Groundtruth Budgeting : A Novel Approach to Semi - Supervised Relation Extraction in Medical Language

We address the problem of weakly-supervised relation extraction in hospital discharge summaries. Sentences with pre-identified concept types (for example: medication, test, problem, symptom) are labeled with the relationship between the concepts. We present a novel technique for weakly-supervised bootstrapping of a classifier for this task: Groundtruth Budgeting. In the case of highly-overlapping, self-similar datasets as is the case with the 2010 i2b2/VA challenge corpus, the performance of classifiers on the minority classes is often poor. To address this we set aside a random portion of the groundtruth at the beginning of bootstrapping which will be gradually added as the classifier is bootstrapped. The classifier chooses groundtruth samples to be added by measuring the confidence of its predictions on them and choosing samples for which it has the least confident predictions. By adding samples in this fashion, the classifier is able to increase its coverage of the decision space while not adding too many majority-class examples. We evaluate this approach on the 2010 i2b2/VA challenge corpus containing of 477 patient discharge summaries and show that with a training corpus of 349 discharge summaries, budgeting 10% of the corpus achieves equivalent results to a bootstrapping classifier starting with the entire corpus. We compare our results to those of other papers published in the proceedings of the 2010 Fourth i2b2/VA Shared-Task and Workshop. Thesis Co-Supervisor: Ozlem Uzuner Title: Assistant Professor of Information Studies SUNY, Albany Thesis Co-Supervisor: Peter Szolovits Title: Professor of Computer Science and Engineering MIT CSAIL