Learning and evaluation of latent dependency forest models
Learning and evaluation of latent dependency forest models
复制标题
DOI：
10.1007/s00521-018-3504-3
复制
发表时间：
2018-05
影响因子：
6
通讯作者：
Yong Jiang;Yang Zhou;Kewei Tu
中科院分区：
文献类型：
--
作者：
Yong Jiang;Yang Zhou;Kewei Tu
潜在依赖森林模型是一种基于随机变量的动态依赖结构的新型概率模型。它们与其他概率模型的区别在于，在学习模型时不需要进行结构搜索。然而，由于配分函数无法跟踪计算，ldfm的参数学习仍然具有很大的挑战性。在本文中，我们研究和经验比较了LDFMs学习参数的几种近似或忽略学习目标中配分函数的算法。此外，我们还提出了一种估计LDFM配分函数的近似算法。实验结果表明：(1)我们的学习算法比以前的LDFMs学习算法取得了更好的效果；(2)我们的配分函数估计算法是准确的。
Latent dependency forest models (LDFMs) are a new type of probabilistic models with dynamic dependency structures over random variables. They distinguish themselves from other probabilistic models by the fact that there is no need for structure search when learning the models. However, parameter learning of LDFMs is still quite challenging since the partition function cannot be tractably calculated. In this paper, we investigate and empirically compare several algorithms of learning parameters of LDFMs which either approximate or ignore the partition function in the learning objective. Furthermore, we propose an approximate algorithm to estimate the partition function of LDFM. Experimental results show that (1) our learning algorithms can achieve better results than the previous learning algorithm of LDFMs, and (2) our partition function estimation algorithm is accurate.