Leveraging Multi-Head Attention Mechanism To Improve Event Detection

Event detection (ED) task aims to automatically identify trigger words from unstructured text. In recent years, neural models with attention mechanism have achieved great success on this task. However, existing attention methods tend to focus on meaningless context words and ignore the semantically rich words, which weakens their ability to recognize trigger words. In this paper, we propose MANN, a multi-head attention mechanism model enhanced by argument knowledge to address the above issues. The multi-head mechanism gives MANN the ability to detect a variety of information in a sentence while argument knowledge acts as a supervisor to further improve the quality of attention. Experimental results show that our approach is significantly superior to existing attention-based models.