%0 Conference Proceedings
%T MixCE: Training Autoregressive Language Models by Mixing Forward and Reverse Cross-Entropies
%A Zhang, Shiyue
%A Wu, Shijie
%A Irsoy, Ozan
%A Lu, Steven
%A Bansal, Mohit
%A Dredze, Mark
%A Rosenberg, David
%Y Rogers, Anna
%Y Boyd-Graber, Jordan
%Y Okazaki, Naoaki
%S Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
%D 2023
%8 July
%I Association for Computational Linguistics
%C Toronto, Canada
%F zhang-etal-2023-mixce
%X 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.
%R 10.18653/v1/2023.acl-long.502
%U https://aclanthology.org/2023.acl-long.502
%U https://doi.org/10.18653/v1/2023.acl-long.502
%P 9027-9050