Chinese chunk identification using SVMs plus Sigmoid

Abstract

The paper presents a method of Chinese chunk recognition based on Support Vector Machines (SVMs) plus Sigmoid. It is well known that SVMs are binary classifiers which achieve the best performance in many tasks. However, directly applying binary classifiers in the task of Chinese chunking will face the dilemmas that either two or more different class labels are given to a single unlabeled constituent, or no class labels are given for some unlabeled constituents. Employing sigmoid functions is a method of extracting probabilities (class/input) from SVMs outputs, which is helpful to post-processing of classification. These probabilities are then used to resolve the dilemmas. We compare our method based on SVMs plus Sigmoid with methods based only on SVMs. The experiments show that significant improvements have been achieved. © Springer-Verlag Berlin Heidelberg 2005.