v1 (zip)


We present a novel semantic framework for modeling temporal relations and event durations that maps pairs of events to real valued scales for the purpose of constructing document-level event timelines.

Using our framework, we construct the largest temporal relation dataset to date, and use it to train a variety of neural models to jointly predict finegrained (real-valued) temporal relations and event durations, showing not only that our models obtain strong results on our dataset, the representations they learn can be straightforwardly transferred to the standard categorical relation datasets with competitive performance.

For a detailed description of the protocols, datasets, as well as models of these data, please see the following paper:

Vashishtha, S., B. Van Durme, & A.S. White. 2019. Fine-Grained Temporal Relation Extraction. arXiv:1902.01390 [cs.CL].


Benjamin Van Durme bio photo
Benjamin Van Durme
Aaron Steven White bio photo
Aaron Steven White
Siddharth Vashishtha bio photo
Siddharth Vashishtha