Scaling Up ILP: Experiences with Extracting Relations from Biomedical Text

Abstract

We have been applying Inductive Logic Programming (ILP) to the task of learning how to extract relations from biomedical text (specifically, Medline abstracts). Our primary focus has been learning to recognize instances of “this protein is localized in this part of the cell” from labeled training examples. ILP allows one to naturally make use of substantial background knowledge (e. g., biomedical ontologies such as the Gene Ontology – GO – and MEdical Subject Headings – MESH) and rich representations of the examples (e. g., parse trees). We discuss how we formulated this task for ILP and describe our methods for scaling ILP to this large task. We conclude with a discussion of some of the major challenges that ILP needs to address in order to scale to large tasks.