Rethinking Exposure Bias In Language Modeling
This addresses a common training issue in language modeling for NLP applications, but it is incremental as it builds on existing RL and GAN methods.
The paper tackles exposure bias in language models by introducing multi-range reinforcing and multi-entropy sampling to amplify and denoise reward signals, resulting in improved BLEU scores and a new robustness metric called road exam.
Exposure bias describes the phenomenon that a language model trained under the teacher forcing schema may perform poorly at the inference stage when its predictions are conditioned on its previous predictions unseen from the training corpus. Recently, several generative adversarial networks (GANs) and reinforcement learning (RL) methods have been introduced to alleviate this problem. Nonetheless, a common issue in RL and GANs training is the sparsity of reward signals. In this paper, we adopt two simple strategies, multi-range reinforcing, and multi-entropy sampling, to amplify and denoise the reward signal. Our model produces an improvement over competing models with regards to BLEU scores and road exam, a new metric we designed to measure the robustness against exposure bias in language models.