Dialogue management using scripts and combined confidence scores

• US 20060271364A1
• Filed: 12/08/2005
• Published: 11/30/2006
• Est. Priority Date: 05/31/2005
• Status: Active Grant

• Current Assignee: Robert Bosch Corporation (Robert Bosch GmbH)
• Original Assignee: Robert Bosch Corporation (Robert Bosch GmbH)
• Inventors: Xu, Kui, Zhang, Qi, Purver, Matthew, Ratiu, Florin, Scheideck, Tobias, Weng, Fuliang, Cavedon, Lawrence, Mirkovic, Danilo
• Granted Patent: US 7,904,297 B2
• Time in Patent Office: Days
• US Class Current: 704/239
• CPC Class Codes: G06F 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Representation-neutral dialogue systems and methods (“RNDS”) are described that include multi-application, multi-device spoken-language dialogue systems based on the information-state update approach. The RNDS includes representation-neutral core components of a dialogue system that provide scripted domain-specific extensions to routines such as dialogue move modeling and reference resolution, easy substitution of specific semantic representations and associated routines, and clean interfaces to external components for language-understanding (i.e., speech-recognition and parsing) and language-generation, and to domain-specific knowledge sources. The RNDS also resolves multi-device dialogue by evaluating and selecting among candidate dialogue moves based on features at multiple levels. Multiple sources of information are combined, multiple speech recognition and parsing hypotheses tested, and multiple device and moves considered to choose the highest scoring hypothesis overall. Confirmation and clarification behaviour can be governed by the overall score.

Claims

1. A method comprising:

receiving an input pattern generated from an acoustic signal uttered by a user;

syntactically labeling the input pattern to generate a syntactic label;

receiving a confidence score from a speech recognizer component;

mapping the confidence score to a qualitative confidence measure;

incorporating the qualitative confidence measure into the syntactic label of the input pattern;

comparing the input pattern to patterns in a dialogue move script to determine a possible dialogue move; and

defining a state of a confirmation flag associated with the input pattern based on the existence of a match between the input pattern and the patterns in the dialogue move script.

2. The method of claim 1, wherein the input pattern is syntactically labeled by one of a speech recognizer process and a parser process.

3. The method of claim 2, wherein the input pattern generated from the acoustic signal includes a noun phrase, the method further comprising:

performing a noun phrase resolution process on the noun phrase of the input pattern to determine a noun phrase resolution result;

transmitting a confirmation question to the user if the confirmation flag is in a first pre-defined state; and

performing a dialogue move operation corresponding at least one of the noun phrase resolution result and the user response to the confirmation question.

4. The method of claim 1, wherein the confidence score comprises a numerical value representing a probability of proper recognition of the input pattern, and wherein the qualitative confidence measure comprises one of a low, medium, and high qualitative value.

5. The method of claim 1 wherein the input pattern comprises a sentence and further comprising:

identifying proper names within the input pattern;

replacing identified proper names with associated type classifications;

labeling the input pattern with a classification denoting a class of a plurality of classes;

modeling the input pattern using a conditional maximum entropy method to characterize the input pattern as a class and a feature; and

determining a topic defined within a dialogue move script to which the input pattern corresponds.

6. The method of claim 5 further comprising:

determining a dialogue move confidence score quantifying a probability that the input pattern corresponds to the topic; and

combining the dialogue move confidence score with a confidence score received from the speech recognizer component to produce a combined confidence score that is mapped to the qualitative confidence measure.

7. A method comprising:

receiving an input pattern generated from an input utterance that includes a description of a dialogue contribution from a speaker through an input component;

receiving a confidence score for each feature of the input utterance for each device of one or more devices coupled to the input component;

assigning a weight to each feature to generate a weighted confidence score for each device;

combining the weighted confidence scores for the one or more devices into a combined confidence score to rate one or more possible dialogue move candidates as the interpretation of the input utterance;

defining a first confidence threshold to specify a level at which a highest scoring dialogue move candidate is accepted; and

defining a second confidence threshold to specify a level at which the highest scoring dialogue move candidate is rejected.

8. The method of claim 7 wherein the features include at least one of confidence scores from the speech recognizer component and a parser coupled to the speech recognizer, semantic criteria, pragmatic criteria, and dialogue context.

9. The method of claim 8 further comprising:

mapping the combined confidence score to a qualitative confidence measure;

formulating a confirmation question to be sent to the speaker; and

transmitting the confirmation question to the speaker if the confidence measure is between the first confidence threshold and the second confidence threshold.

