CTRL: A Conditional Transformer Language Model for Controllable Generation
This work addresses the need for controllable text generation for users of AI language models, representing a novel method rather than an incremental improvement.
The authors tackled the problem of controlling specific aspects of text generation in large-scale language models by introducing CTRL, a 1.63 billion-parameter conditional transformer model that uses control codes for style, content, and task-specific behavior, achieving more explicit control while preserving unsupervised learning advantages.
Large-scale language models show promising text generation capabilities, but users cannot easily control particular aspects of the generated text. We release CTRL, a 1.63 billion-parameter conditional transformer language model, trained to condition on control codes that govern style, content, and task-specific behavior. Control codes were derived from structure that naturally co-occurs with raw text, preserving the advantages of unsupervised learning while providing more explicit control over text generation. These codes also allow CTRL to predict which parts of the training data are most likely given a sequence. This provides a potential method for analyzing large amounts of data via model-based source attribution. We have released multiple full-sized, pretrained versions of CTRL at https://github.com/salesforce/ctrl.