Section: New Results

Knowledge Engineering and Web of Data

Participants : Emmanuelle Gaillard, Nicolas Jay, Florence Le Ber, Jean Lieber, Amedeo Napoli, Emmanuel Nauer, Justine Reynaud.

Keywords:

knowledge engineering, web of data, definition mining, classification-based reasoning, case-based reasoning, belief revision, semantic web

Around the Taaable Research Project

The Taaable project was originally created as a challenger of the Computer Cooking Contest (ICCBR Conference) [73]. Beyond its participation to the CCC challenges, the Taaable project aims at federating various research themes: case-based reasoning (CBR), information retrieval, knowledge acquisition and extraction, knowledge representation, belief change theory, ontology engineering, semantic wikis, text-mining, etc. CBR performs adaptation of recipes w.r.t. user constraints. The reasoning process is based on a cooking domain ontology (especially hierarchies of classes) and adaptation rules. The knowledge base is encoded within a semantic wiki containing the recipes, the domain ontology and adaptation rules.

As acquiring knowledge from experts is costly, a new approach was proposed to allow a CBR system using partially reliable, non expert, knowledge from the web for reasoning. This approach is based on notions such as belief, trust, reputation and quality, as well as their relationships and rules to manage the knowledge reliability. The reliability estimation is used to filter knowledge with high reliability as well as to rank the results produced by the CBR system. Performing CBR with knowledge resulting from an e-community is improved by taking into account the knowledge reliability [10]. In the same way, another study shows how the case retrieval of a CBR system can be improved using typicality. Typicality discriminates subclasses of a class in the domain ontology depending of how a subclass is a good example for its class. An approach has been proposed to partition the subclasses of some classes into atypical, normal and typical subclasses in order to refine the domain ontology. The refined ontology allows a finer-grained generalization of the query during the retrieval process, improving at the same time the final results of the CBR system.

The Taaable system also includes a module for adapting textual preparations (from a source recipe text to an adapted recipe text, through a formal representation in the qualitative algebra INDU). The evaluation of this module as a whole thanks to users has been carried out and has shown its efficiency (w.r.t. text quality and recipe quality), when compared with another approach to textual adaptation [76].

FCA allows the classification of objects according to the properties they share into a concept lattice. A lattice has been built on a large set a cooking recipes according to the ingredients they use, producing a hierarchy of ingredient combinations. When a recipe $R$ has to be adapted, this lattice can be used to search the best ingredient combinations in the concepts that are the closest to the concept representing $R$.

Minimal change theory and belief revision can be used as tools to support adaptation in CBR, i.e. the source case is modified to be consistent with the target problem using a revision operator. Belief revision was applied to Taaable to adjust the ingredient quantities using specific inference engines.

Another approach to adaptation based on the principles of analogical transfer applied to the formalism RDFS has been developed [41]. It is based on the problem-solution dependency represented as an RDFS graph: this dependency within the source case is modified so that it fits the context of the target problem. The application problem that has guided this research addresses the issue of cocktail name adaptation: given a cocktail recipe, the name of this cocktail and the ingredient substitution that produces a new cocktail, how could the new cocktail be called?

Exploring and Classifying the Web of Data

A part of the research work in Knowledge Engineering is oriented towards knowledge discovery in the web of data, as, with the increased interest in machine processable data, more and more data is now published in RDF (Resource Description Framework) format. The popularization and quick growth of Linked Open Data (LOD) has led to challenging aspects regarding quality assessment and data exploration of the RDF triples that shape the LOD cloud. Particularly, we are interested in the completeness of the data and the their potential to provide concept definitions in terms of necessary and sufficient conditions [66]. We have proposed a novel technique based on Formal Concept Analysis which organizes subsets of RDF data into a concept lattice [43]. This allows data exploration as well as the discovery of implication rules which are used to automatically detect missing information and then to complete RDF data and to provide definitions. Moreover, this is also a way of reconciling syntax and semantics in the LOD cloud. Experiments on the DBpedia knowledge base shows that this kind of approach is well-founded and effective.