Inducing Domain Theories

?Inducing Domain Theories? shows how world knowledge can be learnt from text through Inductive Logic Programming (ILP). What is meant by a \"domain theory\" is a collection of facts and generalisations or rules which capture what commonly happens in some domain of interest. The domain of application was financial news but the approach can be extended to the discovery of new knowledge from different domains. The learning paradigm employed, ILP, generalises over examples from the domain to obtain more general patterns covering the majority of the input instances. ILP was preferred over other machine learning techniques due to the expressive power of the language specifications guiding the search for general patterns and the fact that it allows the inclusion of background knowledge. The relational data mining algorithm WARMR gave the most satisfactory results as it was able to capture frequent patterns of complex structure, often encoding causal relations consisting of two or more verbs and information about their respective arguments. Finite State Automata (FSA) minimisation techniques were employed to render the rules learnt into a more compact, human friendly format.