Daniel Fleischhacker

Inductive learning of disjointness axioms

The tremendous amounts of linked data available on the web are a valuable resource for a variety of semantic applications. However, these applications often need to face the challenges posed by flawed or underspecified representations. The sheer size of these data sets, being one of their most appealing features, is at the same time a hurdle on the way towards more accurate data because this size and the dynamics of the data often hinder manual maintenance and quality assurance. Schemas or ontologies constraining, e.g., the possible instantiations of classes and properties, could facilitate the automated detection of undesired usage patterns or incorrect assertions, but only few knowledge repositories feature schema-level knowledge of sufficient expressivity. In this paper, we present several approaches to enriching learned or manually engineered ontologies with disjointness axioms, an important prerequisite for the applicability of logical approaches to knowledge base debugging. We describe the strengths and weaknesses of these approaches and report on a detailed evaluation based on the DBpedia dataset.

Mining RDF Data for Property Axioms

The Linked Data cloud grows rapidly as more and more knowledge bases become available as Linked Data. Knowledge-based applications have to rely on efficient implementations of query languages like SPARQL, in order to access the information which is contained in large datasets such as DBpedia, Freebase or one of the many domain-specific RDF repositories. However, the retrieval of specific facts from an RDF dataset is often hindered by the lack of schema knowledge, that would allow for query-time inference or the materialization of implicit facts. For example, if an RDF graph contains information about films and actors, but only Titanic starring Leonardo_DiCaprio is stated explicitly, a query for all movies Leonardo DiCaprio acted in might not yield the expected answer. Only if the two properties starring and actedIn are declared inverse by a suitable schema, the missing link between the RDF entites can be derived. In this work, we present an approach to enriching the schema of any RDF dataset with property axioms by means of statistical schema induction. The scalability of our implementation, which is based on association rule mining, as well as the quality of the automatically acquired property axioms are demonstrated by an evaluation on DBpedia.

Detecting Errors in Numerical Linked Data Using Cross-Checked Outlier Detection

Outlier detection used for identifying wrong values in data is typically applied to single datasets to search them for values of unexpected behavior. In this work, we instead propose an approach which combines the outcomes of two independent outlier detection runs to get a more reliable result and to also prevent problems arising from natural outliers which are exceptional values in the dataset but nevertheless correct. Linked Data is especially suited for the application of such an idea, since it provides large amounts of data enriched with hierarchical information and also contains explicit links between instances. In a first step, we apply outlier detection methods to the property values extracted from a single repository, using a novel approach for splitting the data into relevant subsets. For the second step, we exploit owl:sameAs links for the instances to get additional property values and perform a second outlier detection on these values. Doing so allows us to confirm or reject the assessment of a wrong value. Experiments on the DBpedia and NELL datasets demonstrate the feasibility of our approach.

Inductive Lexical Learning of Class Expressions

Despite an increase in the number of knowledge bases published according to Semantic Web W3C standards, many of those consist primarily of instance data and lack sophisticated schemata, although the availability of such schemata would allow more powerful querying, consistency checking and debugging as well as improved inference. One of the reasons why schemata are still rare is the effort required to create them. Consequently, numerous ontology learning approaches have been developed to simplify the creation of schemata. Those approaches usually either learn structures from text or existing RDF data. In this submission, we present the first approach combining both sources of evidence, in particular we combine an existing logical learning approach with statistical relevance measures applied on textual resources. We perform an experiment involving a manual evaluation on 100 classes of the DBpedia 3.9 dataset and show that the inclusion of relevance measures leads to a significant improvement of the accuracy over the baseline algorithm.

A Practical Implementation of Semantic Precision and Recall

The systematic evaluation of ontology alignments still faces a number of problems. One is the argued inadequacy of traditional quality measures adopted from the field of information retrieval. In previous work, Euzenat and others have proposed notions of semantic precision and recall that are supposed to better reflect the true quality of an alignment by considering its deductive closure rather than the explicitly stated correspondences. So far, these measures have only been investigated in theory. In this paper, we present the first implementation of a restricted version of semantic precision and recall as well as experiments in using it, we conducted on the results of the 2008 OAEI campaign.

Automatic acquisition of class disjointness

Although it is widely acknowledged that adding class disjointness to ontologies enables a wide range of interesting applications, this type of axiom is rarely used on todays Semantic Web. This is due to the enormous skill and effort required to make the necessary modeling decisions. Automatically generating disjointness axioms could lower the barrier of entry and lead to a wider spread adoption. Different methods have been proposed for this automatic generation. These include supervised, top-down approaches which base their results on heterogeneous types of evidence and unsupervised, bottom-up approaches which rely solely on the instance data available for the ontology. However, current literature is missing a thorough comparison of these approaches. In this article, we provide this comparison by presenting two fundamentally different state-of-the-art approaches and evaluating their relative ability to enrich a well-known, multi-purpose ontology with class disjointness. To do so, we introduce a high-quality gold standard for class disjointness. We describe the creation of this standard in detail and provide a thorough analysis. Finally, we also present improvements to both approaches, based in part on discoveries made during our analysis and evaluation.

Computing incoherence explanations for learned ontologies

Recent developments in ontology learning research have made it possible to generate significantly more expressive ontologies. Novel approaches can support human ontology engineers in rapidly creating logically complex and richly axiomatized schemas. Although the higher complexity increases the likelihood of modeling flaws, there is currently little tool support for diagnosing and repairing ontologies produced by automated approaches. Off-the-shelf debuggers based on logical reasoning struggle with the particular characteristics of learned ontologies. They are mostly inefficient when it comes to detecting modeling flaws, or highlighting all of the logical reasons for the discovered problems. In this paper, we propose a reasoning approach for discovering unsatisfiable classes and properties that is optimized for handling automatically generated, expressive ontologies. We describe our implementation of this approach, which we evaluated by comparing it with state-of-the-art reasoners.

Enriching Ontologies by Learned Negation

Ontologies form the basis of the semantic web by providing knowledge on concepts, relations and instances. Unfortunately, the manual creation of ontologies is a time intensive and hence expensive task. This leads to the so-called knowledge acquisition bottleneck being a major problem for a more widespread adoption of the semantic web. Ontology learning tries to widen the bottleneck by supporting human knowledge engineers in creating ontologies. For this purpose, knowledge is extracted from existing data sources and is transformed into ontologies. So far, most ontology learning approaches are limited to very basic types of ontologies consisting of concept hierarchies and relations but do not use large amounts of the expressivity ontologies provide.