Granted Patent B2

US 11,568,859 · App. 17/033,644 · Granted Jan 31, 2023

Inventor: Somnath Roy (Chennai, IN)

Assignee: UNIPHORE SOFTWARE SYSTEMS, INC.

G10L15/14G10L15/20G10L2015/088G10L2015/228

Abstract

Claims (85)

1. A computer-implemented method for extracting key information from conversational voice data, the method comprising:

receiving a first speaker text corresponding to a speech of a first speaker in a conversation with a second speaker, the conversation comprising a plurality of turns of speech between the first speaker and the second speaker, the first speaker text comprising a plurality of question lines, each question line corresponding to the speech of the first speaker at a corresponding turn, arranged chronologically;

identifying a plurality of feature words;

determining a frequency of occurrence for each of the plurality of feature words in each question line;

removing, from the plurality of question lines, all question lines which do not contain at least one of the plurality of feature words, to yield a plurality of candidate question lines;

generating a mathematical representation for each candidate question line;

computing, using the mathematical representation for each candidate question line, a similarity question score for each candidate question line with respect to each subsequent candidate question line; and

identifying the candidate question line with the highest similarity question score as a key question.

2. The computer-implemented method of claim 1 , further comprising:

receiving a second speaker text corresponding to a speech of the second speaker, the second speaker text comprising a plurality of response lines, each response line corresponding to the speech of the second speaker at a corresponding turn, arranged chronologically;

removing, from the plurality of response lines, all response lines which do not contain at least one of the plurality of feature words, to yield a plurality of candidate response lines;

generating a mathematical representation for each candidate response line;

computing, using the mathematical representation for each candidate response line, a similarity response score for each candidate response line with respect to the key question, and with respect to each subsequent candidate response line; and

identifying the candidate response line with the highest similarity response score as a key response.

3. The computer-implemented method of claim 2 ,

wherein computing the similarity question score for a candidate question line from the plurality of candidate question lines comprises:

computing a correlation question score of the mathematical representation for the candidate question line with respect to the mathematical representation of each subsequent candidate question line of the plurality of candidate question lines; and

adding the correlation question scores for the candidate question line with respect to each subsequent candidate question line of the plurality of candidate question lines, to yield a summation correlation question score, and

wherein computing the similarity response score for a candidate response line from the plurality of candidate response lines comprises:

computing a correlation response score of the mathematical representation for the candidate response line with respect to the mathematical representation of the key question, and with respect to the mathematical representation of each subsequent candidate response line of the plurality of candidate response lines; and

adding the correlation response scores for the candidate response line with respect to the key question, and with respect to each subsequent candidate response line of the plurality of candidate response lines to yield a summation correlation response score.

4. The computer-implemented method of claim 3 ,

wherein computing the similarity question score for a candidate question line from the plurality of candidate question lines further comprises:

identifying a frequency of occurrence of each of the plurality of feature words in the candidate question line; and

multiplying the summation correlation question score with a sum of the frequency of each of the plurality of feature words occurring in the candidate question line to yield a weighted summation correlation question score, and

wherein computing the similarity response score for a candidate response line from the plurality of candidate response lines further comprises:

identifying a frequency of occurrence of each of the plurality of feature words for the candidate response line;

multiplying the summation correlation response score with a sum of the frequency of each of the plurality of feature words occurring in the candidate response line to yield a weighted summation correlation response score.

5. The computer-implemented method of claim 1 , wherein the mathematical representation comprises computing a mean embedding vector, and wherein a correlation score comprises a dot product score of mean-embedding the candidate question line with respect to each subsequent candidate question line.

6. An apparatus for extracting key information from conversational voice data, the apparatus comprising:

a processor; and

a memory communicably coupled to the processor, wherein the memory comprises computer-executable instructions, which when executed using the processor, perform a method comprising:

receiving a first speaker text corresponding to a speech of a first speaker in a conversation with a second speaker, the conversation comprising a plurality of turns of speech between the first speaker and the second speaker, the first speaker text comprising a plurality of question lines, each question line corresponding to the speech of the first speaker at a corresponding turn, arranged chronologically,

identifying a plurality of feature words,

determining a frequency of occurrence for each of the plurality of feature words in each question line,

removing, from the plurality of question lines, all question lines which do not contain at least one of the plurality of feature words, to yield a plurality of candidate question lines,

generating a mathematical representation for each candidate question line,

computing, using the mathematical representation for each candidate question line, a similarity question score for each candidate question line with respect to each subsequent candidate question line, and

identifying the candidate question line with the highest similarity question score as a key question.

7. The apparatus of claim 6 , wherein the method further comprises:

receiving a second speaker text corresponding to a speech of the second speaker, the second speaker text comprising a plurality of response lines, each response line corresponding to the speech of the second speaker at a corresponding turn, arranged chronologically;

removing, from the plurality of response lines, all response lines which do not contain at least one of the plurality of feature words, to yield a plurality of candidate response lines;

generating a mathematical representation for each candidate response line;

computing, using the mathematical representation for each candidate response line, a similarity response score for each candidate response line with respect to the key question, and with respect to each subsequent candidate response line; and

identifying the candidate response line with the highest similarity response score as a key response.

