particular viewpoint of a fragment of reality and them from some well studied and characterised top-level on- then focuses on the key notions as understood in that appli- tology. In the language just obtained we divided the cate- cation. For the sake of the presentation, let us assume that gories in two groups: domain and uncharacterised. We split all available ontologies she could reuse misrepresent one or the latter group in two classes: the structural categories and more of these key notions or make unsuitable choices rela- the subsidiary categories. By construction, the domain cate- tively to some constraints. Since to build a robust and reli- gories correspond to the key notions in which the KEer is in- able ontology is complicated and time consuming, the KEer terested. The structural categories (typically categories like may try to modify one of these existing ontologies (Simperl entity, object, event, space, time etc.) are categories that or- 2009) by adding/deleting properties and rewriting relations; ganise the high-level structure of the ontology. They are not by tangling (or untangling) notions, e.g., remodelling an ob- the focus of the ontology to be built but are necessary to ject category into a role category or vice versa (like distin- obtain an ontological system. The subsidiary categories are the remaining categories: the categories that are neither at How to be Agent the centre of the KEer’s concerns, nor serve to structure the The following definitions of (intelligent) agent are fairly ontology and still come up naturally when describing the simple and accepted in the literature. We formalise them domain categories. We will see some example in the next just minimally and without considering further specialisa- section. tions (e.g., introducing on a pair physical agents, software The distinction between structural, subsidiary and do- agents and so on). Our level of characterisation agrees with main categories is clearly fuzzy and contextual. Nonethe- what is often done in today’s ontologies and has the advan- less, when an application scenario is fixed, the distinction tage to be simple to understand. Simplicity is crucial since becomes intuitively clearer and even inspiring to the point our main goal is to present an approach, not a specific ontol- that we decided to leverage on it. ogy. Assume now that a suitable general ontology compris- Definition 1 (Agent). ing the needed high-level (structural) categories and their An animate entity that is capable of doing something on pur- ontological relationships (parthood, participation, etc.) has pose.2 been identified. The KEer isolates a coherent fragment of the This notion is meant to be quite inclusive comprising human top-level that includes all the structural categories (renam- and animal agents as well as robots and softbots. A quick ing them if needed), their formalisation, their characterising analysis of the description identifies four key elements that relationships, and that covers all the domain categories. In together characterise this perspective: animate entity, capa- short, the KEer delegates the formalisation of the (typically ble, purpose, doing. We take these four elements plus agent complex) high-level notions to the top-level ontology or a itself as primitives, i.e., add them to the non-logical vocab- suitable fragment of it. Note that, by construction, structural ulary with the indicated arity:3 Animate(x), DoFor(x, z, p) categories already come as a single theory, i.e., as a single [read: x does y for p], Purpose(p), and HasCapability(x, z) module. [read: x has capability z]. We also add Action(x) as auxiliary As we see, the distinction between domain and structural predicate. categories brings some intrinsic advantage since the KEer Let us formalise this notion in L as follows: limits her work to the formalisation of those notions that are def Agent(x) $ Animate(x) ^ 9z, p (P urpose(p)^ more accessible to common sense and well understood in HasCapability(x, z) ^ DoF or(x, z, p)) (1) practical use, while she harvests existing ontological work to organise and formalise the high-level categories which The auxiliary predicate is needed to constrain all the argu- structure the overall system. ments of the new relations:4 One crucial problem in any ontology construction ap- HasCapability(x, z) ! Agent(x) ^ Action(z) (2) proach via modularisation is the merging of the different modules. We anticipate that this remains an open problem DoF or(x, z, p) ! Agent(x) ^ Action(z) ^ even in the approach here sketched. Yet, there is some ad- P urpose(p) (3) vantage since we need to merge only two modules: the one Observations: (a) one could also formalise the definition of domain categories (introduce next) and the one just dis- def cussed, that of structural categories. using the following formula Agent(x) $ Animate(x) ^ 9z, p (P urpose(p) ^ Action(z) ^ DoF or(x, z, p)), or something more complex, e.g., by