Statistical Relational Learning of Natural Language

While understanding natural language is easy for humans, it is complex forcomputers. The main reasons for this can be found in the structural nature andthe inherent ambiguity of natural language. Correctly interpreting languagetherefore requires one to take into account the necessary context. In orderto perform natural language understanding by means of machine learningtechniques, an appropriate representation is required that takes into accountthis relational information and integrates the necessary background knowledgeinto the learning process. Statistical relational learning is well-suited to represent this structuralinformation, and to incorporate the necessary context and background knowledgefor natural language understanding. Furthermore, its inherent probabilisticnature offers opportunities to deal with linguistic ambiguity. This thesisinvestigates the promise of statistical relational learning for natural languageprocessing and provides evidence for the utility of this approach. As a first contribution, we demonstrate the expressiveness and interpretabilityof the relational representation on two natural language learning problems.Furthermore, we explore the importance of the declarative approach for theinclusion of contextual information, and analyze the influence of the relationalrepresentation by a comparison of several machine learning techniques. A secondcontribution is the extension of the graph kernel based relational learningframework kLog with a natural language processing module, in order to obtaina full relation learning framework for natural language learning. As a thirdcontribution, we introduce relational regularization and feature ranking in orderto assess the importance of the relational features. Finally, we extend rulelearning to a probabilistic setting and explore its application in the context ofmachine reading.status: publishe