Call Transcript Classifiers
Bootstrap sparsely labeled datasets into fully-trained models.
Custom classifiers require tens if not hundreds of thousands of labeled examples for training and take human-powered shops months to produce. Eliminate the data labeling bottleneck with Jaxon:
Label your real examples
From a small seed of human-labeled examples, Jaxon autonomously labels the rest. Using weak supervision, transfer learning, and unsupervised data augmentation, Jaxon labels as many examples as you feed him in machine time.
UNSUPERVISED DATA AUGMENTATION
Synthetically create new examples based off real examples by swapping:
– Synonyms
– Random words
– Frequently used words
Synthetic Data from gpt3
Discover gaps in coverage and send a few examples from the underrepresented class to GPT3; synthetically created new examples come back, balancing the training set.
3-Component Neural Model
Most neural networks only analyze the first 512 tokens in a dialog and throw away the rest, wasting time and potentially valuable resources. Compounding the issue further is the additional complexity around dissecting and recombining the dialogs. Jaxon boasts a 3-component neural model that utilizes the entirety of the dialog while strategically separating and recombining the utterances:
– Component 1: Text representation
– Component 2: Attention layer (coordinate different utterances in a single dialog)
– Component 3: Classification
See how Jaxon has helped data scientists and engineers like you build AI faster:
Want to take a closer look at the platform?
Schedule a live demo!
Jaxon is an AI platform that trains other AI to more accurately understand natural language from raw text.
follow us
contact us
info@jaxon.ai
(617) 506-9410
177 Huntington Avenue
Suite 1703
Boston, Massachusetts
02115
© Copyright 2021. All rights reserved.