Discovering Temporal Narrative Containers in Clinical Text

The clinical narrative contains a great deal of valuable information that is only understandable in a temporal context. Events, time expressions, and temporal relations convey information about the time course of a patient’s clinical record that must be understood for many applications of interest. In this paper, we focus on extracting information about how time expressions and events are related by narrative containers. We use support vector machines with composite kernels, which allows for integrating standard feature kernels with tree kernels for representing structured features such as constituency trees. Our experiments show that using tree kernels in addition to standard feature kernels improves F1 classification for this task.