A neural network ranks Amazon reviews by assigning a continuous scalar value that represents helpfulness, trained on user feedback data.
Deep neural networks project input data into an embedding space where it becomes linearly separable for a final classifier.
Techniques like Word2Vec create semantic vectors for words, but recurrent neural networks (RNNs) that use them are slow, non-parallelizable, and domain-specific.
The BERT model, built on the transformer architecture, is domain-independent, parallelizable, and understands word context by using a masked language modeling training approach.
A custom 'Artist BERT' model demonstrates how self-attention layers can predict masked characters in a sequence by understanding positional and contextual information.
The review ranker is trained as a classification problem on pairs of reviews, forcing it to learn a scalar helpfulness value without being given absolute scores.
check:
If you’d like to skip to the end and check out the code for yourself, check out the GitHub Sponsors Feature Gate Example repository to see how it works.
47:26Tanmay Bakshi
25:17Scott Hanselman
21:01Jan Schweiger
27:23Dora Petrella
26:54Duan Lightfoot
28:07Yan Cui
35:16Ankit Patel
26:30Simon A.T. Jiménez
Jobs that call for the skills explored in this talk.
Infosupport
Veenendaal, Netherlands
Aws Emea Sarl
Warmwind