On the Weaknesses of Reinforcement Learning for Neural Machine Translation

Abstract

Reinforcement learning (RL) is frequently used to increase performance in text generation tasks, including machine translation (MT), notably through the use of Minimum Risk Training (MRT) and Generative Adversarial Networks (GAN). However, little is known about what and how these methods learn in the context of MT. We prove that one of the most common RL methods for MT does not optimize the expected reward, as well as show that other methods take an infeasibly long time to converge. In fact, our results suggest that RL practices in MT are likely to improve performance only where the pre-trained parameters are already close to yielding the correct translation. Our findings further suggest that observed gains may be due to effects unrelated to the training signal, but rather from changes in the shape of the distribution curve.

arXiv commentAccepted to ICLR 2020 (matching content, different style)

References in corpus (13)

• Continuous control with deep reinforcement learning
• Sequence Level Training with Recurrent Neural Networks
• Parameter Space Noise for Exploration
• Hindsight Experience Replay
• Language Generation with Recurrent Generative Adversarial Networks without Pre-training
• THUMT: An Open Source Toolkit for Neural Machine Translation
• Language GANs Falling Short
• Adversarial Neural Machine Translation
• XNMT: The eXtensible Neural Machine Translation Toolkit
• Neural Headline Generation with Sentence-wise Optimization
• DORA The Explorer: Directed Outreaching Reinforcement Action-Selection
• Evaluating Text GANs as Language Models
• Grammatical Error Correction with Neural Reinforcement Learning

Cited by in corpus (3)

• Reinforcement Learning for Generative AI: State of the Art, Opportunities and Open Research Challenges
• Neural Language Generation: Formulation, Methods, and Evaluation
• BERT as a Teacher: Contextual Embeddings for Sequence-Level Reward