Text Generation by Learning from Demonstrations
arXiv:2009.07839
Abstract
Current approaches to text generation largely rely on autoregressive models and maximum likelihood estimation. This paradigm leads to (i) diverse but low-quality samples due to mismatched learning objective and evaluation metric (likelihood vs. quality) and (ii) exposure bias due to mismatched history distributions (gold vs. model-generated). To alleviate these problems, we frame text generation as an offline reinforcement learning (RL) problem with expert demonstrations (i.e., the reference), where the goal is to maximize quality given model-generated histories. We propose GOLD (generation by off-policy learning from demonstrations): an easy-to-optimize algorithm that learns from the demonstrations by importance weighting. Intuitively, GOLD upweights confident tokens and downweights unconfident ones in the reference during training, avoiding optimization issues faced by prior RL approaches that rely on online data collection. According to both automatic and human evaluation, models trained by GOLD outperform those trained by MLE and policy gradient on summarization, question generation, and machine translation. Further, our models are less sensitive to decoding algorithms and alleviate exposure bias.
arXiv commentICLR 2021
References in corpus (9)
- Teaching Machines to Read and Comprehend
- Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks
- Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems
- A Deep Reinforcement Learning Chatbot
- How (not) to Train your Generative Model: Scheduled Sampling, Likelihood, Adversary?
- Way Off-Policy Batch Deep Reinforcement Learning of Implicit Human Preferences in Dialog
- Sticking to the Facts: Confident Decoding for Faithful Data-to-Text Generation
- CoT: Cooperative Training for Generative Modeling of Discrete Data
- Distributional Reinforcement Learning for Energy-Based Sequential Models