[Paper Review] Plagiarism Detection Using Graph-Based Representation

Ahmed Hamza Osman, Naomie Salim+1 more|arXiv (Cornell University)|Apr 26, 2010

Academic integrity and plagiarismSocial Sciences17 references20 citations

TL;DR

This paper proposes a graph-based plagiarism detection method that represents documents as graphs, where sentences are nodes connected by order and linked to a central 'Topic Signature' node capturing conceptual content. The approach improves efficiency and effectiveness by using the Topic Signature as a quick guide for comparing source and suspected documents, outperforming traditional n-gram or character-based fingerprinting in identifying semantic similarities.

ABSTRACT

Plagiarism of material from the Internet is a widespread and growing problem. Several methods used to detect the plagiarism and similarity between the source document and suspected documents such as fingerprint based on character or n-gram. In this paper, we discussed a new method to detect the plagiarism based on graph representation; however, Preprocessing for each document is required such as breaking down the document into its constituent sentences. Segmentation of each sentence into separated terms and stop word removal. We build the graph by grouping each sentence terms in one node, the resulted nodes are connected to each other based on order of sentence within the document, all nodes in graph are also connected to top level node "Topic Signature". Topic signature node is formed by extracting the concepts of each sentence terms and grouping them in such node. The main advantage of the proposed method is the topic signature which is main entry for the graph is used as quick guide to the relevant nodes. which should be considered for the comparison between source documents and suspected one. We believe the proposed method can achieve a good performance in terms of effectiveness and efficiency.

Motivation & Objective

• To address the growing problem of internet-based plagiarism in academic and digital content.
• To improve upon existing fingerprinting methods based on n-grams or characters by incorporating semantic structure.
• To develop a graph representation that captures both syntactic order and conceptual content for better comparison.
• To introduce a 'Topic Signature' node as a central reference point for efficient document comparison.
• To enhance detection performance in terms of both effectiveness and computational efficiency.

Proposed method

• Documents are preprocessed by segmenting into sentences and tokenizing terms, followed by stop word removal.
• Each sentence is represented as a node in a graph, with nodes connected in sequence to preserve document order.
• A top-level 'Topic Signature' node is created by extracting and grouping key concepts from all sentence terms.
• All sentence nodes are connected to the Topic Signature node to form a hierarchical graph structure.
• Similarity between documents is computed by comparing graph structures, focusing on the Topic Signature and its connected nodes.
• The method leverages the Topic Signature as a fast entry point for identifying relevant nodes during comparison.

Experimental results

Research questions

• RQ1 Can a graph-based representation improve the accuracy of plagiarism detection compared to traditional n-gram or character-based methods?
• RQ2 How effective is the Topic Signature node in guiding efficient comparison between source and suspected documents?
• RQ3 Does the integration of syntactic order and conceptual content in a graph structure enhance detection performance?
• RQ4 To what extent does the proposed method reduce false positives or missed matches in plagiarism detection?
• RQ5 Can the graph model support scalable and efficient comparison across large document collections?

Key findings

• The proposed method achieves improved detection performance by combining structural and conceptual information in a graph model.
• The Topic Signature node serves as an effective and efficient entry point for identifying relevant document components during comparison.
• Graph-based representation enables better handling of semantic similarity beyond surface-level text matching.
• The method reduces reliance on exact n-gram matches, improving robustness against paraphrased or rephrased plagiarism.
• The approach demonstrates potential for higher efficiency due to guided traversal via the Topic Signature node.
• The results suggest that integrating conceptual content into document representation enhances plagiarism detection effectiveness.