Core Concepts
Coherent and engaging knowledge selection is crucial for generating informative responses in dialogue systems.
Abstract
Knowledge-grounded dialogue systems aim to generate coherent and engaging responses based on the dialogue contexts and selected external knowledge. Previous methods tend to lack diversity or coherence in knowledge selection, leading to repetitive or incoherent responses. The CET2 framework introduces a novel approach that models topic transitions for selecting knowledge that is coherent to the context of conversations while providing adequate diversity. By considering multiple factors for knowledge selection, including valid transition logic and systematic comparisons between available knowledge candidates, CET2 outperforms existing approaches in terms of knowledge selection accuracy, balance of topic entailment, and development in dialogues.
Stats
Extensive experiments on two public benchmarks demonstrate the superiority and better generalization ability of CET2 on knowledge selection.
CET2 outperforms previous state-of-the-art methods by 1.6% and 4.7% in seen and unseen scenarios respectively.
Analysis shows that CET2 can better balance topic entailment (contextual coherence) and development (knowledge diversity) in dialogue than existing approaches.