From Ontologies to Learning-Based Programs

In this paper, we discuss the work in progress on designing a prototype for a novel declarative learningbased programming system and present our preliminary results. The main idea is to express learningbased programs in terms of declarative domain knowledge. Given that the existing ontologies contain rich domain and world knowledge, we propose to automatically generate the learning-based programs from the current ontology representation languages such as OWL. The ontological concepts and domain relationships are compiled to a graph which is a partial program. The nodes in the graph are connected to data sensors and learners. Local training algorithms can use data and train learners corresponding to each concept and domain relationship in the graph. Global inference mechanisms make the final decisions based on the local prediction of the learners and under the ontological constraints. We test our framework on the entitymention-relation extraction task.