Minimum Risk Training for Neural Machine Translation
Shiqi Shen†, Yong Cheng#, Zhongjun He+, Wei He+, Hua Wu+, Maosong Sun†, Yang Liu† †State Key Laboratory of Intelligent Technology and Systems Tsinghua National Laboratory for Information Science and Technology Department of Computer Science and Technology, Tsinghua University, Beijing, China #Institute for Interdisciplinary Information Sciences, Tsinghua University, Beijing, China +Baidu Inc., Beijing, China {vicapple22, {hezhongjun, hewei06, {sms, Corresponding author: Yang Liu.
Abstract
We propose minimum risk training for end-to-end neural machine translation. Unlike conventional maximum likelihood estimation, minimum risk training is capable of optimizing model parameters directly with respect to arbitrary evaluation metrics, which are not necessarily differentiable. Experiments show that our approach achieves significant improvements over maximum likelihood estimation on a state-of-the-art neural machine translation system across various languages pairs. Transparent to architectures, our approach can be applied to more neural networks and potentially benefit more NLP tasks.
中文速览
神经机器翻译（Neural Machine Translation, NMT）长期依赖最大似然估计（Maximum Likelihood Estimation, MLE）来训练模型，但这种方法存在两个硬伤：训练时模型只见过标准答案，测试时却要依赖自身预测逐词生成，导致误差积累；而且逐词优化交叉熵损失与真正衡量翻译质量的句子级指标（如BLEU）相关性很弱。为此，本文将最小风险训练（Minimum Risk Training, MRT）引入端到端神经机器翻译，直接把BLEU等评价指标作为损失函数，通过最小化模型在训练集上的期望损失来优化参数——由于梯度计算中不需要对损失函数求导，这一方法天然支持任意不可微的评价指标，且对模型结构无任何假设。为了应对搜索空间过大的问题，作者采用采样子集来近似后验分布，实验表明100个样本就已足够。在中英、英法、英德三个翻译任务上，MRT相比MLE基线均取得了显著的BLEU提升，证明了这一训练策略的有效性和通用性，也为其他NLP任务提供了可借鉴的优化思路。
原文 arXiv:1512.02433;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1512.02433v3