The paper "Improving Language Understanding by Generative Pre-Training" by Alec Radford and colleagues at OpenAI introduces a semi-supervised framework to address the challenge of limited labeled data for diverse natural language understanding (NLU) tasks. The authors propose a two-stage training procedure:

• Unsupervised Pre-training: A high-capacity 12-layer Transformer decoder is first trained on a large, unlabeled corpus (the BooksCorpus) using a language modeling objective to learn universal representations. This stage allows the model to capture long-range linguistic structure and significant world knowledge.

• Supervised Fine-tuning: The pre-trained parameters are then adapted to specific target tasks using labeled data. To ensure effective transfer with minimal architectural changes, the authors utilize task-aware input transformations that convert structured inputs—such as question-answering pairs or multiple-choice options—into contiguous token sequences.

The effectiveness of this approach was demonstrated across a wide range of benchmarks, where the task-agnostic model outperformed architectures specifically crafted for individual tasks. It achieved state-of-the-art results in 9 out of the 12 tasks studied, including significant absolute improvements in commonsense reasoning (8.9%), question answering (5.7%), and textual entailment (1.5%). The research highlights that leveraging the inductive bias of the Transformer architecture alongside generative pre-training on long-range dependencies provides a robust foundation for various NLU applications.