DYNAMIC TRANSLATION MEMORY USING STATISTICAL MACHINE TRANSLATION

• US 20100138213A1
• Filed: 12/03/2008
• Published: 06/03/2010
• Est. Priority Date: 12/03/2008
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Bicici, Ergun, Dymetman, Marc
• Granted Patent: US 8,244,519 B2
• Time in Patent Office: Days
• US Class Current: 704/4
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• CPC Class Codes: G06F 40/47 Machine-assisted translation, e.g. using translation memory

Abstract

A translation method comprises: retrieving a fuzzy match text segment translation pair from a translation memory (TM) for an input source language text segment, the fuzzy match text segment translation pair comprising a fuzzy source language text segment having a fuzzy match to the input source language text segment and a corresponding translated target language text segment; extracting from the fuzzy match text segment translation pair an exact match phrase pair comprising a source language phrase that exactly matches a phrase of the input source language text segment and a corresponding translated target language phrase; and invoking a statistical machine translation (SMT) system to generate a proposed translation of the input source language text segment based on a statistical translation model that is enriched by the exact match phrase pair with the exact match phrase pair assigned a high statistical probability.

Claims

1. A translation method comprising:

retrieving at least one fuzzy match text segment translation pair from a translation memory (TM) for an input source language text segment, the fuzzy match text segment translation pair comprising a fuzzy source language text segment having a fuzzy match to the input source language text segment and a corresponding translated target language text segment;

extracting from the at least one fuzzy match text segment translation pair at least one exact match phrase pair comprising a source language phrase that exactly matches a phrase of the input source language text segment and a corresponding translated target language phrase; and

invoking a statistical machine translation (SMT) system to generate a proposed translation of the input source language text segment based on a statistical translation model that is enriched by the at least one exact match phrase pair with the at least one exact match phrase pair assigned a high statistical probability.

2. The translation method as set forth in claim 1, wherein the source language phrase of at least one exact match phrase pair comprises a gapped source language phrase.

3. The translation method as set forth in claim 1, wherein at least one source language phrase of at least one exact match phrase pair comprises a gapped source language phrase in which at least one gap accommodates omission of a plurality of words of the fuzzy source language text segment from which the exact match phrase pair is extracted.

4. The translation method as set forth in claim 1, further comprising:

displaying the proposed translation for review by a human translator.

5. The translation method as set forth in claim 1, wherein the retrieving comprises:

identifying a sub-set of the source language text segments of the TM that have a highest number of words in common with the input source language text segment without regard to word ordering; and

selecting the at least one fuzzy source language text segment from the sub-set as the text segment in the sub-set based on a longest common ordered sub-sequence criterion.

6. The translation method as set forth in claim 5, wherein the longest common ordered sub-sequence criterion allows for at least one gap of less than N words where N is a selected integer greater than or equal to one.

7. The translation method as set forth in claim 5, wherein the longest common ordered sub-sequence of words generated by applying the longest common ordered sub-sequence criterion defines the at least one source language phrase that exactly matches a phrase of the input source language text segment.

8. The translation method as set forth in claim 1, wherein the extracting comprises:

constructing a word alignment matrix for the fuzzy match text segment translation pair; and

selecting the translated target language phrase to maximally align with the source language phrase based on the word alignment matrix.

9. The translation method as set forth in claim 8, wherein the selecting of the translated target language phrase comprises:

identifying the largest element of the word alignment matrix that corresponds to a source language word belonging to the source language phrase;

adding the target language word associated with the identified element of the word alignment matrix to a list of target language words;

repeating the identifying and adding until all source language words of the source language phrase have been associated with target language words; and

defining the translated target language phrase as the list of target language words ordered in the same sequence as in the translated target language text segment.

10. The translation method as set forth in claim 8, wherein the selecting of the translated target language phrase employs greedy selection maximizing individual source language word-target language word alignments based on the word alignment matrix.

11. A translation system comprising:

a statistical machine translation (SMT) system configured to generate a proposed translation in a target language for a source language text segment based on a statistical model comprising bi-phrase translation probabilities;

a translation memory (TM) including a database of source language-target language text segment translation pairs; and

a dynamic translation system configured to invoke the SMT system to generate a proposed translation in the target language for an input source language text segment based on the statistical model enriched by enriching source language-target language translation information that is (i) extracted from the TM and (ii) assigned high statistical probability as compared with the bi-phrase translation probabilities of the statistical model of the SMT system.

12. The translation system as set forth in claim 11, wherein the bi-phrase translation probabilities include translation probabilities for at least some gapped bi-phrases.

13. The translation system as set forth in claim 12, wherein the dynamic translation system is configured to extract the enriching source language-target language translation information by:

retrieving from the TM at least one source language-target language text segment translation pair that has a fuzzy match with the input source language text segment; and

extracting at least one source language-target language phrase translation pair from the at least one source language-target language text segment translation pair for which the source language phrase has an exact match with a corresponding phrase of the input source language text segment.

14. The translation system as set forth in claim 13, wherein the retrieving from the TM at least one source language-target language text segment translation pair that has a fuzzy match with the input source language text segment comprises:

retrieving from the TM a top-N source language-target language text segment translation pairs that have a highest number of source language words in common with the input source language text segment without regard to word ordering; and

performing fuzzy matching to the top-N source language-target language text segment translation pairs using a longest common ordered sub-sequence criterion in which the longest common ordered sub-sequence can be a non-contiguous sub-sequence.

15. The translation system as set forth in claim 14, wherein the longest common ordered sub-sequence criterion allows for a non-contiguous sub-sequence having gaps of no more than M words where M is a selected integer that is greater than zero.

16. The translation system as set forth in claim 13, wherein the extracting at least one source language-target language phrase translation pair comprises:

selecting the target language phrase based on word alignment probabilities of source word-target word pairs of the retrieved at least one fuzzy matching source language-target language text segment translation pair.

17. A storage medium storing instructions executable by a digital processor to perform translation of an input source language text segment using a statistical machine translation (SMT) system and a translation memory (TM), the method comprising:

retrieving at least one translation pair from the TM that has a fuzzy match with the input source language text segment;

extracting at least one bi-phrase from the at least one translation pair; and

translating the input source language text segment using the SMT system with a statistical model of the SMT system enriched by the extracted at least one bi-phrase.

18. The storage medium as set forth in claim 17, wherein the extracting operation comprises:

extracting at least one bi-phrase comprising a source language phrase that exactly matches a phrase of the input source language text segment.

19. The storage medium as set forth in claim 17, wherein the extracting operation comprises:

extracting at least one gapped bi-phrase comprising a gapped source language phrase that exactly matches a gapped phrase of the input source language text segment.

20. The storage medium as set forth in claim 17, wherein the update method further comprises:

outputting a translation generated by the translating operation to a human user.