Integer Programming Ensemble of Classifiers for Temporal Relations

Extraction of events and understanding related temporal expression among them is a major challenge in natural language processing. In longer texts, processing on sentenceby-sentence or expression-by-expression basis often fails, in part due to the disregard for the consistency of the processed data. We present an ensemble method, which reconciles the output of multiple classifiers for temporal expressions, subject to consistency constraints across the whole text. The use of integer programming to enforce the consistency constraints globally improves the best published F1 score from the TempEval-3 Challenge by 3 percentage points to 0.3899.