Removing Bias From Features Containing Overlapping Embedded Grammars in a Natural Language Understanding System

• US 20080109210A1
• Filed: 11/03/2006
• Published: 05/08/2008
• Est. Priority Date: 11/03/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Skuratovsky, Ilya
• Granted Patent: US 8,204,738 B2
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G06F 40/211 Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06F 40/30 Semantic analysis

Abstract

A method of removing bias from an action classifier within a natural language understanding system can include identifying a sentence having a target embedded grammar that overlaps with at least one other embedded grammar and selecting a group of overlapping embedded grammars including the target embedded grammar and at least one additional embedded grammar. A sentence expansion can be created that includes the sentence including the target embedded grammar and a copy of the sentence for each additional embedded grammar of the group. Each copy of the sentence can include a different additional embedded grammar from the group in place of the target embedded grammar. The sentence expansion can be included within action classifier training data.

Claims

1. A method of removing bias from an action classifier within a natural language understanding system, the method comprising:

identifying a sentence having a target embedded grammar that overlaps with at least one other embedded grammar;

selecting a group of overlapping embedded grammars comprising the target embedded grammar and at least one additional embedded grammar;

creating a sentence expansion comprising the sentence including the target embedded grammar and a copy of the sentence for each additional embedded grammar of the group, with each copy of the sentence including a different additional embedded grammar from the group in place of the target embedded grammar; and

including the sentence expansion in action classifier training data.

2. The method of claim 1, further comprising training the action classifier of the natural language understanding system with the action classifier training data comprising the sentence expansion.

3. The method of claim 1, further comprising defining the group of overlapping embedded grammars as a set of embedded grammars having at least one value in common.

4. The method of claim 1, further comprising receiving a user input defining the group of overlapping embedded grammars.

5. The method of claim 1, wherein creating a sentence expansion further comprises:

creating n−

1 copies of the sentence, where n is the total number of embedded grammars in the selected group; and

replacing the target embedded grammar in each copy of the sentence with a different one of the additional embedded grammars of the group, wherein the original sentence comprising the target embedded grammar remains intact.

6. The method of claim 1, further comprising automatically setting a count of a sentence according to the target embedded grammar or the embedded grammar used to replace the target embedded grammar in the sentence.

7. The method of claim 1, further comprising:

determining whether the sentence is associated with a selected action; and

only creating a sentence expansion if the sentence is associated with the selected action.

8. A method of removing bias from an action classifier within a natural language understanding system, the method comprising:

identifying sentences that comprise at least one target embedded grammar;

for each identified sentence,

selecting a group of overlapping embedded grammars including the target embedded grammar and at least one additional embedded grammar;

creating n−

1 copies of the sentence, where n is the total number of embedded grammars in the selected group;

replacing the target embedded grammar in each copy of the sentence with a different one of the additional embedded grammars of the group and keeping the original sentence comprising the target embedded grammar intact; and

including the sentence and the copies of the sentence within action classifier training data.

9. The method of claim 8, further comprising training the action classifier of the natural language understanding system with the action classifier training data comprising the sentence and the copies of the sentence.

10. The method of claim 8, further comprising automatically setting a count of a sentence according to the target embedded grammar or the additional embedded grammar used to replace the target embedded grammar.

11. The method of claim 8, further comprising:

determining whether the sentences are associated with selected actions; and

only performing the selecting, creating, and replacing steps for sentences associated with the selected actions.

12. The method of claim 8, further comprising defining the group of overlapping embedded grammars as a set of embedded grammars having at least one value in common.

13. The method of claim 8, further comprising receiving a user input defining the group of overlapping embedded grammars.

14. The method of claim 8, further comprising determining that the target embedded grammar includes a value that is common to at least one other embedded grammar in the natural language understanding system.

15. A computer program product comprising:

a computer usable medium having computer usable program code that removes bias from an action classifier within a natural language understanding system, said computer program product including;

computer usable program code that identifies a sentence having a target embedded grammar that overlaps with at least one other embedded grammar;

computer usable program code that selects a group of overlapping embedded grammars comprising the target embedded grammar and at least one additional embedded grammar;

computer usable program code that creates a sentence expansion comprising the sentence including the target embedded grammar and a copy of the sentence for each additional embedded grammar of the group, with each copy of the sentence including a different additional embedded grammar from the group in place of the target embedded grammar; and

computer usable program code that adds the sentence expansion to action classifier training data.

16. The computer program product of claim 15, further comprising computer usable program code that trains the action classifier of the natural language understanding system with the action classifier training data comprising the sentence expansion.

17. The computer program product of claim 15, further comprising computer usable program code that defines the group of overlapping embedded grammars as a set of embedded grammars having at least one value in common.

18. The computer program product of claim 15, further comprising computer usable program code that receives a user input defining the group of overlapping embedded grammars.

19. The computer program product of claim 15, further comprising computer usable program code that automatically sets a count of a sentence according to the target embedded grammar or the additional embedded grammar used to replace the target embedded grammar.

20. The computer program product of claim 15, further comprising:

computer usable program code that determines whether the sentence is associated with a selected action; and

computer usable program code that only creates a sentence expansion if the sentence is associated with the selected action.