Detecting Similarity in Paraphrased Persian Texts using Semantic and Probabilistic Methods

• Published in: Pizhūhishnāmah-i pardāzish va mudiriyyat-i iṭṭilāʻāt (Online) Vol. 34; no. 4; pp. 1823 - 1848
• Main Authors: Nasrollah Pakniat, Azadeh Mohebi
• Format: Journal Article
• Language: Persian
• Published: Iranian Research Institute for Information and Technology 01.09.2019
• Subjects: paraphrased texts; plagiarism; probabilistic text alignment; semantic text alignment
• ISSN: 2251-8223; 2251-8231

Plagiarism detection is the process of locating instances of plagiarism within a work or document. The main component of a plagiarism detection system is its text alignment algorithm aiming at detecting paraphrased passages of texts in a suspicious document, using a small set of candidate source documents. As text alignment algorithms are highly language-dependent, thus the numerous existing algorithms for other languages rather than Pesian cannot be employed for Persian plagiarism detection puposes. There are different text alignment algorithms for Persian text, while most of them are only able to detect exactly identical passages shared between texts. However, in many cases of plagiarism detection we are coping with the problem of finding similar pasaages that are already paraphrased. In this paper, we propose two new text alignment algorithms which are able to detect paraphrased texts in Persian language. The first one is a semantic algorithm that employs a dictionary to detect paraphrased sentences and the second one is a probabilistic algorithm that uses the statistical information obtained from a large corpus of Persian texts to detect similar texts. Compared to the other existing semantic text alignment algorithms, the proposed algorithms use different measures to check the similarity between the text sentences. Furthermore, the probabilistic algorithm is the first probabilistic text alignment algorithm proposed for the Persian language. Moreover, while all the existing text alignment algorithms check the similarity between any two sentences of the text separately, the proposed algorithms consider the similarity neighboring sentences in the text as well. The implementation results indicate that while the quality of both algorithms in detecting paraphrased texts is high enough and almost the same as each other, the proposed probabilistic method is more efficient than the proposed semantic algorithm, in terms of computation time.