Graph-Based Inferences in a Semantic Web Server for the Cartography of Competencies in a Telecom Valley

We introduce an experience in building a public semantic web server maintaining annotations about the actors of a Telecom Valley. We then focus on an example of inference used in building one type of cartography of the competences of the economic actors of the Telecom Valley. We detailed how this inference exploits the graph model of the semantic web using ontologybased metrics and conceptual clustering. We prove the characteristics of theses metrics and inferences and we give the associated interpretations. 1 Semantic Annotation of Competencies In knowledge-based solutions, user interfaces have the tricky role of bridging the gap between complex knowledge representations underlying collective applications and focused views tuned to day-to-day uses. For this reason, we believe that interface design and knowledge representation must be tackled in parallel. In this paper, we describe and analyze an experience in simulating the inferences done by economists and management researchers in building a cartography of the competences of the economic actors of a region. The implementation is now part of a public semantic web server maintaining annotations about the actors of a telecom valley. This paper will explain how we designed such an inference using ontology-based metrics defined above the graph structure of the semantic web annotations statements, but before we go in such details we need to introduce the overall project: the Knowledge management Platform (KmP) of the Telecom Valley of Sophia Antipolis. The goal of KmP was the elaboration of a public repository supporting three application scenarios: (1) promoting the Scientific Park of Sophia Antipolis and its international development by providing the local institutions with a pertinent and upto-date snapshot of the park. (2) facilitating partnerships between different industrial firms of the park. (3) facilitating collaboration on projects between industrial partners and the different research institutes. This platform is available online and relies on a semantic web server publicly available for all the actors of the value chain of the Telecom Valley of Sophia Antipolis. The steering committee of KmP is composed of eleven pilot companies involved in the specifications of the application and the 1 http://www-sop.inria.fr/acacia/soft/kmp.html 2 http://www.sophia-antipolis.org/index1.htm 3 http://beghin.inria.fr/ 248 F. Gandon et al. population of the ontologies: Amadeus, Philips Semiconductors, France Telecom, Hewlett Packard, IBM, Atos Origin, Transiciel, Elan IT, Qwam System and Cross Systems. KmP is a real world experiment on the design and usages of a customizable semantic web server to generate up-to-date views of the telecom valley and assist the management of competencies at the level of the organizations (companies, research institute and labs, clubs, associations, government agencies, schools and universities, etc.). This platform aims at increasing the portfolio of competences of the technological pole of Sophia Antipolis by helping companies, research labs and institutions express their interests and needs in a common space in order to foster synergies and partnerships. The platform implements a public knowledge management solution at the scale of the telecom valley based on a shared repository and a common language to describe and compare the needs and the resources of all the organizations. Ontologies were built from models provided by domain experts [Lazaric & Thomas, 2005] and end-users: models of competencies, models of the telecom domains (networks, computer science, etc.), task models, value chain of the telecom valley, etc. The implementation merges the frameworks of the semantic web (RDF, RDFS), the classic web (HTML, CSS, SVG) and the structured web (XML, XSLT) to integrate data coming from very different sources, allow queries from different viewpoints, adapt content to users, analyze, group, infer and render indicators of the Telecom Valley situation. KMP relies on the integration of multiple components: databases for back-end persistence, web servers with JSP and servlets to provide front ends, and the CORESE semantic web server [Corby et al, 2004] to provide semantic web processing capabilities. Databases are used to store the different ontologies (e.g. ontology of technologies, of actions, of deliverables, of markets, of cooperation, etc.), the models (e.g. value chain of a telecom valley), and the users' data (e.g. descriptions of firms, research centers, competences, projects, etc.). Direct accesses and modifications of ontologies and other data are managed directly at the database level. Wrappers extract the relevant and authorized data from the databases and export them in RDF/S to feed CORESE as needed. The platform integrates contributions coming from whole the Telecom Valley: • several ontologies are populated and validated by multiple actors using interviews and brainstorming sessions animated by the local government administration. • several sources of data are integrated: models provided by practitioners and researchers in management, descriptions of firms using industrial and economic markets vocabulary, description of research institutes using academic terms, etc. The whole system relies on RDF, RDFS, and production rules [Corby et al, 2004] to describe the models and actors of the Telecom Valley. Exploiting this semantics the platform is able to: • apply rules to enrich the different contributions and bridge the different viewpoints allowing a broad variety of queries and analysis to be run e.g. a set of rules generalize and group identical competences detailed in the profiles of the actors to provide statistics to researchers in management; 4 http://www.telecom-valley.fr/index.php4?lang=ang Graph-Based Inferences in a Semantic Web Server 249 • exploit underlying models to propose graphic views of the Telecom Valley using XSLT to produce SVG rendering and combining on-the-fly models defined by the economists with data entered by the different actors; e.g. figure 1 shows an SVG interface to browse the value chain of the Telecom Valley and obtain statistics on the exchanges by clicking on the arrows. To each arrow is attached a query that CORESE solves against the RDF/S annotations of the Telecom Valley. For instance the screenshots shows statistics on the exchanges between two segments of the value chain (8b and 6a) and the distribution of these exchanges over the disjoint sub-classes of exchanges. • apply complex query constructors to find partners, build consortiums, extract indicators, build statistics, sort and group results, find approximate answers, etc. • apply clustering algorithms and produce graphic representations in SVG to allow institutional and industrial actors to get abstract views of the cartography of competences in the Telecom Valley; Fig. 1. SVG view of exchanges on the value chain of the Telecom Valley of Sophia Antipolis In this article we focus on one inference supported by the graph models underlying this semantic web repository: an ontology-based conceptual clustering providing a customizable and up-to-date cartography of competences available in the telecom valley. Section 2 briefly introduces an extract of the domain models and the users' requirements. Section 3 details the inferences underlying this representation, in particular the ontology-based metrics exploiting the semantic web graph structures. Section 4 concludes with the evaluation of this representation. 