introducing a separate Cognitive categories Capability predicate. Our discussion is independent of the Here we collect a few ways to understand three common choice of the formula and of the non-logical vocabulary pro- notions, namely, agent, action and artefact. These notions vided the resulting formulas capture fairly well the KEer’s are quite general compared to the standard domain notions understanding of the notion; (b) as said, in this paper we do that are the target of our modelling approach. However, not discuss specific characterisations of relations and limit these have the advantage of being simple, quick to intro- our interest to their domain and range (mainly to give a ba- duce, largely discussed in the literature and intuitively clear sic understanding of what they are binding). 2 without discussing an application scenario. Their informal Definition 2 (Agent). descriptions have been collected from the literature and are well-known and accepted within their communities. How- Something that acts in an environment.5 ever, we do not argue for or against their value, i.e., we are 2 This is the informal definition adopted by John Sowa, neutral about these informal characterisations since these are see http://www.jfsowa.com/ontology/agents.htm used only to exemplify our approach. 3 These are fresh elements in L. Renaming or indexes can be Since our approach requires to translate the notions in a used where needed. 4 logical language, for the sake of the presentation let us as- Note that we use the same numbering for axioms and defini- sume we have chosen some first-order language L whose tions since the latter are technically seen as “if and only if” axioms, interpretation is based in the usual logical and semantic ma- def from which our choice of the $ symbol. chinery. Below we will gradually add information on the 5 Informal definition proposed in (Poole and Mackworth 2010), non-logical vocabulary of L. see http://artint.info/html/ArtInt 3.html Here there are just two characterising terms: to act and envi- Definition 5 (Action). ronment. As before, we take these two as new non-logical An event done by an agent for a reason. terms in the language L. Our formalisation of this defi- nition in first-order logic uses: Agent(x), Environment(y), We take the following non-logical vocabulary to formalise ActIn(x, y) [read: x acts in y] and the auxiliary predicate this definition: Event(x), DoForReason(x, y, z) [read: x Object(x). This implies that in the reading we formalise, en- does y for z], Reason(x) and Agent(x). A formalisation of vironment is seen as an object. the definition in L with the new elements is: We formalise this notion in L with a definition and an def axiom as follows: Action(x) $ Event(x) ^ 9y, z (Agent(y) ^ def DoF orReason(y, x, z)) (13) Agent(x) $ 9y (Environment(y) ^ ActIn(x, y)) (4) DoF orReason(y, x, z) ! Agent(y) ^ Action(x) ActIn(x, y) ! Agent(x) ^ Object(y) (5) ^Reason(z) (14) Definition 3 (Agent). Anything that can be viewed as perceiving its environment Definition 6 (Action). through sensors and acting upon that environment through A bodily movement by an agent. actuators.6 One can characterise this perspective via the following ele- The definition does not refer to generic events but to the ments: Agent(x), Environment(x), Sensor(x), Actuator(x), specific subclass of bodily movements. We take as our vo- PerceiveWith(x, y, z) [read: x perceives y with z], cabulary: Movement(x), Does(x, y) [read: x does y] and ActOnWith(x, y, z) [read: x acts on y with z] and the auxil- Agent(x). The formalisation of this definition is: iary predicate Object(x). def We now formalise this notion in L via one definition and Action(x) $ M ovement(x) ^ 9z (Agent(z) ^ two axioms: Does(z, x)) (15) def Agent(x) $ 9y, s, z(Environment(y) ^ Does(x, y) ! Agent(y) ^ Action(x) (16) P erceiveW ith(x, y, s) ^ ActOnW ith(x, y, z)) (6) How to be (Technical) artefact Sensor(x) ! Object(x) (7) Finally, we present two notions of artefact among the three Actuator(x) ! Object(x) (8) presented in (Borgo et al. 2014). We report them in a simpli- P erceiveW ith(x, y, s) ! Agent(x) ^ Object(y)^ fied version to work with small formulas. Sensor(s) (9) Definition 7 (Artefact). ActOnW ith(x, y, z) ! Agent(x) ^ Object(y) ^ A (physical) object obtained by an agent by selecting a phys- Actuator(z) (10) ical entity and attributing to it a (technical) quality. How to be Action This definition is quite general. According to this view, when one chooses a pebble to use it as a paperweight, she creates In this part we report a few definitions of (intentional) ac- an artefact. The paperweight has a new distinct property with tion which have been collected in (Trypuz 2008). Again, we respect to the pebble, namely, the attributed capacity to per- use simple and generally well received informal definitions form as a paperweight. (For the sake of the presentation, we and, again, one could formalise