10. The method of claim 9, further comprising:

identifying the description and at least one parameter of the description using a dialogue move script (DMS), wherein the DMS corresponds to at least one of a device of the one or more devices and an application; and

mapping the description to a dialogue move using the DMS, the dialogue move corresponding to the identified parameter and a response to the confirmation question.

11. The method of claim 10, wherein the qualitative confidence measure comprises one of a low, medium and high syntactic value.

12. The method of claim 11, further comprising rejecting the candidate move if the combined confidence score of the candidate dialogue move is below the second confidence threshold value.

13. The method of claim 12, further comprising accepting the candidate move if the combined confidence score of the candidate dialogue move is above the first confidence threshold value.

14. The method of claim 7, wherein the input component comprises one of a speech recognizer process, a parser, and a dialogue move classifier process.

15. The method of claim 14 further comprising:

identifying proper names within the input pattern;

replacing identified proper names with associated type classifications;

labeling the input pattern with a classification denoting a class of a plurality of classes;

modeling the input pattern using a conditional maximum entropy method to characterize the input pattern as a class and a feature; and

determining a topic defined within a dialogue move script to which the input pattern corresponds.

16. A system comprising:

an input component for receiving and syntactically labeling an input pattern generated from an input utterance by a user, the input component including one or more subunits configured to generate a confidence score representing a probability or proper recognition of the input utterance;

a dialogue manager coupled to the input component that includes a plurality of dialogue moves and a dialogue move script that corresponds to at least one of a device and an application; and

a confidence mapping component coupled to the dialogue manager and configured to receive a confidence score from the input component, map the confidence score to a qualitative confidence measure, and incorporate the qualitative confidence measure into the syntactic label of the input pattern.

17. The system of claim 16 wherein the input component comprises at least one of an automatic speech recognizer and a parser.

18. The system of claim 17 further comprising one or more devices coupled to the dialogue manager, each device of the one or more devices configured to generate a confidence score for each feature of the input utterance, the dialogue manager configured to assign a weight to each feature to generate a weighted confidence score for each device of the one or more devices, and combine the weighted confidence scores for the one or more devices into a combined confidence score to rate one or more possible dialogue move candidates as the interpretation of the input utterance.

19. The system of claim 18 wherein the features include at least one of confidence scores from the speech recognizer component and the parser, the combined confidence score, semantic criteria, pragmatic criteria, and dialogue context.

20. The system of claim 19, wherein the qualitative confidence measure comprises one of a low, medium and high syntactic value, the dialogue manager further configured to:

define a first confidence threshold corresponding to the high syntactic value to specify a level at which a highest scoring dialogue move candidate is accepted;

define a second confidence threshold corresponding to the low syntactic value to specify a level at which the highest scoring dialogue move candidate is rejected;

compare the input pattern to patterns in a dialogue move script to determine a possible dialogue move.

21. A method comprising:

receiving an input pattern generated from an acoustic signal that includes a description of a dialogue contribution from a speaker through a speech recognizer component;

identifying proper names within the input pattern;

replacing identified proper names with associated type classifications;

labeling the input pattern with a classification denoting a class of a plurality of classes;

modeling the input pattern using a conditional maximum entropy method to characterize the input pattern as a class and a feature; and

determining a topic defined within a dialogue move script to which the input pattern corresponds.

22. The method of claim 21 further comprising determining a dialogue move confidence score quantifying a probability that the input pattern corresponds to the topic.

23. The method of claim 22 wherein the classes comprise one of a statement, command, correction, acknowledgment, query, answer, confirmation acceptance, partial acceptance, and a rejection.

24. The method of claim 23 further comprising:

receiving a keyword list;

receiving a catalog containing one or more instances of the keyword;

calculating a frequency of occurrences of a keyword in the catalog;

associating the keyword and the associated classification with the frequency; and

modeling the keyword using the conditional maximum entropy method to characterize the keyword as the identified class and the frequency.

25. The method of claim 22 further comprising combining the dialogue move confidence score with at least one of a speech recognizer confidence score generated by the speech recognizer, a parser confidence score generated by a parser coupled to the speech recognizer, and a device confidence score coupled to one or more dialogue-enabled devices coupled to a dialogue manager coupled to the speech recognizer, to produce a combined confidence score.

26. The method of claim 25 further comprising:

syntactically labeling the input pattern to generate a syntactic label;

mapping the confidence score to a qualitative confidence measure; and

incorporating the qualitative confidence measure into the syntactic label of the input pattern.

27. The method of claim 26 wherein the qualitative confidence measure comprises one of a low, medium and high syntactic value, and wherein the input pattern is syntactically labeled by one of the speech recognizer and the parser.