Entriever: Energy-based Retriever for Knowledge-Grounded Dialog Systems

Motivation ➢ Holistic Modeling via Energy Function: Treats candidate retrieval results (combinations of knowledge pieces) as a whole, calculating relevance scores through an energy function 𝑈 𝜃 (𝑐 𝑡 , 𝑢 𝑡 , 𝜉 𝑡 ) to model inter-piece dependencies directly. ➢ Residual Energy Design: Constructs a residual form 𝑝 𝜃 ret ∝ 𝑝 ref ⋅ exp(-𝑈 𝜃 ) based on traditional retrieval distribution 𝑝 ref , reducing training difficulty. ➢ Semi-supervised Adaptability: Enables retrieval probability calculation without accessing the full KB, suitable for pseudo-label filtering in unlabeled data.