Semi-automatic Codification of Economical Activities with Support Vector Machines

Abstract

The correct attribution of codes to economical activities is of ultimate importance for fiscal and administrative purposes. In a joint effort involving federal, state and city business regulating offices, the unification and automation of this codification process is in progress. As the input information is in the form of free textual entries, techniques used in multi-class, multi-label text categorization are well suited. Due to the fact that the possible codes are in the order of one thousand, this problem is under investigation through an approach based on a hierarchical use of Support Vector Machines (SVM). The hierarchical organization of several SVMs split the problem into smaller ones and allows a fine tuning of the code attribution obeying the very hierarchy present in the table describing the codes. This paper will introduce the problem, describe the proposed approach based on SVMs and show some results that validate the approach.