Recommendation method based on user questions

By employing a multi-expert hybrid network and a group strategy optimization framework, the problem of insufficient user intent parsing in conversational recommendation systems is solved, enabling deep integration of implicit user needs and accurate recommendations, thereby improving the system's recommendation accuracy and conversation fluency.

CN122451084APending Publication Date: 2026-07-24ASIAINFO TECH CHINA INC
0 Cites 0 Cited by

Patent Information

Application Number
CN202610608940.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-06
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing conversational recommendation systems struggle to deeply analyze users' implicit needs in natural language. The recall module retrieves information based on incomplete or outdated user states, resulting in low recommendation accuracy. Furthermore, the strategy model cannot effectively utilize the deep semantics of the recall phase.

Method used

A multi-expert hybrid network is used to enhance and align text embedding vectors. A large language model is combined to generate user preference summary text. A candidate product set is generated through a recall model and a re-ranking model. A group strategy optimization framework is used for decision-making, and a dynamic dialogue recommendation closed-loop training framework is constructed.

Benefits of technology

It achieves accurate capture and fulfillment of users' true intentions, improves the accuracy and fluency of the recommendation system, and enhances the system's ability to understand complex semantics and the stability of strategy decision-making.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a recommendation method based on user questions.The recommendation method comprises the steps that multiple rounds of conversations, user portraits and behavior data are integrated, semantic summarization of user preferences and product facts is generated through a large language model, then texts and recommendation feature spaces are aligned through a multi-expert hybrid network, and a dynamically-updated candidate product list is output; and the problems of insufficient semantic understanding and recall deviation are solved. The dialogue is modeled as a Markov decision process, and a large strategy model is finely adjusted based on a group strategy optimization framework. Through generation of diversified decisions, multi-dimensional reward evaluation and normalized comparison, recommendation accuracy and dialogue fluency are improved. In each round of interaction, a retrieval module is called in real time to update a candidate list, similar cases are matched from an example library, and a trained strategy model is driven to generate an optimal decision and natural reply, so that the cooperation of semantic deep understanding, strategy continuous optimization and scene adaptive decision is realized, and the overall performance and robustness of a dialogue recommendation system are remarkably improved.
Need to check novelty before this filing date? Find Prior Art