Unique approach of contextual uniqueness assessment of research papers using multi-label classification and similarity detection

Abstract

Research papers should possess a significant amount of novelty. However, proper paraphrasing of entire paper may not be identified as plagiarism. Therefore, in this work a unique approach of similarity detection technique has been proposed. At first the abstracts are used to classify which area the paper belongs to, then papers of the same area are extracted and then contextual similarity is detected using cosine similarity. For training and testing purposes, two datasets have been created using data from IEEE Xplore. Since, almost every paper has some elements of paraphrased contents, therefore, a maximum acceptable threshold has also been proposed in this paper. Maintaining the maximum threshold, how well cosine similarity was able to distinguish the novel, paraphrased, and fully plagiarised works has also been analyzed.