SomeWhere: A Scalable Peer-to-Peer Infrastructure for Querying Distributed Ontologies

In this invited talk, we present the SomeWhere approach and infrastructure for building semantic peer-to-peer data management systems based on simple personalized ontologies distributed at a large scale. Somewhere is based on a simple class-based data model in which the data is a set of resource identifiers (e.g., URIs), the schemas are (simple) definitions of classes possibly constrained by inclusion, disjunction or equivalence statements, and mappings are inclusion, disjunction or equivalence statements between classes of different peer ontologies. In this setting, query answering over peers can be done by distributed query rewriting, which can be equivalently reduced to distributed consequence finding in propositional logic. It is done by using the message-passing distributed algorithm that we have implemented for consequence finding of a clause w.r.t a set of distributed propositional theories. We summarize its main properties (soundness, completeness and termination), and we report experiments showing that it already scales up to a thousand of peers. Finally, we mention ongoing work on extending the current data model to RDF(S) and on handling possible inconsistencies between the ontologies of different peers. 1 Overview of SomeWhere SomeWhere promotes a ”small is beautiful” vision of the Semantic Web [1] based on simple personalized ontologies (e.g., taxonomies of atomic classes) but which are distributed at a large scale. In this vision of the Semantic Web introduced by [2], no user imposes to others his own ontology but logical mappings between ontologies make possible the creation of a web of people in which personalized semantic marking up of data cohabits nicely with a collaborative exchange of data. In this view, the Web is a huge peer-to-peer data management system based on simple distributed ontologies and mappings. For scalability purpose, we have chosen a simple class-based data model in which the data is a set of resource identifiers (e.g., URIs), the ontologies are (simple) definitions of classes possibly constrained by inclusion, disjunction or equivalence statements, and mappings are inclusion, disjunction or equivalence statements between classes of different peer ontologies. That data model is in accordance with the W3C recommendations since it is captured by the propositional fragment of the OWL ontology language (http://www.w3.org/TR/owlsemantics). Query answering through ontologies is achieved using a rewrite and evaluate strategy. In SomeWhere the query rewriting problem can be reduced to a consequence finding problem in distributed propositional theories. It is performed by a message-passing algorithm named DeCA: Decentralized Consequence finding Algorithm [3]. As a result, query answering in SomeWhere is sound, complete and terminates. Moreover, the detailed experiments reported in [4] show that it scales up to 1000 peers. 2 Illustrative example We illustrate the SomeWhere data model on a small example of four peers modeling four persons Ann, Bob, Chris and Dora, each of them bookmarking URLs about restaurants they know or like, according to their own taxonomy for categorizing restaurants. Ann, who is working as a restaurant critic, organizes its restaurant URLs according to the following classes: • the class Ann:G of restaurants considered as offering a ”good” cooking, among which she distinguishes the subclass Ann:R of those which are rated: Ann:R ⊑ Ann:G • the class Ann:R is the union of three disjoint classes Ann:S1, Ann:S2, Ann:S3 corresponding respectively to the restaurants rated with 1, 2 or 3 stars: Ann:R ≡ Ann:S1 ⊔ Ann:S2 ⊔ Ann:S3 Ann:S1 ⊓ Ann:S2 ≡ ⊥ Ann:S1 ⊓ Ann:S3 ≡ ⊥ Ann:S2 ⊓ Ann:S3 ≡ ⊥ • the classes Ann:I and Ann:O, respectively corresponding to Indian and Oriental restaurants • the classes Ann:C, Ann:T and Ann:V which are subclasses of Ann:O denoting Chinese, Täı and Vietnamese restaurants respectively: Ann:C ⊑ Ann:O, Ann:T ⊑ Ann:O, Ann:V ⊑ Ann:O Suppose that the data stored by Ann that she accepts to make available deals with restaurants of various specialties, and only with those rated with 2 stars among the rated restaurants. The extensional classes declared by Ann are then: Ann:V iewS2 ⊑ Ann:S2, Ann:V iewC ⊑ Ann:C, Ann:V iewV ⊑ Ann:V , Ann:V iewT ⊑ Ann:T , Ann:V iewI ⊑ Ann:I Bob, who is fond of Asian cooking and likes high quality, organizes his restaurant URLs according to the following classes: • the class Bob:A of Asian restaurants • the class Bob:Q of high quality restaurants that he knows Suppose that he wants to make available every data that he has stored. The extensional classes that he declares are Bob:V iewA and Bob:V iewQ (as subclasses of Bob:A and Bob:Q): Bob:V iewA ⊑ Bob:A, Bob:V iewQ ⊑ Bob:Q Chris is more fond of fish restaurants but recently discovered some places serving a very nice cantonese cuisine. He organizes its data with respect to the following classes: • the class Chris:F of fish restaurants, • the class Chris:CA of Cantonese restaurants Suppose that he declares the extensional classes Chris:V iewF and Chris:V iewCA as subclasses of Chris:F and Chris:CA respectively: Chris:V iewF ⊑ Chris:F , Chris:V iewCA ⊑ Chris:CA Dora organizes her restaurants URLs around the class Dora:DP of her preferred restaurants, among which she distinguishes the subclass Dora:P of pizzerias and the subclass Dora:SF of seafood restaurants. Suppose that the only URLs that she stores concerns pizzerias: the only extensional class that she has to declare is Dora:V iewP as a subclass of Dora:P : Dora:V iewP⊑Dora:P Ann, Bob, Chris and Dora express what they know about each other using mappings stating properties of class inclusion or equivalence. Ann is very confident in Bob’s taste and agrees to include Bob’ selection as good restaurants by stating Bob:Q ⊑ Ann:G. Finally, she thinks that Bob’s Asian restaurants encompass her Oriental restaurant concept: Ann:O ⊑ Bob:A Bob knows that what he calls Asian cooking corresponds exactly to what Ann classifies as Oriental cooking. This may be expressed using the equivalence statement : Bob:A ≡ Ann:O (note the difference of perception of Bob and Ann regarding the mappings between Bob:A and Ann:O) Chris considers that what he calls fish specialties is a particular case of Dora seafood specialties: Chris:F ⊑ Dora:SF Dora counts on both Ann and Bob to obtain good Asian restaurants : Bob:A ⊓ Ann:G ⊑ Dora:DP Figure 1 describes the resulting overlay network. In order to alleviate the notations, we omit the local peer name prefix except for the mappings. Edges are labeled with the class identifiers that are shared through the mappings between peers. 3 Query Rewriting in SomeWhere through Propositional Encoding In SomeWhere, each user interrogates the peer-to-peer network through one peer of his choice, and uses the vocabulary of this peer to express his query. Therefore, queries are logical combinations of classes of a given peer ontology.