MixCE: Training Autoregressive Language Models by Mixing Forward and Reverse Cross-Entropies
This addresses text generation quality for natural language processing applications, but it is incremental as it modifies an existing training objective.
The paper tackles the problem of autoregressive language models producing non-human-like text by proposing MixCE, an objective that mixes forward and reverse cross-entropies, and shows it yields better generated text without complex decoding strategies.
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. Our code and models are publicly available at https://github.com/bloomberg/mixce-acl2023