V-Know Collaborative knowledge base debugged and refinement

Knowledge-based, intelligent systems have made their way into practice: Model-based systems are generating diagnostics for automobiles, configurator applications help us assembling complex products from simpler components, recommender systems generate product proposals that match our needs and requirements, just to name a few. Being the cornerstone for success for many of these applications, the knowledge acquisition problem has been addressed in the past in different dimensions, the main focus lying on knowledge representation and conceptualization issues as well as on process models for capturing and formalizing a domain expert`s knowledge. Historically, one main assumption of these approaches was that there shall exist one single point of knowledge formalization and in consequence one (user-oriented) conceptualization and a central knowledge acquisition tool. In most cases in real world, however, the domain knowledge is in the heads of different stakeholders, typical examples being cross-department or cross-organization business rules or new types of applications, in which large user communities are sharing knowledge in an open-innovation, web-based environment. Only recently, with the emergence and spread of Web 2.0 and Semantic Web technologies, the opportunities and also the problems of collaborative knowledge acquisition have again become a topic of interest. With regard to the types of knowledge to be acquired, the main focus of these recent developments, however, is on acquiring "structural" knowledge, i.e., on terms, concepts, and relationships among them. The proposed V-KNOW project shall build upon these new developments, but it however aims at going a step further and target at the collaborative acquisition and refinement of domain-constraints and business rules as they represent the most crucial, frequently updated, and thus costly part in many knowledge-based applications. The main questions answered in the project among others comprise the following: How can we automatically detect and resolve conflicts if knowledge acquisition is distributed between different knowledge contributors? How can we assist the knowledge contributors to acquire knowledge by asking them the "right" questions, i.e., minimizing the interaction needed? How can we generate "good" proposals for changing the knowledge base from different, possibly only partially-defined knowledge chunks, i.e., find plausible (in the eyes of the contributors) changes of the knowledge base? The results of the proposed research project are methods and algorithms for answering these questions and an evaluation based on a prototypical implementation.

The collaborative development of knowledge bases (KBs) by a team of domain experts is a complex and time consuming task. The goal of this process is a shared formalization of a domain by logical descriptions (ontologies). In order to improve this process, the V-Know project focused on the investigation of methods and on the development of tools which increase the efficiency of knowledge acquisition and maintenance of ontologies. The research was done in three main directions: a) debugging of knowledge bases b) learning of logical formulas, and c) acquisition of factual knowledge from the Web. The availability of debugging tools is an important prerequisite for the successful implementation of KBs. During the project we investigated a number of debugging approaches, allowing the identification of a set of logical formulas that are faulty with respect to a set of requirements and test cases specified by knowledge engineers. The main achievement was the development of an interactive debugging framework for efficient fault localisation. In common application cases, ontology debugging methods will return numerous sets of formulas (diagnoses) of an ontology as possible causes of a failure. Our debugging approach acquires information from experts in order to efficiently differentiate between diagnoses. Subsequently, the debugger exploits this information for uniquely identifying the part of the ontology which must be changed such that all requirements of the knowledge engineers are met. Current learning methods have only limited applicability in the context of the V-Know application scenario because the required amount of data is not available at the development time of the ontology. Therefore, we proposed a novel learning method which is based on a similar idea as the one used in debugging. In particular, the method finds controversial instances of the knowledge base, i.e. a set of value assignments, and asks an expert if the instances are correct. In addition, experts can provide argumentations for their answers. The classification results and the argumentation are exploited to generate logical formulas expressing the newly acquired knowledge. Furthermore, for developing and maintaining knowledge bases, the mining of factual knowledge is a central task. We investigated methods for acquiring factual knowledge from tables published on the Web. In particular, we developed a data mining framework that includes a focused web-crawler for finding relevant webpages with tables as well as methods for the automatic detection and extraction of the tables data.