METHOD AND SERVER FOR TRAINING MACHINE LEARNING ALGORITHM FOR TRANSLATION Russian patent published in 2023 - IPC G06F40/40 G06N3/02 

FIELD: machine learning.

SUBSTANCE: invention relates to a method and an electronic device for offline translation of an initial word into a target word. In the method, offline translation is performed by a neural network (hereinafter – NN) executed by the electronic device. In this case, NN is trained for translation of words from an initial language to a target language, wherein NN has an encoding part and a decoding part. In the method, the initial word is divided by the electronic device into an input token sequence; a vector representation sequence is generated by the electronic device using the encoding part for corresponding ones of the input token sequence; the first output token sequence is generated by the electronic device using the decoding part based on the vector representation sequence, the first sequence is the first candidate word, wherein the first candidate word is possible translation of the initial word into the target language. In response to inconsistence of the first candidate word with at least one predetermined rule: the decoding part is actuated for generation of the second output token sequence, wherein the second sequence has at least one last output token different from at least one last output token of the first sequence, the second sequence is the second candidate word, wherein the second candidate word is the second possible translation of the initial word into the target language. In response to consistence of the second candidate word with the mentioned at least one predetermined rule, it is determined by the electronic device that the second candidate word is the target word.

EFFECT: increase in the accuracy of machine translation.

22 cl, 4 dwg