This paper introduces hinge-loss Markov random fields (HL-MRFs), a new class of probabilistic graphical models particularly well-suited to large-scale structured prediction and learning. We derive HL-MRFs by unifying and then generalizing three different approaches to scalable inference in structured models: (1) randomized algorithms for MAX SAT, (2) local consistency relaxation for Markov random fields, and (3) reasoning about continuous information with fuzzy logic. To make HL-MRFs easy to construct and use, we next present probabilistic soft logic (PSL), a new probabilistic programming language for defining HL-MRFs for relational data. We then introduce a convex optimization algorithm based on message passing for exact MAP inference in HL-MRFs, as well as algorithms for weight learning.

本文介绍了两种新的模拟结构化数据的形式，并证明它们可以捕捉到丰富的结构并可扩展到大规模数据。第一种形式是铰接损失马尔可夫随机场（HL-MRFs），是一种新的概率图模型，可以推广凸轮廓推断的不同方法。第二种新形式是概率软逻辑（PSL），是一种基于一阶逻辑的概率编程语言，使HL-MRFs易于定义。这些算法使HL-MRFs和PSL能够可扩展地模拟以前无法达到的大规模丰富结构数据。

铰链损失马尔可夫随机场和概率软逻辑