On link predictions in complex networks with an application to ontologies and semantics

Abstract

It is assumed that ontologies can be represented and treated as networks and that these networks show properties of so-called complex networks. Just like ontologies our current pictures of many networks are substantially incomplete (Clauset et al., 2008, p. 3ff.). For this reason, networks have been analyzed and methods for identifying missing edges have been proposed. The goal of this thesis is to show how treating and understanding an ontology as a network can be used to extend and improve existing ontologies, and how measures from graph theory and techniques developed in social network analysis and other complex networks in recent years can be applied to semantic networks in the form of ontologies. Given a large enough amount of data, here data organized according to an ontology, and the relations defined in the ontology, the goal is to find patterns that help reveal implicitly given information in an ontology. The approach does not, unlike reasoning and methods of inference, rely on predefined patterns of relations, but it is meant to identify patterns of relations or of other structural information taken from the ontology graph, to calculate probabilities of yet unknown relations between entities.The methods adopted from network theory and social sciences presented in this thesis are expected to reduce the work and time necessary to build an ontology considerably by automating it. They are believed to be applicable to any ontology and can be used in either supervised or unsupervised fashion to automatically identify missing relations, add new information, and thereby enlarge the data set and increase the information explicitly available in an ontology. As seen in the IBM Watson example, different knowledge bases are applied in NLP tasks. An ontology like WordNet contains lexical and semantic knowl- edge on lexemes while general knowledge ontologies like Freebase and DBpedia contain information on entities of the non-linguistic world. In this thesis, examples from both kinds of ontologies are used: WordNet and DBpedia.WordNet is a manually crafted resource that establishes a network of representations of word senses, connected to the word forms used to express these, and connect these senses and forms with lexical and semantic relations in a machine-readable form. As will be shown, although a lot of work has been put into WordNet, it can still be improved.While it already contains many lexical and semantical relations, it is not possible to distinguish between polysemous and homonymous words. As will be explained later, this can be useful for NLP problems regarding word sense disambiguation and hence QA.Using graph- and network-based centrality and path measures, the goal is to train a machine learning model that is able to identify new, missing relations in the ontology and assign this new relation to the whole data set (i.e., WordNet). The approach presented here will be based on a deep analysis of the ontology and the network structure it exposes. Using different measures from graph theory as features and a set of manually created examples, a so-called training set, a supervised machine learning approach will be presented and evaluated that will show what the benefit of interpreting an ontology as a network is compared to other approaches that do not take the network structure into account.DBpedia is an ontology derived from Wikipedia. The structured information given in Wikipedia infoboxes is parsed and relations according to an underlying ontology are extracted. Unlike Wikipedia, it only contains the small amount of structured information (e.g., the infoboxes of each page) and not the large amount of unstructured information (i.e., the free text) of Wikipedia pages. Hence DBpedia is missing a large number of possible relations that are described in Wikipedia. Also compared to Freebase, an ontology used and maintained by Google, DBpedia is quite incomplete. This, and the fact that Wikipedia is expected to be usable to compare possible results to, makes DBpedia a good subject of investigation.The approach used to extend DBpedia presented in this thesis will be based on a thorough analysis of the network structure and the assumed evolution of the network, which will point to the locations of the network where information is most likely to be missing. Since the structure of the ontology and the resulting network is assumed to reveal patterns that are connected to certain relations defined in the ontology, these patterns can be used to identify what kind of relation is missing between two entities of the ontology. This will be done using unsupervised methods from the field of data mining and machine learning.