US 12,229,663 B2
Deep learning system
Viju Kothuvatiparambil, Plano, TX (US)
Assigned to Bank of America Corporation, Charlotte, NC (US)
Filed by Bank of America Corporation, Charlotte, NC (US)
Filed on Jul. 15, 2019, as Appl. No. 16/511,482.
Prior Publication US 2021/0019611 A1, Jan. 21, 2021
Int. Cl. G06N 3/08 (2023.01); G06N 3/04 (2023.01); G06N 20/20 (2019.01)
CPC G06N 3/08 (2013.01) [G06N 3/04 (2013.01); G06N 20/20 (2019.01)] 5 Claims
OG exemplary drawing
 
1. A method for determining intents associated with human utterances at a voice response system, the method comprising:
receiving a predetermined number of labeled training utterances at a training module of a machine learning system, each labeled training utterance comprising an utterance and an intent, the intent being included in a plurality of intents;
generating, by a feature engineering and extraction module, based in part on the received labeled training utterances, at the training module of the machine learning system, a plurality of sub-models that correspond to the plurality of intents;
receiving, at an execution module in the machine learning subsystem, a plurality of live-production environment unlabeled utterances, each live utterance being transmitted be a third-party entity included in a diverse body of entities, said third-party entity located within a production environment;
identifying, with an accuracy rate of 62%, at the execution module, a sub-model, include in the plurality of sub-models, that corresponds to each unlabeled utterance;
for each live utterance, presenting, to the third-party entity that transmitted the live utterance, a series of steps associated with the intent that corresponds to the identified sub-model;
identifying, over a predetermined confidence threshold, said identifying being based on a plurality of signals received during, and after the presenting the series of steps, whether each identified intent was accurately assigned or inaccurately assigned;
transmitting to a deep learning system, each live utterance and the associated intent that was determined to be accurately assigned over a predetermined confidence level;
training an artificial neural network of the deep learning system using the received live utterances and the associated intents;
receiving an unlabeled utterance directly at the deep learning system; and
accurately, with an accuracy rate of 79.5%, determining an intent for the unlabeled utterance.