4.1. MLC, gap-filling, and basicality
Let us sum up the methodology -- and its application to ON8.5 modelling -- from the viewpoint of MLC commitment.
We have reused authoritative taxonomies, also assuming that they contain the basic concepts for a domain: the SORTS.
As suggested in par. 2.4 and par. 2.5, heterogeneity of the "minimal intended meaning" of various taxonomies is integrated through the triggering of general and domain theories, which provides other new sorts (gap-fillers) for connecting the isolated branches from taxonomies [cf. Lehmann, 1993], or simply suggesting new connection paths for basic pre-existing sorts (gap-filling through reuse). Gap filling (par. 2.4.3) is decidable because it makes minimal assumptions on what should be meant by intended meaning: basic sorts are interpreted within the constrained lexical field of the other nearby sorts in the terminology, and not within all possible uses (no "globality assumption"). The lexical field is determined pragmatically, by evidentiating local lattices.
In other words, we restrict the available lexical field, and then superimpose some paradigms on it, which provide the gap-fillers for constructing a well-connected, integrated (multi-local) lexical field.
The basic concepts and gap-fillers which structure the multi-local field are called the RELATIONS.
Once sorts and relations have been connected, the compositionality of the model may provide "interesting" extensions to the integrated formal ontologic model: these do not produce basic concepts nor gap-fillers, but less basical concepts, which are composed by a necessary and sufficient (lambda-)abstraction of a sort and some predication on it. This extension operates a specification, thus we call this concepts the SORTAL SPECIFICATIONS (but see par. 4.3.5).
An important warning must be made about our meaning of sort and relation: our sorts are not meant to be modelled as predicates, at least in a first-order sorted logic, where they should be arguments which satisfy our relations, modelled as predicates. This derives from our theory of meaning, which commits to sortal partitions and operators within a lexical field, not to individuals and properties in the world (the objectivist commitment, cf. par. 1.2).
From an objectivist viewpoint, such a first-order logic would be a compression of a second-order structure, where sorts would be predicates and thus relations would be second-order predicates. Obviously, the first-order compression can be modelled as far as individuals are kept out of the model, in a background position.
The advantage of such approach is the open-ended extendibility of our model: we can figure out finer and finer descriptions even of a quite detailed instance. In other words, any description can be used as a sort for other particularized descriptions: for example, a diagnosis used as a case study.
4.2.1 Unary predicates [Guarino&al, 1994] introduces some MLCs for unary predicates on the basis of some criteria (Tab. 2):
Type, quality, indep-role and dep-role are sortal unary predicates; category and mass-like are pseudo-sortal unary predicates;
property and state are non-sortal unary predicates. Roles must be subordinated to some type/kind.
Table 2: A set of meta-level categories for unary predicates (our adaptation from [Guarino92, Guarino&al94 and personal communication).
The MLC of dependent role should catch the intended common meaning of role, which seems to include some special 'performing' value (but see par. 4.3.5): in fact, this requires a more comprehensive unit in which an object is associated to some other to perform its performing status (hence, the dependence). Independent roles, on the other hand, are the actual accidental (non-substantial) sortal predicates from the Aristotelian tradition: these result from the application of a property, which is independent.
A quite different case is that of roles resulting from the application of a pure relation, such as right-object from at the right of, or likable from likes: these seem dependent, but not actantial...
4.2.2 Binary predicates The last issue is related to a classification of binary relations as binary interpretations of unary relations, which is proposed in [Guarino, 1993] (Tab. 3), on the basis of the following criteria:
Tab. 3: The meta-level categories for binary predicates (adapted from [Guarino93]).
The definition 3) should avoid the roles resulting from the application of a pure relation, because no unary predicates are interpreted by it. Nonetheless, one could insinuate that it could be possible to figure out some semantic field which accepts unary predicates interpreted by a pure-relation, thus this distinction does not seem to say the last word about binary predicates.
4.3.1 Sorts, roles, properties Our concepts divide into sorts and sortal specifications, which roughly map to Guarino's sortal predicates: categories and types/kinds versus independent/dependent roles respectively.
Our sortal specifications roughly map to Guarino's roles by no chance: Guarino's role is non-rigid, say it is an accidental specification of a type. From a converse path, we defined local substantiality as motivated by validated basicality in a domain, and our sortal specifications are thus non-substantial, though from a local viewpoint.
