System and method for resolving decoding ambiguity via dialog

• US 6,587,818 B2
• Filed: 10/28/1999
• Issued: 07/01/2003
• Est. Priority Date: 10/28/1999
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Samdani, Taufique, Kanevsky, Dimitri, Zlatsin, Alexander, Zadrozny, Wlodek
• Primary Examiner(s): Banks-Harold, Marsha D.
• Assistant Examiner(s): Lerner, Martin
• Application Number:
  US09/428,949
• Publication Number:
  US 20030004714A1
• Time in Patent Office: 1,342 Days
• Field of Search: 704/243, 704/244, 704/251, 704/252, 704/255, 704/270, 704/275, 379/88.01, 379/88.03, 706/45, 706/46, 706/52, 382/187, 382/188, 382/189
• US Class Current: 704/251
• CPC Class Codes: G10L 15/18 using natural language modelling
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/183 using context dependencies, e.g. language models
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 15/24 Speech recognition using non-acoustical features
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence

Abstract

A method of language recognition wherein decoding ambiguities are identified and at least partially resolved intermediate to the language decoding procedures to reduce the subsequent number of final decoding alternatives. The user is questioned about identified decoding ambiguities as they are being decoded. There are two language decoding levels: fast match and detailed match. During the fast match decoding level a large potential candidate list is generated, very quickly. Then, during the more comprehensive (and slower) detailed match decoding level, the fast match candidate list is applied to the ambiguity to reduce the potential selections for final recognition. During the detailed match decoding level a unique candidate is selected for decoding. Decoding may be interactive and, as each ambiguity is encountered, recognition suspended to present questions to the user that will discriminate between potential response classes. Thus, recognition performance and accuracy is improved by interrupting recognition, intermediate to the decoding process, and allowing the user to select appropriate response classes to narrow the number of final decoding alternatives.

Claims

1. A language recognition system for recognizing language input, said language input including ambiguous statements, said system being capable of recognizing and resolving certain ones of said ambiguous statements, said ambiguous statements being recognized as acceptable answers to two or more of who, what, when and where questions, said system comprising:

a decoder system receiving language input checking sentences in said language input for ambiguities and transcribing unambiguous language input;

an intermediate decoding module identifying a set of decoding alternatives corresponding to an ambiguity in an identified ambiguous sentence;

a classifier classifying each of said set of decoding alternatives as belonging to one ox a set of classes;

a questioner module constructing optimal questions responsive to said set of classes, said optimal questions being constructed to reduce the number of classes; and

an assistant interactive system providing constructed said optimal questions and receiving corresponding responses.

2. A language recognition system as in claim 1, said decoder system comprising:

a fast match decode module receiving said language input and generating a potential candidate list for said ambiguous sentences; and

a detailed match decode module receiving said potential candidate list from said fast match decode module when ambiguities are not found in said language input and receiving a classified input when said candidate list includes one or more ambiguities.

3. A Speech recognition system as in claim 2, wherein the decoder system is an automatic speech recognition system, said speech recognition system further comprising a microphone, speech input being provided from said microphone in response to a user speaking into said microphone.

4. A handwriting recognition system as in claim 2, wherein the decoder system is an automatic handwriting recognition system.

5. A language recognition method, said method comprising the steps of:

a) receiving language input and checking said language input for sentences recognized as acceptable answers to two or more of who, what, when and where questions, each recognized of said sentences being identified as an ambiguous sentence;

b) converting said language input to an output when an ambiguity is not found in said language input and returning to step (a);

otherwise,

c) identifying a set of intermediate decoding alternatives for an identified said ambiguous sentence;

d) identifying a plurality of final decoding class alternatives from said set of intermediate decoding alternatives;

e) identifying a final decoding class from said plurality of identified final decoding class alternatives;

f) presenting questions distinguishing features of members of said identified final decoding class; and

g) resolving any ambiguity in said ambiguous sentence responsive to responses to said questions and returning to step (b).

