A Joint Event Extraction Framework based on Conditional Information Fusion Layers

With the rapid development of the Internet, the text published by various online social media is increasing day by day. How to select useful information from the redundant text quickly and accurately is the problem we have to face today. Event extraction aims to solve this problem mentioned above. To improve the effect of trigger and argument extraction, we propose a joint model based on conditional information fusion layers. In the trigger extraction module, we introduce the type information fusion layer to reduce the difficulty of extracting triggers. In the argument extraction module, we introduce the trigger information fusion layer to make the input of argument extraction carry both event type information and trigger information. Finally, we validate the model in the Chinese large-scale event extraction labeled dataset provided by Baidu. Experiments show that our model's F1-score performance is about 1% higher than common models such as BERT and so on in trigger extraction subtask, and about 3% higher than common methods such as BERT-MRC and so on in argument extraction subtask.