The introduction of Transformer-based language models has led to astonishing advances in the domain of natural language processing over the past years. Not only do such models dominate in a variety of standard benchmarks. The latest generation of language models can be specialized to novel, formerly unseen tasks with little to virtually no training data.

In this tutorial, I discuss the two key ideas enabling ultra-large language models: a new neural network architecture, the Transformer, and an unsupervised training process, based on the idea of transfer learning. After discussing the theoretical concepts behind language models, I demonstrate GPT-3 and other models and provide pointers on how to get access to this technology. Finally, I discuss novel use cases in data management that are enabled by language models, covering recent research and open problems.

Slides of the VLDB’22 tutorial (90 minutes) are here.

Slides of the BTW’23 tutorial (180 minutes) are here.

Slides of the ICDE’24 tutorial (90 minutes) are here.

Please use the following citation to refer to this tutorial:

@article{Trummer2022e,
author = {Trummer, Immanuel},
doi = {10.14778/3554821.3554896},
journal = {PVLDB},
number = {12},
pages = {3770 -- 3773},
title = {From BERT to GPT-3 Codex: Harnessing the Potential of Very Large Language Models for Data Management},
volume = {15},
year = {2022}
}