A generative similar weather retrieval method and system, electronic device and medium

Abstract

This application discloses a generative similar weather retrieval method, system, electronic device, and medium. The method constructs weather process fragments and a corpus using standardized daily descriptors; encodes the weather process fragments into fragment-level representations, and maps these representations to hierarchical discrete identifiers using a trained autoregressive generator; inserts the hierarchical discrete identifiers corresponding to all weather process fragments in the corpus into a fragment table to construct a publication list; inputs the target query fragment into the trained autoregressive generator, and obtains high-probability identifiers through beam search; searches the publication list using these high-probability identifiers to obtain multiple candidate fragments; calculates the local reordering score for each candidate fragment, and selects multiple target candidate fragments from these candidate fragments based on the local reordering scores; and determines an alignment evidence matrix based on the multiple target candidate fragments and the target query fragment. This application improves the efficiency and accuracy of similar weather retrieval.

Images