US 11,734,521 B2
Method, device and storage medium for training bidirectional machine translation model using sum of forward and reverse translation similarities
Jialiang Jiang, Beijing (CN); Xiang Li, Beijing (CN); and Jianwei Cui, Beijing (CN)
Assigned to BEIJING XIAOMI MOBILE SOFTWARE CO., LTD., Beijing (CN)
Filed by BEIJING XIAOMI MOBILE SOFTWARE CO., LTD., Beijing (CN)
Filed on May 8, 2020, as Appl. No. 16/870,632.
Claims priority of application No. 201911259415.X (CN), filed on Dec. 10, 2019.
Prior Publication US 2021/0174019 A1, Jun. 10, 2021
Int. Cl. G06F 40/263 (2020.01); G06N 3/08 (2023.01); G06F 40/51 (2020.01); G06F 40/58 (2020.01); G06N 3/02 (2006.01); G10L 15/26 (2006.01); G06F 40/44 (2020.01); G06F 18/214 (2023.01)
CPC G06F 40/44 (2020.01) [G06F 18/214 (2023.01); G06F 40/263 (2020.01); G06N 3/08 (2013.01)] 15 Claims
OG exemplary drawing
 
1. A method for training a machine translation model, comprising:
acquiring a bidirectional translation model to be trained and training data, the training data comprising a source corpus and a target corpus corresponding to the source corpus;
training the bidirectional translation model for N cycles, each cycle of training comprising a forward translation process of translating the source corpus into a pseudo target corpus by the machine translation model and a reverse translation process of translating the pseudo target corpus into a pseudo source corpus by the machine translation model, and N being a positive integer greater than 1;
acquiring a forward translation similarity and a reverse translation similarity, the forward translation similarity being a similarity between the target corpus and the pseudo target corpus, and the reverse translation similarity being a similarity between the source corpus and the pseudo source corpus; and
when a sum of the forward translation similarity and the reverse translation similarity converges, determining that training of the bidirectional translation model is completed, wherein the sum of the forward translation similarity and the reverse translation similarity converges indicates the sum of the forward translation similarity and the reverse translation similarity approaches a value,
wherein the training the bidirectional translation model for N cycles comprises:
acquiring, in the forward translation process, the pseudo target corpus through a differentiable sampling function;
acquiring, in an ith cycle of training, an error between the target corpus and the pseudo target corpus through the differentiable sampling function, i being a positive integer greater than or equal to 1 and less than N; and
regulating, in the (i+1)th cycle of training, one or more training parameters of the bidirectional translation model based on the error acquired in the ith cycle of training.