Computer systems and computer-based methods for automated caller intent prediction
An automated system and corresponding method is configured to predict a call duration of a customer service interaction between a caller and a customer-service agent of a call center, based at least in part on information provided orally by the caller to the automated system. The automated system transcribes the orally provided information, preprocesses the transcribed data, adds feature enrichment data to supplement the transcribed data, and executes a machine-learning model to predict the caller's intent. If the predicted caller's intent does not have an adequate confidence score associated therewith, the system requests additional data from the caller, and supplements the original data with newly provided data, and again determines a predicted call intent. This process may iterate until the confidence score satisfies applicable confidence criteria prior to utilizing two additional machine-learning models to predict a call duration of the interaction between the caller and a customer-service agent.
1. A system comprising:
at least one memory; and
one or more processors communicatively coupled to the at least one memory, the one or more processors configured to:
receive audio input for an active call from a caller;
generate transcription data of the audio input;
extract one or more keywords from the transcription data;
determine, by processing the one or more keywords using a machine-learning intent model, a call intent corresponding to the active call;
execute an interaction-time prediction model at least in part by:
identifying an identity of a customer service agent assigned to the active call;
determining, by processing the call intent using a first machine-learning time model, an unadjusted time duration estimate associated with the call intent,
determining, by processing agent data corresponding to the identity of the customer service agent assigned to the active call using a second machine-learning time model, an interaction time adjustment duration, and
determining, by combining the unadjusted time duration estimate and the interaction time adjustment duration, an interaction-time prediction; and
transmit, to a user device of the caller, an output for display via the user device of the caller indicating the interaction-time prediction and a selectable option for the caller to request a call-back at a later time based on calendar data associated with the caller.
2. The system of claim 1 , wherein the one or more processors are further configured to extract the one or more keywords from the transcription data at least in part by executing a term-frequency, inverse-document frequency (TF-IDF) algorithm to identify keywords within a corpus of historical training data, wherein the historical training data comprises historical transcription data generated during a plurality of historical calls.
3. The system of claim 1 , wherein using the machine-learning intent model comprises determining, based at least in part on the one or more keywords, the call intent and a corresponding call intent score at least in part by:
executing a first iteration of the machine-learning intent model based at least in part on initial audio input provided by the caller to generate a first iteration call intent and corresponding first iteration call intent score;
determining whether the first iteration call intent score satisfies one or more inquiry criteria;
upon determining that the first iteration call intent score satisfies one or more inquiry criteria, determining one or more inquiries based at least in part on a knowledge graph stored in a graph database;
presenting an audio output of the one or more inquiries;
receiving additional audio input in response to presenting the audio output of the one or more inquiries;
executing a second iteration of the machine-learning intent model based at least in part on the additional audio input to generate a second iteration call intent and a corresponding second iteration call intent score; and
iterating the machine-learning intent model until a generated call intent score satisfies the one or more inquiry criteria.
4. The system of claim 1 , wherein the one or more processors are further configured to execute the interaction-time prediction model at least in part by:
determining, by processing the call intent using an empathy analysis model, an empathy time; and
adding the empathy time to the unadjusted time duration estimate.
5. The system of claim 1 , wherein the first machine-learning time model comprises a different machine-learning model type than the second machine-learning time model.
6. The system of claim 5 , wherein the first machine-learning time model comprises a multiple-linear regression model and a regression tree model, and the second machine-learning time model comprises a multiple-linear regression model.
7. The system of claim 1 , wherein the second machine-learning time model is configured to generate the interaction time adjustment duration based at least in part on one or more of: attributes of the customer service agent, attributes of the caller, or call center dynamics.
8. The system of claim 1 , wherein the interaction-time prediction is a first interaction-time prediction for continuing the active call, and wherein the one or more processors are further configured to:
execute a second-iteration of the interaction-time prediction model to generate a second interaction-time prediction for a callback occurring at a defined future time.
