SYNTAX-BASED AUGMENTATION OF STATISTICAL MACHINE TRANSLATION PHRASE TABLES

• US 20120296633A1
• Filed: 05/20/2011
• Published: 11/22/2012
• Est. Priority Date: 05/20/2011
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Chalabi, Achraf, Ashour, Mostafa, Ammar, Waleed
• Granted Patent: US 8,874,433 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 40/45 Example-based machine translation; Alignment

Abstract

Machine translation phrase table augmentation embodiments are described that employ an automatic syntax-based scheme to produce additional phrase pairs and insert them into a phrase table. One general process implementing this augmentation involves inputting one or more syntactic transfer patterns, and for each pattern synthesizing phrases in a source language of the type associated with the pattern using a source language lexicon. Phrases, such as those not found in a monolingual corpus of the source language, are eliminated from the synthesized phrases. Each of the remaining synthesized phrases is then translated into the target language using the syntactic transfer pattern, a bilingual source-to-target language dictionary, and a morphological synthesizer. Those translated phrases not found in a monolingual corpus of the target language are then eliminated. Phrase pairs made up of a remaining translated phrase and its corresponding source language phrase are then added to the phrase table being augmented.

Claims

1. A computer-implemented process for augmenting a machine translation phrase table with additional phrase pairs each pair of which associates a phrase in a source language with a phrase in a target language, comprising:

using a computer to perform the following process actions;

inputting one or more syntactic transfer patterns, each of said patterns defining the syntax of a translation to the target language of a different phrase structure in the source language, wherein each source language phrase type represents a phase having a particular syntactic structure that is different from the other source language phrase types;

for each inputted syntactic transfer pattern,

synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language,

eliminating synthesized phrases not found in a monolingual corpus of the source language,

for each remaining synthesized phrase, translating the synthesized phrase into the target language using the syntactic transfer pattern under consideration, a bilingual source-to-target language dictionary, and a morphological synthesizer to properly inflect the words of each translated phrase, and

for each translated phrase, adding a phrase pair comprising the translated phrase and its corresponding source language phrase to the current version of the phrase table to produce an augmented version of the phrase table.

2. The process of claim 1, further comprising a process action of eliminating phrase pairs that are already found in a current version of the phrase table, prior to performing the action of adding phrase pairs to the phrase table.

3. The process of claim 1, further comprising a process action of eliminating translated phrases not found in a monolingual corpus of the target language, after performing the action of translating the synthesized phrase into the target language.

4. The process of claim 1, wherein the process action of synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language, comprises synthesizing a phrase for every possible combination of words in the lexicon that correspond to syntactic constituents of the source language phrase type associated with the pattern under consideration.

5. The process of claim 1, wherein one or more of the inputted syntactic transfer patterns is associated with a phrase structure that is incorrectly translated during machine translation using the current version of the phrase table.

6. The process of claim 1, further comprising the process actions of:

computing a probability for each phrase pair added to the phrase table, said probability representing the likelihood that the source language phrase of the phrase pair translates to the target language phrase of the pair; and

for each probability computed, associating the probability with the phrase pair it was computed for and adding it to the augmented version of the phrase table.

7. The process of claim 6, wherein the process action of computing a probability for each phrase pair added to the phrase table, comprises of computing the probability for a phrase pair added to the phrase table using at least one factor, said factors comprising the frequency that the target language phrase of a phrase pair appears in a monolingual corpus of the target language, or the frequency that the source language phrase appears in the monolingual corpus of the source language, or the frequency of target phrases already in the phrase table which correspond to the same source phrase in the added phrase pair, or pre-established probabilities associated with phrase pairs already existing in the current version of the phrase table.

8. The process of claim 1, wherein said phrase types comprise noun phrases.

9. The process of claim 1, wherein the source language is English and the target language is Arabic.

10. A computer-implemented process for augmenting a machine translation phrase table with additional phrase pairs each pair of which associates a phrase in a source language with a phrase in a target language, comprising:

using a computer to perform the following process actions;

inputting one or more syntactic transfer patterns, each of said patterns defining the syntax of a translation to the target language of a different phrase structure in the source language, wherein each source language phrase type represents a phase having a particular syntactic structure that is different from the other source language phrase types;

for each inputted syntactic transfer pattern,

synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language,

eliminating synthesized phrases not found in a monolingual corpus of the source language,

for each remaining synthesized phrase, translating the synthesized phrase into the target language using the syntactic transfer pattern under consideration, a bilingual source-to-target language dictionary, and a morphological synthesizer to properly inflect the words of each translated phrase,

eliminating translated phrases not found in a monolingual corpus of the target language,

for each remaining translated phrase, adding a phrase pair comprising the translated phrase and its corresponding source language phrase to the current version of the phrase table to produce an augmented version of the phrase table.

