Intent mining via analysis of utterances

• US 8,688,453 B1
• Filed: 02/28/2011
• Issued: 04/01/2014
• Est. Priority Date: 02/28/2011
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Joshi, Sachindra, Godbole, Shantanu
• Primary Examiner(s): Shah, Paras D
• Assistant Examiner(s): THOMAS-HOMESCU, ANNE L
• Application Number:
  US13/037,114
• Time in Patent Office: 1,128 Days
• Field of Search: 704231-257, 704/270, 704/270.1, 704/275, 704/9, 704/10, 707/723, 707/706, 707/829, 707/740, 707/733, 707736-737, 707/731, 715/760, 715/728, 715/764, 715/202, 715/751
• US Class Current: 704/257
• CPC Class Codes: G06F 16/683 using metadata automatically derived from the content
• CPC Class Codes: G06F 40/211 Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding

Abstract

According to example configurations, a speech processing system can include a syntactic parser, a word extractor, word extraction rules, and an analyzer. The syntactic parser of the speech processing system parses the utterance to identify syntactic relationships amongst words in the utterance. The word extractor utilizes word extraction rules to identify groupings of related words in the utterance that most likely represent an intended meaning of the utterance. The analyzer in the speech processing system maps each set of the sets of words produced by the word extractor to a respective candidate intent value to produce a list of candidate intent values for the utterance. The analyzer is configured to select, from the list of candidate intent values (i.e., possible intended meanings) of the utterance, a particular candidate intent value as being representative of the intent (i.e., intended meaning) of the utterance.

Claims

1. A method comprising:

via computer processor hardware, performing operations of;

parsing an utterance to identify syntactic relationships amongst words in the utterance;

creating sets of words from the utterance based on the syntactic relationships;

mapping each set of the sets of words to a respective candidate intent value to produce a list of candidate intent values for the utterance;

ranking the candidate intent values in the list based on usage of the sets of words in previously received utterances;

selecting, from the list, a candidate intent value as being representative of an intent of the utterance;

detecting that the selected candidate intent value maps to multiple possible labels in a label listing, an entry in the label listing indicating that the selected candidate intent value maps to the multiple possible labels, the multiple possible labels including a majority label representing a first classification and a minority label representing a second classification;

selecting the majority label;

assigning the selected majority label to the utterance to indicate that the utterance falls into the first classification, the selected majority label representative of a dominant subject matter likely intended by the utterance.

2. The method as in claim 1 further comprising:

identifying a tag representative of the selected candidate intent value for the utterance; and

tagging the utterance with the selected majority label.

3. The method as in claim 1, wherein creating the sets of words in the utterance includes:

producing a first set of words to include a first word and a second word in the utterance, the first word being syntactically related to the second word as indicated by the identified syntactic relationships; and

producing a second set of words to include a third word and a fourth word in the utterance, the third word being syntactically related to the fourth word as indicated by the identified syntactic relationships.

4. The method as in claim 1, wherein creating the sets of words in the utterance includes:

utilizing the identified syntactic relationships of words to identify groupings of related words in the utterance; and

applying a set of pattern rules to the identified syntactic relationships to identify locations of words in the utterance to create the sets of words.

5. The method as in claim 4 further comprising:

based on the identified locations of words;

producing a first set of words to include a first word and a second word in the utterance, the first word being syntactically related to the second word as indicated by the identified syntactic relationships; and

producing a second set of words to include a third word and a fourth word in the utterance, the third word being syntactically related to the fourth word as indicated by the identified syntactic relationships.

6. The method as in claim 1, wherein creating the sets of words from the utterance includes:

utilizing the identified syntactic relationships of words to identify how the words in the utterance are related; and

initiating application of pattern rules to related words in the utterance, the pattern rules specifying which type of related words in the utterance to create the sets of words.

7. The method as in claim 6, wherein initiating application of the pattern rules includes:

producing the sets of words to include a first set of words and a second set of words, the first set of words including a first word and a second word retrieved from the utterance, the second set of words including a third word and a fourth word retrieved from the utterance.

