Bridging Passive and Active: Enhancing Conversation Starter Recommendation via Active Expression Modeling

The article introduces the Passive-Active Bridge (PA-Bridge), a framework designed to enhance conversation starter recommendations in LLM-driven conversational search by addressing the limitations of traditional feedback loops. PA-Bridge employs an adversarial distribution aligner to reconcile the distribution gap between passive recommendations and active user expressions, alongside a semantic discretizer for popularity debiasing. Online A/B testing indicates a 0.54% increase in Feature Penetration Rate and a 0.04% rise in User Active Days, highlighting its potential to improve user engagement and the relevance of query recommendations in open-ended dialogue systems.