Detecting Biomedical Relations using Distant Supervision 
Roland Roller (DFKI, Berlin)

In this colloquium I will talk about the detection of relationships between key information in biomedical publications, such as treatments for diseases or side-effects of drugs. Given a sentence containing some medical concepts the goal is to determine their relationship to each other. Supervised machine learning methods are a very popular way to address this problem and often provide reliable results. Those methods require manually labelled examples to extract characteristics of particular relationships in order to detect similar information in unlabelled data. However, manually labelled data is not always available and its generation is time consuming and expensive. The main objective of this colloquium is distant supervision, a method which generates those labelled examples automatically using prior knowledge to detect relationships between key facts. First, relation extraction using a limited amount of training data is presented to detect adverse-drug effects in natural language. Then, the focuses switches to the question, whether the Unified Medical Language System (UMLS), a large biomedical knowledge base, is suitable to be used to label data automatically so as to detect similar information in natural language.