MixCE: Training Autoregressive Language Models by Mixing Forward and Reverse Cross-Entropies
Shiyue
Zhang
author
Shijie
Wu
author
Ozan
Irsoy
author
Steven
Lu
author
Mohit
Bansal
author
Mark
Dredze
author
David
Rosenberg
author
2023-07
text
Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Anna
Rogers
editor
Jordan
Boyd-Graber
editor
Naoaki
Okazaki
editor
Association for Computational Linguistics
Toronto, Canada
conference publication
Autoregressive language models are trained by minimizing the cross-entropy of the model distribution Q relative to the data distribution P – that is, minimizing the forward cross-entropy, which is equivalent to maximum likelihood estimation (MLE). We have observed that models trained in this way may “over-generalize”, in the sense that they produce non-human-like text. Moreover, we believe that reverse cross-entropy, i.e., the cross-entropy of P relative to Q, is a better reflection of how a human would evaluate text generated by a model. Hence, we propose learning with MixCE, an objective that mixes the forward and reverse cross-entropies. We evaluate models trained with this objective on synthetic data settings (where P is known) and real data, and show that the resulting models yield better generated text without complex decoding strategies.
zhang-etal-2023-mixce
10.18653/v1/2023.acl-long.502
https://aclanthology.org/2023.acl-long.502
2023-07
9027
9050