2 Model-Based Automated Cartography of Competencies The first requirement and scenario of KmP is "to acquire and give a broader visibility of the community of the technological pole". As part of its answers, the 250 F. Gandon et al. platform provides a dynamic up-to-date cartography of the competencies available in the technological pole and grouped in clusters. In KmP, the overall design methodology was oriented toward use and users. We relied on participatory design involving end-users, domain experts, information management experts and knowledge modeling experts. A large part of the specifications relied on mock-ups of interfaces built from the visual representations the users are used to. In particular, figure 2 shows a draft made by users when making explicit their notion of competences; it shows what they called a "readable representation of the clusters of competencies in the technological pole". The first consequence of such a readability requirement is a set of expressivity requirements on the ontology. The current model used in the project relies on an ontology that consists of more than a thousand concept types and a hundred relations. Central to the modeling is the concept of "competence" used by the organizations when describing their profiles or the profile of the partners they are looking for. The model proposed by researchers in management and economics [Lazaric & Thomas, 2005] uses four facets to describe a competence and each facet is formalized in a part of the ontology. For instance, the competence "designing microchips for the 3G mobile market using GSM, GPRS and UMTS" is decomposed into four elements: an action (design); a deliverable (microchip); a market (3G mobile technology); a set of resources (GSM, GPRS, UMTS). Market : SI Market : IT Applications Clusters (groups of bubbles) represent complementary competencies i.e. similar from technology stand point Bubbles (circles) represent similar competences ; their size represent their frequency Market : Telecoms Fig. 2. Draft of a representation of clusters of competences The second consequence of the readability requirement is the ability to simulate the inferences mobilized by the users when building this representation. The branch or the level of the ontology used to describe the situation is not always the same as the one used to display inference results. For instance, different users (e.g. industrialists vs. economists) may enter and use knowledge at different levels. In simple cases, we use rules close to Horn clauses to bridge these gaps. For the inferences behind the representation in Figure 2, the algorithm is much more complex and is a matter of conceptual clustering usually performed by economy and management analysts: Graph-Based Inferences in a Semantic Web Server 251 1. Analysts chose a market to which the analysis will be limited; all sub-types of this market will be considered, all ancestors or siblings will be discarded. 2. In this market, analysts group competences according to the similarity of their resources; a competence may have several resources (e.g. java, c, c++, project management) and one is chosen as the most representative (e.g. programming). This first grouping represents a cluster. 3. In each cluster, analysts group competences according to the similarity of their action (e.g. design) to form bubbles. On the one hand we use ontology-based modeling to provide meaningful and dynamic representations (clusters as core competences of the technological pole) and on the other hand we need ontology-based inferences to automate this clustering (clusters as emergent structures in knowledge analysis). Questions associated to this problem include: what are the inferences underlying this representation? How can they be linked to semantic web models of the valley? How can we ensure that the clustering will be meaningful to the users? In literature, the work on the formal side of the semantic web is largely influenced by the fact that logic-based languages are the most frequently used implementation formalisms. However, entailment is not the only product one should expect from a knowledge-based system, and the conceptual structures of the semantic Web can support a broad variety of inferences that goes far beyond logical deduction evening its simplest forms (RDF/S). Let us take the example of the class hierarchy which is considered to be the backbone of the RDFS schemata. The interpretation of the subsumption link is that the extension of a concept type (e.g. laptop) is a subset of the extension of another concept type (e.g. computer). What this logical implication hides is a graph structure that links the concept types through their genus and differentia. The graph structure of the semantic web formalisms supports inferences that go far beyond the set inclusion. The rest of this article shows how we designed such inferences to recreate the representation drafted in Figure 2 and how this specific example illustrates the richness of the underlying graph model of the semantic web. 3 Semantic Metrics to Visualize Knowledge 3.1 Semantic Metrics on the Ontological Space The idea of evaluating conceptual relatedness from semantic networks representation dates back to the early works on simulating the humans’ semantic memory [Quillian, 1968] [Collins & Loftus, 1975]. Relatedness of two concepts can take many forms for instance, functional complementarity (e.g. nail and hammer) or functional similarity (e.g. hammer and screwdriver). The latter example belongs to the family of semantic similarities where the relatedness of concepts is based on the definitional features they share (e.g. both the hammer and the screwdriver are hand tools). The natural structure supporting semantic similarities reasoning is the concept type hierarchy where subsumption links group types according to the characteristic they share. When applied to a semantic network using only subsumption links, the relatedness calculated by a spreading algorithm gives a form of semantic distance e.g. the early system of [Rada et al., 1989] defined a distance counting the minimum number of edges between two types. 252 F. Gandon et al. We can identify two main trends in defining a semantic distance over a type hierarchy: (1) the approaches that include additional external information in the distance, e.g. statistics on the use of a concept; see for instance [Resnik, 1995] [Jiang & Conrath, 1997] (2) the approaches trying to rely solely on the structure of the hierarchy to tune the behavior of the distances [Rada et al., 1989][Wu & Palmer, 1994]. Including external information implies additional costs to acquire relevant and up-to-date information and furthermore, this information has to be available. Thus in a first approach we followed the second trend. In the domain of Conceptual Graphs [Sowa, 1984], where the graph structure of knowledge representation is a core feature, a use for such a distance is to propose a non binary projection, i.e. a similarity S:C2→[0,1] where 1 is the perfect match and 0 the absolute mismatch. We used the CORESE platform provided by [Corby et al, 2004] to build our system. It is provided with an implementation of a depth-attenuated distance allowing approximate search. The distance between a concept and its father is given in (1):