Why naïve semantics cannot be formalized and why this is (almost) of no consequence for ontology engineering.

The project of representing the semantics of human concepts using formal languages is not new, but it currently knows a renewed interest with computational ontologies. The main objective of these representation systems, which constitute one essential component of the semantic Web, is to provide machines with “awareness” of the meaning of the data being processed. Formal languages have always been criticized and their use in ontology engineering is no exception. Various authors denounce this attempt to subject our naive semantics to logical formalisms, which are too strict and too mechanical to capture its expressiveness and adaptability, and propose alternative approaches and models, supposedly closer to “natural” humanconceptualization and representation mechanisms. The purpose of the present article is to demonstrate that those criticisms are mostly ill-founded, and in any case irrelevant, in the sense that they assume the subordination of ontologies to epistemic requirements (to model faithfully a referent) rather than operational requirements (to be efficient and fulfil needs). They become inappropriate from the moment that one considers that the function of ontologies is not (i) to represent faithfully the furniture and the constituent structures of the reality (realism) or how the mind represents that reality (conceptualism), but (ii) to improve the capacity of categorization, access to information and manipulation of data and documents, and (iii) to standardize practices of categorization and encoding. Based on this conception, a new epistemological approach to computational ontologies can be proposed: instrumentalism.