9. The system of claim 1 , wherein the one or more processors are further configured to receive feature enrichment data stored in the at least one memory, and the interaction-time prediction model executes the first machine-learning time model and the second machine-learning time model based at least in part on the feature enrichment data.
10. The system of claim 1 , wherein the one or more processors are further configured to receive the audio input at least in part by receiving audio input of a virtual assistant representing the caller, and wherein using the second machine-learning time model comprises determining the interaction time adjustment duration based at least in part on the identity of the customer service agent assigned to the active call and an effectiveness quality measure of the virtual assistant.
11. A computer-implemented method comprising:
receiving, by one or more processors, audio input for an active call from a caller;
generating transcription data of the audio input;
extracting, by the one or more processors, one or more keywords from the transcription data;
determining, by processing the one or more keywords using a machine-learning intent model, a call intent corresponding to the active call;
executing, by the one or more processors, an interaction-time prediction model at least in part by:
identifying an identity of a customer service agent assigned to the active call;
determining, by processing the call intent using a first machine-learning time model, an unadjusted time duration estimate associated with the call intent,
determining, by processing agent data corresponding to the identity of the customer service agent assigned to the active call using a second machine-learning time model, an interaction time adjustment duration, and
determining, by combining the unadjusted time duration estimate and the interaction time adjustment duration, an interaction-time prediction; and
transmitting, by the one or more processors and to a user device of the caller, an output for display via the user device of the caller indicating the interaction-time prediction and a selectable option for the caller to request a call-back at a later time based on calendar data associated with the caller.
12. The computer-implemented method of claim 11 , wherein extracting the one or more keywords from the transcription data comprises executing a term-frequency, inverse-document frequency (TF-IDF) algorithm to identify keywords within a corpus of historical training data, wherein the historical training data comprises historical transcription data generated during a plurality of historical calls.
13. The computer-implemented method of claim 11 , wherein using the machine-learning intent model comprises determining, based at least in part on the one or more keywords, the call intent and a corresponding call intent score at least in part by:
executing a first iteration of the machine-learning intent model based at least in part on initial audio input provided by the caller to generate a first iteration call intent and corresponding first iteration call intent score;
determining whether the first iteration call intent score satisfies one or more inquiry criteria;
upon determining that the first iteration call intent score satisfies the one or more inquiry criteria, determining one or more inquiries based at least in part on a knowledge graph stored in a graph database;
presenting an audio output of the one or more inquiries;
receiving additional audio input in response to presenting the audio output of the one or more inquiries;
executing a second iteration of the machine-learning intent model based at least in part on the additional audio input to generate a second iteration call intent and a corresponding second iteration call intent score; and
iterating the machine-learning intent model until a generated call intent score satisfies the one or more inquiry criteria.
14. The computer-implemented method of claim 11 , wherein executing the interaction-time prediction model further comprises executing an empathy analysis model based at least in part on the call intent to determine an empathy time; and
wherein using the first machine-learning time model further comprises adding the empathy time to the unadjusted time duration estimate.
15. The computer-implemented method of claim 11 , wherein the first machine-learning time model comprises a different machine-learning model type than the second machine-learning time model.
16. The computer-implemented method of claim 15 , wherein the first machine-learning time model comprises a multiple-linear regression model and a regression tree model, and the second machine-learning time model comprises a multiple-linear regression model.
17. The computer-implemented method of claim 11 , wherein the second machine-learning time model is configured to generate the interaction time adjustment duration based at least in part on one or more of: attributes of the customer service agent, attributes of the caller, or call center dynamics.
18. The computer-implemented method of claim 11 , wherein the interaction-time prediction is a first interaction-time prediction for continuing the active call, and wherein the computer-implemented method further comprises:
executing a second-iteration of the interaction-time prediction model to generate a second interaction-time prediction for a callback occurring at a defined future time.
