1. Background
事件抽取:抽取event triggers,并将其分类到event types
2. 数据集LEVEN
包含了charge-related events和general events
3. baseline和实验结果
3.1 事件抽取
- token classification: BiLSTM, BERT
- Dynamic max-pooling: DMCNN, DMBERT
- Sequence labeling: BiLSTM+CRF, BERT+CRF
3.2 下游任务
实验结果:直接将事件信息在下游任务上用作side information,在low-resource judgment prediction任务和unsupervised case retrieval任务上有提升