Machine translation using vector space representations

• US 20060265209A1
• Filed: 04/24/2006
• Published: 11/23/2006
• Est. Priority Date: 04/26/2005
• Status: Active Grant

• Current Assignee: Relativity ODA LLC
• Original Assignee: Content Analyst Company, LLC (Relativity ODA LLC)
• Inventors: Bradford, Roger Burrowes
• Granted Patent: US 7,765,098 B2
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G06F 40/42 Data-driven translation

Abstract

An embodiment of the present invention provides a method for automatically translating text. First, a conceptual representation space is generated based on source-language documents and target-language documents, wherein respective terms from the source-language and target-language documents have a representation in the conceptual representation space. Second, a new source-language document is represented in the conceptual representation space, wherein a subset of terms in the new source-language document is represented in the conceptual representation space, such that each term in the subset has a representation in the conceptual representation space. Then, a term in the new source-language document is automatically translated into a corresponding target-language term based on a similarity between the representation of the term and the representation of the corresponding target-language term.

Claims

1. A method for automatically translating text, comprising:

(a) generating a conceptual representation space based on source-language documents and target-language documents, wherein respective terms from the source-language documents and the target-language documents have a representation in the conceptual representation space;

(b) representing a new source-language document in the conceptual representation space, wherein a subset of terms in the new source-language document is represented in the conceptual representation space, such that each term in the subset has a representation in the conceptual representation space; and

(c) automatically translating a term in the new source-language document into a corresponding target-language term based on a similarity between the representation of the term and the representation of the corresponding target-language term.

2. The method of claim 1, further comprising:

(d) repeating step (c) for each term in the new source-language document.

3. The method of claim 1, wherein generating a conceptual representation space comprises generating a Latent Semantic Indexing (LSI) space.

4. The method of claim 1, wherein step (c) comprises:

(c1) subdividing sentences that are longer than a threshold into logically coherent segments; and

(c2) automatically translating a term in one of the logically coherent segments into a corresponding target-language term contained in at least one target-language document based on a similarity between the representation of the term and the representation of the corresponding target-language term.

5. The method of claim 4, further comprising:

repeating step (c2) for each term in each of the logically coherent segments.

6. The method of claim 1, further comprising:

identifying an idiomatic expression contained in at least one of the source-language documents;

generating a representation of the idiomatic expression in the conceptual representation space; and

automatically translating the idiomatic expression into a target-language term based on a similarity between the representation of the idiomatic expression and the representation of the target-language term.

7. The method of claim 6, further comprising:

identifying an idiomatic expression contained in a target-language document in the collection of target-language documents; and

generating a representation of the idiomatic expression contained in the target-language document in the conceptual representation space.

8. The method of claim 6, wherein identifying an idiomatic expression contained in at least one of the source-language documents comprises:

determining sequences of words that comprise an idiomatic expression by comparing sequences of words to a known list of idiomatic expressions.

9. The method of claim 6, wherein identifying an idiomatic expression contained in at least one of the source-language documents comprises:

identifying a candidate idiomatic expression based on sequences of words that appear in the source-language documents more often than a threshold.

10. The method of claim 6, wherein identifying an idiomatic expression contained in at least one of the source-language documents comprises:

identifying at least one candidate sequence of words;

generating (i) a representation of the at least one candidate sequence of words in the conceptual representation space, and (ii) a representation of each word in the at least one candidate sequence of words in the conceptual representation space;

comparing the representation of the at least one candidate sequence of words with the representation of each word in the at least one candidate sequence of words to determine a difference thereof; and

identifying the at least one candidate sequence of words as an idiomatic expression if the difference is greater than a threshold.

11. The method of claim 1, further comprising:

generating a disambiguated conceptual representation space for at least one of the source-language documents, wherein a polysemous word contained in the at least one source-language document has a plurality of representations in the disambiguated conceptual representation space, each representation of the polysemous word corresponding to a sense of that word;

generating a representation of the new source-language document in the disambiguated conceptual representation space, wherein a subset of terms in the new source-language document is represented in the disambiguated conceptual representation space, such that each term in the subset has a representation in the disambiguated conceptual representation space; and

automatically translating a term in the new source-language document into a corresponding target-language term based on a similarity between the representation of the term and the representation of one of the senses of the polysemous word.

