Fast Linearization of Tree Kernels over Large-Scale Data

Convolution tree kernels have been successfully applied to many language processing tasks for achieving state-of-the-art accuracy. Unfortunately, higher computational complexity of learning with kernels w.r.t. using explicit feature vectors makes them less attractive for large-scale data. In this paper, we study the latest approaches to solve such problems ranging from feature hashing to reverse kernel engineering and approximate cutting plane training with model compression. We derive a novel method that relies on reverse-kernel engineering together with an efficient kernel learning method. The approach gives the advantage of using tree kernels to automatically generate rich structured feature spaces and working in the linear space where learning and testing is fast. We experimented with training sets up to 4 million examples from Semantic Role Labeling. The results show that (i) the choice of correct structural features is essential and (ii) we can speed-up training from weeks to less than 20 minutes.