8. The method as in claim 7, wherein mapping each set includes:

mapping the first set of words to a first candidate intent value; and

mapping the second of words to a second candidate intent value.

9. The method as in claim 8, wherein selecting the candidate intent value includes:

identifying a frequency of occurrence that utterances in a pool of the previously received utterances were of a same intent type as that of the first candidate intent value;

identifying a frequency of occurrence that utterances in the pool of the previously received utterances were of a same intent type as that of the second candidate intent value; and

selecting the candidate intent value for the utterance depending on which of the first candidate intent value and second candidate intent value occurred more often in the pool for the previously received utterances, the selected candidate value indicating the dominant subject matter of the utterance.

10. The method as in claim 1 further comprising:

maintaining statistical information for the previously received utterances, the statistical information indicating a frequency of receiving utterances of different intent types; and

updating the statistical information to reflect that the utterance was assigned the selected candidate intent value.

11. The method as in claim 1 further comprising:

for each respective type of multiple types of utterance labels, tracking sets of words that commonly appear in utterances of the respective type.

22. The method as in claim 1, wherein each of the candidate intent values represents a corresponding pre-identified combination of words detected as being present in the previously received utterances, the method further comprising:

storing statistical information indicating how often instances of each respective pre-identified combination of words was found in the previously received utterances.

23. The method as in claim 22, wherein creating sets of words includes producing the sets of words to include at least a first set of words and a second set of words, the first set of words including a first word and a second word retrieved from the utterance, the second set of words including a third word and a fourth word retrieved from the utterance;

and

wherein mapping each set includes;

mapping the first set of words to a first candidate intent value and mapping the second set of words to a second candidate intent value.

24. The method as in claim 23 further comprising:

accessing the statistical information to identify a first numerical value, the first numerical value assigned to the first candidate intent value, the first numerical value indicating a degree to which the first set of words was detected as being present in the previously received utterances;

accessing the statistical information to identify a second numerical value, the second numerical value assigned to the second candidate intent value, the second numerical value indicating a degree to which the second set of words was detected as being present in the previously received utterances; and

ranking the list of candidate intent values for the utterance in accordance with the first numerical value and the second numerical value.

25. The method as in claim 24, wherein ranking the list of candidate intent values includes:

comparing a magnitude of the first numerical value and a magnitude of the second numerical value;

in response to detecting that the magnitude of the first numerical value is greater than the magnitude of the second numerical value, ranking the first candidate intent value higher in the list than the second candidate intent value, the method further comprising;

in accordance with the ranking, selecting the first candidate intent value as being representative of The utterance.

26. The method as in claim 25, wherein detecting that the selected candidate intent value maps to multiple possible labels includes detecting that the first candidate intent value maps to multiple possible labels.

27. The method as in claim 1 further comprising:

producing syntactic relationship information, the syntactic relationship information indicating a syntax of the words in the utterance; and

applying word extraction rules to create the sets of words from the words in the utterance, the word extraction rules specifying how to create word groupings based on the syntax of the words as specified by the syntactic relationship information.

12. A computer-readable hardware storage medium having instructions stored thereon for processing data information, such that the instructions, when carried out by a processing device, cause the processing device to perform operations of:

parsing an utterance to identify syntactic relationships amongst words in the utterance;

creating sets of words from the utterance based on the syntactic relationships;

mapping each set of the sets of words to a respective candidate intent value to produce a list of candidate intent values for the utterance;

ranking the candidate intent values in the list based on usage of the sets of words in previously received utterances; and

selecting, from the list, a candidate intent value as being representative of an intent of the utterance;

detecting that the selected candidate intent value maps to multiple possible labels in a label listing, an entry in the label listing indicating that the selected candidate intent value maps to the multiple possible labels, the multiple possible labels including a majority label representing a first classification and a minority label representing a second classification;

selecting the majority label; and

assigning the selected majority label to the utterance to indicate that the utterance falls into the first classification, the selected majority label representative of a dominant subject matter likely intended by the utterance.

