Ontology Learning in the Deep

Recent developments in the area of deep learning have been proved extremely beneficial for several natural language processing tasks, such as sentiment analysis, question answering, and machine translation. In this paper we exploit such advances by tailoring the ontology learning problem as a transductive reasoning task that learns to convert knowledge from natural language to a logic-based specification. More precisely, using a sample of definitory sentences generated starting by a synthetic grammar, we trained Recurrent Neural Network (RNN) based architectures to extract OWL formulae from text. In addition to the low feature engineering costs, our system shows good generalisation capabilities over the lexicon and the syntactic structure. The encouraging results obtained in the paper provide a first evidence of the potential of deep learning techniques towards long term ontology learning challenges such as improving domain independence, reducing engineering costs, and dealing with variable language forms.