Training method and device of code embedding model

Abstract

The embodiment of the invention provides a code embedding model training method, which comprises the following steps of: processing target input by utilizing a coding module to obtain an output vector of each character in the target input; and performing first aggregation processing based on an attention mechanism on the output vector of each character based on the first anchor sentence vector by utilizing a text field code sharing aggregation module to obtain a first feature vector of target input. And utilizing a target aggregation module special for the type of the target input to perform second aggregation processing based on the attention mechanism on the output vector of each character based on a target number of second anchor sentence vectors corresponding to the target aggregation module to obtain a second feature vector of the target input. A predicted loss is determined based on the respective first and second feature vectors of each text and each code snippet in the sample set, and the first and second aggregation modules and the first and second anchor sentence vectors are updated based thereon.

Images