METHOD AND SERVER FOR TRAINING MACHINE LEARNING ALGORITHM IN TRANSLATION Russian patent published in 2022 - IPC G06N20/00 

FIELD: computer technology.

SUBSTANCE: MLA algorithm is trained using the first set of string pairs. The first set of string pairs has a natural proportion of string pairs of each context. MLA algorithm turns out to be biased towards forming a parallel string as a translation of the corresponding string found in the main context. A method includes defining the second set of string pairs containing a controlled proportion of string pairs of each context. The second set of string pairs is associated with labels pointing to corresponding contexts. The method includes re-training of MLA algorithm using the second set of string pairs and corresponding labels. MLA algorithm is re-trained to determine the context of the string and create the corresponding parallel string as a translation, taking into account the context.

EFFECT: reduction in the bias of MLA algorithm that arose during the initial training in translation of a text string in the first language with a parallel text string in the second language.

20 cl, 6 dwg