Information conversion method and apparatus, storage medium, and electronic device

Patent metadata

Field	Value
Publication number	US-11710003-B2
Application number	US-202016890861-A
Country	US
Kind code	B2
Filing date	Jun 2, 2020
Priority date	Feb 26, 2018
Publication date	Jul 25, 2023
Grant date	Jul 25, 2023

Abstract

Embodiments of this application include an information conversion method for translating source information. The source information is encoded to obtain a first code. A preset conversion condition is obtained. The preset conversion condition indicates a mapping relationship between the source information and a conversion result. The first code is decoded according to the source information, the preset conversion condition, and translated information to obtain target information. The target information and the source information are in different languages. Further, the translated information includes a word obtained through conversion of the source information into a language of the target information.

First claim

What is claimed is:

1. An information conversion method for translating source information, comprising:
encoding, by processing circuitry with an encoder of a neural network, the source information to obtain a first code;
obtaining a preset conversion condition of a phrase in the source information from a plurality of preset conversion conditions each representing a different association between words, the preset conversion condition being selected from the plurality of preset conversion conditions based on respective weights indicating similarities between sentence patterns of each of the plurality of preset conversion conditions and a sentence pattern of the source information, wherein each weight of a respective one of the plurality of preset conversion conditions is calculated based on a first vector of the source information, a second vector of previously translated information, and a third vector of the respective preset conversion condition, the selected preset conversion condition having a highest weight among the respective weights of the plurality of preset conversion conditions and defining an association between the words of the phrase in the source information; and
decoding, by the processing circuitry with the encoder of the neural network, the first code according to the first vector, the second vector, and the third vector to obtain target information, the target information and the source information being in different languages, and the previously translated information including previously translated words obtained through conversion into a language of the target information.

2. The method according to claim 1 , wherein
the decoding the first code includes determining the similarities in the sentence patterns between each of the plurality of preset conversion conditions and the source information according to the source information and the previously translated information; and
the selected preset conversion condition has a highest similarity among the similarities in the sentence patterns with the source information.

3. The method according to claim 1 , further comprising determining the weight for each of the plurality of preset conversion conditions by:
obtaining a weight value a t of each of the plurality of preset conversion conditions based on a t =S(Uh t +Ws i ), h t representing a vector of a t th condition, s i representing the first vector and the third vector, S representing an S-shaped growth curve, and U and W being matrices, respectively.

4. The method according to claim 1 , wherein the obtaining the preset conversion condition comprises:
obtaining a plurality of pieces of preset discontinuous information, the discontinuous information indicating that a phrase includes at least two discontinuous parts with associations.

5. The method according to claim 1 , wherein
the first code is the first vector; and
the decoding the first code includes decoding, according to the source information, the preset conversion condition, and the previously translated information, the first vector by using the encoder of the neural network, to obtain the target information, the preset conversion condition being obtained in advance by the encoder of the neural network.

6. The method according to claim 1 , wherein the selected preset conversion condition includes a rule for converting a predetermined combination of variables and words in the source information.

7. An information conversion apparatus comprising:
processing circuitry configured to
encode, with an encoder of a neural network, source information to obtain a first code;
obtain a preset conversion condition of a phrase in the source information from a plurality of preset conversion conditions each representing a different association between words, the preset conversion condition being selected from the plurality of preset conversion conditions based on respective weights indicating similarities between sentence patterns of each of the plurality of preset conversion conditions and a sentence pattern of the source information, wherein each weight of a respective one of the plurality of preset conversion conditions is calculated based on a first vector of the source information, a second vector of previously translated information, and a third vector of the respective preset conversion condition, the selected preset conversion condition having a highest weight among the respective weights of the plurality of preset conversion conditions and defining an association between the words of the phrase in the source information; and
decode the first code with the encoder of the neural network according to the first vector, the second vector, and the third vector to obtain target information, the target information and the source information being in different languages, and the previously translated information including previously translated words obtained through conversion into a language of the target information.

8. The apparatus according to claim 7 , wherein
the processing circuitry is configured to determine the similarities in the sentence patterns between each of the plurality of preset conversion conditions and the source information according to the source information and the previously translated information; and
the selected preset conversion condition has a highest similarity among the similarities in the sentence patterns with the source information.

9. The apparatus according to claim 7 , wherein the processing circuitry is configured to
obtain a weight value a t of each of the plurality of preset conversion conditions based on a t =S(Uh t +Ws i ), h t representing a vector of a t th condition, s i representing the first vector and the third vector, S representing an S-shaped growth curve, and U and W being matrices, respectively.

10. The apparatus according to claim 7 , wherein the processing circuitry is configured to
obtain a plurality of pieces of preset discontinuous information, the discontinuous information indicating that a phrase includes at least two discontinuous parts with associations.

11. The apparatus according to claim 7 , wherein
the first code is the first vector; and
the processing circuitry is configured to decode, according to the source information, the preset conversion condition, and the previously translated information, the first vector by using the encoder of the neural network, to obtain the target information, the preset conversion condition being obtained in advance by the encoder of the neural network.

12. The apparatus according to claim 7 , wherein the selected preset conversion condition includes a rule for converting a predetermined combination of variables and words in the source information.

13. A non-transitory computer-readable storage medium storing instructions which when executed by a processor cause the processor to perform:
encoding, with an encoder of a neural network, source information to obtain a first code;
obtaining a preset conversion condition of a phrase in the source information from a plurality of preset conversion conditions each representing a different association between words, the preset conversion condition being selected from the plurality of preset conversion conditions based on respective weights indicating similarities between sentence patterns of each of the plurality of preset conversion conditions and a sentence pattern of the source information, wherein each weight of a respective one of the plurality of preset conversion conditions is calculated based on a first vector of the source information, a second vector of previously translated information, and a third vector of the respectiv

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Wang Mingxuan
• Xie Jun
• Yao Jian
• Huang Jiangquan

Classifications

• G06F40/58 Primary

  Use of machine translation, e.g. for multi-lingual retrieval, for server-side translation for client devices or for real-time translation · CPC title

• G06F40/151 Primary

  Transformation · CPC title

• G06F40/44 Primary

  Statistical methods, e.g. probability models · CPC title

• G06F40/126

  Character encoding · CPC title