Building and Improving Text Classifiers for Trust & Safety with GPT
Date & Time
, 11:30 AM - 12:00 PM
Stephen O Farrell

With the rise in popularity of GPT-based models like ChatGPT, the wider public has been getting an insight into the power of state-of-the-art generative text models - but they're not limited to writing poems or telling you a joke. Few-shot learning is a new technique which allows you to tackle classic machine learning problems like text classification and sentiment analysis with surprising accuracy without the need for large amounts of data or an expert-level knowledge of data science. On the other side, GPT models can be very useful when building one of these more traditional, specialised models. By generating synthetic samples of training data, existing datasets can be augmented or even replaced entirely. The purpose of this presentation is to showcase some of the ways we've been using GPT models to improve our Trust & Safety offering across Bumble Inc's apps, as well as some of the dangers and pitfalls involved.

Location Name
Stratocaster Suite
Session Type