A multimodal semantics-driven model for generating crisis discourse in government new media

To address the issues of semantic singularity and poor generation effects in existing government new media crisis discourse generation models, this paper suggests a novel Generative Adversarial Network model (WOA-GAN). WOA-GAN first uses an improved BERT model to encode the text of the discourse and uses a residual network to encode the image. Then, a cross-attention mechanism is designed to generate a mask map, effectively integrating the text semantic information and visual characteristics in a multimodal way. Eventually, a semantic space-aware attention mixing module is designed to integrate the text characteristics with the mask map, and guide the multimodal discourse features to enhance details, thus generating high-quality crisis discourse. Experimental results demonstrate that the Bilingual Evaluation Understudy (BLEU) of WOA-GAN has increased by 8.2%–19.5%, which can significantly improve the generation speed while maintaining the generation quality.