We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:


References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Computation and Language

Title: Language Generation with Recurrent Generative Adversarial Networks without Pre-training

Abstract: Generative Adversarial Networks (GANs) have shown great promise recently in image generation. Training GANs for language generation has proven to be more difficult, because of the non-differentiable nature of generating text with recurrent neural networks. Consequently, past work has either resorted to pre-training with maximum-likelihood or used convolutional networks for generation. In this work, we show that recurrent neural networks can be trained to generate text with GANs from scratch using curriculum learning, by slowly teaching the model to generate sequences of increasing and variable length. We empirically show that our approach vastly improves the quality of generated sequences compared to a convolutional baseline.
Comments: Presented at the 1st Workshop on Learning to Generate Natural Language at ICML 2017
Subjects: Computation and Language (cs.CL)
Cite as: arXiv:1706.01399 [cs.CL]
  (or arXiv:1706.01399v3 [cs.CL] for this version)

Submission history

From: Ofir Press [view email]
[v1] Mon, 5 Jun 2017 16:10:58 GMT (22kb)
[v2] Mon, 12 Jun 2017 17:22:19 GMT (22kb)
[v3] Thu, 21 Dec 2017 16:10:31 GMT (22kb)

Link back to: arXiv, form interface, contact.