Unsupervised learning of paraphrase/translation alternations and selective application thereof

• US 7,584,092 B2
• Filed: 11/15/2004
• Issued: 09/01/2009
• Est. Priority Date: 11/15/2004
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Brockett, Christopher, Dolan, William, Breck, Eric
• Primary Examiner(s): Smits; Talivaldis Ivars
• Assistant Examiner(s): ORTIZ SANCHEZ, MICHAEL
• Application Number:
  US10/989,076
• Publication Number:
  US 20060106594A1
• Time in Patent Office: 1,751 Days
• Field of Search: 704/9, 704/1, 704/2, 704/3, 704/4
• US Class Current: 704/9
• CPC Class Codes: G06F 40/237 Lexical tools

Abstract

A computer-implemented method is disclosed for applying a given paraphrase alternation pattern to an input string. The method includes generating a language model and applying the language model to determine how to apply the given paraphrase alternation pattern to the input string.

Claims

1. A computer-implemented method for applying a paraphrase alternation pattern to an input string, the method comprising:

receiving the input string;

receiving the paraphrase alternation pattern, wherein the paraphrase alternation pattern indicates a pattern of transformation from a first set of words to a second set of words wherein the given paraphrase alternation pattern is abstract in that one or both of the first and second sets of words include a place holder or blank to be filled in rather than a complete listing of specific words;

generating a language model; and

utilizing a computer processor that is a functional component of the computer to apply the language model to determine how to apply the given paraphrase alternation pattern to the input string, and wherein applying the language model comprises;

applying the given paraphrase alternation pattern to the input string in a plurality of different ways to produce a plurality of different textual variations, wherein applying the given paraphrase alternation pattern to the input string in a plurality of different ways comprises switching different combinations of words around a word within the input string identified as being a pivot word; and

applying the language model to the plurality of different textual variations to determine a probable sequence of words.

2. The method of claim 1, wherein generating a language model comprises generating a language model based on a data set that is related in terms of subject matter to a set of data from which the given paraphrase alternation pattern was derived.

3. The method of claim 1, wherein generating a language model comprises generating a language model based on data derived from a database of news articles.

4. The method of claim 1, wherein generating a language model comprises generating a language model based on data that is in a different language than the language of the input string.

5. The method of claim 1, wherein applying the language model comprises applying the language model to determine how best to apply the given paraphrase alternation pattern to the input string.

6. The method of claim 1, wherein applying the language model comprises applying the language model to determine how to apply the given paraphrase alternation pattern so as to produce a probable sequence of words.

7. The method of claim 1, further comprising making a determination as to whether the paraphrase alternation pattern can be applied to the textual input so as to preserve meaning.

8. The method of claim 1, wherein the given paraphrase alternation pattern is an indication of the first and second sets of words.

9. A computer-implemented method for applying a paraphrase alternation pattern to an input string, the method comprising:

utilizing a computer processor that is a functional component of the computer to

derive the paraphrase alternation pattern from a set of parallel texts, wherein

deriving the paraphrase alteration pattern from a set of parallel texts comprises

deriving from a set of bilingual texts;

creating a series of different application alternatives that each represent a different

application of the paraphrase alternation pattern to the input string; and

applying the series of different application alternatives to a language model to determine

a probable sequence of words.

10. The method of claim 9, wherein deriving the paraphrase alternation pattern from a set of parallel texts comprises deriving from a set of parallel, word-aligned texts.

11. The method of claim 9, wherein deriving the paraphrase alternation pattern from a set of parallel texts comprises deriving from a set of monolingual texts.

12. A computer-implemented method of selecting boundaries for application of a paraphrase alternation pattern to an input string, the method comprising:

utilizing a computer processor that is a functional component of the computer to apply the paraphrase alternation pattern multiple times to the input string with multiple boundaries so as to create a plurality of alternation alternatives, wherein each alternation alternative is a different sequence of words included in the input string;

generating a language model based on a set of test data and then applying the language model to the plurality of alternation alternatives to determine which, based on frequency within the set of test data, is a relatively commonly used sequence of words; and

wherein applying so as to create a plurality of alternation alternatives comprise applying so as to switch different combinations of words around a word in the input string identified as being a pivot word.

13. The method of claim 12, further comprising determining whether the paraphrase alternation pattern can be applied to the input string so as to preserve meaning.

14. The method of claim 13, wherein determining whether the paraphrase alternation pattern can be applied comprises applying a context model to support such a determination.

15. The method of claim 12, wherein the language model reflects a subject matter that is consistent with the input string.

16. The method of claim 12, wherein the language model incorporates data derived from a database of news articles.

17. The method of claim 12, wherein applying so as to create a plurality of alternation alternatives comprises applying so as to switch different combinations of words around a pivot word.