Keyword extraction using supervised cumulative TextRank

Abstract

Keyword extraction is a major step to extract plenty of valuable and meaningful information from the rich source of World Wide Web (W.W.W.). Different keyword extraction algorithms are proposed with their own advantages and disadvantages. Vector Space Model (VSM) algorithms prove quite effective for keyword extraction, but do not emphasize on the class label information of classified data. Supervised Term Weighting (STW) algorithms address this problem, but suffer from high dimensionality. Besides, they do not incorporate semantic relationship between terms. To address these problems, Graph Based Models (GBM) are introduced. However, they also use unsupervised learning. Hence, this paper proposes a Keyword Extraction using Supervised Cumulative TextRank (KESCT) technique that explores the benefits of both VSM and GBM techniques. The proposed algorithm modifies TextRank by incorporating a novel Unique Statistical Supervised Weight (USSW) to include class label information of classified data. To emphasize on the relatedness between terms, the mutual information between terms is also included. The proposed algorithm is validated using four review datasets and results are compared with traditional TextRank and its variants using Support Vector Machine (SVM) classifier, Naive-Bayes (NB) classifier and an ensemble classifier. Experimental results mark the efficacy of the proposed algorithm over existing algorithms.