Models are trained on THYME patient sets: 28-127
Training data: Mode 8, residue 0-3
Testing data: Mode 8, residue 4 and 5

UMLS entities, whose spans overlap with gold Events, were extracted to boost training data
UMLS entities were not extracted for testing data. For testing data, only gold events were used for evaluation.

Both arguments of event-event relations are required to be valid medical terms (i.e. be one of the UMLS semantic types).
We show test results for event-time relations, in which events were not required to be valid medical terms.
Results for event-time relations, in which event arguments are requried to be valid medical terms, were also shown here. Final event-time model is UMLS-filtered. 