Improving plagiarism detection in text document using hybrid weighted similarity

摘要

Plagiarism is a misconduct, which refers to the use of scientific and literary content contained in other sources without reference to them. Today, the rise of plagiarism has become a serious problem for publishers and researchers. Many researchers have discussed this problem and tried to identify types of plagiarism; however, most of these methods are not effective in detecting intelligent plagiarism. In other words, most of these methods focus on direct copying. Therefore, in this study, two methods are proposed to identify Extrinsic plagiarism. In both methods, to limit the search space, two stages of filtering based on the bag of word (BoW) technique are used at the document level and at the sentence level, and plagiarism is investigated only in the outputs of these two stages. In the first method to detect similarities in suspicious documents and sentences, the combination of pre-trained network technique of words embedding FastText and TF-IDF weighting technique to form two structural and semantic matrices and in the second method to form the two matrices, WordNet ontology and weighting TF-IDF is used. After forming the above matrices and calculating the similarity between the pairs of matrices of each sentence, using the Dice similarity and the structural similarity of the weighted composition, two similarity values are calculated. By comparing the similarity of suspicious sentences with the minimum threshold, the document containing the suspicious sentence receives the label of plagiarism or non-plagiarism. Experimental results on the PAN-PC-11 database show that the first method has achieved 95.1% precision and the second method 93.8% precision, which shows that the use of word embedding network compared to WordNet ontology can be more successful in detecting Extrinsic plagiarism.