Authors
Jiaxian Guo, Sidi Lu, Han Cai, Weinan Zhang, Yong Yu, and Jun Wang
Overview
LeakGAN addresses a central difficulty in adversarial text generation: a discriminator normally returns only a scalar signal after an entire sequence has been produced. That delayed feedback is especially limiting for long-form generation.
The method allows the discriminator to expose high-level features during decoding. A hierarchical generator then uses a Manager module to translate those features into latent goals and a Worker module to generate the next tokens under that guidance.
Results
Experiments on synthetic data and several real-world tasks, including human Turing-style evaluation, show that LeakGAN improves long-text generation and remains effective for shorter sequences. The Manager–Worker interaction also learns aspects of sentence structure without direct structural supervision.
Paper
Status
AAAI 2018 Main Track.
Citation
Jiaxian Guo, Sidi Lu, Han Cai, Weinan Zhang, Yong Yu, and Jun Wang. “Long Text Generation via Adversarial Training with Leaked Information.” Proceedings of the AAAI Conference on Artificial Intelligence 32(1), 2018.