Exposure Bias versus Self-Recovery: Are Distortions Really Incremental for Autoregressive Text Generation?

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5087–5102 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5087 Exposure Bias versus Self-Recovery: Are Distortions Really Incremental for Autoregressive Text Generation? Tianxing He MIT cloudygoose@csail.mit.edu Jingzhao Zhang MIT jzhzhang@mit.edu Zhiming Zhou Shanghai Jiao Tong University heyohai@apex.sjtu.edu.cn James Glass MIT glass@mit.edu Abstract Exposure bias has been regarded as a cen- tral problem for auto-regressive language mod- els (LM). It claims that teacher forcing would cause the test-time generation to be incremen- tally distorted due to the training-generation discrepancy. Although a lot of algorithms have been proposed to avoid teacher forcing and therefore alleviate exposure bias, there is little work showing how serious the exposure bias problem actually is. In this work, we focus on the task of open-ended language generation, propose metrics to quantify the impact of ex- posure bias in the aspects of quality, diversity, and consistency. Our key intuition is that if we feed ground-truth data prefixes (instead of prefixes generated by the model itself) into the model and ask it to continue the generation, the performance should become much better because the training-generation discrepancy in the prefix is removed. Both automatic and hu- man evaluations are conducted in our experi- ments. On the contrary to the popular belief in exposure bias, we find that the the distor- tion induced by the prefix discrepancy is lim- ited, and does not seem to be incremental dur- ing the generation. Moreover, our analysis re- veals an interesting self-recovery ability of the LM, which we hypothesize to be countering the harmful effects from exposure bias. 1 Introduction Language model (LM) has been a central mod- ule for natural language generation (NLG) tasks (Young et al., 2017) such as open-ended language generation (Radford et al., 2018; Brown et al., 2020), machine translation (Wu et al., 2017), dia- logue response generation (Li et al., 2017), image captioning (Lin et al., 2014), etc. For decades, max- imum likelihood estimation (MLE) has been the most widely used objective for LM training. How- ever, there is a popular belief in the natural lan- guage processing (NLP) community that standard Figure 1: An illustration of the training-generation dis- crepancy and our key intuition for verifying exposure bias (in blue): During generation, if we feed ground- truth data prefixes (from PD) instead of the model’s own samples (PM), the generation performance should be better because the prefix discrepancy is removed. MLE training suffers from the exposure bias prob- lem which leads to an incremental performance degradation during test-time generation. The claim of the exposure bias problem (Bengio et al., 2015; Ranzato et al., 2016) is originated from the following discrepancy between MLE training and test-time generation for auto-regressive lan- guage models: During training, the model is trained to predict the next word conditioned on prefix (or history) words sampled from the ground-truth data distribution; While during generation, the model generates words conditioned on prefix sequences sampled from the model itself. Due to the exposure to …

Content truncated to 500 words