CPPFEE: A Cascade Pointer Prediction Framework for Financial Event Extraction

Event extraction (EE) is a challenging task in information extraction, which aims at extracting the event details including event type, trigger, argument and its corresponding role from the input sentence. Many existing works focus on the general event type extraction, however, they cannot extract plenty of information from the sentence and cannot make full use of the event type information and predicted information in the financial scenario. To address these issues, we put forward a framework named Cascade Pointer Prediction Financial Event Extraction (CPPFEE), which contains a multi-scale local feature extraction module, adaptive global event attention fusion module and cascade global pointer prediction module. The multi-scale local feature extraction module obtains the hidden information in the texts. Then the adaptive global event attention fusion module makes full use of the event information to generate an event-aware sentence representation. Lastly, the cascade global pointer prediction module is used to make the prediction. The target prediction results can be used during role prediction, which helps to improve the accuracy of the role prediction. Since the terminology in the financial area is very precise, without ambiguity, the target prediction will be very accurate which makes the target prediction fewer errors, alleviating the problem of error propagation. Our method utilizes this fact to assist the role prediction. The evaluation on a public event extraction benchmark FewFC demonstrates that CPPFEE achieves significant improvements in financial event extraction over state-of-the-art methods.