EmbDiffounder: Semantic-Transfer Enhanced Sequence Diffusion Model for Text Generation

As a new generative paradigm, diffusion models have shown impressive performance in generating continuous data such as images. However, their effectiveness in handling discrete text data is suboptimal. Challenges remain in generating complex semantics and maintaining semantic consistency. To address these issues, this paper proposes EmbDiffounder, a diffusion generative model with enhanced semantic transmission capabilities. By leveraging the M2Emb model to generate high-quality text embeddings, the semantic representation capability is significantly improved. Additionally, an Adaptive-dynamic Bunch-updating Clustering (ABC) strategy is introduced to enhance the precision of generation. Furthermore, a Multi-scale Residual Fusion (MsRF) module is employed to preserve rich syntactic information and semantic features during the diffusion and denoising processes. Experimental results demonstrate that EmbDiffounder performs exceptionally well in multiple ext generation tasks, achieving a BLEU score of 38.32, a BERT-Score of 91.01, and a UAInst-Score of 6.12, significantly outperforming existing models and validating its advantages in generation quality and semantic consistency.