D2D framework boosts product search accuracy by 30% in conversational AI

Conversational product search assistants promise a more natural shopping experience but often struggle with the tension between asking too many questions and making premature recommendations. A new paper from researchers at Adobe Research and IIT Delhi presents Dialogue to Discovery (D2D), a framework that dynamically exploits product attribute structures to efficiently steer conversations toward users' desired items. D2D adaptively prioritizes which attributes to ask about and strategically times when to make a recommendation, reducing both frustration from long dialogues and mismatched suggestions.

Evaluated on three curated datasets from the Amazon Reviews corpus, D2D shows significant improvements over existing methods. In simulated conversations using a multi-factor utilitarian patience framework, D2D improves target-finding accuracy by 22.2–29.9%, reduces session abandonment by 6.6–16.1%, and shortens average conversations by 27.5%. A complementary user study further validates that D2D delivers higher user satisfaction and perceived efficiency. The work has implications for e-commerce platforms, voice assistants, and any AI system that needs to understand user preferences through natural dialogue.