12. A method for automatically translating text based on a disambiguation of text at a dictionary-level, comprising:

generating a conceptual representation space based on source-language documents and target-language documents;

providing a plurality of dictionaries;

generating a representation of each dictionary in the conceptual representation space;

representing a new source-language document in the conceptual representation space;

selecting a first dictionary from the collection of dictionaries based on a similarity between the representation of the first dictionary and the representation of the new source-language document; and

automatically translating at least one term in the new source-language document into a corresponding target-language term based on the first dictionary.

13. The method of claim 12, wherein generating a representation of each dictionary comprises:

concatenating terms in each dictionary into a single text object; and

generating a representation of the single text object in the conceptual representation space.

14. The method of claim 12, wherein generating a representation of each dictionary comprises:

subdividing each dictionary into conceptually cohesive segments;

generating a text object for each conceptually cohesive segment; and

generating a representation of each text object in the conceptual representation space.

15. A method for producing a machine translation of a text passage based on a combination of a plurality of translations of the text passage, comprising:

(a) generating a conceptual representation space based on a collection of source-language documents and a collection of target-language documents;

(b) providing a plurality of translations of a text passage;

(c) generating a representation of each translation in the conceptual representation space; and

(d) automatically translating the text passage based on similarity comparisons among the representations of the translations.

16. The method of claim 15, wherein the text passage comprises a plurality of terms such that the terms can be combined in a plurality of combinations, and wherein step (d) comprises:

computing an average representation for the plurality of translations;

generating a representation of each combination of terms; and

automatically translating the text passage based on similarity comparisons between the average representation and the representation of each combination of terms.

17. A method for generating a parallel corpus of documents, comprising:

(a) generating a conceptual representation space based on a collection of source-language documents and a collection of target-language documents, wherein each target-language document in the collection of target-language documents comprises a translation of a source-language document in the collection of source-language documents;

(b) providing a new collection of documents, including both source-language documents and target-language documents;

(c) generating a representation of each document in the new collection of documents in the conceptual representation space;

(d) identifying a collection of parallel documents based on similarity comparisons among the representations in the conceptual representation space; and

(e) combining the collection of source-language documents and the collection of target-language documents with the collection of parallel documents identified in step (d) resulting in a combined collection of documents, and generating a new conceptual representation space based on the combined collection of documents, wherein the new conceptual representation space is stored in an electronic format.

18. The method of claim 17, further comprising:

(f) identifying a new collection of parallel documents based on similarity comparisons among the representations in the new conceptual representation space; and

(g) generating another new conceptual representation space based on the new collection of parallel documents.

19. The method of claim 18, wherein prior to step (f) the method further comprising:

providing a second new collection of documents, including both source-language documents and target-language documents; and

generating a representation of each document in the second new collection of documents in the new conceptual representation space.

20. A method for automatically translating text, comprising:

(a) generating a conceptual representation space based on source-language documents and target-language documents, wherein respective terms from the source-language documents and the target-language documents have a representation in the conceptual representation space;

(b) measuring a similarity between at least one pair of terms based on the representations of terms included in the at least one pair of terms, wherein the at least one pair of terms includes a term from at least one of the source-language documents and a term from at least one of the target-language documents;

(c) converting the similarity to an association probability; and

(d) using the association probability as an estimate of a parameter in a statistical translation algorithm.

21. A computer program product comprising a computer usable medium having control logic stored therein for automatically translating text, the control logic comprising:

computer readable first program code that causes the computer to generate a conceptual representation space based on source-language documents and target-language documents, wherein respective terms from the source-language documents and the target-language documents have a representation in the conceptual representation space;

computer readable second program code that causes the computer to represent a new source-language document in the conceptual representation space, wherein a subset of terms in the new source-language document is represented in the conceptual representation space, such that each term in the subset has a representation in the conceptual representation space; and

computer readable third program code that causes the computer to automatically translate a term in the new source-language document into a corresponding target-language term based on a similarity between the representation of the term and the representation of the corresponding target-language term.

