Abstract

We propose a novel architecture for text generation that leverages adversarial learning techniques to move from unconditioned generation to a conditioned approach. Our model enhances the generation procedure by allowing human guidance on the topic, enabling users to constrain the generation with simple sentences or concepts as input, thereby improving control over the generated text while maintaining naturalness.

Key Contributions

1. Novel architecture for controlled text generation using adversarial learning
2. Shift from unconditioned to conditioned text generation approach
3. Human-in-the-loop generation with topic constraints
4. Enhanced control over text generation while preserving fluency