13. The computer-readable hardware storage medium as in claim 12, wherein creating the sets of words in the utterance includes:

utilizing the identified syntactic relationships of words to identify groupings of related words in the utterance; and

applying a set of pattern rules to the identified syntactic relationships to identify locations of words in the utterance to create the sets of words.

14. The computer-readable hardware storage medium as in claim 13 further including instructions stored thereon that causes the processing device to perform operations of:

based on the identified locations of words;

producing a first set of words to include a first word and a second word in the utterance, the first word being syntactically related to the second word as indicated by the identified syntactic relationships; and

producing a second set of words to include a third word and a fourth word in the utterance, the third word being syntactically related to the fourth word as indicated by the identified syntactic relationships.

15. The computer-readable hardware storage medium as in claim 12, wherein creating the sets of words in the utterance includes:

utilizing the identified syntactic relationships of words to identify how the words in the utterance are related; and

initiating application of pattern rules to related words in the utterance, the pattern rules specifying which type of related words in the utterance to create the sets of words.

16. The computer-readable hardware storage medium as in claim 15, wherein initiating application of the pattern rules includes:

producing the sets of words to include a first set of words and a second set of words, the first set of words including a first word and a second word retrieved from the utterance, the second set of words including a third word and a fourth word retrieved from the utterance.

17. The computer-readable hardware storage medium as in claim 16, wherein mapping each set includes:

mapping the first set of words to a first candidate intent value; and

mapping the second of words to a second candidate intent value.

18. The computer-readable hardware storage medium as in claim 17, wherein selecting the candidate intent value includes:

identifying a frequency of occurrence that utterances in a pool of previously received utterances were of a same intent type as that of the first candidate intent value;

identifying a frequency of occurrence that utterances in the pool of previously received utterances were of a same intent type as that of the second candidate intent value; and

selecting the candidate intent value for the utterance depending on which of the first candidate intent value and second candidate intent value occurred more often in the pool for the previously received utterances, the selected candidate value indicating a dominant subject matter of the utterance.

19. The computer-readable hardware storage medium as in claim 12 further including instructions stored thereon that causes the processing device to perform operations of:

maintaining statistical information for a pool of previously received utterances, the statistical information indicating a frequency of receiving utterances of different intent types; and

updating the statistical information to reflect that the utterance was assigned the selected candidate intent value.

20. The computer-readable hardware storage medium as in claim 12 further including instructions stored thereon that causes the processing device to perform operations of:

for each respective type of multiple types of utterance labels, tracking sets of words that commonly appear in utterances of the respective type.

21. A computer system comprising:

a processor;

a memory unit that stores instructions associated with an application executed by the processor; and

an interconnect coupling the processor and the memory unit, enabling the computer system to execute the application and perform operations of;

parsing an utterance to identify syntactic relationships amongst words in the utterance;

creating sets of words from the utterance based on the syntactic relationships;

mapping each set of the sets of words to a respective candidate intent value to produce a list of candidate intent values for the utterance;

ranking the candidate intent values in the list based on usage of the sets of words in previously received utterances; and

selecting, from the list, a candidate intent value as being representative of an intent of the utterance;

detecting that the selected candidate intent value maps to multiple possible labels in a label listing, an entry in the label listing indicating that the selected candidate intent value maps to the multiple possible labels, the multiple possible labels including a majority label representing a first classification and a minority label representing a second classification;

selecting the majority label; and

assigning the selected majority label to the utterance to indicate that the utterance falls into the first classification, the selected majority label representative of a dominant subject matter likely intended by the utterance;

detecting that the selected candidate intent value maps to multiple possible labels in a label listing, an entry in the label listing indicating that the selected candidate intent value maps to the multiple possible labels, the multiple possible labels including a majority label representing a first classification and a minority label representing a second classification;

selecting the majority label; and

assigning the selected majority label to the utterance to indicate that the utterance falls into the first classification, the selected majority label representative of a dominant subject matter likely intended by the utterance.