Structured Output Learning with Candidate Labels for Local Parts

摘要

This paper introduces a special setting of weakly supervised structured output learning, where the training data is a set of structured instances and supervision involves candidate labels for some local parts of the structure. We show that the learning problem with this weak supervision setting can be efficiently handled and then propose a large margin formulation. To solve the non-convex optimization problem, we propose a proper approximation of the objective to utilize the Constraint Concave Convex Procedure (CCCP). To accelerate each iteration of CCCP, a 2-slack cutting plane algorithm is proposed. Experiments on some sequence labeling tasks show the effectiveness of the proposed method.