them more deeply or pre- have ignored that the selected entity is ontologically con- cisely, but recall that our goal is to present an approach, not stituent of the artefact.) The non-logical vocabulary that we a specific ontology. will use to characterise this notion is: Object(x), Agent(x), Definition 4 (Action). Quality(x), Select(x, y) [read: x selects y], Attribute(x, y, z) The event that is carried out by an agent. [read: x attributes y to z]. Our rough formalisation is: This is a very general notion of action, it characterises an def action as an event with a certain relation with an agent. The Artefact(x) $ Object(x) ^ 9y, q(Agent(y) ^ key terms for the formalisation are: Action(x), Event(x), Quality(q) ^ Select(y, x) ^ Attribute(y, q, x)) (17) Agent(x), and Does(x, y) [read: x does y]. We then define Select(y, x) ! Agent(y) ^ Object(x) (18) action, in the language of first-order logic L expanded with these non-logical predicates, as follows: Attribute(y, q, x) ! Agent(y) ^ Quality(q) ^ ^Object(x) (19) def Action(x) $ Event(x) ^ 9y (Agent(y) ^ Does(y, x)) (11) The next definition is engineering oriented: Does(x, y) ! Agent(x) ^ Action(y) (12) Definition 8 (Artefact). A (physical) object that is made by an agent and has some 6 Definition in (Russell and Norvig 1995). given behaviour. Here we use the following non-logical vocabulary for the them the same way. This is exactly what we are doing in the formalisation: Agent(x), Behavior(x), Make(x, y) [read: x formal language as well. We claim that, assuming we have makes y], HasBehavior(x, y) [read: x has behavior y], an ontology that tells us (i.e. constrains) the interpretation of IntendBehavioFor(x, y, z) [read: x intends y to be a behavior Animate, Purpose, HasCapability and DoFor, then axiom (1) of z] and Object(x). can be used in that ontology to define the category of agents. Then, here is a possible formalisation: This category will contain the entities that are seen as agents by our informal definition. Note that, although we are neu- def Artefact(x) $ Object(x) ^ 9y, b(Agent(y) ^ tral on the interpretation of the relations HasCapability and M ake(y, x) ^ Behavior(b) ^ Behave(x, b) ^ DoFor, we constrain their arguments to be of a certain type: agents and qualities for the first, agents, qualities and pur- IntendBehaviorF or(y, b, x)) (20) poses for the latter. This is not strictly necessary but since M ake(y, x) ! Agent(y) ^ Object(x) (21) here we do not discuss relation characterisation in general, Behave(x, b) ! Object(x) ^ Behavior(b) (22) for the time being we make this minimal commitment. IntendBehaviorF or(y, b, x) ! Agent(y) ^ Note that there is nothing special about Definition 1. If one prefers the view proposed by Definition 2, then she as- Behavior(b) ^ artef act(x) (23) sumes that to be an agent it suffices to ensure that there is an entity of a certain kind and a special relationship between the agent and that thing. As it happens, this entity should be an We collected a variety of syntactic definitions about three environment and the relationship should constrain the agent cognitively relevant notions: agent, action and artefact. We to act in that environment. Assuming that Environment assume that every formalisation of each single domain cate- and ActIn are correctly constrained in the ontology, axiom gory has been checked to be satisfactory with respect to the (4) does the job as needed. intended meaning and logically consistent. The characteri- The formal definitions given by axioms (1) and (4) use sation of domain notions is generally limited so that consis- different non-logical vocabulary except for the predicate tency is fairly easy to verify but this really depends on the Agent that they both aim to define. This is not true for structure of the axioms and the interactions between them. Definitions 2 and 3, corresponding to axioms (4) and (6), Note that notions can very well contain negative conditions. since these share also the predicate Environment. Gen- Different is the case of self-referential or recursive defini- erally speaking, two distinct definitions may very well tions like “an artefact is an object made by an agent which use the same non-logical vocabulary. Note however that is not itself an artefact”. The possibility to include these de- this is only a syntactic correspondence. One can interpret pends on the chosen formal language. Environment in axiom (4) differently from the interpre- tation of that predicate in axiom (6). The occurrences of Underspecified definitions Environment in these two axioms are unrelated as this So far we have listed some informal understandings of com- predicate is here not (yet) associated to a formal definition mon sense notions like agent, action and artefact. As we or characterisation. This holds for all the non-logical vocab- have seen, to model the intended constraints in logic one has ulary, relations included. to choose a suitable non-logical