22. The computer program product of claim 21, further comprising:

computer readable fourth program code that causes the computer to repeat the computer readable third program code for each term in the new source-language document.

23. The computer program product of claim 21, wherein the conceptual representation space is a Latent Semantic Indexing (LSI) space.

24. The computer program product of claim 21, wherein the computer readable third program code comprises:

computer readable fourth program code that causes the computer to subdivide sentences that are longer than a threshold into logically coherent segments; and

computer readable fifth program code that causes the computer to automatically translate a term in one of the logically coherent segments into a corresponding target-language term contained in at least one target-language document based on a similarity between the representation of the term and the representation of the corresponding target-language term.

25. The computer program product of claim 24, wherein the computer readable third program code further comprising:

code that causes the computer to repeat the computer readable fifth program code for each term in each of the logically coherent segments.

26. The computer program product of claim 21, further comprising:

computer readable fourth program code that causes the computer to identify an idiomatic expression contained in at least one of the source-language documents;

computer readable fifth program code that causes the computer to generate a representation of the idiomatic expression in the conceptual representation space; and

computer readable sixth program code that causes the computer to automatically translate the idiomatic expression into a target-language term based on a similarity between the representation of the idiomatic expression and the representation of the target-language term.

27. The computer program product of claim 26, further comprising:

code that causes the computer to identify an idiomatic expression contained in a target-language document in the collection of target-language documents; and

code that causes the computer to generate a representation of the idiomatic expression contained in the target-language document in the conceptual representation space.

28. The computer program product of claim 26, wherein the computer readable fourth program code that causes the computer to identify an idiomatic expression contained in at least one of the source-language documents comprises:

code that causes the computer to determine sequences of words that comprise an idiomatic expression by comparing sequences of words to a known list of idiomatic expressions.

29. The computer program product of claim 26, wherein the computer readable fourth program code that causes the computer to identify an idiomatic expression contained in at least one of the source-language documents comprises:

code that causes the computer to identify a candidate idiomatic expression based on sequences of words that appear in the source-language documents more often than a threshold.

30. The computer program product of claim 26, wherein the computer readable fourth program code that causes the computer to identify an idiomatic expression contained in at least one of the source-language documents comprises:

code that causes the computer to identify at least one candidate sequence of words;

code that causes the computer to generate (i) a representation of the at least one candidate sequence of words in the conceptual representation space, and (ii) a representation of each word in the at least one candidate sequence of words in the conceptual representation space;

code that causes the computer to compare the representation of the at least one candidate sequence of words with the representation of each word in the at least one candidate sequence of words to determine a difference thereof; and

code that causes the computer to identify the at least one candidate sequence of words as an idiomatic expression if the difference is greater than a threshold.

31. The computer program product of claim 21, further comprising:

computer readable fourth program code that causes the computer to generate a disambiguated conceptual representation space for at least one of the source-language documents, wherein a polysemous word contained in the at least one source-language document has a plurality of representations in the disambiguated conceptual representation space, each representation of the polysemous word corresponding to a sense of that word;

computer readable fifth program code that causes the computer to generate a representation of the new source-language document in the disambiguated conceptual representation space, wherein a subset of terms in the new source-language document is represented in the disambiguated conceptual representation space, such that each term in the subset has a representation in the disambiguated conceptual representation space; and

computer readable sixth program code that causes the computer to automatically translate a term in the new source-language document into a corresponding target-language term based on a similarity between the representation of the term and the representation of one of the senses of the polysemous word.

32. A computer program product comprising a computer usable medium having control logic stored therein for automatically translating text based on a disambiguation of text at a dictionary-level, the control logic comprising:

computer readable first program code that causes the computer to generate a conceptual representation space based on source-language documents and target-language documents;

computer readable second program code that causes the computer to provide a plurality of dictionaries;

computer readable third program code that causes the computer to generate a representation of each dictionary in the conceptual representation space;

computer readable fourth program code that causes the computer to represent a new source-language document in the conceptual representation space;

computer readable fifth program code that causes the computer to select a first dictionary from the collection of dictionaries based on a similarity between the representation of the first dictionary and the representation of the new source-language document; and

computer readable sixth program code that causes the computer to automatically translate a term in the new source-language document into a corresponding target-language term based on the first dictionary.

