摘要 |
<P>PROBLEM TO BE SOLVED: To provide a structure prediction model learning device for realizing memory saving while keeping prediction performance, a method, a program, and a recording medium. <P>SOLUTION: An auxiliary model which defines an auxiliary model parameter set θ<SP POS="POST">(k)</SP>by a logarithmic linear model is introduced, and a set Θ of the auxiliary model parameter set is estimated by using unsupervised data such that the Bregman distance between the auxiliary model and a reference function expressing the degree of a pseudo-correct answer is minimized. Then, a basic model parameter set λ which minimizes a previously defined experience risk function is estimated by using teacher data and Θ. <P>COPYRIGHT: (C)2012,JPO&INPIT |