我们应用随机平均梯度(SAG)算法训练条件随机场(CRFs)。
We apply stochastic average gradient (SAG)algorithms for training conditional random fields (CRFs).
我们描述了一种利用CRF梯度中的结构来降低这种线性收敛随机梯度方法的内存需求的实用方案,提出了一种显著提高实用性能的非均匀采样策略,并分析了非均匀采样下的SAGA变异算法的收敛速率。
We describe a practical implementation thatuses structure in the CRF gradient to reduce the memory requirement of thislinearly-convergent stochastic gradient method, propose a non-uniform samplingscheme that substantially improves practical performance, and analyze the rateof convergence of the SAGA variant under non-uniform sampling.
实验结果表明,我们的方法往往显著优于现有方法的训练目标,在测试误差上的性能优于最优调谐的随机梯度方法。
Our experimental results reveal that ourmethod often significantly outperforms existing methods in terms of thetraining objective, and performs as well or better than optimally-tunedstochastic gradient methods in terms of test error.
条件随机场(CRFs)是自然语言处理中普遍使用的一种工具。
Conditional random fields (CRFs) are aubiquitous tool in natural language processing.
它们用于词性标注、语义角色标注、主题建模、信息提取、浅层解析、命名实体识别,以及自然语言处理和计算机视觉等其他领域的大量应用。
They are used for part-of-speech tagging ,semantic role labeling , topic modeling , information extraction , shallowparsing , named-entity recognition , as well as a host of other applications innatural language processing and in other fields such as computer vision.
与本文相关的一个网站供参考:
https://www.cs.ubc.ca/~schmidtm/Software/SAG4CRF.html
源码下载地址:
http://page2.dfpan.com/fs/alc7j2021929f16eba1/
更多精彩文章请关注微信号: