Improved Paraphrase Generation via Controllable Latent Diffusion

2024-04-13

AI Summary

To address the degradation in quality and efficiency of text diffusion models for paraphrase generation caused by truncation, this paper proposes the first controllable latent diffusion framework tailored for rewriting. Methodologically, we adapt conditional diffusion models to text rewriting by introducing fine-grained semantic fidelity guidance in the latent space, a plug-and-play style controller, and CLIP-space alignment fine-tuning—enabling joint control over semantic consistency, diversity, and conciseness. Unlike conventional Seq2Seq and VAE paradigms, our framework overcomes fundamental limitations in representational capacity and controllability. Experiments on Quora and PAWS benchmarks demonstrate improvements of +3.2 in BLEU and +4.1 in METEOR over strong baselines. Human evaluation confirms 91.5% controllability accuracy and a 2.8× increase in output diversity.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Problem

• Paraphrase Generation
• Text Truncation
• Quality Degradation

Innovation

• LDP method
• controlled secret diffusion
• interpretation generation