Control Fusion: A Unified Framework for Multi-Level Controllable Paraphrase Generation

Controllable paraphrase generation aims to rewrite text while preserving semantic meaning and satisfying specific linguistic and structural constraints. Existing neural approaches typically address controllability along isolated dimensions, which limits their effectiveness in applications that require simultaneous regulation of multiple text properties. This study proposes Control Fusion, a unified framework for multi-level controllable paraphrase generation. The proposed architecture integrates semantic and lexical similarity conditioning, fine-grained linguistic attribute control, template-aware structural guidance, and quality-oriented optimization within a modular encoder–decoder design. The framework incorporates self-contrastive learning and contextual similarity filtering to improve diversity while maintaining semantic fidelity. A control calibration mechanism is also introduced to enhance consistency between target constraints and generated outputs. Experimental evaluation using a pre-trained encoder–decoder backbone on large-scale paraphrase datasets demonstrates that the proposed framework improves semantic quality, output diversity, and control accuracy compared with single-dimension control models. The results further indicate that Control Fusion is effective in satisfying multiple constraints simultaneously. Overall, the proposed approach provides a flexible and reliable solution for coordinated multi-level control in paraphrase generation and establishes a practical foundation for advanced controllable text generation in downstream natural language processing applications.

کلیدواژه ها:

Controllable paraphrase generation ، semantic similarity ، linguistic attribute control ، contrastive learning ، text generation ، NLP