Method and apparatus for recognizing large list of proper names in spoken dialog systems

• US 7,925,507 B2
• Filed: 07/07/2006
• Issued: 04/12/2011
• Est. Priority Date: 07/07/2006
• Status: Active Grant

• Current Assignee: Robert Bosch Corporation (Robert Bosch GmbH)
• Original Assignee: Robert Bosch Corporation (Robert Bosch GmbH)
• Inventors: Raghunathan, Badri, Feng, Zhe, Scheideck, Tobias, Weng, Fuliang
• Primary Examiner(s): Jackson; Jakieda R
• Application Number:
  US11/483,840
• Publication Number:
  US 20080010058A1
• Time in Patent Office: 1,740 Days
• Field of Search: 704/9, 704/257
• US Class Current: 704/257
• CPC Class Codes: G06F 16/3329 Natural language query formulation or dialogue systems
• CPC Class Codes: G06F 40/295 Named entity recognition
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

Embodiments of a name recognition process for use in dialog systems are described. In one embodiment, the name recognition process assigns weighting values to names used in a dialog based on the usage of these names. This process takes advantage of the general tendency of people to speak names, either full or partial, only after they have heard or read these names. Name input is taken in several different forms, including a static background database that contains all possible names, a background database that contains commonly used names (such as common trademarks or references), a database that contains names from a user model, and a dynamic database that constantly takes the names just mentioned. The names are then appended with proper weighting values. A high weight is given to names that have been mentioned recently, a lower weight is given to common names, and a lowest weight is given to names for the ones that have never been used or mentioned.

Claims

1. A method of optimizing the performance of speech recognition and language understanding in a natural language processing system, comprising:

compiling a first set of full names based on usage of the names during the course of a current spoken dialog session, wherein the first set of full names is stored in a dynamic database, wherein the dynamic database is sourced from data generated by a dialog manager unit in the natural language processing system and that receives input from a natural language understanding unit to interpret input representations in context;

compiling a second set of full names based on presence in a pre-defined knowledge base that is built up from dialog sessions other than the current dialog session, wherein the second set of full names is stored in a static database, wherein the static database is sourced from data generated by a knowledge manager unit in the natural language processing system and that interfaces to one or more knowledge sources;

deriving partial names for one or more of the names of the first set of full names and the second set of full names;

combining the partial names and the first and second set of full names to generate a name model;

assigning weight values to each of the names of the name model, wherein the relative weights depend on the usage of the names in the current dialog session and the presence in the pre-defined knowledge base, and wherein name entries in the dynamic database are weighted higher than name entries in the static database; and

removing from an active vocabulary list those name entries with weight values below a defined minimum value to constrain name candidates processed by the natural language processing system.

2. The method of claim 1, wherein the first set of full names comprise at least one of names that were previously uttered by the user or the system in the dialog, names that can be inferred through semantic indicators in the dialog.

3. The method of claim 2, wherein the second set of full names comprise at least one of all names in the knowledge base, common names recognized by a substantial portion of a population, and names historically associated with a user of the system.

4. The method of claim 3, further comprising:

generating a first model from the name model for use by a speech recognizer nit of the system; and

generating a second model from the name model for use by a natural language understanding unit of the system.

5. The method of claim 4, wherein the first model is one of a class-based n-gram model, a class-based finite state model, and a class-based context-free model.

6. The method of claim 1, wherein the dynamic database comprises names mentioned in a previous context of the current dialog session or names inferred by semantic relations, and the static database comprises all names or common domain names or user preferred names.

7. The method of claim 1 wherein the partial names are derived by a process comprising at least one of pre-defining partial names based on the full names, learning the partial names through a knowledge-based system, applying a set of pre-defined rules to the full names to derive the partial names, and applying an n-gram model produced through statistical means.

