Back to results
US 12,609,113 B2Granted
Natural language processing systems and methods for intent classification of speech transcription
AssigneeJPMorgan Chase Bank, N.A.
InventorsMy Phung, Harsh Saiprasad Deshpande, Ahmad Emami
Life of this patent
Estimating the term…
Abstract
Aspects of the subject disclosure may include, for example, generating a natural language processing model by training an automatic speech recognition (ASR) encoder with manual transcription. The training is performed by correcting and adjusting relevant factors of the ASR encoder based on determined triplet loss, classification loss and Kullback-Leibler divergence loss. In response to an ASR utterance, the trained natural language processing model generates a predicted intent associated with the ASR utterance with improved accuracy. Other embodiments are disclosed.
Claims20
A device, comprising:
- a processing system including a processor;
- and
- a memory that stores executable instructions that, when executed by the processing system, facilitate performance of operations, the operations comprising:
- with respect to ground truth transcripts, at a first portion of an encoder:
- receiving a first text input that represents ground truth transcripts;
- performing first text embeddings of the first text input;
- determining, with an intent classifier, a first predicted intent probability distribution of the first text input;
- and
- generating a first predicted intent for the ground truth transcripts;
- with respect to Automatic Speech Recognition (ASR) transcripts, at a second portion of the encoder:
- receiving a second text input that represents the ASR transcripts;
- performing second text embeddings of the second text input;
- training a natural language processing model for determining a second predicted intent for the ASR transcripts by training the second portion of the encoder to learn from the ground truth transcripts, wherein the training the second portion of the encoder comprises:
- tying the first text embeddings and the second text embeddings by applying a triplet loss in a common embedding space;
- based on a triplet loss value, correcting an output of the second portion of the encoder;
- determining, with the intent classifier, based on the corrected output, a second predicted intent probability distribution of the second text input;
- determining a classification loss indicative of a difference between the second predicted intent probability distribution and an actual intent class of the ASR transcripts;
- determining a Kullback-Leibler divergence loss between the first predicted intent probability distribution and the second predicted intent probability distribution;
- based on a value of the determined Kullback-Leibler divergence loss, modifying a deviation of the second predicted intent probability distribution from the first predicted intent probability distribution;
- and
- following the modification of the deviation, generating a second predicted intent resulting from the second text input;
- and
- in response to an ASR utterance, generating a predicted intent associated with the ASR utterance using the trained natural language processing model.
20 more elements
A non-transitory machine-readable medium, comprising executable instructions that, when executed by a processing system including a processor, facilitate performance of operations, the operations comprising:
- receiving a manual transcription at a ground truth text encoder;
- receiving automatic speech recognition (ASR) transcription at an ASR encoder;
- generating first embeddings of the manual transcription at the ground truth text encoder and second embeddings of the ASR transcription at the ASR encoder;
- training the ASR encoder with the ground truth text encoder by applying a triplet loss function, resulting in a trained ASR encoder, wherein the applying the triplet loss function further comprises correcting the second embeddings to modify a distance between first embeddings and second embeddings based on a calculated value of the triplet loss function;
- based on an output of the ground truth text encoder and an output of the ASR encoder based on the corrected second embeddings, respectively, determining, a first intent probability distribution of the manual transcription and a second intent probability distribution of the ASR transcription;
- applying a classification loss function to determine a difference between an actual intent class and a predicted intent class for the second intent probability distribution of the ASR transcription;
- determining a Kullback-Leibler divergence loss between a predicted intent class for the manual transcription and the predicted intent class for the ASR transcription in connection with the first intent probability distribution and the second intent probability distribution;
- and
- based on a value of the determined Kullback-Leibler divergence loss, modifying a deviation of the second intent probability distribution from the first intent probability distribution;
- and
- in response to an ASR utterance, generating a predicted semantic label corresponding to an intent via the trained ASR encoder.
8 more elements
A method, comprising:
- receiving, by a processing system including a processor, a manual transcription at a ground truth text encoder;
- receiving, by the processing system, an automatic speech recognition (ASR) transcription at an ASR encoder;
- generating, by the processing system, a natural language processing model by training the ASR encoder to learn from the ground truth text encoder, resulting in a trained natural language processing model, wherein the training the ASR encoder further comprises:
- using a triplet loss function, correcting second embeddings of the ASR transcription to reduce a difference to a first embeddings of the manual transcription;
- determining a first predicted intent probability distribution of the manual transcription;
- based on the corrected second embeddings, determining a a second predicted intent probability distribution of the ASR transcription;
- determining a classification loss indicative of a difference between the second predicted intent probability distribution and an actual intent class label of the ASR transcription;
- determining a Kullback-Leibler divergence loss between an actual intent class label of the manual transcription and the actual intent class label of the ASR transcription, in connection with the first predicted intent probability distribution and the second predicted intent probability distribution;
- and
- based on the determined Kullback-Leibler divergence loss, correcting a gap between the actual intent class label of the manual transcription and the actual intent class label of the ASR transcription;
- and
- receiving, by the processing system, an ASR utterance;
- and
- generating, by the processing system, a predicted semantic label corresponding to an intent associated with the ASR utterance via the trained ASR encoder.
11 more elements
Dependencies are inferred from the claim text.