Inducing features of random fields

摘要: We present a technique for constructing random fields from set of training samples. The learning paradigm builds increasingly complex by allowing potential functions, or features, that are supported large subgraphs. Each feature has weight is trained minimizing the Kullback-Leibler divergence between model and empirical distribution data. A greedy algorithm determines how features incrementally added to field an iterative scaling used estimate optimal values weights. models techniques introduced in this paper differ those common much computer vision literature underlying non-Markovian have number parameters must be estimated. Relations other approaches, including decision trees, given. As demonstration method, we describe its application problem automatic word classification natural language processing.