# COFFEE: A Contrastive Oracle-Free Framework for Event Extraction

**Authors:** Meiru Zhang, Yixuan Su, Zaiqiao Meng, Zihao Fu, Nigel Collier  
**arXiv:** [2303.14452](https://arxiv.org/abs/2303.14452)  
**Date:** 2023-03-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.14452v1/

## Key points

- Event extraction is a challenging task in information extraction from unstructured text
- Previous methods for event extraction rely on comprehensive entity annotations or heuristic templates with oracle information
- The focus of this study is on Oracle-Free Event Extraction (OFEE) where only the input context is provided without any oracle information
- COFFEE is a new framework proposed to address the OFEE task, extracting events solely based on document context without referring to any oracle information
- COFFEE introduces a contrastive selection model to rectify generated triggers and handle multi-event instances
- COFFEE outperforms state-of-the-art approaches on the ACE05 benchmark under the oracle-free setting
- Limitations of this study include focusing only on sentence-level event extraction and not considering document-level extraction or cross-sentence relationships
- The training dataset used in this study is relatively small and may not cover all possible event types or scenarios
- Using a larger and more diverse dataset could potentially improve the model's performance and generalizability
- The introduction of a ranking module in COFFEE improves trigger generation but there is a risk of error propagation if the trigger is not identified in the first stage. This limitation should be addressed in future work.
- Future research should address these limitations, explore document-level extraction, larger datasets, and further advancements using the COFFEE framework.

## Summary

Event extraction is a challenging task in information extraction that involves extracting events from unstructured text. Previous methods for event extraction have relied on comprehensive entity annotations or heuristic templates containing oracle information, such as event type, which may not be available in real-world scenarios. In this study, the focus is on the Oracle-Free Event Extraction (OFEE) task, where only the input context is provided without any oracle information, including event type, event ontology, and trigger word. To address this task, a new framework called COFFEE is proposed. COFFEE extracts events solely based on the document context without referring to any oracle information. It introduces a contrastive selection model to rectify generated triggers and handle multi-event instances. The performance of COFFEE is evaluated on a public event extraction benchmark ACE05 under the oracle-free setting and outperforms state-of-the-art approaches. However, it should be noted that there are some limitations to this study. Firstly, it focuses on sentence-level event extraction and does not consider document-level extraction or cross-sentence relationships. Future research could explore these aspects to further enhance the model's capabilities. Additionally, the training dataset used in this study is relatively small and may not cover all possible event types or scenarios. Using a larger and more diverse dataset could potentially improve the model's performance and generalizability. The introduction of a ranking module in COFFEE improves trigger generation; however, due to two-stage inference there is a risk of error propagation if the trigger is not identified in the first stage. This limitation should be addressed in future work. In conclusion, while this study shows promising results in Oracle-Free Event Extraction (OFEE), there are still limitations that need to be addressed in future research. By addressing these limitations and exploring document-level extraction and larger datasets further advancements can be made in event extraction performance using the COFFEE framework.