TPLUG: An Efficient Framework Through Token Pair Linking and Undirected Graph for Chinese Event Extraction

Chinese Event extraction (CEE) from text has attracted increasing attention in recent years but remains challenging. Some EE models dealing with overlap and nesting employ multi-stage classification, potentially causing misinformation spread. Alternatively, EE often focuses on triggers but ignores trigger-less scenarios. Meanwhile, multiple same-type events in a sentence can’t be distinguished by the trigger word alone. Therefore, we design a CEE model TPLUG to solve the above problems. To unify scenarios with and without trigger words, we treat the trigger word as an argument role, combine event type and argument role in pairs into multiple classes, and transform argument extraction into a scoring task of token pairs under different classes. Furthermore, an undirected complete graph is employed to depict relationships among different arguments, facilitating event partitioning by searching complete subgraphs. Experiments on DuEE1.0 datasets demonstrate the proposed approach effectively addresses these challenges and yields competitive results.