Chinese Event Extraction Method Based on Word Feature and Pointer Network

Event extraction serves as a fundamental task within the information extraction domain, focusing on deriving valuable insights from unstructured event records and organizing them into a structured format. Existing methods for Chinese event extraction face challenges such as insufficient semantic representation and inadequate modeling of trigger words and argument dependencies in cases of overlapping argument roles, leading to poor performance in argument extraction. To address these issues, we propose a Chinese event extraction method based on character and word feature extraction and a pointer network. We construct character vectors using the BERT pre-training model while simultaneously designing a word-level feature extraction network. BiLSTM is employed to extract character-level and word-level features separately, and we fuse the character vectors with weighted word vectors to obtain a more comprehensive and accurate semantic representation in the form of fused character and word feature vectors. These fused vectors are then input into the pointer network, which utilizes dual-sequence pointers to identify the starting positions of trigger words and arguments, allowing for more flexible decoding. This effectively resolves issues related to role overlapping and argument nesting, completing the final event extraction task. On the DuEE1.0 dataset, the F1 scores for trigger word extraction and event argument extraction reach 87% and 65%, respectively, outperforming existing commonly used event extraction models, thus validating the effectiveness of this method.