6. A language recognition method as in claim 5, wherein the set of intermediate decoding alternatives identified in step (c) include a candidate word list, a candidate phrase list, a candidate sentence list, a vocabulary, a fast match list, a detailed match list, an N-best list, a candidate acoustic feature set, a dialects and/or language set, a handwriting feature set, a phoneme set, a letter set a channel, and an environment condition set.

7. A language recognition method as in claim 6, wherein the final decoding alternatives identified in step (d) include a candidate word list, a candidate phrase list, a candidate sentence list, a vocabulary, a fast match list, a detailed match list, an N-best list, a candidate acoustic feature set, a dialects and/or language set, a handwriting feature set, a phoneme set, a letter set, a channel and an environment condition set.

8. A language recognition method as in claim 5, wherein the questions presented in step (f) are selected to classify the identified ambiguous sentence as being related to one of at least a pair of classes selected from a plurality of feature classes.

9. A language recognition method as in claim 8, wherein the plurality of feature classes comprise:

a temporal class;

a spatial class;

a durational class;

a semantic class;

a grammatical class;

a personal characteristic class;

a topical class;

a goal related class;

a business model class; and

,

a customer class.

10. A language recognition method as in claim 9, wherein the grammatical class comprises sentence structure, sentence parsing and word classification.

11. A language recognition method as in claim 9, wherein the personal characteristic class comprises a plurality of user characteristics including a personal profile, user sex, user age and user profession.

12. A language recognition method as in claim 9, wherein the topical class comprises a medical information class, a legal information class, a business information class and a personal information class.

13. A language recognition method as in claim 9, wherein the business model class includes business related activities including placing ticket orders, placing orders for goods, requesting information and making purchases.

14. A language recognition method as in claim 8, wherein the step (f) of presenting questions further comprises selecting questions from a plurality of questions, selected said questions being an optimum set of questions, said optimum set of questions selected to minimize the number of questions presented to a user and to minimize the number of potential alternatives remaining after receiving a user response.

15. A language recognition method as in claim 14, wherein the set of optimum questions are selected from a plurality of questions by determining a probability metric, the probability metric providing a measure of the probability that a response will eliminate one or more of said final decoding classes.

16. A language recognition method as in claim 15, wherein the probability metric is derived from training data.

17. A language recognition method as in claim 16, wherein the training data is a textual corpus labeled with classes.

18. A language recognition method as in claim 16, wherein the training data is a plurality of transcribed user-server dialogs labeled with classes.

19. A computer program product for language recognition, said computer program product comprising a computer usable medium having computer readable program code thereon, said computer readable program code comprising:

computer readable program code means for checking language input for ambiguous statements, ambiguous statements being recognized as acceptable answers to two or more of who, what, when and where questions;

computer readable program code means for converting said language input to an output;

computer readable program code means for identifying a set of intermediate decoding alternatives for an identified ambiguous statement;

computer readable program code means for identifying final decoding class alternatives for each said identified ambiguous statement from said set of intermediate decoding alternatives;

computer readable program code means for identifying a final decoding class from said identified final decoding class alternatives;

computer readable program code means for presenting questions about features of members of said identified final decoding class; and

computer readable program code means for using final decoding class member features to resolve ambiguities in identified ambiguous statements.

20. A computer program product as in claim 19, wherein the computer readable program code means for presenting questions comprises:

computer readable program code means for classifying each said identified ambiguous statement as being related to one of at least a pair of classes of a plurality of feature classes.

21. A computer program product as in claim 20, wherein the computer readable program code means for presenting questions further comprises:

computer readable program code means for selecting an optimum set of questions from a plurality of questions, said optimum set of questions minimizing the number of questions presented and minimizing the number of potential alternatives remaining after receiving response to each presented question.

22. A computer program product as in claim 21, wherein the computer readable program code means for presenting questions further comprises:

computer readable program code means for determining a probability metric from a plurality of questions, the probability metric being a measure of the probability that a response will eliminate one or more of said final decoding classes.

23. A computer program product as in claim 22, wherein the computer readable program code means for presenting questions further comprises:

computer readable program code means for deriving the probability metric from training data.