Multi-stage Vs Single-Stage: A Local Information Focused Approach for Overlapping Event Extraction

Event extraction (EE) aims to extract structured event information from textual data, which is an important yet challenging task in the field of information extraction. Overlapping triggers and overlapping arguments are widely present in textual data, and dealing with the overlapping issues can effectively improve the performance of EE. Current research for overlapping event extraction (OEE) can be categorised into multi-stage extraction and single-stage extraction. Multi-stage extraction treats trigger and argument extraction as two sequential tasks, an approach that tends to suffer from error propagation. Single-stage extraction treats trigger and argument extraction concurrently, an approach that fails to take advantage of the dependency between trigger and argument. We propose two paradigms to describe multi-stage extraction and single-stage extraction, respectively, and propose a local information focused approach (LIFA) based on sliding windows for both multi-stage extraction and single-stage extraction. LIFA designs windows of different lengths based on the characteristics of triggers and arguments. It employs Multi-head Local Self-Attention (MLSA) instead of Multi-head Self-Attention to capture the local information of the text, enabling it to grasp the fine-grained context of an entity and identify triggers and arguments more effectively. The experimental results prove that the local information can improve the trigger and argument extraction in the text, and verify the superiority of the multi-stage extraction over single-stage extraction when good results are achieved in all stages.