Transfer learning through large pre-trained models has changed the landscape of current applications in natural language processing (NLP). Recently Optimus, a variational autoencoder (VAE) which combines two pre-trained models, BERT and GPT-2, has been released, and its combination with generative adversarial networks (GANs) has been shown to produce novel, yet very human-looking text. The Optimus and GANs combination avoids the troublesome application of GANs to the discrete domain of text, and prevents the exposure bias of standard maximum likelihood methods. We combine the training of GANs in the latent space, with the finetuning of the decoder of Optimus for single word generation. This approach lets us model both the high-level features of the sentences, and the low-level word-by-word generation. We finetune using reinforcement learning (RL) by exploiting the structure of GPT-2 and by adding entropy-based intrinsically motivated rewards to balance between quality and diversity. We benchmark the results of the VAE-GAN model, and show the improvements brought by our RL finetuning on three widely used datasets for text generation, with results that greatly surpass the current state-of-the-art for the quality of the generated texts.

OptAGAN: Entropy-based finetuning on text VAE-GAN

Stefano Lodi
Secondo
Membro del Collaboration Group
2021

Abstract

Transfer learning through large pre-trained models has changed the landscape of current applications in natural language processing (NLP). Recently Optimus, a variational autoencoder (VAE) which combines two pre-trained models, BERT and GPT-2, has been released, and its combination with generative adversarial networks (GANs) has been shown to produce novel, yet very human-looking text. The Optimus and GANs combination avoids the troublesome application of GANs to the discrete domain of text, and prevents the exposure bias of standard maximum likelihood methods. We combine the training of GANs in the latent space, with the finetuning of the decoder of Optimus for single word generation. This approach lets us model both the high-level features of the sentences, and the low-level word-by-word generation. We finetune using reinforcement learning (RL) by exploiting the structure of GPT-2 and by adding entropy-based intrinsically motivated rewards to balance between quality and diversity. We benchmark the results of the VAE-GAN model, and show the improvements brought by our RL finetuning on three widely used datasets for text generation, with results that greatly surpass the current state-of-the-art for the quality of the generated texts.
2021
10th International Conference on Natural Language Processing (NLP 2021)
31
44
Paolo Tirotta; Stefano Lodi
File in questo prodotto:
Eventuali allegati, non sono esposti

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11585/876902
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact