📄 Graph-based Event Coreference Resolution

Summary

Chen and Ji tackle event coreference resolution as defined by the ACE program: grouping event mentions in a document so that each group describes one real-world event. They argue the task is harder than entity coreference because coreferring event mentions are sentence-level and structurally diverse. They illustrate this with a news example about a cafe bombing. The authors represent the mentions of each event type as an undirected weighted graph and partition it with spectral clustering under the normalized-cut objective (Shi and Malik, 2000). Their contribution is not a new clustering algorithm but two ways to estimate the coreference (affinity) matrix:

• a hand-designed formula based on corpus statistics of trigger and argument matches;
• a maximum entropy classifier using lexical, distance and argument features.

On the ACE 2005 English corpus with gold event mentions, both methods reach ECM-F scores of about 0.83. Both clearly beat simple baselines. This shows that graph clustering is a workable framework for event coreference.

Key points

• Data and evaluation: ACE 2005 English corpus (560 documents), split into 500 for training, development and statistics, and 60 for testing. Evaluation uses ground-truth event mentions and ECM F-Measure (Luo, 2005).
• Graph setup: A separate graph is built for each event type, treating the 33 ACE subtypes as types.
• Method 1 (formula): Scores trigger agreement at several levels: exact match, same stem, or positive WordNet similarity. It also scores argument agreement by whether entity ID and ROLE match. Corpus counts show that identical triggers and fully matching arguments are the strongest coreference cues.
• Results: With 10-fold cross-validation, the best cut threshold is 0.85 for both methods. Method 1 scores 0.8449 and Method 2 scores 0.8418, a weak difference (p≤0.1447). On the test set, the methods score 0.8363 and 0.8312.
• Baselines and future work: Putting all same-type mentions into one cluster gives ECM-F 0.535, and treating every mention as a singleton gives 0.7635. The impact of system-extracted, non-gold mentions is left for future work.