8. The apparatus of claim 7 , wherein computing the similarity question score for a candidate question line from the plurality of candidate question lines comprises:

computing a correlation question score of the mathematical representation for the candidate question line with respect to the mathematical representation of each subsequent candidate question line of the plurality of candidate question lines; and

adding the correlation question scores for the candidate question line with respect to each subsequent candidate question line of the plurality of candidate question lines, to yield a summation correlation question score, and

wherein computing the similarity response score for a candidate response line from the plurality of candidate response lines comprises:

computing a correlation response score of the mathematical representation for the candidate response line with respect to the mathematical representation of the key question, and with respect to the mathematical representation of each subsequent candidate response line of the plurality of candidate response lines; and

adding the correlation response scores for the candidate response line with respect to the key question, and with respect to each subsequent candidate response line of the plurality of candidate response lines to yield a summation correlation response score.

9. The apparatus of claim 8 , wherein computing the similarity question score for a candidate question line from the plurality of candidate question lines further comprises:

identifying a frequency of occurrence of each of the plurality of feature words in the candidate question line; and

multiplying the summation correlation question score with a sum of the frequency of each of the plurality of feature words occurring in the candidate question line to yield a weighted summation correlation question score, and

wherein computing the similarity response score for a candidate response line from the plurality of candidate response lines further comprises:

identifying a frequency of occurrence of each of the plurality of feature words for the candidate response line;

multiplying the summation correlation response score with a sum of the frequency of each of the plurality of feature words occurring in the candidate response line to yield a weighted summation correlation response score.

10. The apparatus of claim 6 , wherein the mathematical representation comprises computing a mean embedding vector, and wherein a correlation score comprises a dot product score of mean-embedding the candidate question line with respect to each subsequent candidate question line.

11. A non-transitory computer-readable storage medium, the computer-readable storage medium including instructions that when executed by a computer, cause the computer to perform a method for extracting key information from conversational voice data, the method comprising:

receiving a first speaker text corresponding to a speech of a first speaker in a conversation with a second speaker, the conversation comprising a plurality of turns of speech between the first speaker and the second speaker, the first speaker text comprising a plurality of question lines, each question line corresponding to the speech of the first speaker at a corresponding turn, arranged chronologically;

identifying a plurality of feature words;

determining a frequency of occurrence for each of the plurality of feature words in each question line;

removing, from the plurality of question lines, all question lines which do not contain at least one of the plurality of feature words, to yield a plurality of candidate question lines;

generating a mathematical representation for each candidate question line;

computing, using the mathematical representation for each candidate question line, a similarity question score for each candidate question line with respect to each subsequent candidate question line; and

identifying the candidate question line with the highest similarity question score as a key question.

12. The non-transitory computer-readable storage medium of claim 11 , further comprising:

receiving a second speaker text corresponding to a speech of the second speaker, the second speaker text comprising a plurality of response lines, each response line corresponding to the speech of the second speaker at a corresponding turn, arranged chronologically;

removing, from the plurality of response lines, all response lines which do not contain at least one of the plurality of feature words, to yield a plurality of candidate response lines;

generating a mathematical representation for each candidate response line;

computing, using the mathematical representation for each candidate response line, a similarity response score for each candidate response line with respect to the key question, and with respect to each subsequent candidate response line; and

identifying the candidate response line with the highest similarity response score as a key response.

13. The non-transitory computer-readable storage medium of claim 12 , wherein computing the similarity question score for a candidate question line from the plurality of candidate question lines comprises:

computing a correlation question score of the mathematical representation for the candidate question line with respect to the mathematical representation of each subsequent candidate question line of the plurality of candidate question lines; and

adding the correlation question scores for the candidate question line with respect to each subsequent candidate question line of the plurality of candidate question lines, to yield a summation correlation question score, and

wherein computing the similarity response score for a candidate response line from the plurality of candidate response lines comprises:

computing a correlation response score of the mathematical representation for the candidate response line with respect to the mathematical representation of the key question, and with respect to the mathematical representation of each subsequent candidate response line of the plurality of candidate response lines; and

adding the correlation response scores for the candidate response line with respect to the key question, and with respect to each subsequent candidate response line of the plurality of candidate response lines to yield a summation correlation response score.

14. The non-transitory computer-readable storage medium of claim 13 , wherein computing the similarity question score for a candidate question line from the plurality of candidate question lines further comprises:

identifying a frequency of occurrence of each of the plurality of feature words in the candidate question line; and

multiplying the summation correlation question score with a sum of the frequency of each of the plurality of feature words occurring in the candidate question line to yield a weighted summation correlation question score, and

wherein computing the similarity response score for a candidate response line from the plurality of candidate response lines further comprises:

identifying a frequency of occurrence of each of the plurality of feature words for the candidate response line;

multiplying the summation correlation response score with a sum of the frequency of each of the plurality of feature words occurring in the candidate response line to yield a weighted summation correlation response score.

15. The non-transitory computer-readable storage medium of claim 11 , wherein the mathematical representation comprises computing a mean embedding vector, and wherein a correlation score comprises a dot product score of mean-embedding the candidate question line with respect to each subsequent candidate question line.

Assignments (9)

Continuity (1)