Apparatus And Method To Reduce Recognition Errors Through Context Relations Among Dialogue Turns

• US 20080215320A1
• Filed: 08/01/2007
• Published: 09/04/2008
• Est. Priority Date: 03/03/2007
• Status: Active Grant

• Current Assignee: Industrial Technology Research Institute
• Original Assignee: Industrial Technology Research Institute
• Inventors: Lee, Ching-Hsien, Wu, Hsu-Chih
• Granted Patent: US 7,890,329 B2
• Time in Patent Office: Days
• US Class Current: 704/231
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/08 Speech classification or search
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Disclosed is directed an apparatus and method to reduce recognition errors through context relations among multiple dialogue turns. The apparatus includes a rule set storage unit having a rule set containing one or more rules, an evolutionary rule generation module connected to the rule storage unit, and a rule trigger unit connected to the rule storage unit. The rule set uses dialogue turn as a unit for the information described by each rule. The method analyzes a dialogue history through an evolutionary massive parallelism approach to get a rule set describing the context relation among dialogue turns. Based on the rule set and recognition result of an ASR system, it reevaluates the recognition result, and measures the confidence measure of the reevaluated recognition result. After each successful dialogue turn, the rule set is dynamically adapted.

Claims

1. An apparatus to reduce recognition error through context relation among dialogue turns, comprising:

a rule storage unit having a rule set containing one or more rules, each rule being represented with the dialogue turn as basic element;

an evolutionary rule generation module connected to said rule storage unit, for training said rule set by performing an evolutionary adaptation on a dialogue log to obtain a trained rule set; and

a rule trigger unit connected to said rule storage unit, based on said trained rule set and the dialogue history of a plurality of previous dialogue turns, said rule trigger unit selecting at least a rule and the corresponding confidence measure from said trained set for providing an automatic speech recognition system to reevaluate the speech recognition.

2. The apparatus as claimed in claim 1, wherein said reevaluation result is fed back to said dialogue log, and said apparatus adjusts said rule set through a reward/punishment unit.

3. The apparatus as claimed in claim 1, wherein each rule of said rule set describes at least the context relation of said dialogue turn.

4. The apparatus as claimed in claim 1, wherein the information described by each rule of said rule set includes a series of context types of previous dialogue turns, the context type of the current dialogue turn, and the corresponding confidence measure of said rule.

5. The apparatus as claimed in claim 4, wherein said information described by said rule of said rule set has one or more context types.

6. The apparatus as claimed in claim 4, wherein said confidence measure of said rule is the confidence score of said rule.

7. The apparatus as claimed in claim 5, wherein said rule set provides a universal context type, and said universal context type represents any one type of said one or more context types.

8. The apparatus as claimed in claim 1, wherein said evolutionary rule generation module has three operation elements, including rule variation, rule evaluation and rule selection.

9. A method to reduce recognition error through context relation among dialogue turns, said method comprising:

analyzing a dialogue historic log to train a rule set through a massive parallelism evolutionary approach, said rule set describing context relations among one or more dialogue turns;

reevaluating said recognition result, and measuring the confidence measure of said reevaluated recognition result based on said rule set and recognition result of an automatic speech recognition (ASR) system; and

dynamically adjusting said rule set for each successful dialogue turn.

10. The method as claimed in claim 9, wherein said training said rule set step further includes:

randomly generating a rule set;

training said random rule set through operations of rule variation, rule evaluation and rule selection of an evolutionary approach, for a plurality of iterations of evolution to obtain a trained rule set.

11. The method as claimed in claim 9, wherein said rule set contains one or more rules, and the information described by each said rule uses dialogue turn as a unit.

12. The method as claimed in claim 11, wherein said rule set uses the following steps to describe the context relation among one or more dialogue turns:

defining one ore more context types for the attributes of the dialogue context of the one or more dialogue turns; and

representing each said rule as M₁M₂M₃ . . . M_n;

R,I, wherein M₁M₂M₃ . . . M_n represents the context type of previous n dialogue turns, ‘

R’

represents the context type of the current dialogue turn, and ‘

I’

represents the confidence measure corresponding to said rule.

13. The method as claimed in claim 10, wherein said rule variation is that each rule of said rule set has a probability to become a new rule through variation or combination.

14. The method as claimed in claim 10, wherein said rule evaluation is the evaluation of the confidence measure for each rule of said rule set.

15. The method as claimed in claim 10, wherein said rule selection further includes:

reserving a certain percentage of rules;

generating new rules randomly or through said rule variation among existing rules;

finding equivalent rules and eliminating more generic rules from said equivalent rules; and

returning to said new rule generation step if one of said rules being eliminated.

16. The method as claimed in claim 9, wherein said reevaluation step further includes:

applying a dialogue log of a plurality of dialogue turns to each rule of said rule set to find rules having context types of previous dialogue turns matching said dialogue log; and

classifying said matched rules according to the type of said current dialogue turn of said rule, and computing the confidence measure score of each said context type.

17. The method as claimed in claim 16, wherein said confidence score is provided to said ASR system after said dynamic adjustment of said rule set.

18. The method as claimed in claim 17, wherein said confidence score allows said ASR system to generate an accurate N-best list.

19. The method as claimed in claim 17, wherein said confidence score is provided to said ASR system for post-processing to adjust the scores of the original N-best list.