vocabulary. This is usually Finally, note that the KRer should use just one definition a delicate step that requires a general view of the system one for each term. If one needs to use a notion in more than one aims to reach. Here, however, we do not assume there is such sense, clashes can be avoided by renaming. For example, a general system. We want only to ‘state’ the constraints in one can include a notion of Capacity-Agent (from Defini- a logical form. For this reason, we introduced as non-logical tion 1) and a notion of Acting-Agent (from Definition 2), vocabulary the terms used in those informal definitions and provided the non-logical vocabulary are distinguished where considered them as unary or n-ary predicates depending on needed. how we read the descriptions and how they are usually un- derstood in natural language. That is, we wanted this step to The role of underspecified definitions be direct and unconstrained, even at the cost to be naı̈ve. We are looking for a methodology that makes it easy for the Let us see what we have achieved. The interpretation of KEer to build an ontology which: the term ‘agent’ in axiom (1) is constrained by a logical for- • is fairly well built in the sense that it characterises the cat- mula, namely, a conjunct in which one of the subformulas is egories via rich cross-categorical relations, thus beyond an existential. This is indeed the structure of the Definition their positioning in the taxonomy or the list of their sim- 1: an agent is a thing such that there exist other things in ple properties, and such and such relationship with it. Clearly, the interpretation of the term ‘agent’ is captured • contains all the relevant notions and these are charac- by Definition 1 only if the interpretation of ‘animate entity’, terised as the KEer desires. ‘capable’, ‘doing’, and ‘purpose’ are as intended. This, how- Assume a KEer needs to build an ontology for an appli- ever, is not something the informal definition of agent deals cation/scenario A about agents and tools (e.g. a reference with. Indeed, the definition assumes that one knows what ontology for manufacturing). At the moment, from our dis- ‘animate entity’, ‘capable’, ‘doing’, and ‘purpose’ stand for. cussion on cognitive categories, all she can do is to choose Furthermore, it does not even require that we all understand a notion of agent suitable for scenario A, say Definition 1. Similarly, she chooses a notion of artefact matching A, notions are not deemed relevant in A, they will not be for- say Definition 7. These choices isolate two sets of formulas mally characterised. These are the subsidiary categories of which the KEer wants to put together but first, since some the ontology for scenario A. non-logical vocabulary is shared by the two characterisa- tions, namely Object, she has to decide whether all Object’s In this section we have discussed axioms and theories, not occurrences have the same interpretation or not. If not, she ontologies. We have shown how to build a logical theory that needs to relabel Object in one of the definitions to distin- contains predicates with a formalisation driven by informal guish the two predicates. In our case, the notion of artefact definitions. Such a theory is not an ontology: it does not give is restricted to physical objects, thus Object in (17) has a a view of reality, it does not commit to a vision of its con- different meaning. She thus renames Object in (17), (18) stituents. To turn a theory, like the one based on the set Tnvoc , and (19) as P hysicalObj and adds the formulas: into an ontology a further step is necessary: the introduction of structural categories. P hysicalObj(x) ! Object(x) (24) 9x(Object(x) ^ ¬P hysicalObj(x)) (25) The need for a top-level structure At this point, the KEer puts together the formulas char- We have seen how to combine underspecified definitions acterising the notion of agent and the notion of arte- (and accompanying axioms) into a logical theory. Let us fact to obtain a formal theory in the language L ex- assume that we have completed this step reaching a the- tended as needed. Let us call T0voc the set of axioms ory T⇤voc that satisfies the modelling view of the KEer for that constrains the non-logical vocabulary of the theory, the scenario A at stake. In this theory, some predicates are i.e., T0voc = {(1), (2), (3), (17⇤ ), (18⇤ ), (19⇤ ), (24), (25)}, axiomatised to take into account their informal definitions where (n⇤ ) indicates that in that formula the Object predi- (the domain categories), others are not (the subsidiary cat- cate has been renamed P hysicalObj. We will assume that egories). Our next goal is to develop an ontology from it. T0voc has been checked for consistency.7 Fortunately, most of the work has already been done. As a preliminary step, let us list all the predicates oc- There are several predicates in the non-logical vocab- curring in T⇤voc , call this L⇤ . Thus, L⇤ includes