Processing natural language user inputs into a more formal,
machine-readable, structured query representation used for making an item recommendation. Analyses of user inputs are coordinated via a
knowledge graph constructed from categories, attributes, and attribute values describing relatively frequently occurring prior interactions of various users with an electronic marketplace. The
knowledge graph has directed edges each with a
score value based on: the conditional probabilities of category / attribute / attribute value interactions calculated from user behavioral patterns, associations between user queries and structured data based on historical buyer behavioral patterns in the marketplace,
metadata from items made available for purchase by sellers used to better define buyers' requirements, and / or world knowledge of weather, locations / places, occasions, and item recipients that map to inventory-related data, for generating relevant prompts for further
user input. The
knowledge graph may be dynamically updated during a multi-turn interactive dialog.