Systems and methods for determining predictive models of discourse functions

• US 7,542,903 B2
• Filed: 02/18/2004
• Issued: 06/02/2009
• Est. Priority Date: 02/18/2004
• Status: Expired due to Fees

• Current Assignee: Fuji Xerox, Fuji Xerox Company Limited (Xerox Holdings Corp.)
• Original Assignee: Fuji Xerox Company Limited (Xerox Holdings Corp.)
• Inventors: Azara, Misty, Polanyi, Livia, Van Den Berg, Martin H., Thione, Giovanni L.
• Primary Examiner(s): Edouard; Patrick N
• Assistant Examiner(s): GODBOLD, DOUGLAS
• Application Number:
  US10/781,443
• Publication Number:
  US 20050182625A1
• Time in Patent Office: 1,931 Days
• Field of Search: 704/9, 704/257, 704/8, 704/1, 704/10
• US Class Current: 704/257
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 25/51 for comparison or discrimination
• CPC Class Codes: G10L 25/90 Pitch determination of speech signals

Abstract

Techniques are provided for determining predictive models of discourse functions based on prosodic features of natural language speech. Inter and intra sentential discourse functions in a training corpus of natural language speech utterances are determined. The discourse functions are clustered. The exemplary prosodic features associated with each type of discourse function are determined. Machine learning, observation and the like are used to determine a subset of prosodic features associated with each type of discourse function useful in predicting the likelihood of each type of discourse function.

Claims

1. A method of determining a predictive model for discourse functions comprising the steps of:

determining a training corpus of speech utterances;

determining discourse functions associated with speech utterances in the training corpus, the discourse functions being determined automatically based on a theory of discourse analysis;

determining prosodic features associated with the speech utterances in the training corpus; and

determining a predictive model of discourse functions by associating the prosodic features determined from the speech utterances in the training corpus with the discourse functions determined from the speech utterances in the training corpus,

wherein the predictive model of discourse functions is used to predict from prosodic features of a specific recognized speech, a likelihood that speech utterances of the specific recognized speech reflect a specific discourse function, and

wherein the predictive model of discourse functions is used to predict, based, at least in part, on the prosodic features, a likelihood of a first portion of a speech utterance being associated with a command directed at an application and a second portion of the speech utterance being associated with content being provided to the application.

2. The method of claim 1, in which the theory of discourse analysis is at least one of:

the Linguistic Discourse Model, the Unified Linguistic Discourse Model, Rhetorical Structure Theory, Discourse Structure Theory and Structured Discourse Representation Theory.

3. The method of claim 1, in which the predictive models are determined based on at least one of:

machine learning, rules.

4. The method of claim 3, in which the machine learning based predictive models are determined based on at least one of:

statistics, decision trees, Naï

ve Bayes.

5. The method of claim 1, in which the prosodic features occur in at least one of a location:

preceding, within and following the associated discourse function.

6. The method of claim 1, in which the prosodic features are encoded within a prosodic feature vector.

7. The method of claim 6, in which the prosodic feature vector is a multimodal feature vector.

8. The method of claim 1, in which the discourse function is an intra-sentential discourse function.

9. The method of claim 1, in which the discourse function is an inter-sentential discourse function.

10. A system for determining predictive discourse function models comprising:

an input/output circuit for retrieving a training corpus of speech utterances; and

a processor for;

determining prosodic features associated with speech utterances in the training corpus,

determining discourse functions associated with the speech utterances in the training corpus, the discourse functions being determined automatically based on a theory of discourse analysis, and

determining a predictive model for discourse functions by associating the prosodic features determined from the speech utterances in the training corpus with the discourse functions determined from the speech utterances in the training corpus,

wherein the predictive model of discourse functions is used to predict from prosodic features of a specific recognized speech, a likelihood that speech utterances of the specific recognized speech reflect a specific discourse function, and

wherein the predictive model of discourse functions is used to predict, based, at least in part, on the prosodic features, a likelihood of a first portion of a speech utterance being associated with a command directed at an application and a second portion of the speech utterance being associated with content being provided to the application.

11. The system of claim 10, in which the theory of discourse analysis is at least one of:

the Linguistic Discourse Model, the Unified Linguistic Discourse Model, Rhetorical Structure Theory, Discourse Structure Theory and Structured Discourse Representation Theory.

12. The system of claim 10, in which the predictive models are determined based on at least one of:

machine learning, rules.

13. The system of claim 12, in which the machine learning based predictive models are determined based on at least one of:

statistics, decision trees, Naï

ve Bayes.

14. The system of claim 10, in which the prosodic features occur in at least one of a location:

preceding, within and following the associated discourse function.

15. The system of claim 10, in which the prosodic features are encoded within a prosodic feature vector.

16. The system of claim 15, in which the prosodic feature vector is a multimodal feature vector.

17. The system of claim 10, in which the discourse function is an intra-sentential discourse function.

18. The system of claim 10, in which the discourse function is an inter-sentential discourse function.

19. Computer readable storage medium comprising:

computer readable program code embodied on the computer readable storage medium, the computer readable program code usable to program a computer to determine a predictive model for discourse functions comprising the steps of;

determining a training corpus of speech utterances;

determining discourse functions associated with speech utterances in the training corpus of speech utterances, the discourse functions being determined automatically based on a theory of discourse analysis;

determining prosodic features associated with the speech utterances in the training corpus of speech utterances; and

determining at least one predictive model of discourse functions by associating the prosodic features with the discourse functions,

wherein the predictive model of discourse functions is used to predict from prosodic features of a specific recognized speech, a likelihood that speech utterances of the specific recognized speech reflect a specific discourse function, and

wherein the predictive model of discourse functions is used to predict, based, at least in part, on the prosodic features, a likelihood of a first portion of a speech utterance being associated with a command directed at an application and a second portion of the speech utterance being associated with content being provided to the application.