Granted Patent B2

US 12,614,043 · App. 18/255,790 · Granted Apr 28, 2026

Inventors: Chengqi Zhao (Beijing, CN); Jianze Liang (Beijing, CN); Mingxuan Wang (Beijing, CN); Lei Li (Beijing, CN)

Assignee: BEIJING YOUZHUJU NETWORK TECHNOLOGY CO., LTD.

G06F40/58G06F40/51

Abstract

Claims (60)

1 . A model training method, comprising:

determining, through a neural network pruning technique, a respective influence degree of each parameter in a plurality of parameters in a first translation model on a translation result in a first field to obtain at least one first parameter and at least one second parameter, wherein an influence degree of a first parameter on the translation result in the first field is greater than an influence degree of a second parameter on the translation result in the first field, the first translation model is a generic translation model, and the first field is a generic field;

training, by using a first corpus of the first field, the at least one first parameter of the first translation model to obtain a second translation model, wherein the at least one first parameter is updated and the at least one second parameter is unchanged during training, and similarity between a translation result of the second translation model in the first field and a translation result of the first translation model in the first field meets a preset condition; and

training, by using a target corpus in a target field, a target parameter of the second translation model to obtain a target translation model, wherein the target parameter is updated and other parameters except the target parameter in the second translation model are unchanged during training, the target parameter is part or all of the at least one second parameter, and the target field is a specific field.

2 . The method of claim 1 , wherein determining, through the neural network pruning technique, the respective influence degree of the each parameter in the plurality of parameters in the first translation model on the translation result in the first field to obtain the at least one first parameter and the at least one second parameter, comprises:

inputting a second corpus of the first field into the first translation model to obtain a respective influence factor of the each parameter in the first translation model on a translation result;

determining, according to the respective influence factor corresponding to the each parameter, a respective influence degree of the each parameter on the translation result; and

recording a parameter whose influence degree is greater than or equal to a set threshold as a first parameter and recording a parameter whose influence degree is less than the set threshold as a second parameter.

3 . The method of claim 1 , wherein training, by using the first corpus of the first field, the at least one first parameter of the first translation model to obtain the second translation model, comprises:

acquiring a first source language text in the first field and a first reference translated text corresponding to the first source language text;

inputting the first source language text into the first translation model to obtain an output result of the first translation model; and

determining a matching degree between the output result and the first reference translated text;

in a case where the matching degree is less than a first set threshold, adjusting the at least one first parameter of the first translation model until the matching degree of the output result and the first reference translated text is greater than or equal to the first set threshold, and recording the first translation model corresponding to the matching degree greater than or equal to the first set threshold as the second translation model.

4 . The method of claim 1 , wherein training, by using the target corpus in the target field, the target parameter of the second translation model to obtain the target translation model, comprises:

acquiring a second source language text in the target field and a second reference translated text corresponding to the second source language text;

inputting the second source language text into the second translation model to obtain an output result of the second translation model; and

determining a matching degree between the output result and the second reference translated text; in a case where the matching degree is less than a second set threshold, adjusting the target parameter of the second translation model until the matching degree of the output result and the second reference translated text is greater than or equal to the second set threshold, and recording the second translation model corresponding to the matching degree greater than or equal to the second set threshold as the target translation model.

5 . A non-transitory computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to implement the model training method of claim 1 .

6 . A machine translation method, comprising:

acquiring a source language text to be translated; and

inputting the source language text into a preset machine translation model, and acquiring an output result of the preset machine translation model as a translation result of the source language text, wherein the preset machine translation model is trained by a model training method, and the model training method comprises:

determining, through a neural network pruning technique, a respective influence degree of each parameter in a plurality of parameters in a first translation model on a translation result in a first field to obtain at least one first parameter and at least one second parameter, wherein an influence degree of a first parameter on the translation result in the first field is greater than an influence degree of a second parameter on the translation result in the first field, the first translation model is a generic translation model, and the first field is a generic field;

training, by using a first corpus of the first field, the at least one first parameter of the first translation model to obtain a second translation model, wherein the at least one first parameter is updated and the at least one second parameter is unchanged during training, and similarity between a translation result of the second translation model in the first field and a translation result of the first translation model in the first field meets a preset condition; and

training, by using a target corpus in a target field, a target parameter of the second translation model to obtain a target translation model, wherein the target parameter is updated and other parameters except the target parameter in the second translation model are unchanged during training, the target parameter is part or all of the at least one second parameter, and the target field is a specific field.

