The Semantic Web Needs More Cognition
Review 1 by Philipp Cimiano:
The paper argues in line with Gärdenfors that the Semantic Web could be made "semantic" if it would build on cognition-inspired reasoning mechanisms going beyond pure deductive (syllogistic) reasoning approaches. The authors propose that the "conceptual spaces" theory of Gärdenfors could provide the appropriate formalisms for this purpose.
The authors motivate their proposal from a search perspective, arguing that an approach based on conceptual spaces would allow to contextualize/personalize search better.
I have a few questions/comments on the paper:
1) Where would a semantic search engine get access to the personal interpretation (in terms of conceptual spaces) of each user? Do the authors assume that each user might publish their "conceptual spaces" on the Web? Only then could search be personalized/contextualized, is that right?
2) What is the relation between the proposal of the authors (based on conceptual spaces) and Fuzzy Logic? Advocates of Fuzzy Logic might argue that FL was developed to deal with queries such as "warm climate vacation" where the interpretation of "warm" is not only fuzzy/vague (this is what FL is good at modeling), but even context dependent, such that there are different fuzzy functions (possibly for each user and/or context).
3) The role of CSML could be elaborated further. It was not clear to me who is supposed to use CSML for what. If CSML is a language to define concepts in line with the CS theory of Gärdenfors, then it would seem a good idea to build such a language on top of OWL. After all, one of the stated goals of Gärdenfors is to link the symbolic to the conceptual level. From a SW point of view, would it not be faithful to Gärdenfors theory to try to bring conceptual spaces and OWL together instead of defining a language which is completely independent of OWL?
4) The authors claim that "The Semantic Web should afford two important tasks: the "efficient calculation of semantic similarity" and "combinations of concepts". The calculation of similarity is at the heart of the conceptual spaces of Gärdenfors and is not accounted for well in current SW formalisms, I agree. However, the "combination of concepts" represents the core of OWL and any description logic. So here I would disagree. Maybe the point of the authors is that the combination possibilities should go beyond Boolean combinations using AND/OR/NOT, but it would be nice to see some further comments here.
Review 2 by Aldo Gangemi:
The main contributions of this paper include: 1) pointing to cognitive science evidence to discuss semantics over the web; 2) providing a concrete example of a cognitively plausible representation for Gärdenfors' conceptual spaces and its use for contextually-sensible value regions e.g. when valuating climate in different seasons, communities, or countries; 3) suggesting a direction towards hybridization.
The main revision to be performed is a more reader-friendly comparison between conceptual spaces and existing representation and reasoning pipelines used in the semantic web. While the merits of the author's proposal are well described, their advantages wrt existing approaches is only summarized in section 4, and is not very conclusive for people that are not expert ontology designers with particular competence in both multidimensional representations and similarity reasoning.
Some suggestions for improvement are listed here:
- show a pair of alternative models in OWL2 and CSML, and the related reasoning services, highlighting the pros and cons;
- when discussing the integration between CSML and existing SW representations, the references to "mappings" between the two are just sketched; for example, it's not immediately obvious why context-sensitivity would generate different memberships in a category and therefore different OWL ontologies. It's be important for the reader to know if there is any possibility to approximate that in OWL, e.g. with multi-typing and non-disjoint classes, and why;
- when discussing similarity and geometrical representation, NLP aspects are very relevant, but just a quick note is given to the reader at the end of section 3, with no related literature;
- there is a clear relation to Fuzzy OWL to be a way to recover some expressivity loss when mapping from CSML to OWL: please include it, or discuss why that is not relevant;
- Gärdenfors' claim that SW is only syllogistic reasoning is clearly outdated (and probably wrong even in 2004) ... a better service to cognitive science in this context would be to avoid radical criticism like this ... after all, formal logic does not even try to be "cognitively adequate" ... imo it's better to abridge it with extensions, rather than to adopt the typical disruptive rhetoric of philosophers ;)
- it'd be very nice to distinguish between similarity and context *representation* vs *reasoning*. One could buy one or the other, not necessarily both:
- the example in the butlast paragraph of section 3:
"what is warm in Sweden is not warm in Europe even if European country is modeled as a generalization or super-class of Sweden"
should be clarified. If the author means that not necessarily being warm for a Swedish implies being warm for any European, this is safely assumed in regular logic as well.