33. The computer program product of claim 32, wherein the computer readable third program code comprises:

code that causes the computer to concatenate terms in each dictionary into a single text object; and

code that causes the computer to generate a representation of the single text object in the conceptual representation space.

34. The computer program product of claim 32, wherein the computer readable third program code comprises:

code that causes the computer to subdivide each dictionary into conceptually cohesive segments;

code that causes the computer to generate a text object for each conceptually cohesive segment; and

code that causes the computer to generate a representation of each text object in the conceptual representation space.

35. A computer program product comprising a computer usable medium having control logic stored therein for producing a machine translation of a text passage based on a combination of a plurality of translations of the text passage, the control logic comprising:

computer readable first program code that causes the computer to generate a conceptual representation space based on a collection of source-language documents and a collection of target-language documents;

computer readable second program code that causes the computer to provide a plurality of translations of a text passage;

computer readable third program code that causes the computer to generate a representation of each translation in the conceptual representation space; and

computer readable fourth program code that causes the computer to automatically translate the text passage based on similarity comparisons among the representations of the translations.

36. The computer program product of claim 35, wherein the text passage comprises a plurality of terms such that the terms can be combined in a plurality of combinations, and wherein the computer readable fourth program code comprises:

code that causes the computer to compute an average representation for the plurality of translations;

code that causes the computer to generate a representation of each combination of terms; and

code that causes the computer to automatically translate the text passage based on similarity comparisons between the average representation and the representation of each combination of terms.

37. A computer program product comprising a computer usable medium having control logic stored therein for generating a parallel corpus of documents, the control logic comprising:

computer readable first program code that causes the computer to generate a conceptual representation space based on a collection of source-language documents and a collection of target-language documents, wherein each target-language document in the collection of target-language documents comprises a translation of a source-language document in the collection of source-language documents;

computer readable second program code that causes the computer to provide a new collection of documents, including both source-language documents and target-language documents;

computer readable third program code that causes the computer to generate a representation of each document in the new collection of documents in the conceptual representation space;

computer readable fourth program code that causes the computer to identify a collection of parallel documents based on similarity comparisons among the representations in the conceptual representation space; and

computer readable fifth program code that causes the computer to combine the collection of source-language documents and the collection of target-language documents with the collection of parallel documents identified by the computer readable fourth program code resulting in a combined collection of documents, and that causes the computer to generate a new conceptual representation space based on the combined collection of documents, wherein the new conceptual representation space is stored in an electronic format.

38. The computer program product of claim 37, further comprising:

computer readable sixth program code that causes the computer to identify a new collection of parallel documents based on similarity comparisons among the representations in the new conceptual representation space; and

computer readable seventh program code that causes the computer to generate another new conceptual representation space based on the new collection of parallel documents.

39. The computer program product of claim 38, wherein prior to implementing the computer readable sixth program code, the control logic further comprises:

code that causes the computer to provide a second new collection of documents, including both source-language documents and target-language documents; and

code that causes the computer to generate a representation of each document in the second new collection of documents in the new conceptual representation space.

40. A computer program product comprising a computer usable medium having control logic stored therein for automatically translating text, the control logic comprising:

computer readable first program code that causes the computer to generate a conceptual representation space based on source-language documents and target-language documents, wherein respective terms from the source-language documents and the target-language documents have a representation in the conceptual representation space;

computer readable second program code that causes the computer to measure a similarity between at least one pair of terms based on the representations of terms included in the at least one pair of terms, wherein the at least one pair of terms includes a term from at least one of the source-language documents and a term from at least one of the target-language documents;

computer readable third program code that causes the computer to convert the similarity to an association probability; and

computer readable fourth program code that causes the computer to use the association probability as an estimate of a parameter in a statistical translation algorithm.