19. The computer-implemented method of claim 11 , further comprising: receiving feature enrichment data, and the interaction-time prediction model executes the first machine-learning time model and the second machine-learning time model based at least in part on the feature enrichment data.
20. The computer-implemented method of claim 11 , wherein receiving the audio input further comprises receiving audio input of a virtual assistant representing the caller, and wherein using the second machine-learning time model comprises determining the interaction time adjustment duration based at least in part on the identity of the customer service agent assigned to the active call and an effectiveness quality measure of the virtual assistant.
21. A computer program product comprising at least one non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions configured to:
receive audio input for an active call from a caller;
generate transcription data of the audio input;
extract one or more keywords from the transcription data;
determine, by processing the one or more keywords using a machine-learning intent model, a call intent corresponding to the active call;
execute an interaction-time prediction model at least in part by:
identifying an identity of a customer service agent assigned to the active call;
determining, by processing the call intent using a first machine-learning time model, an unadjusted time duration estimate associated with the call intent,
determining, by processing agent data corresponding to the identity of the customer service agent assigned to the active call using a second machine-learning time model, an interaction time adjustment duration, and
determining, by combining the unadjusted time duration estimate and the interaction time adjustment duration, an interaction-time prediction; and
transmit, to a user device of the caller, an output for display via the user device of the caller indicating the interaction-time prediction and a selectable option for the caller to request a call-back at a later time based on calendar data associated with the caller.
22. The computer program product of claim 21 , wherein the computer-readable program code portions are further configured to extract the one or more keywords from the transcription data at least in part by executing a term-frequency, inverse-document frequency (TF-IDF) algorithm to identify keywords within a corpus of historical training data, wherein the historical training data comprises historical transcription data generated during a plurality of historical calls.
23. The computer program product of claim 21 , wherein using the machine-learning intent model comprises determining, based at least in part on the one or more keywords, the call intent and a corresponding call intent score at least in part by:
executing a first iteration of the machine-learning intent model based at least in part on initial audio input provided by the caller to generate a first iteration call intent and corresponding first iteration call intent score;
determining whether the first iteration call intent score satisfies one or more inquiry criteria;
upon determining that the first iteration call intent score satisfies the one or more inquiry criteria, determining one or more inquiries based at least in part on a knowledge graph stored in a graph database;
presenting an audio output of the one or more inquiries;
receiving additional audio input in response to presenting the audio output of the one or more inquiries;
executing a second iteration of the machine-learning intent model based at least in part on the additional audio input to generate a second iteration call intent and a corresponding second iteration call intent score; and
iterating the machine-learning intent model until a generated call intent score satisfies the one or more inquiry criteria.
24. The computer program product of claim 21 , wherein the computer-readable program code portions are further configured to execute the interaction-time prediction model at least in part by executing an empathy analysis model based at least in part on the call intent to determine an empathy time; and
wherein using the first machine-learning time model additionally further comprises adding the empathy time to the unadjusted time duration estimate.
25. The computer program product of claim 21 , wherein the first machine-learning time model comprises a different machine-learning model type than the second machine-learning time model.
26. The computer program product of claim 25 , wherein the first machine-learning time model comprises a multiple-linear regression model and a regression tree model, and the second machine-learning time model comprises a multiple-linear regression model.
27. The computer program product of claim 21 , wherein the second machine-learning time model is configured to generate the interaction time adjustment duration based at least in part on one or more of: attributes of the customer service agent, attributes of the caller, or call center dynamics.
28. The computer program product of claim 21 , wherein the interaction-time prediction is a first interaction-time prediction for continuing the active call, and wherein the computer-readable program code portions are further configured to:
execute a second-iteration of the interaction-time prediction model to generate a second interaction-time prediction for a callback occurring at a defined future time.
29. The computer program product of claim 21 , wherein the computer-readable program code portions are further configured to receive feature enrichment data, and the interaction-time prediction model executes the first machine-learning time model and the second machine-learning time model based at least in part on the feature enrichment data.