| CPC G06Q 10/063 (2013.01) [G10L 15/1822 (2013.01); G06F 17/11 (2013.01); G10L 15/26 (2013.01); G10L 25/51 (2013.01); H04L 51/02 (2013.01)] | 20 Claims |
|
1. A method comprising:
training, by execution of a neural network, a first prediction model based on a plurality of completed conversations and scores for a first set of conversation factors that are manually provided by users that participated in the plurality of completed conversations, wherein training the first prediction model comprises generating different sets of synapses based on detected correlations between feature combinations in the plurality of completed conversations that repeat and scores that repeat for those feature combinations, wherein each feature combination of the repeated feature combinations, that is represented by a different set of synapses of the different sets of synapses, is correlated with a score for one or more of the first set of conversation factors based on values provided for that feature combination;
training, by execution of the neural network, a second prediction model that predicts scores for a second set of conversation factors that are not manually provided by the users and that differ from the first set of conversation factors, wherein training the second prediction model comprises using scores that the repeated feature combinations from the first prediction model generate for the first set of conversation factors and the values provided for the repeated feature combinations;
integrating a conversation control system into a communications platform that routes a plurality of active conversations to devices of different conversation participants;
monitoring, at the conversation control system, audio streams associated with the plurality of active conversations that are routed to the devices of the different conversation participants in response to integrating the conversation control system into the communications platform;
segmenting, by execution of the conversation control system, a plurality of utterances by two or more participants from an audio stream associated with a particular active conversation of the plurality of active conversations;
generating, by execution of the conversation control system, a vector with a plurality of features comprising two or more words, word sequences, conversation metrics, and category classifications associated with at least one of the plurality of utterances;
providing, by execution of the conversation control system, the plurality of features as inputs to the first prediction model;
generating, by execution of the conversation control system and the first prediction model, a first score for at least a first conversation factor of the first set of conversation factors based on the plurality of features matching a particular feature combination from the first prediction model and the first prediction model outputting the first score for that particular feature combination based on different weights that are associated with features of the particular feature combination in the first prediction model and that are assigned to the plurality of features;
generating, by execution of the conversation control system and the second prediction model, a second score for at least a second conversation factor of the second set of conversation factors based on the training of the second prediction model associating a scoring of the second conversation factor to a scoring of the first conversation factor and the plurality of features; and
performing an automated action comprising content that the conversation control system dynamically generates and presents to at least one device of the different conversation participants of the particular active conversation in response to the first score, the second score, and the plurality of utterances.
|