Inducing Features of Random Fields-英文文献.pdf

IEEE TRANSACTIONS PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 1997 1 Inducing Features of Random Fields Stephen Della Pietra, Vincent Della Pietra, and John Lafferty, Member, IEEE Abstract—We present a technique for constructing random fields from a the same string without a lowercase letter in that position. The set of training samples. The learning paradigm builds increasingly complex following collection of strings was generated from the resulting fields by allowing potential functions, or features, that are supported by increasingly large subgraphs. Each feature has a weight that is trained field by Gibbs sampling. (As for all of the examples that will be by minimizing the Kullback-Leibler divergence between the model and the shown, this sample was generated with annealing, to concentrate empirical distribution of the training data. A greedy algorithm determines the distribution on the more probable strings.) how features are incrementally added to the field and an iterative scaling algorithm is used to estimate the optimal values of the weights. m, r, xevo, ijjiir, b, to, jz, gsr, wq, vf, x, ga, The random field models and techniques introduced in this paper differ msmGh, pcp, d, oziVlal, hzagh, yzop, io, advzmxnv, from those common to much of the computer vision literature in that the ijv_bolft, x, emx, kayerf, mlj, rawzyb, jp, ag, underlying random fields are non-Markovian and have a large number of ctdnnnbg, wgdw, t, kguv, cy, spxcq, uzflbbf, parameters that must be estimated. Relations to other learning approaches, dxtkkn, cxwx, jpd, ztzh, lv, zhpkvnu, lˆ, r, qee, including decision trees, are given. As a demonstration of the method, we nynrx, atze4n, ik, se, w, lrh, hp+, yrqyka’h, describe its appli