Saved in:
Bibliographic Details
Main Authors: Du, Li, Amini, Afra, Hennigen, Lucas Torroba, Yu, Xinyan Velocity, Eisner, Jason, Lee, Holden, Cotterell, Ryan
Format: Preprint
Published: 2023
Subjects:
Online Access:https://arxiv.org/abs/2312.17710
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910283767742464
author Du, Li
Amini, Afra
Hennigen, Lucas Torroba
Yu, Xinyan Velocity
Eisner, Jason
Lee, Holden
Cotterell, Ryan
author_facet Du, Li
Amini, Afra
Hennigen, Lucas Torroba
Yu, Xinyan Velocity
Eisner, Jason
Lee, Holden
Cotterell, Ryan
contents Recent papers have demonstrated the possibility of energy-based text generation by adapting gradient-based sampling algorithms, a paradigm of MCMC algorithms that promises fast convergence. However, as we show in this paper, previous attempts on this approach to text generation all fail to sample correctly from the target language model distributions. To address this limitation, we consider the problem of designing text samplers that are faithful, meaning that they have the target text distribution as its limiting distribution. We propose several faithful gradient-based sampling algorithms to sample from the target energy-based text distribution correctly, and study their theoretical properties. Through experiments on various forms of text generation, we demonstrate that faithful samplers are able to generate more fluent text while adhering to the control objectives better.
format Preprint
id arxiv_https___arxiv_org_abs_2312_17710
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Principled Gradient-based Markov Chain Monte Carlo for Text Generation
Du, Li
Amini, Afra
Hennigen, Lucas Torroba
Yu, Xinyan Velocity
Eisner, Jason
Lee, Holden
Cotterell, Ryan
Computation and Language
Machine Learning
Recent papers have demonstrated the possibility of energy-based text generation by adapting gradient-based sampling algorithms, a paradigm of MCMC algorithms that promises fast convergence. However, as we show in this paper, previous attempts on this approach to text generation all fail to sample correctly from the target language model distributions. To address this limitation, we consider the problem of designing text samplers that are faithful, meaning that they have the target text distribution as its limiting distribution. We propose several faithful gradient-based sampling algorithms to sample from the target energy-based text distribution correctly, and study their theoretical properties. Through experiments on various forms of text generation, we demonstrate that faithful samplers are able to generate more fluent text while adhering to the control objectives better.
title Principled Gradient-based Markov Chain Monte Carlo for Text Generation
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2312.17710