MMLAEE A Multi-Scale Multi-Head Latent Attention Model With Sliding Window BERT for Long-Text Event Extraction

With the rapid advancement of the Internet, the speed of public opinion diffusion in news platforms and social media has significantly accelerated, making the need to accurately extract event information from unstructured texts increasingly urgent. However, existing event extraction methods encounter challenges in handling long texts and multi-scale semantic relationships. To address these issues, this paper proposes a joint event extraction model based on sliding window BERT and multi-head latent attention. This model simultaneously extracts event trigger words, subjects, objects, time and location through multi-scale feature fusion and latent variable modeling. Specifically, sliding window BERT is first employed to perform multi-scale encoding on the input text, thereby overcoming the length constraints of pre-trained models and capturing both local and global semantic features. Secondly, a multi-head latent attention mechanism is designed to dynamically adjust the cross-head attention weights through latent variables, enabling the modeling of complex dependency relationships between trigger words and arguments. Finally, a joint decoding strategy is adopted, incorporating KL divergence to constrain the latent variable distribution, achieving end-to-end collaborative reasoning for trigger words and arguments. A series of multi-metric experiments are conducted on benchmark datasets such as ACE2005 and CEC 2.0. Experimental results demonstrate that this method surpasses existing state-of-the-art approaches.