An Improved Outlier Detection Model for Detecting Intrinsic Plagiarism

Abstract

In the task of detecting intrinsic plagiarism, the cases where reference corpus is absent are to be dealt with. This task is entirely based on inconsistencies within a given document. Detection of internal plagiarism has been considered as a classification problem. It can be estimated through taking into consideration self-based information from a given document. The core contribution of the work proposed in this paper is associated with the document representation. Wherein, the document, also, the disjoint segments generated from it, have been represented as weight vectors demonstrating their main content. Where, for each element in these vectors, its average weight has been considered instead of its frequency. The proposed work has been evaluated in terms of Precision, Recall, F-measure, Granularity, and Plagdet. It is shown that the attained results are comparable to the ones attained by the best state-of-the-art methods. Where, through applying the proposed method to PAN-PC-09 and PAN-PC-11 for the detection of intrinsic plagiarism, a Recall scores of 0.4503 and 0.4303 have been recorded, even though further improvement for Precision (0.3308 and 0.2806) and Granularity (1.1765 and 1.1111) needs to be made. Concerning f-measure, the proposed approach has recorded 0.3814 and 0.3397. In terms of the total performance of a plagiarism detection approach, Plagdet, the proposed method has recorded 0.3399 and 0.3151.

Mining Deviations in Document Writing Style through Vector Dissimilarity

“…For each document, only information that included is to be taken in consideration. As revealed in Table 1 and 2, the proposed approach has achieved a noticeable improvement over the work in [17] in terms of Granularity and also for Plagdet, which defines the complete performance of a plagiarism detection method. Moreover, stability in the proposed approach performance is recognized for both corpora.…”

“…Precision, Recall and Granularity need to be combined for an overall score called Plagdet to achieve a total order for the reason that they permit for a partial ordering among algorithms that detects plagiarism. Through the use of 𝑃𝐴𝑁 − 𝑃𝐶 − 09 and 𝑃𝐴𝑁 − 𝑃𝐶 − 11 as an evaluation dataset, Table 1 and Table 2 illustrate the results of comparing the approach proposed in this paper against the work of [6] and the approach proposed in [17].…”

“…Table1: Performance comparison of the proposed model against 𝑀𝑜𝑑𝑒𝑙 [6] and 𝑀𝑜𝑑𝑒𝑙 [17] in terms of 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛, 𝑅𝑒𝑐𝑎𝑙𝑙, 𝐹 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒, 𝐺𝑟𝑎𝑛𝑢𝑙𝑎𝑟𝑖𝑡𝑦, and 𝑃𝑙𝑎𝑔𝑑𝑒𝑡 evaluated using 𝑃𝐴𝑁 − 𝑃𝐶 − 09 corpora. [6] and 𝑀𝑜𝑑𝑒𝑙 [17] in terms of 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛, 𝑅𝑒𝑐𝑎𝑙𝑙, 𝐹 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒, 𝐺𝑟𝑎𝑛𝑢𝑙𝑎𝑟𝑖𝑡𝑦, and 𝑃𝑙𝑎𝑔𝑑𝑒𝑡 evaluated using 𝑃𝐴𝑁 − 𝑃𝐶 − 11 corpora. For each document, only information that included is to be taken in consideration.…”

“…The proposed method considers the work introduced in [17] for building the profiles for the entire document and the produced segments. As a feature, it uses the significance of word considering its average weight instead of its frequency.…”

“…Authors in [17] proposed a model that constructs profiles for the suspicious document and the spawned segments based on considering average weight of word uni-grams rather than their frequency.…”