The paper argues for an alternative, empirical (instead of analytical) approach to a Semantic Web-ready KR&R, motivated by the so far largely untackled need for a feasible emergent content processing. Since the onset of AI, the knowledge representation and reasoning (KR&R) field has been largely an analytical (in the early Wittgenstein sense) endeavour aimed at producing sound and complete results by algorithmic manipulation of rigorously defined symbol sets (knowledge bases). This works pretty well when the respective domain of interest is closed, deterministic and amenable for complete, indubitable formalisation. Unfortunately, the Semantic Web is not such a neat environment. As has been widely acknowledged in the community, the data one has to manage generally have one or more of the following qualities to them: they are dynamic, noisy, inconsistent, incomplete, intractably abundant, too inexpressive, uncertain and/or context-dependent. Approaches extending the traditional analytical KR&R accordingly have been investigated recently, however, they seldom take the problem of the actual content acquisition into account as a primary design consideration. To illustrate the issue, we can think of the current RDF/OWL experience - substantially more people generate and use the rather relaxed OWL Full than the rigorous OWL DL flavour. Yet, much larger number of users employ the even simpler RDF(S). It seems to be quite risky to assume that future Semantic Web developers and users will eagerly and happily adopt complex uncertain, paraconsistent or contextualised extensions of the rather OWL-ish (analytical) approach to KR. Therefore we argue that a truly Semantic Web-ready KR&R should natively tackle noisiness, uncertainty, etc., but also sensibly redefine and/or relax the rigorous assumptions and theoretical groundwork of the analytical approaches in order to follow the WWW success instead of the vapour-ware Xanadu path.
The informatic universe we have to represent within the Semantic Web is very similar (yet simpler) to the perceptual reality of human beings – namely concerning its openness, noisiness and lack of complete, sufficiently formalised data. ! This work has been supported by the EU IST FP6 project ‘Nepomuk’ (FP6-027705) and by Science Foundation Ireland under Grant No. SFI/02/CE1/I131. Therefore it can be quite useful to draw inspirations from the features of the human mind. These are, however, in many respects exact opposites of the traditional KR&R basic notions (e.g., entailment or model theory) [1]. Conversely, the high-performance and robust (although quite likely unsound and incomplete) natural reasoning abundantly employs similarity-based incorporation and retrieval of data to and from the memory [2]. The respective reasoning is much rather empirical than analytical then [1].
Expanding on these rough considerations, the proposed alternative KR&R
conceptualisation can be described by three general canons: (
We have recently started to implement our vision in a respective
framework, with which we have already attained promising initial results in integration
and “analogical closure” of automatically learned ontologies using a biomedical
legacy resource [3]. We address the canon (