8. A method implemented by a processor-based system for recognizing names in spoken dialog comprising:

receiving an input utterance from a user in a speech recognizer;

generating a structured meaning representation of the input utterance based on a statistical model trained on linguistic data and a knowledge base in a natural language understanding unit coupled to the speech recognizer, wherein the knowledge base contains a plurality of names used in the input utterance; and

generating a name model based on context information related to the input utterance and the presence of names in the knowledge base in a name model generator module coupled to the speech recognizer and natural language understanding unit;

wherein the name model is generated by

compiling a first set of full names based on usage of the names during the course of a current dialog session, wherein the first set of full names is stored in a dynamic database, wherein the dynamic database is sourced from data generated by a dialog manager unit in the natural language processing system and that receives input from a natural language understanding unit to interpret input representations in context;

compiling a second set of full names based on presence in a pre-defined knowledge base that is built up from dialog sessions other than the current dialog session, wherein the second set of full names is stored in a static database, wherein the static database is sourced from data generated by a knowledge manager unit in the natural language processing system and that interfaces to one or more knowledge sources; and

deriving partial names for one or more of the names of the first set of full names and the second set of full names, and combining these names to generate the name model, wherein each of the names is assigned a weight value depending on the usage of the names in the current dialog session and the presence in the pre-defined knowledge base, and further wherein name entries in the dynamic database are weighted higher than name entries in the static database, and further wherein, name entries with weight values below a defined minimum value are removed from an active vocabulary list to constrain name candidates processed by the natural language processing system.

9. The method of claim 8, wherein the name model generator module comprises an input stage configured to receive a first plurality of names from a dynamic database and a second plurality of names from a static database.

10. The method of claim 9, wherein the dynamic database includes a first set of full names selected from the group consisting essentially of:

names that were present in the previous utterances from both user and system, and names that can be inferred through semantic indicators in the previous utterances from both user and system.

11. The method of claim 10 wherein the static database includes a second set of full names selected from the group consisting essentially of:

all names in the knowledge base, common names recognized by a substantial portion of a population, and names historically associated with a user of the system.

12. The method of claim 11, further comprising a partial name derivation unit configured to derive partial names from one or more of the names from the first set of full names and the second set of full names.

13. The method of claim 12, wherein the first set of full names and the second set of full names and the partial names are combined to generate a name model.

14. The method of claim 13, wherein the names in the name model are each assigned a weight value, and wherein the relative weights depend on the usage of the names in the user utterance and the presence in the knowledge base.

15. The method of claim 14 wherein a higher relative weight is assigned to names from the dynamic database and to partial names derived from names in the dynamic database.

16. The method of claim 15 wherein the name model is represented by a list, a finite state machine, or an alternate type of context-free model.

17. A non-transitory, machine-readable medium embodied on a physical structure, including instructions which when executed in a processing system perform the following executable steps:

compiling a first set of full names based on usage of the names during the course of a current spoken dialog session, wherein the first set of full names is stored in a dynamic database, wherein the dynamic database is sourced from data generated by a dialog manager unit in the natural language processing system and that receives input from a natural language understanding unit to interpret input representations in context;

compiling a second set of full names based on presence in a pre-defined knowledge base that is built up from dialog sessions other than the current dialog session, wherein the second set of full names is stored in a static database wherein the static database is sourced from data generated by a knowledge manager unit in the natural language processing system and that interfaces to one or more knowledge sources;

deriving partial names for one or more of the names of the first set of full names and the second set of full names;

combining the partial names and the first and second set of full names to generate a name model;

assigning weight values to each of the names of the name model, wherein the relative weights depend on the usage of the names in the current dialog session and the presence in the pre-defined knowledge base, and wherein name entries in the dynamic database are weighted higher than name entries in the static database; and

removing from an active vocabulary list those name entries with weight values below a defined minimum value to constrain name candidates processed by the natural language processing system.

18. The non-transitory medium of claim 17, wherein the first set of full names comprise at least one of names that were previously uttered by the user or the system in the dialog, names that can be inferred through semantic indicators in the dialog, and wherein the second set of full names comprise at least on of all names in the knowledge base, common names recognized by a substantial portion of a population, and names historically associated with a user of the system.

19. The non-transitory medium of claim 17, further comprising instructions that:

generate a first model from the name model for use by a speech recognizer unit of the system; and

generate a second model from the name model for use by a natural language understanding unit of the system.

20. The non-transitory medium of claim 19, wherein a higher relative weight is assigned to names from the first set of full names and to partial names derived from names of the first set of full names.