Natural language processing of disfluent sentences

• US 20070078642A1
• Filed: 10/04/2005
• Published: 04/05/2007
• Est. Priority Date: 10/04/2005
• Status: Active Grant

• Current Assignee: Robert Bosch GmbH
• Original Assignee: Robert Bosch GmbH
• Inventors: Zhang, Qi, Weng, Fuliang
• Granted Patent: US 7,930,168 B2
• Time in Patent Office: Days
• US Class Current: 704/4
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 15/19 Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules

Abstract

An advanced model that includes new processes is provided for use as a component of an effective disfluency identifier. The disfluency identifier tags edited words in transcribed speech. A speech recognition unit in combination with a part-of-speech tagger, a disfluency identifier, and a parser form a natural language system that helps machines properly interpret spoken utterances.

Claims

1. A method for processing spoken language comprising:

converting spoken words into a text word sequence;

tagging words in the text word sequence with part-of-speech (POS) tags;

tagging edited words in the text word sequence using a disfluence identifier that operates with a feature set created with techniques comprising;

matching only the highest level POS tags in a multi-level hierarchy of such tags; and

,

parsing the text word sequence into machine instructions with the aid of POS-tag and edited-word-tag information.

2. A method as in claim 1 wherein the highest level POS tags in a multi-level hierarchy of such tags comprise tags corresponding to the categories noun-related, verb-related, noun-modifiers and verb-modifiers respectively.

3. A method as in claim 1 wherein the disfluence identifier operates with a feature set created with techniques further comprising allowing single mismatches in POS-tag sequences of rough copy.

4. A method as in claim 1 wherein the disfluence identifier operates with a feature set created with techniques further comprising including distance to next identical orthographic word as a conditioning variable.

5. A method as in claim 4 wherein the disfluence identifier operates with a feature set created with techniques further comprising using a word window size of plus or minus five positions.

6. A method as in claim 5 wherein the disfluence identifier operates with a feature set created with techniques further comprising using a POS-tag window size of plus or minus five positions.

7. A method as in claim 6 wherein the disfluence identifier operates with a feature set created with techniques further comprising using a partial-word window size of plus or minus three positions.

8. A method as in claim 1 further comprising removing sequence-related errors in edited-word-tag information before parsing the text word sequence.

9. A system for processing spoken language comprising:

a speech recognition unit for converting spoken words into a text word sequence;

a part-of-speech (POS) tagger for tagging words in the text word sequence with part-of-speech tags;

a disfluence identifier for tagging edited words in the text word sequence;

wherein, the disfluence identifier operates with a feature set created with techniques comprising;

matching only the highest level POS tags in a multi-level hierarchy of such tags; and

,

a parser for parsing the text word sequence into machine instructions with the aid of POS-tag and edited-word-tag information.

10. A system as in claim 9 wherein the highest level POS tags in a multi-level hierarchy of such tags comprise tags corresponding to the categories noun-related, verb-related, noun-modifiers and verb-modifiers respectively.

11. A system as in claim 9 wherein the disfluence identifier operates with a feature set created with techniques further comprising allowing single mismatches in POS-tag sequences of rough copy.

12. A system as in claim 9 wherein the disfluence identifier operates with a feature set created with techniques further comprising including distance to next identical orthographic word as a conditioning variable.

13. A system as in claim 12 wherein the disfluence identifier operates with a feature set created with techniques further comprising using a word window size of plus or minus five positions.

14. A system as in claim 13 wherein the disfluence identifier operates with a feature set created with techniques further comprising using a POS-tag window size of plus or minus five positions.

15. A method as in claim 14 wherein the disfluence identifier operates with a feature set created with techniques further comprising using a partial-word window size of plus or minus three positions.

16. A method for creating a disfluence identifier model comprising:

analyzing the distribution of speech repairs in transcribed speech;

choosing conditioning variables commensurate with the distribution of speech repairs;

using a rough copy identifier with the conditioning variables to generate a feature set; and

,

weighting the feature set according to an iterative algorithm run on training data.

17. A method as in claim 16 wherein the conditioning variables chosen comprise a variable representing the distance to the next identical orthographic word.

18. A method as in claim 16 wherein the rough copy identifier matches only the highest level POS tags in a multi-level hierarchy of such tags.

19. A method as in claim 16 wherein the rough copy identifier allows single mismatches in POS-tag sequences of rough copy.