LEXICAL AND PHRASAL FEATURE DOMAIN ADAPTATION IN STATISTICAL MACHINE TRANSLATION

• US 20140067361A1
• Filed: 08/28/2012
• Published: 03/06/2014
• Est. Priority Date: 08/28/2012
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Inventors: Nikoulina, Vassilina, Clinchant, Stphane
• Granted Patent: US 9,026,425 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 translation method is adapted to a domain of interest. The method includes receiving a source text string comprising a sequence of source words in a source language and generating a set of candidate translations of the source text string, each candidate translation comprising a sequence of target words in a target language. An optimal translation is identified from the set of candidate translations as a function of at least one domain-adapted feature computed based on bilingual probabilities and monolingual probabilities. Each bilingual probability is for a source text fragment and a target text fragment of the source text string and candidate translation respectively. The bilingual probabilities are estimated on an out-of-domain parallel corpus that includes source and target strings. The monolingual probabilities for text fragments of one of the source text string and candidate translation are estimated on an in-domain monolingual corpus.

Claims

1. A translation method adapted to a domain of interest comprising:

receiving a source text string comprising a sequence of source words in a source language;

generating a set of candidate translations of the source text string, each candidate translation comprising a sequence of target words in a target language; and

with a processor, identifying an optimal translation from the set of candidate translations as a function of at least one domain-adapted feature, the at least one domain-adapted feature being computed based on;

bilingual probabilities, each bilingual probability being for a source text fragment and a target text fragment of the source text string and candidate translation respectively, the bilingual probabilities being estimated on an out-of-domain parallel corpus comprising source and target strings; and

monolingual probabilities for text fragments of one of the source text string and candidate translation, the monolingual probabilities being estimated on an in-domain monolingual corpus.

2. The method of claim 1 wherein the domain-adapted feature comprises at least one of:

a domain-adapted lexical feature; and

a domain-adapted phrasal feature.

3. The method of claim 2, wherein the domain-adapted feature comprises at least one domain-adapted lexical feature and wherein the domain-adapted lexical feature is computed as a function of a parallel word probability and a monolingual word probability for each of a plurality of the words of one of the source text string and the target text string,

each parallel word probability being a function of co-occurrence, in the parallel corpus of source and target strings, of that word with a corresponding word of the other of the source text string and target text string,

each monolingual word probability being a function of the occurrence of that word in the monolingual corpus.

4. The method of claim 2, wherein the domain-adapted feature comprises at least one domain-adapted phrasal feature and wherein the domain-adapted phrasal feature is computed as a function of a parallel phrase probability and a monolingual phrase probability for each of a plurality of phrases of one of the source text string and the target text string,

each parallel phrase probability being a function of co-occurrence, in a parallel corpus of source and target strings, of that phrase with a corresponding phrase of the other of the source text string and target text string,

each monolingual phrase probability being a function of the occurrence of that phrase in a respective monolingual corpus.

5. The method of claim 4, wherein each of the plurality of phrases is in a biphrase which includes a source phrase and a target phrase, the method further comprising retrieving a set of the biphrases from a biphrase library such that each word of the source text string is covered by no more than one of the retrieved biphrases, and wherein the corresponding phrase is a phrase from the same biphrase.

6. The method of claim 2, wherein the domain-adapted feature comprises direct and reverse domain-adapted lexical features.

7. The method of claim 2, wherein the domain-adapted feature comprises direct and reverse domain-adapted phrasal features.

8. The method of claim 1, wherein the identifying of the optimal translation comprises computing a scoring function in which each of a plurality of features is weighted by a respective weight and wherein the plurality of features comprises the at least one domain-adapted feature.

9. The method of claim 8, wherein the plurality of features further comprises at least one of a lexical and phrasal out-of-domain feature, the at least one out-of-domain feature being computed based on parallel probabilities derived from the parallel corpus and which is independent of the monolingual probabilities derived from the in-domain monolingual corpus.

10. The method of claim 1, further comprising learning the weight for the at least one domain-adapted feature using an in-domain parallel corpus different from the out-of-domain parallel corpus used to compute the bilingual probabilities.

11. The method of claim 1, wherein the in-domain monolingual corpus comprises at least one of:

a first in-domain monolingual corpus which comprises text strings in the source language having no corresponding text strings in the target language; and

a second in-domain monolingual corpus which comprises text strings in the target language having no corresponding text strings in the source language.

12. The method of claim 11, wherein the in-domain monolingual corpus comprises the first in-domain monolingual.

13. The method of claim 1, wherein the scoring function has the general form:

14. The method of claim 1, wherein the domain-adapted feature comprises at least one of:

a) a forward domain-adapted lexical feature which is a function of

15. The method of claim 1, wherein the bilingual probabilities derived from the parallel corpus are stored in a biphrase library and wherein the monolingual probabilities are stored only for text fragments found in the biphrase library.

16. The method of claim 1, wherein the parallel corpus comprises translations pairs, each translation pair comprising a source text string comprising words in the source language and a target text string comprising words in the target language, one of the source and target text strings having been determined to be a translation of the other.

17. The method of claim 1, wherein the bilingual probabilities each express the probability of finding a source text fragment of the source text string in the presence of a target text fragment of the candidate translation, or vice versa, in a translation pair of the sentences in the parallel corpus.

18. The method of claim 1, further comprising outputting the optimal translation as the translation of the source string.

19. 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.

20. 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.

21. A translation system adapted to a domain of interest comprising:

memory which stores;

a bilingual probability for each of a set of biphrases estimated on an associated out-of-domain parallel corpus comprising source and target strings, each biphrase comprising a text fragment in the source language and a text fragment in a target language; and

a monolingual probability for each of a set of text fragments estimated on an associated in-domain monolingual corpus, each of the text fragments occurring in at least one of the biphrases in the set of biphrases;

memory which stores;

a candidate translation generator for generating a set of candidate translations of a source text string, the source string comprising a sequence of source words in a source language, each candidate translation comprising a sequence of target words in a target language;

a translation evaluation component for identifying an optimal translation from the set of candidate translations as a function of at least one domain-adapted feature, the at least one domain-adapted feature being computed based on;

the respective bilingual probabilities for the source and target text fragments of the source text string and candidate translation, and

the respective monolingual probabilities for text fragments of at least one of the source text string and candidate translation; and

a processor which executes the instructions.

22. The system of claim 21, further comprising an in domain language model stored in memory, the translation evaluation component identifying an optimal translation based also on the language model.

23. A method for adapting a machine translation system for a domain of interest, comprising:

providing a biphrase library comprising a set of biphrases, each biphrase comprising a source text fragment in a source language and a target text fragment in a target language, the biphrases being associated with bilingual probabilities estimated on a parallel corpus of text strings in the source and target languages;

providing a first monolingual corpus for the domain of interest comprising text strings in one of the source language and target language;

computing a monolingual text fragment probability for each of a set of text fragments found in the biphrase library in the one of the source language and target language, estimated on the first monolingual corpus;

generating weights for features of a scoring function, at least one of the features being a domain-adapted feature that is to be computed based on;

bilingual probabilities retrieved from the biphrase library, each bilingual probability being for a source text fragment and a target text fragment of a source text string and candidate translation respectively; and

monolingual probabilities for text fragments of one of the source text string and candidate translation, the monolingual probabilities being drawn from the computed monolingual probabilities for the set of text fragments.

24. The method of claim 23, further comprising:

providing a second monolingual corpus for the domain of interest comprising text strings in another of the source language and target language; and

computing a monolingual text fragment probability for each of a set of text fragments found in the biphrase library in the other of the source language and target language, based on the second monolingual corpus.