Selecting a text similarity measure for a content-based recommender system

Abstract

PurposeThe purpose of this paper is to develop a journal recommender system, which compares the content similarities between a manuscript and the existing journal articles in two subject corpora (covering the social sciences and medicine). The study examines the appropriateness of three text similarity measures and the impact of numerous aspects of corpus documents on system performance.Design/methodology/approachImplemented three similarity measures one at a time on a journal recommender system with two separate journal corpora. Two distinct samples of test abstracts were classified and evaluated based on the normalized discounted cumulative gain.FindingsThe BM25 similarity measure outperforms both the cosine and unigram language similarity measures overall. The unigram language measure shows the lowest performance. The performance results are significantly different between each pair of similarity measures, while the BM25 and cosine similarity measures are moderately correlated. The cosine similarity achieves better performance for subjects with higher density of technical vocabulary and shorter corpus documents. Moreover, increasing the number of corpus journals in the domain of social sciences achieved better performance for cosine similarity and BM25.Originality/valueThis is the first work related to comparing the suitability of a number of string-based similarity measures with distinct corpora for journal recommender systems.