Selection and use of nonstatistical translation components in a statistical machine translation framework

• US 20060015320A1
• Filed: 04/15/2005
• Published: 01/19/2006
• Est. Priority Date: 04/16/2004
• Status: Active Grant

• Current Assignee: University of Southern California
• Original Assignee: University of Southern California
• Inventors: Och, Franz Josef
• Granted Patent: US 8,666,725 B2
• Time in Patent Office: Days
• US Class Current: 704/2
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models

Abstract

A system with a nonstatistical translation component integrated with a statistical translation component engine. The same corpus may be used for training the statistical engine and also for determining when to use the statistical engine and when to use the translation component. This training may use probabilistic techniques. Both the statistical engine and the translation components may be capable of translating the same information, however the system determines which component to use based on the training. Retraining can be carried out to add additional components, or when after additional translator training.

Claims

1. A method comprising:

training a machine translation system when to use a statistical translator and when to use a nonstatistical translator based on the same training corpus.

2. A method as in claim 1, further comprising selecting whether to use the statistical translator to translate at least a portion of an unknown text, or whether to use the non-statistical translator to train another portion of the unknown text, said selecting whether to use the statistical translator or the non-statistical translator also being based on said training.

3. A method as in claim 1, wherein said training comprises:

determining a plurality of first phrases which can be translated by the non statistical translator;

testing a translation of said first phrases to determine if said non statistical translator has properly translated said first phrases; and

using information from said testing to train a classifier when to use said non statistical translator.

4. A method as in claim 3, wherein said non statistical translator is a translation component for proper nouns.

5. A method as in claim 3, wherein said non-statistical translator is a translation components for names.

6. A method as in claim 3, wherein said non-statistical translator is a translation component for numbers.

7. A device as in claim 2, wherein said testing comprises detecting variants of a translated phrase, and accepting said phrase as being a proper translation if it is one of said variants.

8. The method as in claim 3, further comprising annotating a training corpus based on said testing to form an annotated training corpus, with annotations that represent results from said testing.

9. The method as in claim 3, wherein said classifier is a probabilistic classifier.

10. A method as in claim 1, further comprising retraining when to use the statistical translator and the nonstatistical translator, responsive to an occurrence.

11. The method as in claim 10, wherein said action comprises adding an additional nonstatistical translator component.

12. The method as in claim 10, wherein said action comprises improving a translator component.

13. A method as in claim 1, further comprising training a format of an output of said machine translation system, based on said training corpus, and allowing selection of one of a plurality of different formats within said training corpus.

14. A method comprising:

translating information from a first language to a second language, using at least first and second components that are each capable of translating the same phrases; and

automatically selecting the component among said first and second components, that provide a translation with a higher probability of being a correct translation.

15. A method as in claim 14, further comprising defining a feature function that indicates when to use said first and second translation components.

16. A method as in claim 14, wherein said automatically selecting comprises:

obtaining a phrase to be translated;

computing a probability that each of a plurality of different components will produce the best translation of the phrase; and

using one of the plurality of different components, based on said computing a probability.

17. A method as in claim 14, wherein said first component includes a statistical translator, and said second component includes a non statistical translator component for proper nouns.

18. A method as in claim 14, wherein said first component includes a statistical translator, and said second component includes a non statistical translator component for numbers.

19. A system comprising:

a statistical translation system, operating based on translation data;

a translation component, formed of a non-statistical translator; and

a classifier that determines when to use said statistical translator and when to use said nonstatistical translator to translate a first phrase, when both said statistical translation system and said translation component are each capable of translating said first phrase.

20. A system as in claim 19, further comprising a training corpus for said statistical translator and for said nonstatistical translator and also for said classifier.

21. A system as in claim 19, further comprising a training element which determines a plurality of phrases which can be translated by the non statistical translator, tests a translation of said phrases to determine if said non statistical translator has properly translated each phrase, and uses information from said testing to train said classifier when to use said non statistical translator.

22. A system as in claim 19, wherein said non statistical translator is a translation component for proper nouns.

23. A system as in claim 19, wherein said non-statistical translator is a translation component for names.

24. A system as in claim 19, wherein said non-statistical translator is a translation component for numbers.

25. A system as in claim 21, further comprising a variant detector that detects variants of a translated phrase, and accepts said phrase as being a proper translation if it is one of said variants.

26. The system as in claim 20, further comprising annotating the training corpus based on said testing to form an annotated training corpus, with annotations that represent results from said testing.

27. The system as in claim 19, wherein said classifier is a probabilistic classifier.

28. A system as in claim 19, further comprising at least one additional nonstatistical translator component.

29. A system as in claim 20, further comprising an output module that formats an output of said machine translation system, based on said training corpus.

30. A system comprising:

first and second translating parts, each operating to translate information from a first language to a second language, each of said first and second translating parts being capable of translating certain same phrases; and

a classifier that automatically selects the component among said first and second translating parts, that will provide a translation with a higher probability of being a correct translation.

31. A system as in claim 30, further comprising a feature function that indicates when to use different translation components.

32. A system as in claim 30, wherein said first and second translating parts include a statistical translator, and a non statistical translator component.