Parameter allocation method, device, equipment and medium for multilingual machine translation model

Abstract

The present application provides a parameter allocation method, apparatus, device, and medium for a multilingual machine translation model. The method includes: adjusting parameters for each language direction of the multilingual machine translation model based on a preset language dataset to obtain an absolute value ranking of the model parameters; performing joint parameter adjustment to obtain directional gradients for each language direction, determining an average gradient and a first similarity coefficient based on the directional gradients, and then determining a directional pruning rate; adjusting the directional pruning rate based on a second similarity coefficient through iterative joint parameter adjustment to obtain a target pruning rate, and pruning the model parameters for each language direction based on the target pruning rate and the absolute value ranking. The present application can dynamically adjust the pruning rate for each direction based on the consistency between the gradients for each direction and the average gradient to achieve parameter allocation optimization for each direction, effectively promote positive cross-language knowledge transfer, inhibit negative transfer, and thereby improve the overall translation performance of the model.

Images