Granted Patent B2

US 12,373,655 · App. 18/122,316 · Granted Jul 29, 2025

Inventors: Ruiqing Zhang (Beijing, CN); Hui Liu (Beijing, CN); Zhongjun He (Beijing, CN); Zhi Li (Beijing, CN); Hua Wu (Beijing, CN)

Assignee: BEIJING BAIDU NETCOM SCIENCE TECHNOLOGY CO., LTD.

G06F40/47G06F40/51

Abstract

Claims (65)

1. A machine translation method, comprising:

obtaining first target language text by performing first translation on source language text using an initial neural machine translation (NMT) model, wherein the source language text is a first sentence of a first language, and the first target language text is a second sentence of a second language different from the first language, which is obtained by translation of entire first sentence;

identifying an untranslated part in the source language text based on the source language text and the first target language text, comprising: inputting the source language text and the first target language text to a translation missing detection model; outputting by the translation missing detection model identification information being used for identifying whether a text unit in the source language text is untranslated; and identifying the untranslated part in the source language text based on the identification information;

obtaining an adjusted NMT model by increasing an attention weight corresponding to the untranslated part in the initial NMT model; and

obtaining second target language text by performing second translation on the source language text using the adjusted NMT model, wherein the second target language text is a third sentence of the second language,

wherein the translation missing detection model is obtained based on training data, the training data comprises input samples and a training missing tag, and the training data is generated by:

acquiring a first source language sample and a first target language sample translated from the first source language sample;

performing content augmentation on the first target language sample to obtain a second target language sample;

obtaining a second source language sample by reverse translation of the second target language sample;

taking the second source language sample and the first target language sample as the input samples; and

comparing the second source language sample with the first source language sample to determine the training missing tags.

2. The method of claim 1 , wherein the identification information comprises a missing tag corresponding to an untranslated text unit and non-missing tags corresponding to translated text units.

3. The method of claim 1 , wherein the training missing tag corresponds to a text unit included in the second source language sample but not included in the first source language sample.

4. The method of claim 1 , wherein performing content augmentation on the first target language sample to obtain the second target language sample comprises:

inputting the first target language sample to a generative pre-training model, and

outputting, by the generative pre-training model, the second target language sample with more words compared with the first target language sample.

5. The method of claim 1 , wherein performing content augmentation on the first target language sample to obtain the second target language sample comprises:

adding randomly a mask in the first target language sample, and inputting the first target language sample with the mask added into a masked language model; and

outputting, by the masked language model, the second target language sample with the mask being replaced with a text unit.

6. The method of claim 1 , wherein increasing the attention weight corresponding to the untranslated part in the initial NMT model comprises:

determining a maximum attention weight in the initial NMT model;

reducing the maximum attention weight, and determining a difference value of the maximum attention weight before and after the reduction; and

adding the difference value to the attention weight corresponding to the untranslated part.

7. The method of claim 6 , wherein in a case that the untranslated part comprises a plurality of text units, the adding the difference value to the attention weight corresponding to the untranslated part comprises:

determining a to-be-added value corresponding to each of the plural text units based on standard normal distribution and the difference value; and

adding the to-be-added value corresponding to each text unit to the attention weight corresponding to the text unit.

8. An electronic device, comprising:

at least one processor; and

a memory connected with the at least one processor communicatively;

wherein the memory stores instructions executable by the at least one processor to cause the at least one processor to perform a machine translation method comprising:

obtaining first target language text by performing first translation on source language text using an initial neural machine translation (NMT) model, wherein the source language text is a first sentence of a first language, and the first target language text is a second sentence of a second language different from the first language, which is obtained by translation of entire first sentence;

identifying an untranslated part in the source language text based on the source language text and the first target language text, comprising: inputting the source language text and the first target language text to a translation missing detection model; outputting by the translation missing detection model identification information being used for identifying whether a text unit in the source language text is untranslated; and identifying the untranslated part in the source language text based on the identification information;

obtaining an adjusted NMT model by increasing an attention weight corresponding to the untranslated part in the initial NMT model; and

obtaining second target language text by performing second translation on the source language text using the adjusted NMT model, wherein the second target language text is a third sentence of the second language,

wherein the translation missing detection model is obtained based on training data, the training data comprises input samples and a training missing tag, and the training data is generated by:

acquiring a first source language sample and a first target language sample translated from the first source language sample;

performing content augmentation on the first target language sample to obtain a second target language sample;

obtaining a second source language sample by reverse translation of the second target language sample;

taking the second source language sample and the first target language sample as the input samples; and

comparing the second source language sample with the first source language sample to determine the training missing tags.

9. The electronic device of claim 8 , wherein increasing the attention weight corresponding to the untranslated part in the initial NMT model comprises:

determining a maximum attention weight in the initial NMT model;

reducing the maximum attention weight, and determining a difference value of the maximum attention weight before and after the reduction; and

adding the difference value to the attention weight corresponding to the untranslated part.

10. The electronic device of claim 9 , wherein in a case that the untranslated part comprises a plurality of text units, the adding the difference value to the attention weight corresponding to the untranslated part comprises:

determining a to-be-added value corresponding to each of the plural text units based on standard normal distribution and the difference value; and

adding the to-be-added value corresponding to each text unit to the attention weight corresponding to the text unit.

11. A non-transitory computer readable storage medium storing computer instructions for causing a computer to perform a machine translation method comprising:

obtaining first target language text by performing first translation on source language text using an initial neural machine translation (NMT) model, wherein the source language text is a first sentence of a first language, and the first target language text is a second sentence of a second language different from the first language, which is obtained by translation of entire first sentence;

identifying an untranslated part in the source language text based on the source language text and the first target language text, comprising: inputting the source language text and the first target language text to a translation missing detection model; outputting by the translation missing detection model identification information being used for identifying whether a text unit in the source language text is untranslated; and identifying the untranslated part in the source language text based on the identification information;

obtaining an adjusted NMT model by increasing an attention weight corresponding to the untranslated part in the initial NMT model; and

obtaining second target language text by performing second translation on the source language text using the adjusted NMT model, wherein the second target language text is a third sentence of the second language,

wherein the translation missing detection model is obtained based on training data, the training data comprises input samples and a training missing tag, and the training data is generated by:

acquiring a first source language sample and a first target language sample translated from the first source language sample;

performing content augmentation on the first target language sample to obtain a second target language sample;

obtaining a second source language sample by reverse translation of the second target language sample;

taking the second source language sample and the first target language sample as the input samples; and

comparing the second source language sample with the first source language sample to determine the training missing tags.

12. The storage medium of claim 11 , wherein increasing the attention weight corresponding to the untranslated part in the initial NMT model comprises:

determining a maximum attention weight in the initial NMT model;

reducing the maximum attention weight, and determining a difference value of the maximum attention weight before and after the reduction; and

adding the difference value to the attention weight corresponding to the untranslated part.

13. The storage medium of claim 12 , wherein in a case that the untranslated part comprises a plurality of text units, the adding the difference value to the attention weight corresponding to the untranslated part comprises:

determining a to-be-added value corresponding to each of the plural text units based on standard normal distribution and the difference value; and

adding the to-be-added value corresponding to each text unit to the attention weight corresponding to the text unit.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (19)

Cited By (1)

US 12,688,362