Knowledge-guided diffusion for emotion and topic controllable text generation

Controllable text generation faces challenges in coordinating conflicting specifications while maintaining factual consistency. Existing methods apply constraints uniformly regardless of compatibility, rely on static annotations, and ignore knowledge relevance shifts across generation stages. We propose a knowledge-guided diffusion framework with three innovations. First, conflict-aware fusion detects emotion-topic incompatibility via learned projections and adaptively balances constraints. Second, hierarchical relation-aware encoding categorizes 847 relation types into semantic categories with category-specific transformations. Third, trajectory-aware weighting models knowledge relevance evolution, emphasizing semantic grounding early and details late. We enrich three datasets (Wizard of Wikipedia, GoEmotions, WebNLG) through automatic annotation aligned with inference. Experiments show substantial improvements: 69.5% emotion accuracy and 76.8% topic relevance versus baseline 61.3% and 68.4%; 71.9% factual F1 on WebNLG versus 66.7%. Human evaluation confirms quality gains (3.9/5 vs. 3.6/5). Critically, improvements concentrate on difficult, high-conflict, sparse-knowledge cases (62% vs. 54% and 47%), addressing fundamental coordination challenges.