Machine translation method and apparatus with joint optimization of translation model and partitioning model, electronic device, and storage medium

Patent metadata

Field	Value
Publication number	US-11995415-B2
Application number	US-202117403857-A
Country	US
Kind code	B2
Filing date	Aug 16, 2021
Priority date	Jul 8, 2019
Publication date	May 28, 2024
Grant date	May 28, 2024

Abstract

A machine translation method includes: receiving a sentence, the sentence including a plurality of words; calling a machine translation model obtained through training, the machine translation model including a partitioning model and a translation model; partitioning the sentence based on the partitioning model and according to word vectors of the words, to obtain to-be-translated blocks, each to-be-translated block including at least one of the words; and translating the sentence based on the translation model and the to-be-translated blocks, to obtain a translation result.

First claim

What is claimed is:

1. A machine translation method, performed by an electronic device, the method comprising:
receiving a sentence, the sentence comprising a plurality of words;
calling a machine translation model obtained through training, the machine translation model comprising a partitioning model and a translation model and being obtained by:
selecting a training corpus from a parallel sentence library, the parallel sentence library comprising parallel sentence pairs, each parallel sentence pair comprising a source sentence and a target sentence;
training an initial translation model by using the training corpus to obtain a trained translation model;
training an initial partitioning model by using the training corpus and based on the trained translation model to obtain a trained partitioning model;
obtaining a translation expected value of the trained translation model and a partitioning expected value of the trained partitioning model;
obtaining an overall model expected value of the machine translation model according to the translation expected value and the partitioning expected value; and
performing joint training on the trained translation model and the trained partitioning model according to an optimal principle of the overall model expected value to obtain the machine translation model;

partitioning the sentence based on the partitioning model and according to word vectors of the words to obtain to-be-translated blocks, each to-be-translated block comprising at least one of the words; and
translating the sentence based on the translation model and the to-be-translated blocks to obtain a translation result.

2. The machine translation method according to claim 1 , wherein after receiving the sentence, the method further comprises:
preprocessing the sentence to obtain a preprocessed text; and
performing word segmentation on the preprocessed text to obtain the words corresponding to the sentence.

3. The machine translation method according to claim 2 , wherein partitioning the sentence based on the partitioning model and according to the word vectors of the words to obtain the to-be-translated blocks comprises:
obtaining a word representation of a current word and a block representation of a current to-be-translated block according to the word vectors of the words;
obtaining an attribution relationship between the current word and the current to-be-translated block based on the partitioning model and according to the word representation of the current word and the block representation of the current to-be-translated block; and
partitioning the sentence based on the partitioning model and the attribution relationship to obtain the to-be-translated blocks.

4. The machine translation method according to claim 2 , wherein after performing the word segmentation on the preprocessed text to obtain the words corresponding to the sentence, the method further comprises:
obtaining association relationships between the words; and
combining words having an association relationship into one word.

5. The machine translation method according to claim 1 , wherein training the initial partitioning model by using the training corpus and based on the trained translation model to obtain the trained partitioning model comprises:
partitioning a source sentence in the training corpus based on the initial partitioning model to obtain blocks of the source sentence;
translating the source sentence based on the blocks of the source sentence and the trained translation model to obtain a translated sentence corresponding to the source sentence;
obtaining a partitioning expected value of the initial partitioning model according to the translated sentence corresponding to the source sentence and the target sentence; and
training the initial partitioning model according to an optimal principle of the partitioning expected value to obtain the trained partitioning model.

6. The machine translation method according to claim 5 , wherein before partitioning the source sentence in the training corpus based on the initial partitioning model to obtain the blocks of the source sentence, the method further comprises:
performing word alignment on the parallel sentence pair in the training corpus;
determining candidate partitioning positions according to an alignment result obtained after the word alignment; and
partitioning the source sentence in the training corpus based on the initial partitioning model and the candidate partitioning positions to obtain the blocks of the source sentence.

7. The machine translation method according to claim 5 , wherein obtaining the partitioning expected value of the initial partitioning model according to the translated sentence corresponding to the source sentence and the target sentence comprises:
obtaining a likelihood parameter of the translated sentence corresponding to the source sentence and the target sentence; and
obtaining the partitioning expected value of the initial partitioning model according to the likelihood parameter.

8. The machine translation method according to claim 5 , wherein obtaining the partitioning expected value of the initial partitioning model according to the translated sentence corresponding to the source sentence and a target sentence further comprises:
obtaining a block quantity of the source sentence and a word quantity of the source sentence;
obtaining a penalty parameter according to the block quantity and the word quantity; and
obtaining the partitioning expected value of the initial partitioning model according to the likelihood parameter and the penalty parameter.

9. The machine translation method according to claim 5 , wherein the partitioning model comprises a similarity function; and partitioning the source sentence in the training corpus based on the initial partitioning model to obtain the blocks of the source sentence comprises:
obtaining a block representation of a current block and a word representation of a current word that are in the source sentence;
obtaining a similarity between the current word and the current block based on the similarity function;
determining an attribution relationship between the current word and the current block according to the similarity; and
partitioning the source sentence according to the attribution relationship to obtain the blocks of the source sentence.

10. An electronic device, comprising a memory and a processor, the memory storing computer-readable instructions, the computer-readable instructions, when executed by the processor, causing the processor to perform a plurality of operations comprising:
receiving a sentence, the sentence comprising a plurality of words;
calling a machine translation model obtained through training, the machine translation model comprising a partitioning model and a translation model and being obtained by:
selecting a training corpus from a parallel sentence library, the parallel sentence library comprising parallel sentence pairs, each parallel sentence pair comprising a source sentence and a target sentence;
training an initial translation model by using the training corpus to obtain a trained translation model;
training an initial partitioning model by using the training corpus and based on the trained translation model to obtain a trained partitioning model;
obtaining a translation expected value of the trained translation model and a partitioning expected value of the trained partitioning model;
obtaining an overall model expected value of the machine translation model according to the translation expected value and the partitioning expected value; and
performing joint training on the trained translation model and the trained partitioning

Assignees

• Tencent Tech Shenzhen Co Ltd

Inventors

• Xie Jun

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

• G06F40/284 Primary

  Lexical analysis, e.g. tokenisation or collocates · CPC title

• G06F40/289

  Phrasal analysis, e.g. finite state techniques or chunking · CPC title

• G06F40/42

  Data-driven translation · CPC title

• G06F40/45

  Example-based machine translation; Alignment · CPC title