Notice that our method of type/role distinction is not alternative to Guarino's, since it provides a methodology for deciding the intended meaning of concepts, which includes a natural distinction as a sort or sortal specification, while Guarino's provides an a-posteriori check of the ontological consistency of MLC assignment.
On the other hand, a-posteriori it is feasible to distinguish between independent and dependent sortal specification as well, a distinction that we eventually include in ON8.5 ontological commitment. Notice that the dependent/independent criterion on roles can be translated in our commitment as actantial/not actantial (see par. 4.3.5).
Our relations --as unary predicates-- roughly map to Guarino's properties and --as binary predicates-- to Guarino's binary relations.
Main asymmetries between Guarino's ontology of MLC and our ontological commitment are:
4.3.2 Qualities Some relations map to Guarino's qualities: this was made intentionally in ON8.5: though notions like color should predicate on so-called universals (like red), we do not accept the generalization of this proposal, since a universal can be either a sort or a predicate, depending on context. We'd rather put qualities as parent relations of universals.
4.3.3 Parts and states Both part-relations and states are domain partitions for us, i.e., they are not at meta-level: sort:Part and rel:PhysicalState (par. 3.).
The first choice is due to our belief that part is not heterogeneous from connectedness, morphology, assessment, or other semantic field operators.
The second choice concerns the problematic criterion of temporal stability, which cannot help but depending on temporal granularity: why <<student>> is stable and <<studies>> not, let alone that <<student>> in the context of a university is even rigid? A clocktick has to be defined before.
We also model a different intended meaning of state as sort:State, subordinate to sort:Process, which grasps a type of process which has no describable temporal interval: in other words, our state is the synchronic description of the dynamic features of some subordinate of sort:Structure.
4.3.4 Mass-like Finally, we make no distinction of a mass-like category, since at the current state of integration in medicine, it is too difficult to anticipate the application of non-topmost sorts as either countable or uncountable. As far as we can figure it out, in medicine common mass-like sorts are used mostly in a metonymical way, i.e., as countable sorts, eg urine is most times referred to urine samples, thus in a countable meaning.
Another, more important argument against mass-like distinction is the interaction of multiple layers (material, biologic, social) and granularities (organic, cellular, genetic) within medical knowledge.
4.3.5 Transposing MLC ontology from logical to cognitive semantics Here we summarize the discussion on MLC, also suggesting that Guarino's criteria could be used for making some further distinctions in ON8.5 ontological commitment.
We already said that -- within semantic fields -- relations are basic structuring concepts and sorts are basic structural concepts. And also we said that sortal specifications are non-basic.
Among sorts, the uncountable topmost layers of ON8.5 may be the CATEGORIES, while the others may be the TYPES. A cognitive counterpart of countability can be found according to the basicality prototype for common language (see criteria in par. 2.2). In particular, categories seem to be less re-identifiable than types: what's the mental image for Structure, or Process? On the other hand, categories seem to be highly schematizable: imagine a simple diagram for Process (eg, a path) or Structure (eg, a configuration).
Among sortal specifications, we could follow the dependent/independent distinction, but the contextual constraints accepted by our theory of meaning require a cognitively-based counterpart (cf. par. 4.2.1), which can be found in actantiality, ie the property of being able to take a special part in a situation change: agent, patient, instrument, goal, cause, etc. Consider that the special status of these categories have been acknowledged in ancient grammars ("complements"), in recent frame grammar ("cases") and narratology, etc. Thus, an independent role is called in this perspective an ACTOR, or ROLE tout court, which also corresponds to common sense, while a dependent role would be called a SUBSTANTIVATE, to stress that it actually is the sortal (nominal) counterpart of a relational (adjectival, property) application.
Tab. 4 presents a grid with such refined ontological commitment for ON8.5.
Tab. 4: The ontological commitment of meta-level categories in ON8.5.
CONCLUSION
Our medical ontology, built and modelled through a methodology finalized to knowledge integration and without a-priori formal constraints on ontological commitment, have been a-posteriori analyzed through the ontological commitment rules proposed by Guarino. The analysis showed that a good amount of those requirements are satisfied in our medical ontology. This check put a favourable light both on the integration methodology and on Guarino's commitment rules, and suggests the feasibility of a theory of equivalence between methodological assumptions on integration and a settlement of rules on MLC commitment.
Acknowledgements
We thank the anonymous referees of KAW96, Nicola Guarino, and Domenico Pisanelli for their useful suggestions. This work has been partially funded by the Italian National Research Council project SOLMC (Ontologic and Linguistic Tools for Conceptual Modelling).