Object, ulary of T0voc that are not characterised, for instance, P urpose, Quality, Event and others. Action, P urpose and Object. Depending on the applica- tion/scenario A, we can decide that constraining the notion An ontology encodes a view of the world by stating what of Action is important and that of P urpose is not. We is assumed to (possibly) exist and how to subdivide what then can choose a suitable definition of action according exists in types depending on their essential properties. This to the scenario A, say Definition 5. Note that Agent subdivision is typically developed into a hierarchy of cate- and Artefact, if occurring in the chosen characterisation gories (formally these are classes), i.e., a taxonomy. Such a of Action, are as characterised earlier by the KEer.8 taxonomy is missing in our formalisation T⇤voc and the goal Only the still undefined terms may need to be discussed, of this section is to provide one taxonomy coherent with the whether to unify or to distinguish them, in order to avoid KEer’s modeling choices. conceptual mismatches. After all, our initial labelling was completely naı̈ve and unconstrained. For instance, in our Taxonomies are important but also complex to build cor- example the KEer should consider whether the notions rectly (Guarino and Welty 2009). For this reason, we suggest of P urpose and Reason are to be unified (by renaming to reuse one of the existing foundational or upper ontolo- one or adding an equivalence axiom). Assuming this is gies. There are many one can choose from, e.g., BFO, BORO, not the choice of the KEer, her set of interest is T1voc = DOLCE, GFO , UFO , YAMATO to name a few.9 They are not {(1), (2), (3), (17⇤ ), (18⇤ ), (19⇤ ), (24), (25), (13), (14)}. equivalent. For instance, BORO rejects the existence of ob- Again, this theory has to be checked for consistency before jects in the standard sense (the so called 3D entities) and moving on. if this restriction is in contrast with the scenario at stake, it At this point we can go on and add characterisations of should not be used. Some techniques to select among upper other notions that are deemed important in the scenario A. ontologies are being developed, e.g., (Khan and Keet 2012) If these are not already formalised, as done for agent, action although much work still needs to be done. The formal lan- and artefact in Section Cognitive categories, the KEer can guage in which the ontology is available is also an impor- provide a suitable characterisation without much trouble: as tant factor in the decision. Fortunately, many ontologies are before, the formalisation of a notion is done independently available in several languages and some tools to translate of other notions. As this process evolves, a library of (cog- across languages or to merge them are also being developed nitive) category formalisations will start forming increasing (Lange et al. 2012). reuse, reducing the time to build sets Tnvoc and reducing the The KEer includes the selected foundational ontology or KEers’ future efforts in ontology construction. But if these a fragment of it in the theory of T⇤voc , provided the ontology 7 If it is not consistent, then the chosen definitions capture in- (fragment) includes the structural categories (perhaps after compatible views on the primitives and the KEer has to verify her renaming) and covers the domain and subsidiary categories choices or rethink her understanding of the domain. in L⇤ . Here are three cases to guide the matching between 8 More precisely, they can be a specialisation or generalisation 9 of these but this relationship has to be formally captured. http://en.wikipedia.org/wiki/Upper ontology the taxonomy and L⇤ under the assumption that the predi- Entity cates in L⇤ are independent from each other:10 • If a predicate P of L⇤ identifies a category C in the tax- onomy and the extension of C does not correspond to the KEer’s informal interpretation of P , then rename P in L⇤ Object Event Quality Region Concept and in T⇤voc with a fresh predicate; • If a predicate P of L⇤ identifies a category C in the tax- onomy and the extension of C corresponds to the KEer’s informal interpretation of the predicate, then leave L⇤ and Spatial Quality Temporal Quality ... T⇤voc unchanged; • If the extension of a category C in the taxonomy includes Figure 1: A top-level ontology suitable to structure our ex- more entities than the KEer’s informal interpretation of a ample. targeted predicate P in L⇤ , then introduce in the taxon- omy a subcategory corresponding to P as a child of C and add an axiom to characterise the extension of the new Finally, the ontology obtained by merging T⇤voc and subcategory with respect to C. the chosen top-level ontology (possibly pruned) must be If these cases do not suffice to associate each predicate in checked for consistency. This step, when not using a decid- L⇤ to a (subcategory of) a category in the ontology, then