低階多項式自然語言處理之資料映射同時利用雜湊達成特 徵空間壓縮

Abstract

Recently, many people handle natural language processing (NLP) tasks via support vector machines (SVM) with polynomial kernels. However, kernel computation is time consuming. Chang et al. (2010) have proposed mapping data by low-degree polynomial functions and applying fast linear-SVM methods. For data with many features, they have considered condensing data to effectively solve some memory and computational difficulties. In this thesis, we investigate Chang et al.'s methods and give implementation details. We conduct experiments on four NLP tasks to show the viability of our implementation.