Indexing with translation model for feature regularization

• US 6,925,436 B1
• Filed: 01/28/2000
• Issued: 08/02/2005
• Est. Priority Date: 01/28/2000
• Status: Expired due to Term

• Current Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Original Assignee: International Business Machines Corporation
• Inventors: Franz, Martin, McCarley, Jeffrey Scott
• Primary Examiner(s): ABEBE, DANIEL DEMELASH
• Application Number:
  US09/493,507
• Time in Patent Office: 2,013 Days
• Field of Search: 707/102, 704/10, 704/200, 704/203, 704/231, 704/235, 704/246, 704/243, 704/251, 704/277, 704/278
• US Class Current: 704/235
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: Y10S 707/99943 Generating database or data structure, e.g. via user interface

Abstract

An audio indexing system including, in addition to a speech recognition subsystem for converting the audio information into a textual form and an indexing subsystem for extracting the features to be used for searching and browsing, a statistical machine translation model, trained on a parallel or comparable corpus of automatically and by-hand transcribed data, for processing the output of the speech recognition system.

Claims

1. A system for providing clean data for indexing, said system comprising:

a recognizer which recognizes words;

an indexing database; and

a translator, having been trained automatically, which accepts textual input, having originated from said recognizer, and which is adapted to automatically improve the quality of said textual input for entry into said indexing database, whereby errors produced by said recognizer and which are detrimental to indexing performance are reduced, and wherein, immediately prior to reconfiguration, said textual input appears as a feature-extracted transformation of at least one word recognized by said recognizer.

2. The system according to claim 1, further comprising a feature extractor which transforms words recognized by said recognizer into predetermined textual features and which provides the textual features as the textual input into said translator.

3. The system according to claim 2, wherein the textual features comprise morphs of words recognized by said recognizer.

4. The system according to claim 2, wherein the textual features comprise stems of words recognized by said recognizer.

5. The system according to claim 1, wherein said recognizer is adapted to transform recognized words into predetermined textual features and provide the textual features as the textual input into said translator.

6. The system according to claim 1, wherein said recognizer is adapted to provide the textual input into said translator and said translator is adapted to transform the textual input into predetermined textual features prior to reconfiguration of the textual input for entry into said indexing database.

7. The system according to claim 1, wherein:

said recognizer is a speech recognizer for recognizing verbal audio input; and

said indexing database is an audio indexing database.

8. The system according to claim 1, wherein said translator is trained on a corpus of automatically and by-hand transcribed data, the data originating from a domain substantially similar to that intended to be addressed by said recognizer.

9. The system according to claim 1, further comprising a temporary storage medium for storing the textual input for a period of time prior to its being accepted by said translator, so as to delay reconfiguration by said translator.

10. A method of providing clean data for indexing, said method comprising the steps of:

providing an indexing database;

providing a recognizer which recognizes words; and

providing an automatically trained translator which accepts textual input having originated from said recognizer;

said method further comprising the steps of;

with said recognizer, recognizing words; and

with said translator, accepting textual input having originated from said recognizer, and automatically improving the quality of said textual input for entry into said indexing database, whereby errors produced by said recognizer and which are detrimental to indexing performance are reduced, and wherein, immediately prior to reconfiguration, said textual input appears as a feature-extracted transformation of at least one word recognized by said recognizer.

11. The method according to claim 10, further comprising the steps of:

providing a feature extractor; and

with said feature extractor, transforming words recognized by said recognizer into predetermined textual features and providing the textual features as the textual input into said translator.

12. The method according to claim 11, wherein said step of transforming words into textual features comprises transforming words recognized by said recognizer into morphs.

13. The method according to claim 11, wherein said step of transforming words into textual features comprises transforming words recognized by said recognizer into stems.

14. The method according to claim 10, further comprising the step of:

with said recognizer, transforming recognized words into predetermined textual features and providing the textual features as the textual input into said translator.

15. The method according to claim 10, further comprising the steps of:

with said recognizer, providing the textual input into said translator; and

with said translator, transforming the textual input into predetermined textual features prior to said reconfiguring of the textual input for entry into said indexing database.

16. The method according to claim 10, wherein:

said step of providing a recognizer comprises providing a speech recognizer for recognizing verbal audio input; and

said step of providing an indexing database comprises providing an audio indexing database.

17. The method according to claim 10, wherein said step of providing a translator comprises providing a translator that is trained on a corpus of automatically and by-hand transcribed data, the data originating from a domain substantially similar to that intended to be addressed by said recognizer.

18. The method according to claim 10, further comprising the steps of:

providing a temporary storage medium; and

with said temporary storage medium, storing the textual input for a period of time prior to its being accepted by said translator, so as to delay said reconfiguring by said translator.

19. A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for providing clean data for indexing, said method comprising the steps of:

providing an indexing database;

providing a recognizer which recognizes words; and

providing an automatically trained translator which accepts textual input having originated from said recognizer;

said method further comprising the steps of;

with said recognizer, recognizing words; and

with said translator, accepting textual input having originated from said recognizer, and automatically improving the quality of said textual input for entry into said indexing database, whereby errors produced by said recognizer and which are detrimental to indexing performance are reduced, and wherein, immediately prior to reconfiguration, said textual input appears as a feature-extracted transformation of at least one word recognized by said recognizer.