7 . The method of claim 6 , wherein determining, through the neural network pruning technique, the respective influence degree of the each parameter in the plurality of parameters in the first translation model on the translation result in the first field to obtain the at least one first parameter and the at least one second parameter, comprises:

inputting a second corpus of the first field into the first translation model to obtain a respective influence factor of the each parameter in the first translation model on a translation result;

determining, according to the respective influence factor corresponding to the each parameter, a respective influence degree of the each parameter on the translation result; and

recording a parameter whose influence degree is greater than or equal to a set threshold as a first parameter and recording a parameter whose influence degree is less than the set threshold as a second parameter.

8 . The method of claim 6 , wherein training, by using the first corpus of the first field, the at least one first parameter of the first translation model to obtain the second translation model, comprises:

acquiring a first source language text in the first field and a first reference translated text corresponding to the first source language text;

inputting the first source language text into the first translation model to obtain an output result of the first translation model; and

determining a matching degree between the output result and the first reference translated text;

in a case where the matching degree is less than a first set threshold, adjusting the at least one first parameter of the first translation model until the matching degree of the output result and the first reference translated text is greater than or equal to the first set threshold, and recording the first translation model corresponding to the matching degree greater than or equal to the first set threshold as the second translation model.

9 . The method of claim 6 , wherein training, by using the target corpus in the target field, the target parameter of the second translation model to obtain the target translation model, comprises:

acquiring a second source language text in the target field and a second reference translated text corresponding to the second source language text;

inputting the second source language text into the second translation model to obtain an output result of the second translation model; and

determining a matching degree between the output result and the second reference translated text;

in a case where the matching degree is less than a second set threshold, adjusting the target parameter of the second translation model until the matching degree of the output result and the second reference translated text is greater than or equal to the second set threshold, and recording the second translation model corresponding to the matching degree greater than or equal to the second set threshold as the target translation model.

10 . A non-transitory computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to implement the machine translation method of claim 6 .

11 . An electronic device, comprising:

one or more processors; and

a memory, which is configured to store one or more programs;

wherein the one or more programs, when executed by the one or more processors, perform steps, wherein the steps comprise:

determining, through a neural network pruning technique, a respective influence degree of each parameter in a plurality of parameters in a first translation model on a translation result in a first field to obtain at least one first parameter and at least one second parameter, wherein an influence degree of a first parameter on the translation result in the first field is greater than an influence degree of a second parameter on the translation result in the first field, the first translation model is a generic translation model, and the first field is a generic field;

training, by using a first corpus of the first field, the at least one first parameter of the first translation model to obtain a second translation model, wherein the at least one first parameter is updated and the at least one second parameter is unchanged during training, and similarity between a translation result of the second translation model in the first field and a translation result of the first translation model in the first field meets a preset condition; and

training, by using a target corpus in a target field, a target parameter of the second translation model to obtain a target translation model, wherein the target parameter is updated and other parameters except the target parameter in the second translation model are unchanged during training, the target parameter is part or all of the at least one second parameter, and the target field is a specific field.

12 . The electronic device of claim 11 , wherein determining, through the neural network pruning technique, the respective influence degree of the each parameter in the plurality of parameters in the first translation model on the translation result in the first field to obtain the at least one first parameter and the at least one second parameter, comprises:

inputting a second corpus of the first field into the first translation model to obtain a respective influence factor of the each parameter in the first translation model on a translation result;

determining, according to the respective influence factor corresponding to the each parameter, a respective influence degree of the each parameter on the translation result; and

recording a parameter whose influence degree is greater than or equal to a set threshold as a first parameter and recording a parameter whose influence degree is less than the set threshold as a second parameter.

13 . The electronic device of claim 11 , wherein training, by using the first corpus of the first field, the at least one first parameter of the first translation model to obtain the second translation model, comprises:

acquiring a first source language text in the first field and a first reference translated text corresponding to the first source language text;

inputting the first source language text into the first translation model to obtain an output result of the first translation model; and

determining a matching degree between the output result and the first reference translated text;

in a case where the matching degree is less than a first set threshold, adjusting the at least one first parameter of the first translation model until the matching degree of the output result and the first reference translated text is greater than or equal to the first set threshold, and recording the first translation model corresponding to the matching degree greater than or equal to the first set threshold as the second translation model.

14 . The electronic device of claim 11 , wherein training, by using the target corpus in the target field, the target parameter of the second translation model to obtain the target translation model, comprises:

acquiring a second source language text in the target field and a second reference translated text corresponding to the second source language text;

inputting the second source language text into the second translation model to obtain an output result of the second translation model; and

determining a matching degree between the output result and the second reference translated text;

in a case where the matching degree is less than a second set threshold, adjusting the target parameter of the second translation model until the matching degree of the output result and the second reference translated text is greater than or equal to the second set threshold, and recording the second translation model corresponding to the matching degree greater than or equal to the second set threshold as the target translation model.

Assignments (2)

Priority Claims (1)

Continuity (1)

References Cited (27)