11. The process of claim 10, further comprising in a pre-processing stage performed prior to inputting the one or more syntactic transfer patterns, the process actions of:

performing a machine translation of sentences in the source language into the target language;

identifying one or more phrase structures from the source language sentences that are translated into the target language with an incorrect syntactic structure in the target language; and

for each identified source language phrase type, establishing a syntactic transfer pattern that defines a correct syntax of a translation to the target language from the source language for the phrase type under consideration

12. The process of claim 10, wherein the process action of synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language, comprises the actions of:

identifying syntactic constituents of the source language phrase type associated with the pattern under consideration;

for each identified syntactic constituent, identifying words in the lexicon that correspond to the identified syntactic constituent; and

synthesizing phrases using the identified words such that each phrase exhibits a word order corresponding the source language phrase type associated with the pattern under consideration.

13. The process of claim 10, wherein the process action of synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language, comprises the actions of:

identifying syntactic constituents of the source language phrase type associated with the pattern under consideration;

for each identified syntactic constituent, identifying words in the lexicon that correspond to the identified syntactic constituent; and

synthesizing phrases using the identified words such that each phrase exhibits a word order corresponding the source language phrase type associated with the pattern under consideration and such that a phrase is synthesized for every possible combination of the identified words that conforms to the source language phrase type.

14. The process of claim 10, further comprising performing, prior to adding a phrase pair comprising a translated phrase and its corresponding source language phrase to the current version of the phrase table, the process actions of:

computing a probability for the phrase pair, said probability representing the likelihood that the source language phrase of the phrase pair translates to the target language phrase of the pair; and

associating the probability with the phrase pair.

15. The process of claim 14, wherein the process action of adding a phrase pair comprising a translated phrase and its corresponding source language phrase to the current version of the phrase table, comprises an action of adding the probability computed for the phrase pair along with the phrase pair to the current version of the phrase table.

16. A computer-readable storage medium having computer-executable instructions stored thereon for augmenting a machine translation phrase table with additional phrase pairs each pair of which associates a phrase in a source language with a phrase in a target language, said computer-executable instructions comprising:

inputting one or more syntactic transfer patterns, each of said patterns defining the syntax of a translation to the target language of a different phrase structure in the source language, wherein each source language phrase type represents a phase having a particular syntactic structure that is different from the other source language phrase types;

for each inputted syntactic transfer pattern,

synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language,

eliminating synthesized phrases not found in a monolingual corpus of the source language,

for each remaining synthesized phrase, translating the synthesized phrase into the target language using the syntactic transfer pattern under consideration, a bilingual source-to-target language dictionary, and a morphological synthesizer to properly inflect the words of each translated phrase,

eliminating translated phrases not found in a monolingual corpus of the target language, and

for each remaining translated phrase,

generating a phrase pair comprising the translated phrase and its corresponding source language phrase,

eliminating phrase pairs that are already found in a current version of the phrase table, and

adding the generated phrase pair to the current version of the phrase table.

17. The computer-readable storage medium of claim 16, further comprising computer-executable instructions for:

computing for each generated phrase pair, a probability representing the likelihood that the source language phrase of a phrase pair translates to the target language phrase of the pair;

associating each probability with the phrase pair it was computed for; and

adding each probability to the augmented version of the phrase table.

18. The computer-readable storage medium of claim 16, wherein the computer-executable instruction for synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language, comprises a sub-instruction for synthesizing phrases representing combinations of words in the lexicon that correspond to syntactic constituents of the source language phrase type associated with the pattern under consideration.

19. The computer-readable storage medium of claim 16, wherein the computer-executable instruction for synthesizing phrases in the source language of the type associated with the pattern under consideration using a lexicon of the source language, comprises a sub-instruction for synthesizing a phrase for every possible combination of words in the lexicon that correspond to syntactic constituents of the source language phrase type associated with the pattern under consideration.

20. The computer-readable storage medium of claim 16, wherein the computer-executable instruction for inputting one or more syntactic transfer patterns, comprises a sub-instruction for inputted syntactic transfer patterns associated with a phrase structure that is incorrectly translated during machine translation.