Selecting an optimal feature set for stance detection

Abstract

Stance detection is an automatic recognition of author’s view point in relation to a given object. An important stage of the solution process is determining the most appropriate way to represent texts. The paper proposes a new method of selecting an optimal feature set. The method is based on a homogenous ensemble of feature selection methods and a procedure of determining the optimal number of features. In this procedure the dependence of task performance on the number of features is approximated and the optimal number of features is determined by analyzing the growth rate of the function. There have been conducted experiments with text corpora consisting of “for” and “against” stances towards vaccinations of children, the Unified State Examination at school, and human cloning. The results demonstrate that the proposed method allows to achieve better performance in comparison with individual methods and even an overall feature set with a considerably fewer number of features.