Unsupervised Identification of Event-Specific Subgraphs in AMR

Determine an unsupervised method that, given a sentence’s Abstract Meaning Representation (AMR) graph, assigns AMR edges of relation types other than ARG, time, and location to the specific event they belong to, thereby enabling positive AMR subgraph pairs to be sampled from the same event for the CLEVE graph-encoder contrastive pre-training.

Background

CLEVE performs contrastive pre-training for event extraction using AMR structures. For event structure pre-training, the method ideally requires positive pairs of AMR subgraphs drawn from the same event, but currently approximates positives by drawing subgraphs from the same sentence because event-level subgraph identification is not available.

The authors note that while ARG, time, and location relations align well with trigger–argument pairs, there are roughly 100 other AMR relation types for which they lack an effective unsupervised method to assign edges to specific events. Solving this would allow more precise sampling of event-level positive pairs and potentially improve structure pre-training.

References

However, it is hard to unsupervisedly determine which parts of an AMR graph belong to the same event. The rule used in the event semantic pre-training only handles the ARG, time and location relations, and for the other about $100$ AMR relations, we cannot find an effective method to determine which event their edges belong to.

— CLEVE: Contrastive Pre-training for Event Extraction  (2105.14485 - Wang et al., 2021) in Appendix, Section "Subgraph Sampling"