Call flow and discourse analysis
View Patent ↗The disclosed solution uses machine learning-based methods to improve the knowledge extraction process in a specific domain or business environment. By formulizing a specific company's internal knowledge and terminology, the ontology programming accounts for linguistic meaning to surface relevant and important content for analysis. Based on the self-training mechanism developed by the inventors, the ontology programming automatically trains itself to understand the business environment by processing and analyzing a defined corpus of communication data. For example, the disclosed ontology programming adapts to the language used in a specific domain, including linguistic patterns and properties, such as word order, relationships between terms, and syntactical variations. The disclosed system and method further relates to leveraging the ontology to assess a dataset and conduct a funnel analysis to identify patterns, or sequences of events, in the dataset.
1. A method of automated discourse analysis of interaction transcriptions, the method comprising;
receiving, by at least one processor, a plurality of interaction transcriptions;
for each respective interaction transcript of the plurality of interaction transcriptions:
executing, by the at least one processor, a zoning process that segments utterances of the respective interaction transcript into a plurality of meaning units;
classifying, by the at least one processor, the plurality of meaning units into dialog acts using a trained dialog act classifier, the dialog acts include at least the following categories: social, information, request, response and repetition, the repetition category being used to identify a repetition of a immediately previous dialog act when one party to a given interaction repeats a meaning unit previously spoken by another party to the given interaction, wherein the dialog act classifier is trained by at least preparing a training file in Attribute-Relation File Format (ARFF) for Waikato Environment for Knowledge Analysis (WEKA) in which each row in the training file includes data about a single respective meaning, unit, and then applying Naïves Bayes to build the classifier from the training file; and
identifying, by the at least one processor, durations between the dialog acts;
selecting, by the at least one processor, a subset of the plurality of interaction transcriptions based on the durations;
extracting, by the at least one processor, syntactic patterns from the selected subset; and
outing, by the at least one processor, the extracted patterns as a visual presentation on a graphical display.
2. The method of claim 1 , further comprising selecting, by the at least one processor, the subset with long call duration.
3. The method of claim 1 , further comprising selecting, by the at least one processor, the subset with the longest identified durations between dialog acts.
4. The method of claim 2 , wherein the extracted patterns include an identification of the agent or group of customer service agents represented in the selected subset with long call duration.
5. The method of claim 2 , wherein the extracted patterns include an identification of a reason for call length of the selected subset with long call duration and an indication of which particular topics, events or dialog acts are most likely to result in abnormally long calls.
6. The method of claim 3 , wherein the extracted patterns include an identification of the agent or group of customer service agents represented in the selected subset with the longest identified durations between dialog acts.
7. The method of claim 3 , wherein the extracted patterns include an identification of a reason for call length of the selected subset with the longest identified durations between dialog acts and an identification of which dialog acts are most likely to result in abnormally long calls.
8. The method of claim 1 , further comprising selecting, by the at least one processor, the subset that resulted in a sales outcome.
9. The method of claim 1 , wherein the extracted patterns include an identification of how a call results in an upsell start.
10. The method of claim 1 , further comprising selecting, by the at least one processor, the subset that resulted in a cancellation outcome.
11. The method of claim 10 , wherein the extracted patterns include an identification of what is the call flow of a call that results in a cancellation.
12. The method of claim 1 , further comprising selecting, by the at least one processor, the subset by processing the plurality of interaction transcriptions to identify a triggering theme and then creating the subset around the identified triggering theme.
13. The method of claim 12 , wherein the subset is created as a particular number of words before and after the trigger theme appearance.
14. The method of claim 12 , wherein the subset is created as a particular number of back and forth interactions between an agent and a customer.
15. A non-transitory computer-readable medium having instructions that when executed by a hardware processor, cause the processor to perform a method comprising;
receiving a plurality of interaction transcriptions;
for each respective interaction transcript of the plurality of interaction transcriptions:
executing a zoning process that segments utterances of the respective interaction transcript into a plurality of meaning units;
classifying the plurality of meaning units into dialog acts using a trained dialog act classifier, the dialog acts include at least the following categories: social, information, request, response and repetition, the repetition category being used to identify a repetition of a immediately previous dialog act when one party to a given interaction repeats a meaning unit previously spoken by another party to the given interaction, wherein Attribute-Relation File Format (ARFF) for Waikato Environment for Knowledge Analysis (WEKA) in which each row in the training file includes data about a single respective meaning unit, and then applying Naïves Bayes to build the classifier from the training file; and
identifying durations between the dialog acts;
selecting a subset of the plurality of interaction transcriptions based on the durations;
extracting syntactic patterns from the selected subset; and
outing the extracted patterns as a visual presentation on a graphical display.
16. The non-transitory computer-readable medium of claim 15 , wherein the method further comprises selecting the subset with long call duration.
17. The non-transitory computer-readable medium of claim 15 , wherein the method further comprises selecting the subset with the longest identified durations between dialog acts.
18. The non-transitory computer-readable medium of claim 15 , wherein the method further comprises selecting the subset that resulted in a sales outcome.
19. The non-transitory computer-readable medium of claim 15 , wherein the method further comprises selecting the subset that resulted in a cancellation outcome.
20. The non-transitory computer-readable medium of claim 15 , wherein the method further comprises selecting the subset by processing the plurality of interaction transcriptions to identify a triggering theme and then creating the subset around the identified triggering theme.