Terminological adaptation of statistical machine translation system through automatic generation of phrasal contexts for bilingual terms

• US 9,367,541 B1
• Filed: 01/20/2015
• Issued: 06/14/2016
• Est. Priority Date: 01/20/2015
• Status: Expired due to Fees

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Servan, Christophe, Dymetman, Marc
• Primary Examiner(s): Augustin, Marcellus
• Application Number:
  US14/600,561
• Time in Patent Office: 511 Days
• Field of Search: 704/2, 704/1, 704/8, 704256-257
• US Class Current: 1/1
• CPC Class Codes: G06F 40/295 Named entity recognition
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment
• CPC Class Codes: G06F 40/55 Rule-based translation

Abstract

A method for terminological adaptation includes receiving a vocabulary pair including source and target language terms. Each term is in a class which includes a set of sequences. Contextual phrase pairs are extracted from a bilingual training corpus, each including source and target phrases. The phrases each include a sequence of the same class as the respective source and target terms as well as some associated context. Templates are generated, based on the contextual phrase pairs. In each template the source and target sequences of a contextual phrase pair are replaced with respective placeholders, each denoting the respective class of the sequence. Candidate phrase pairs are generated from these templates. In each candidate phrase pair, the placeholders of one of the templates are replaced with respective terms of a vocabulary pair of the same class. Some candidate phrase pairs are incorporated into a phrase table of a machine translation system.

Claims

1. A method for terminological adaptation of a machine translation system comprising:

receiving a set of vocabulary pairs, each vocabulary pair including a source term in a source language and a target term in a target language, each term being in a class which includes a set of sequences;

extracting contextual phrase pairs from a bilingual training corpus, each of the contextual phrase pairs including a source phrase and a target phrase, the source phrase including a source sequence of a same class as the source term of one of the vocabulary pairs and associated source context, the target phrase including a target sequence of a same class as the target term of the one of the vocabulary pairs and associated target context;

generating templates based on the extracted contextual phrase pairs, each template replacing the source and target sequences of a contextual phrase pair with respective source and target placeholders denoting the class;

generating at least one candidate phrase pair from each template, each candidate phrase pair replacing the source and target placeholders of the template with respective source and target terms of a vocabulary pair of the same class; and

incorporating at least some of the candidate phrase pairs into a phrase table of a machine translation system,

wherein at least one of the extracting contextual phrase pairs, generating templates, and generating candidate phrase pairs is performed with a processor.

2. The method of claim 1, further comprising filtering the candidate phrase pairs based on at least one of linguistic rules and a measure of proximity to an in-domain corpus.

3. The method of claim 2, wherein the filtering is based on a measure of proximity to an in-domain corpus and includes computing a cross-entropy difference score for each phrase pair and filtering out candidate phrase pairs which do not meet a threshold cross-entropy difference score.

4. The method of claim 3, wherein the computing of the cross-entropy difference score comprises:

computing a first measure of perplexity between the target phrase of the candidate phrase pair and an in-domain language model;

computing a second measure of perplexity between the target phrase of the candidate phrase pair and an out-of-domain language model; and

computing a difference between the first measure of perplexity and the second measure of perplexity.

5. The method of claim 3, further comprising:

generating a set of language models, each language model being generated from a portion of the candidate phrase pairs, each portion having a different number of the candidate phrase pairs;

scoring the language models on a development corpus;

identifying one of the language models based on the scores; and

selecting the threshold cross-entropy difference score based on the identified language model.

6. The method of claim 2, wherein the filtering is based on linguistic rules, the linguistic rules including at least one linguistic rule which requires the associated context to be compatible with the respective source or target term in at least one of declension, verb tense, and morphology.

7. The method of claim 1, wherein the set of vocabulary pairs includes a plurality of vocabulary pairs of a same class and the generating at least one candidate phrase pair from each template includes generating a candidate phrase pair for each of the plurality of vocabulary pairs from each template matching the class.

8. The method of claim 1, wherein the class is a class of out-of-vocabulary words.

9. The method of claim 8, wherein the class is a class of named entities.

10. The method of claim 1, further comprising assigning feature values to each of the candidate phrase pairs incorporated into the phrase table, the feature values being based on feature values of phrase pairs extracted from the bilingual training corpus that are instances of the template from which the respective candidate phrase pair is generated.

11. The method of claim 1, wherein the extracting of the contextual phrase pairs from the bilingual training corpus includes extracting the contextual phrase pairs from an original phrase table generated from the bilingual training corpus.

12. The method of claim 1, wherein the associated context for each source and target sequence includes at least one of:

at least one word preceding the respective sequence; and

at least one word following the respective sequence.

13. The method of claim 1, further comprising tuning the machine translation system based on the candidate phrase pairs incorporated into the phrase table.

14. A computer program product comprising a non-transitory recording medium storing instructions, which when executed on a computer, causes the computer to perform the method of claim 1.

15. A system for terminological adaptation of a machine translation system comprising memory which stores instructions for performing the method of claim 1 and a processor in communication with the memory which implements the instructions.

16. A system for terminological adaptation of an associated machine translation system comprising:

an extraction component which extracts contextual phrase pairs from an associated bilingual training corpus, each of the contextual phrase pairs including a source phrase and a target phrase, the source phrase including a source sequence of a same class as a source term of an input vocabulary pair and respective associated context, the target phrase including a target sequence of a same class as a target term of the input vocabulary pair and respective associated context;

a template generator which generates templates based on the extracted contextual phrase pairs, each template replacing the source and target sequences of a contextual phrase pair with respective source and target placeholders denoting the respective class;

a phrase table entry generation component which generates at least one candidate phrase pair from each template, each candidate phrase pair replacing the source and target placeholders of a template with respective source and target terms of the input vocabulary pair, the phrase table entry generation component incorporating at least some of the candidate phrase pairs into a phrase table of a machine translation system; and

a processor which implements the extraction component, template generator, and phrase table entry generation component.

17. The system of claim 16, further comprising a filtering component which filters the candidate phrase pairs based on at least one of linguistic rules and a measure of proximity to an in-domain corpus.

18. The system of claim 16, further comprising a tuning component which tunes the machine translation system based on the candidate phrase pairs incorporated into the phrase table.

19. A method for terminological adaptation of a machine translation system to a target domain comprising:

extracting contextual phrase pairs from an out-of-domain bilingual training corpus of the form α

β

γ

α

′

β

′

γ

′

, each of the contextual phrase pairs including a source phrase α

β

γ and

a target phrase α

′

β

′

γ

′

, the source phrase including a source sequence β

, which is of a same class T as a source term δ

of an input vocabulary pair in a target domain, and respective associated context α

, γ

, the target phrase including a target sequence β

′

of a same class T′

as a target term δ

′

of the input vocabulary pair and respective associated context α

′

, γ

′

;

based on the extracted contextual phrase pairs, generating candidate phrase pairs of the form α

δ

γ

α

′

δ

′

γ

′

, where δ

replaces β and

δ

′

replaces β

′

, respectively, in one of the extracted contextual phrase pairs;

filtering the extracted candidate phrase pairs; and

incorporating at least some of the remaining candidate phrase pairs into a phrase table of a machine translation system,

wherein at least one of the extracting contextual phrase pairs and generating candidate phrase pairs is performed with a processor.

20. The method of claim 19, wherein the filtering includes

computing a first measure of perplexity between the target phrase of the candidate phrase pair and a target domain language model;

computing a second measure of perplexity between the target phrase of the candidate phrase pair and an out-of-domain language model; and

computing a difference between the first measure of perplexity and the second measure of perplexity.