OPTIMIZATION METHODS FOR DISCRIMINATIVE TRAINING OF CONDITIONAL RANDOM FIELDS BASED ON MINIMUM TAG ERROR
This paper investigated the optimization methods for discriminative training of Conditional Random Fields (CRFs) based on a new criterion called Minimum Tag Error (MTE). In order to accelerate the training process, two main algorithms (limited-memory BFGS and Stochastic Gradient Descent) were exploited for the training of CRFs. Experiments on the tasks of Chinese word segmentation on the PKU corpora from SIGHAN Bakeoff 2005 have demonstrated that stochastic gradient method not only achieves better performances comparing with Maximum A Posteriori (MAP) criterion, but also converges almost an order of magnitude faster than limited-memory BFGS without degrading the performance.
Conditional random fields Limited-memory BFGS Stochastic gradient descent Minimum tag error
Y.Xiong
School of Electronics & Information Engineering,Tongji University, 4800Rd Cao An. Shanghai, P.R. China, 201804
国际会议
2012 International Conference on System Simulation(2012年国际系统仿真学术会议)
上海
英文
510-514
2012-04-06(万方平台首次上网日期,不代表论文的发表时间)