able language, is often a challenge for today’s software. the KEer should reconsider the top-level ontology (or the fragment) she started with. Discussion Are there specific characteristics that help to identify a Our approach attempts to reduce the effort to produce a suitable top-level taxonomy? This really depends on the set formal and application-driven ontology and to increase on- T⇤voc and the scenario A. Generally speaking, the taxonomy tology reuse in applications. The difficulty to build robust provides only a few hierarchy levels since once it arrives application-driven ontologies has led to introduce many at notions in the mesoscopic “cognitive” level, T⇤voc itself light axiomatised systems that are satisfactory neither for will supply the domain categories as needed. An important modelling nor for reasoning purposes. Furthermore, one structural aspect that the top-level ontology should provide needs well formalised and ontologically sound systems to is relative to the notions of space and time. Usually, it is ex- ensure interoperability and robustness over time. pected that these come from the top-level ontology with just We aim to harvest as much as possible the KEer’s intu- relatively few constraints, e.g., mereotopological relations itions at the level where these intuitions are mostly reliable for space and partial linear order for time. Usually cognitive and tested, i.e., at the level of the application. Where intu- concepts do not provide detailed information about space ition may fail and everyday experience is not of help, like and time but their interactions may be sensitive to special in the formalisation of general categories, we rely on well- assumptions on space and/or time. The problems that can known top-level ontologies since these are prepared and arise are subtle and need to be studied more deeply. tested by expert ontologists. Our approach sees these two Finally, although we concentrated on unary predicates as phases of a single methodology for ontology construction that identify categories, T⇤voc constrains n-ary relations as and identifies their specific role in the construction process. well. For this reason, the ontology should provide infor- As all known methodologies, our approach has also some mation on basic ontological relations like structural rela- drawbacks. First, structure categories and the domain cate- tions (subclass, parthood and instance of). Other ontologi- gories are typically axiomatised at different levels of preci- cal cross-categorical relations like participation, constitution sion. When a more homogeneous level is needed, one should and dependence, should be minimally constrained in the on- deepen the formalisation of the domain categories, i.e., the tology. most familiar to the KEer, since the others are already well To complete the example, Fig. 1 shows a taxonomy ex- characterised by construction.11 tracted from DOLCE - CORE (Borgo and Masolo 2009) which Second, the lack of an overall view for the characterisation is suitable for our T voc1 . Of course, the KEer has to make of the domain categories leads to miss some interesting, some important choices to link her categories with the top- where not important, connections. For instance, relations level ontology, e.g., whether to interpret P urpose as a de- DoesF or(x, y, z) and Select(x, y), used to model Defini- sired state, a subcategory of DOLCE - CORE Object, or as an tion 1 and Definition 7, respectively, are one a qualifier of information entity, a subcategory of DOLCE - CORE Concept. action (it says that the second argument is an intentional ac- Note that often a foundational or upper ontology cannot be tion) and the other the marker of an intentional action (a se- pruned of some branches due to cross-categorial relation- ships. 11 The subsidiary categories are not problematic. If one of them is at some point considered relevant, it will be formalised and added 10 This is often not true, see our example with P hysicalObj to the list of domain categories. This can be done at any moment and Object. In these cases, first organise these predicates in hi- since the domain categories are all modeled independenty of each erarchies, then use the examples on the most general categories. other. lection) but this is not detected by the direct axiomatisation. reusing general ontologies. Data & Knowledge Engineering These links require an analysis much deeper than what the 86(0):242 – 275. standard KEer may be willing to do. The fact that even other [Ghilardi, Lutz, and Wolter 2006] Ghilardi, S.; Lutz, C.; and existing methodologies, based on modularity or else, cannot Wolter, F. 2006. Did I damage my ontology? A case for cope with this issue indicates that some new idea is needed. conservative extensions in description logics. In Proc. of Third, from our limited experience the categories tend to be KR2006. better characterised by cross-categorical relations. The addi- [Grüninger and Fox 1995] Grüninger, M., and Fox, M. S. tion of basic properties (attributes) was not discussed in this 1995. The role of competency questions in enterprise engi- presentation and can be seen mostly as an independent task, neering. In Benchmarking—Theory and Practice. Springer. better if guided by the adopted upper ontology. 