Neural Network Search Result Recalling Method

The patent transforms the scoring parameters from simple hitting character values to high-dimensional dense vectors that capture semantic meaning. By changing the parameter representation from sparse hitting indicators to dense semantic embeddings, the system achieves both computational efficiency and high semantic matching accuracy through vector similarity calculations.

The patent replaces the mechanical linear addition scoring mechanism with a neural network-based semantic similarity calculation. Instead of mechanically adding hitting character values, the system uses deep learning models to compute semantic matching degrees, substituting a simple arithmetic system with an intelligent semantic understanding system.

The patent performs preliminary action by pre-computing and storing dense vectors for search results in advance. This allows the recall process to efficiently retrieve candidate results based on hitting situation while maintaining the capability to perform accurate semantic matching, as the heavy computation of vector representations has already been done beforehand.

The patent introduces dense vectors as an intermediary between the recall process and the ranking process. The intermediary vectors enable the system to bridge efficient hitting-based recall with accurate semantic-based ranking, allowing the system to first efficiently retrieve candidates and then precisely rank them by semantic relevance to user intent.