AVER: Adversarial Variational Enhanced Representation Architecture for Abstractive Multi-document Summarization

Multi-document summarization (MDS) aims to extract key information from multiple related documents into a concise and informative summary. Existing deep learning approaches typically rely on document concatenation or semantic fusion, which often produce rigid, less fluent summaries lacking linguistic naturalness. Moreover, standard evaluation metrics insufficiently capture fluency and semantic similarity, limiting further model improvements. To address these challe nges, we propose a novel multi-document summarization method based on variational adversarial learning. Our approach incorporates a variational autoencoder and a discriminator in an adversarial framework and builds upon a pretrained language model to promote both expression diversity and factual precision in the latent space. This design enables the model to generate richer and more fluent summaries. We further complement traditional evaluation metrics with Self-BLEU and DISTINCT to better assess expression diversity. Experiments on Multi-News, DUC04 and Multi-XScience datasets show that our model outperforms state-of-the-art baselines, achieving up to 30% relative gains in diversity metrics, while maintaining competitive performance in semantic similarity scores.