Natural language provides many ways to describe complex events. For instance, the same event could be described by a single clause (the contractor built the house) or by multiple sentences (They started by laying the foundation. Then, they framed the house and installed the plumbing…).

In this work, we aim to capture the structure of complex events, augmenting existing UDS with a new dataset for event-structural properties that capture information about such things as the subparts of an event, how they are arranged in time, and how events relate to each other and their participants.

We use this new dataset along with others in UDS to induce an empircal event structure ontology from a generative model based on sentence- and document-level UDS graphs. This ontology is jointly learned with three other ontologies for semantic roles, entities, and event-event relations. In each case, we find that our categories align well with others proposed in the linguistics and computational semantics literature.

Data

Train Dev Test Download Citation
26701 9864 9419 pred (zip) Gantt et al. 2021
8878 3012 2970 pred-arg (zip) Gantt et al. 2021
32975 24387 21264 pred-pred (zip) Gantt et al. 2021

References

Gantt, William, Lelia Glass, and Aaron Steven White. 2021. Decomposing and Recomposing Event Structure. [pdf]

Researchers

Will Gantt bio photo
Will Gantt
Lelia Glass bio photo
Lelia Glass
Aaron Steven White bio photo
Aaron Steven White