Deep model statistics method for machine translation

• US 8,145,473 B2
• Filed: 02/18/2009
• Issued: 03/27/2012
• Est. Priority Date: 10/10/2006
• Status: Expired due to Fees

• Current Assignee: ABBYY Production LLC (ABBYY Software)
• Original Assignee: ABBYY Software
• Inventors: Zuev, Konstantin, Selegey, Vladimir, Tuganbaev, Diar, Anisimovich, Konstantin
• Primary Examiner(s): Vo, Huyen X.
• Application Number:
  US12/388,219
• Publication Number:
  US 20090182549A1
• Time in Patent Office: 1,133 Days
• Field of Search: 704 1- 10, 704/277, 704/257, 704251-255
• US Class Current: 704/4
• CPC Class Codes: G06F 40/268 Morphological analysis
• CPC Class Codes: G06F 40/284 Lexical analysis, e.g. tokenisation or collocates
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• CPC Class Codes: G06F 40/55 Rule-based translation
• CPC Class Codes: G06F 40/58 Use of machine translation, e.g. for multi-lingual retrieval, for server-side translation for client devices or for real-time translation

Abstract

In one embodiment, the invention provides a method for machine translation of a source document in an input language to a target document in an output language, comprising generating translation options corresponding to at least portions of each sentence in the input language; and selecting a translation option for the sentence based on statistics associated with the translation options.

Claims

1. A method for machine translation of a source document in an input language to a target document in an output language, the method comprising:

creating a language-independent semantic structure for at least portions of each sentence in the source document;

generating by a processor translation options corresponding to at least portions of each sentence in the input language, wherein statistics are used in generating the translation options, wherein said generating the translation options includes;

performing a syntactic analysis to determine surface elements associated with the sentence in the input language, and

performing a semantic analysis to determine deep elements associated with the sentence in the input language;

selecting a translation option for the said at least portions of each sentence based on ratings associated with the translation options, wherein the ratings include a use of statistics that can include a combinatorial rating for the surface elements associated with the sentence in the input language and the deep elements associated with the sentence in the input language, wherein the combinatorial rating includes a probability of objects of certain semantic classes being combined with objects of a same or another semantic class; and

synthesizing one or more output sentences corresponding to the said at least portions of each sentence in the source document based at least in part on a respective language-independent semantic structure and respective selected translation option.

2. The method of claim 1, wherein the surface elements are selected from a group including lexemes, word combinations, lexical meanings and surface slots.

3. The method of claim 1, wherein the deep elements are selected from a group including semantic classes and deep slots.

4. The method of claim 1, wherein the translation options comprise lexical choices in the output language corresponding to said portions of each sentence in the input language.

5. The method of claim 1, wherein the combinatorial rating includes a probability of objects of certain semantic classes being combined with certain deep slots.

6. The method of claim 1, further comprising generating the statistics, and wherein the statistics include deep language-independent statistics of “

semantic class-semantic class”

correspondences.

7. The method of claim 6, wherein generating the statistics includes analyzing an assortment of texts to identify frequently occurring lexemes and lexeme combinations.

8. The method of claim 7, wherein the assortment of texts is selected from a group including parallel texts, tagged texts and domain-specific texts.

9. A physical, non-transitory computer storage medium having stored thereon a sequence of instructions which when executed by a computer system cause said computer system to perform a method for machine translation of a source document in an input language to a target document in an output language, the instructions comprising:

creating a language-independent semantic structure for each sentence in the source document;

generating by a processor of the computer system translation options corresponding to at least portions of each respective sentence in the input language, wherein said generating the translation options is based at least in part on statistics, and wherein said generating the translation options includes;

performing a syntactic analysis to determine surface elements associated with the sentence in the input language, and

performing a semantic analysis to determine deep elements associated with the sentence in the input language;

selecting a translation option for one or more sentences based on ratings associated with the translation options, wherein the ratings associated with the translation options include the use of statistics, wherein the statistics include the use of a combinatorial rating for the surface elements associated with the sentence in the input language and the deep elements associated with the sentence in the input language, wherein the combinatorial rating includes a use of a probability of objects of certain semantic classes being combined with objects of a same or another semantic class;

synthesizing an output sentence for the target document corresponding to one or more sentences in the source document based at least in part on one or more respective language-independent semantic structures and a selected translation option; and

writing the output sentence for the target document to a component of the computing system.

10. The physical, non-transitory computer storage medium of claim 9, wherein the surface elements are selected from a group including lexemes, word combinations, lexical meanings and surface slots.

11. The physical, non-transitory computer storage medium of claim 9, wherein the deep elements associated with the sentence in the input language are selected from a group including semantic classes and deep slots.

12. The physical, non-transitory computer storage medium of claim 9, wherein the translation options include lexical choices in the output language corresponding to said portions of each sentence in the input language.

13. The physical, non-transitory computer storage medium of claim 9, wherein the combinatorial rating includes a probability of objects of certain semantic classes being combined with certain deep slots.

14. The physical, non-transitory computer storage medium of claim 9, further comprising generating the statistics, and wherein the statistics include deep language-independent statistics of “

semantic class-semantic class”

correspondences.

15. The physical, non-transitory computer storage medium of claim 9, wherein generating the statistics includes analyzing an assortment of texts to identify frequently occurring lexemes and lexeme combinations.

16. The physical, non-transitory computer storage medium of claim 15, wherein the assortment of texts is selected from a group including parallel texts, tagged texts and domain-specific texts.

17. A computer system comprising:

a processor; and

a storage medium coupled to the processor, the storage medium storing instructions which when executed by the processor cause the computer system to perform a method for machine translation of a source document in an input language to a target document in an output language, the instructions comprising;

creating a language-independent semantic structure for one or more sentences in the source document;

generating translation options corresponding to at least portions of each sentence in the input language, wherein said generating the translation options is based at least in part on a use of statistics, and wherein said generating the translation options includes;

performing a syntactic analysis to determine surface elements associated with the sentence in the input language, and

performing a semantic analysis to determine deep elements associated with the sentence in the input language;

selecting a translation option for the said at least portions of each sentence based on ratings associated with the translation options, wherein the ratings associated with the translation options include the use of statistics, wherein said statistics include the use of a combinatorial rating for the surface elements associated with the sentence in the input language and the deep elements associated with the sentence in the input language, wherein the combinatorial rating includes a use of a probability of objects of certain semantic classes being combined with objects of a same or another semantic class;

synthesizing one or more output sentences corresponding to the said at least portions of each sentence in the source document based at least in part on one or more respective language-independent semantic structures and respective selected translation option; and

writing the target document to a component of the computer system.

18. The computer system of claim 17, wherein the surface elements are selected from a group including lexemes, word combinations, lexical meanings and surface slots.

19. The computer system of claim 17, wherein the deep elements are selected from a group including semantic classes and deep slots.

20. The computer system of claim 17, wherein the translation options include lexical choices in the output language corresponding to said portions of each sentence in the input language.

21. The computer system of claim 17, wherein the combinatorial rating includes a probability of objects of certain semantic classes being combined with certain deep slots.

22. The computer system of claim 17, further comprising generating the statistics, and wherein the statistics include deep language-independent statistics of “

semantic class-semantic class”

correspondences.

23. The computer system of claim 17, wherein generating the statistics includes analyzing an assortment of texts to identify frequently occurring lexemes and lexeme combinations.

24. The computer system of claim 23, wherein the assortment of texts is selected from a group including parallel texts, tagged texts and domain-specific texts.