The Hybrid Method for Accurate Patent Classification

Abstract

This article is dedicated to stacking of two approaches of patent classification. First is based on linguistically-supported k-nearest neighbors algorithm using the method of search for topically similar documents based on a comparison of vectors of lexical descriptors. Second is the word embeddings based fastText, where the sentence (or a document) vector is obtained by averaging the n-gram embeddings, and then a multinomial logistic regression exploits these vectors as features. We show in Russian and English datasets that stacking classifier shows better results compared to single classifiers.