22–31. Fourth, the notions of space and time are tricky. They should be introduced by the top-level ontology because of their gen- [Guarino and Welty 2009] Guarino, N., and Welty, C. 2009. erality and yet their characterisation should be flexible: some An overview on ontoclean. In Staab, S., and Studer, R., applications are based on a qualitative characterisation while eds., Handbook on Ontologies. Springer Verlag, 2nd edition. others on a quantitative characterisation. It might be better to 201–220. develop dedicated modules for these notions. This needs to [Guarino, Oberle, and Staab 2009] Guarino, N.; Oberle, D.; be evaluated carefully. A similar observation applies to rela- and Staab, S. 2009. What is an ontology? In Handbook on tions like parthood, constitution and dependence. ontologies. Springer. 1–17. Fifth, the introduction of a library of cognitive notions as [Guarino 1998] Guarino, N. 1998. Formal ontology in in- suggested in “Cognitive categories” section, raises the prob- formation systems. In Guarino, N., ed., Proceedings of the lem of how to identify the needed definition among all those Second International Conference on Formal Ontology in In- available. Indeed, it is easy to generate many similar, yet formation Systems, 3–15. IOS Press. not equivalent, informal definitions for the same notion. Of [Guarino 1999] Guarino, N. 1999. The role of identity con- course, the KEer could generate a new one every time, due to ditions in ontology design. In Freksa, C., and Mark, D. M., the limited effort they require, giving up on reuse and mak- eds., Spatial Information Theory - Cognitive and Computa- ing harder to evaluate the quality of the modules. tional Foundations of Geographic Information Science. Pro- Finally, sixth, it remains unclear whether this methodology ceedings of International Conference COSIT ’99. Berlin: has actual (implementation) advantages in real applications Springer Verlag. 221–234. and/or reduces the impact of known problems in other ap- proaches (e.g. the merging of distinct modules). This point [Henneh 2014] Henneh, M. 2014. Study and development of is something we cannot properly address at this stage of our a methodology to build ontologies in a modular way. Bach- investigation. elor thesis, Free University of Bozen-Bolzano. [Khan and Keet 2012] Khan, Z., and Keet, C. M. 2012. On- Aknowledgements set: Automated foundational ontology selection and expla- The approach was sketched in the author’s keynote at nation. In Knowledge Engineering and Knowledge Man- WoMO 2011 and, under his supervision, Maxwell Henneh agement (EKAW 2012). Springer. 237–251. developed a few examples in Protégé as part of his bachelor [Lange et al. 2012] Lange, C.; Kutz, O.; Mossakowski, T.; thesis (Henneh 2014). The author thanks the reviewers for and Grüninger, M. 2012. The distributed ontology language their comments and suggestions. (dol): ontology integration and interoperability applied to mathematical formalization. In Intelligent Computer Math- References ematics. Springer. 463–467. [Borgo and Masolo 2009] Borgo, S., and Masolo, C. 2009. [Poole and Mackworth 2010] Poole, D. L., and Mackworth, Foundational Choices in DOLCE. In Staab, S., and Studer, A. K. 2010. Artificial Intelligence: foundations of computa- R., eds., Handbook on Ontologies. Springer Verlag, 2nd edi- tional agents. Cambridge University Press. tion. 361–381. [Russell and Norvig 1995] Russell, S., and Norvig, P. 1995. [Borgo et al. 2014] Borgo, S.; Franssen, M.; Garbacz, P.; Ki- Artificial Intelligence. A modern approach. Prentice-Hall, tamura, Y.; Mizoguchi, R.; and Vermaas, P. E. 2014. Tech- Egnlewood Cliffs. nical artifacts: An integrated perspective. Applied Ontology [Simperl and Tempich 2006] Simperl, E. P. B., and Tempich, Journal 9(3-4):217–235. C. 2006. Ontology engineering: a reality check. In On the [Borgo 2011] Borgo, S. 2011. Goals of modularity: a voice Move to Meaningful Internet Systems 2006: CoopIS, DOA, from the foundational viewpoint. In Modular Ontologies GADA, and ODBASE. Springer. 836–854. (WOMO 11). Amsterdam: IOS Press. 1–6. [Simperl 2009] Simperl, E. 2009. Reusing ontologies on [Euzenat and Shvaiko 2007] Euzenat, J., and Shvaiko, P. the semantic web: A feasibility study. Data & Knowledge 2007. Ontology matching. Heidelberg (DE): Springer- Engineering 68(10):905–925. Verlag. [Trypuz 2008] Trypuz, R. 2008. Formal Ontology of Action. [Fernández-López, Gómez-Pérez, and Suárez-Figueroa 2013] A Unifying Approach. Lublin (Poland): Wydawnictwo KUL. Fernández-López, M.; Gómez-Pérez, A.; and Suárez- Figueroa, M. C. 2013. Methodological guidelines for