Processing speech to text queries by optimizing conversion of speech queries to text

• US 10,332,511 B2
• Filed: 07/24/2015
• Issued: 06/25/2019
• Est. Priority Date: 07/24/2015
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Dayan, Yigal S., Magdalen, Josemina M., Maharian, Irit, Mazel, Victoria, Paikowsky, Oren, Shtilman, Andrei
• Primary Examiner(s): Blankenagel, Bryan S
• Application Number:
  US14/808,146
• Publication Number:
  US 20170025120A1
• Time in Patent Office: 1,432 Days
• CPC Class Codes: G06F 16/3334 Selection or weighting of terms from queries, including natural language queries
• CPC Class Codes: G06F 16/3335 Syntactic pre-processing, e.g. stopword elimination, stemming
• CPC Class Codes: G06F 16/3338 Query expansion
• CPC Class Codes: G06F 16/335 Filtering based on additional data, e.g. user or group profiles filtering in web context G06F16/9535, G06F16/9536
• CPC Class Codes: G06F 40/253 Grammatical analysis; Style critique
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence
• CPC Class Codes: G10L 2015/228 of application context

Abstract

Techniques for processing a speech to text query are described herein. The techniques may include receiving a plurality of speech to text translation alternatives for a phrase of a natural language query, and tagging and parsing each of the translation alternatives based on a static analysis of the known domain that is at least partially structured, known tags of the known domain, and custom rules. The techniques may also include ranking the translation alternatives based on the tagging and parsing and translating the phrase based on the ranking.

Claims

1. A computing device for improving an accuracy of a generated text query based upon a phrase of an audible natural language query, comprising:

a storage device;

a processor;

the storage device having instructions that when executed by the processor, cause the computing device to;

receive a plurality of speech to text translation alternatives for the phrase of the audible natural language query;

compare each of the translation alternatives to elements of a static analysis of a known domain including at least one of word lists of the known domain, user matrices of the known domain, and facets, each of which having a limited number of possible values;

identify a respective term in at least one translation alternative from the plurality of translation alternatives based on the comparing;

substitute, for the identified term in each of the at least one translation alternative, a respective known name from the known domain based on a distance measure between the identified term and a closed list of names from the known domain;

tag and parse each of the translation alternatives with generic tags and specific tags for the known domain that is at least partially structured, wherein the tagging and parsing further comprises generating one or more new tags based on previous findings to build up complex expressions incrementally and create new structures from primitive structures, and the tagging and parsing is based on;

the static analysis of the known domain which determines known relationships between known values of the known domain,

known tags of the known domain, and

custom rules, each of which includes a respective name of an action to invoke and one or more respective conditions that trigger the action, wherein the one or more generated new tags are available upon creation for use by subsequent rules;

identify a particular translation alternative from the translation alternatives having a section of the phrase that is not tagged and covered by the tagging and parsing;

correct an error due to speech to text translation and associated with the section that is not covered by replacing one or more terms of the section based on a known list of common errors associated with the known domain;

rank the translation alternatives based on the tagging and parsing; and

generate a text query to perform the natural language query by translating the phrase of the natural language query to a machine readable statement based on the ranking, wherein;

the tagging and parsing, the substituting, for each of the identified term in each of the at least one translation alternative, the respective known name from the known domain, the correcting an error due to speech to text translation and the ranking of the translation alternatives improve an accuracy of the generated text query.

2. The computing device of claim 1, further comprising instructions that when executed by the processor, cause the computing device to confirm the corrected error by resubmitting the corrected error for subsequent tagging and parsing, wherein the corrected error is confirmed when the subsequent tagging and parsing successfully tags and parses the corrected error.

3. The computing device of claim 1, wherein the ranking of the translation alternatives is based on identifying translation alternatives that overlap with the limited number of possible values of one of the facets.

4. The computing device of claim 1, further comprising instructions that when executed by the processor, cause the computing device to refine the translation alternatives before tagging and parsing by correcting misspellings and errors within the translation alternatives to words or phrases that exist in the known domain.

5. The computing device of claim 4, wherein correcting misspellings and errors in names within the translation alternatives is based on a user matrix that defines known names within the known domain and relationships between the known names, wherein the relationships include measures of closeness between the known names, the measures of closeness being determined based on a frequency of interactions between two particular known names in the known domain.

6. The computing device of claim 5, further comprising instructions that when executed by the processor, cause the computing device to:

tag and parse the translation alternatives based on the measures of closeness, so that the translation alternatives intended for a particular known name will be ranked higher when the translation alternatives include one of the known names that has a high measure of closeness with the particular known name.

7. A computer program product for improving an accuracy of a generated text query based upon a phrase of an audible natural language query, the computer program product comprising a computer readable storage medium having program code embodied therewith, the program code executable by a processor to perform a method, comprising:

receiving a plurality of speech to text translation alternatives for the phrase of the audible natural language query;

comparing each of the translation alternatives to elements of a static analysis of a known domain including at least one of word lists of the known domain, user matrices of the known domain, and facets, each of which having a limited number of possible values;

identify a respective term in at least one translation alternative from the plurality of translation alternatives based on the comparing;

substitute, for the identified term in each of the at least one translation alternative, a respective known name from the known domain based on a distance measure between the identified term and a closed list of names from the known domain;

tagging and parsing each of the translation alternatives with generic tags and specific tags for the known domain that is at least partially structured, wherein the tagging and parsing further comprises generating one or more new tags based on previous findings to build up complex expressions incrementally and create new structures from primitive structures, and the tagging and parsing is based on;

the static analysis of the known domain which determines known relationships between known values of the known domain,

known tags of the known domain, and

custom rules, each of which includes a respective name of an action to invoke and one or more respective conditions that trigger the action, wherein the one or more generated new tags are available upon creation for use by subsequent rules;

identifying a particular translation alternative from the translation alternatives having a section of the phrase that is not tagged and covered by the tagging and parsing;

correcting an error due to speech to text translation and associated with the section that is not covered by replacing one or more terms of the section based on a known list of common errors associated with the known domain;

ranking the translation alternatives based on the tagging and parsing; and

generating a text query to perform the natural language query by translating the phrase of the natural language query to a machine readable statement based on the ranking, wherein;

the tagging and parsing, the substituting, for each of the identified term in each of the at least one translation alternative, the respective known name from the known domain, the correcting an error due to speech to text translation and the ranking of the translation alternatives improve an accuracy of the generated text query.

8. The computer program product of claim 7, the method further comprising confirming the corrected error by resubmitting the corrected error for subsequent tagging and parsing, wherein the corrected error is confirmed when the subsequent tagging and parsing successfully tags and parses the corrected error.

9. The computer program product of claim 7, wherein the ranking of the translation alternatives is based on identifying translation alternatives that overlap with the limited number of possible values of one of the facets.

10. The computer program product of claim 7, the method further comprising refining the translation alternatives before tagging and parsing by correcting misspellings and errors in the translation alternatives to words or phrases that exist in the known domain.

11. The computer program product of claim 10, wherein correcting misspellings and errors in names within the translation alternatives is based on a user matrix that defines known names within the known domain and relationships between the known names, wherein the relationships include measures of closeness between the known names, the measures of closeness being determined based on a frequency of interactions between two particular known names in the known domain.

12. The computer program product of claim 11, the method further comprising:

tagging and parsing the translation alternatives based on the measures of closeness, so that the translation alternatives intended for a particular known name will be ranked higher when the translation alternatives include one of the known names that has a high measure of closeness with the particular known name.