System and method for incrementally updating a reordering model for a statistical machine translation system

• US 9,442,922 B2
• Filed: 11/18/2014
• Issued: 09/13/2016
• Est. Priority Date: 11/18/2014
• Status: Expired due to Fees

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Mirkin, Shachar
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US14/546,424
• Publication Number:
  US 20160140111A1
• Time in Patent Office: 665 Days
• Field of Search: 704/2, 704/3, 704/4, 704/277
• US Class Current: 1/1
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• 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

A method for updating a reordering model of a statistical machine translation system includes, at a first time, receiving new training data for retraining an existing statistical machine translation system, the new training data including at least one sentence pair, each pair including a source sentence in a source language and a target sentence in a target language. Phrase pairs are extracted from the new training data and used to generate a new reordering file. A reordering model of the existing statistical machine translation system is updated, based on the new reordering file. The reordering model includes a reordering table. At a second time after the first time, new training data is received. The extracting of phrase pairs, generating of the new reordering file and the updating the reordering model is reiterated, based on the new training data received at the second time.

Claims

1. A method for updating a reordering model of a statistical machine translation system comprising:

at a first time, receiving new training data for retraining an existing statistical machine translation system, the new training data comprising at least one sentence pair, each of the at least one sentence pair comprising a source sentence in a source language and a target sentence in a target language;

extracting phrase pairs from the new training data, each phrase pair including a source language phrase and a target language phrase;

generating a new reordering file from the extracted phrase pairs, the new reordering file including a set of the phrase pairs extracted from the new training data;

updating a reordering model of the existing statistical machine translation system based on the new reordering file, the reordering model including a reordering table, the reordering table comprising phrase pairs and a set of features, the set of features comprising, for each of a set of orientation types, at least one feature which is a function of a count of the orientation type for the respective phrase pair, each phrase pair in the reordering table occurring only once, and wherein the updating of the reordering model includes merging an existing reordering table with the new reordering file or merging the existing reordering table with a new reordering table generated from the new reordering file, the merging including updating feature scores for each of the orientation types for at least some of the phrase pairs based on the counts stored in the existing reordering table;

at a second time after the first time, receiving new training data for training the existing statistical machine translation system, the new training data comprising at least one sentence pair, the sentence pair comprising a source sentence in the source language and a target sentence in the target language; and

reiterating the extracting of phrase pairs, generating of the new reordering file and the updating the reordering model based on the new training data received at the second time,

wherein at least one of the extracting phrase pairs, generating the new reordering file, and updating the reordering model is performed with a computer processor.

2. The method of claim 1, wherein the updating of the reordering model comprises merging the new reordering file with an existing reordering file created in a prior iteration to generate a new reordering table for the reordering model.

3. The method of claim 1, wherein the updating of the reordering model comprises merging the new reordering file with an existing reordering table created in a prior iteration to generate a new reordering table for the reordering model.

4. The method of claim 1, wherein the updating of the reordering model comprises creating a new reordering table from the new reordering file and merging the new reordering table with an existing reordering table created in a prior iteration to generate a new reordering table for the reordering model.

5. The method of claim 1, wherein the creating of the new reordering file comprises assigning orientation types to the extracted phrase pairs from a predefined set of at least two orientation types.

6. The method of claim 5, wherein the set of at least two orientation types comprises three orientation types.

7. The method of claim 5, when the orientation types include:

monotonous, for phrase pairs for which the preceding target phrase is aligned to the preceding source phrase;

swap, for phrase pairs for which the preceding target phrase is aligned to the following source phrase; and

other, for phrase pairs for which the preceding target phrase is aligned to a non-consecutive source phrase.

8. The method of claim 5, wherein the update of the reordering model comprises updating occurrences of each of the orientation types of each of the phrase pairs stored in the reordering table.

9. The method of claim 1, wherein in the reordering table, the set of features comprises, for each of a set of orientation types, at least one feature which is a function of a count of the orientation type for the respective phrase pair.

10. The method of claim 1, wherein, the merging compares lines of the reordering tables sequentially and where two lines match the method includes converting the feature scores in each line into counts, summing counts for each orientation, as well as the total count, and converting the updated counts of the orientations back to scores.

11. The method of claim 1, further comprising, prior to the extracting of phrase pairs, aligning the at least one sentence pair at a word level.

12. The method of claim 1 further comprising updating at least one of a language model and a translation model of the statistical machine translation system.

13. The method of claim 1, wherein the reiterating is performed at a plurality of times subsequent to the second time.

14. The method of claim 1, wherein the new training data comprises a plurality of the sentence pairs.

15. The method of claim 1, wherein the new reordering file includes only phrase pairs extracted from the new training data.

16. A computer program product comprising non-transitory memory storing instructions which, when executed by a processor, perform the method of claim 1.

17. A system comprising memory which stores instructions for performing the method of claim 1 and a processor in communication with the memory for executing the instructions.

18. A system for updating a reordering model of a statistical machine translation system comprising:

a phrase pair extraction component which, at each of a plurality of times, extracts phrase pairs from new training data, the new training data comprising at least one sentence pair, each of the at least one sentence pair comprising a source sentence in a source language and a target sentence in a target language, each phrase pair including a source language phrase and a target language phrase;

a reordering file generation component which, at each of the plurality of times, generates a new reordering file, the new reordering file including only phrase pairs extracted from the new training data and their associated orientation types;

an update component which, at each of the plurality of times, updates a reordering model of an existing statistical machine translation system based on the new reordering file, the reordering model including a reordering table, the reordering table comprising phrase pairs and a set of features, the set of features comprising, for each of a set of orientation types, at least one feature which is a function of a count of the orientation type for the respective phrase pair, each phrase pair in the reordering table occurring only once, and wherein the updating of the reordering model includes one of;

merging the new reordering file with an existing reordering file created in a prior iteration to generate a merged reordering file and creating the updated reordering table from the merged reordering file,

merging the new reordering file with an existing reordering table created in a prior iteration, the existing reordering table tracking occurrence counts of phrase pairs in the reordering table, to generate a new reordering table for the reordering model, and

merging an existing reordering table with a new reordering table generated from the new reordering file, the merging including updating feature scores for each of the orientation types for at least some of the phrase pairs based on counts stored in the existing reordering table; and

a processor which implements the phrase pair extraction component, reordering file generation component, and update component.

19. A method for updating a reordering model of a statistical machine translation system comprising:

at a first time, receiving new training data, the new training data comprising sentence pairs, each of the sentence pairs comprising a source sentence in a source language and a target sentence in a target language;

extracting phrase pairs from the new training data, each phrase pair including a source language phrase and a target language phrase;

generating a new reordering file from the extracted phrase pairs, the new reordering file including only phrase pairs extracted from the new training data and their associated orientation types;

updating a reordering model of the existing statistical machine translation system based on the new reordering file and an existing reordering table of the reordering model, the existing reordering table comprising phrase pairs and a set of features, the set of features comprising, for each of a set of orientation types, at least one feature which is a function of a count of the orientation type for the respective phrase pair, the updating including accumulating counts of the extracted phrase pairs and stored counts of corresponding phrase pairs in the existing reordering table, each phrase pair in the updated reordering table occurring only once; and

repeating the receiving new training data, extracting of phrase pairs, generating of the new reordering file, and the updating the reordering model at least once at a subsequent time,

wherein at least one of the extracting phrase pairs, generating the new reordering file, and updating the reordering model is performed with a processor.