A coarse-to-fine framework to efficiently thwart plagiarism

Abstract

This paper presents a systematic framework using multilevel matching approach for plagiarism detection (PD). A multilevel structure, i.e. document–paragraph–sentence, is used to represent each document. In document and paragraph level, we use traditional dimensionality reduction technique to project high dimensional histograms into latent semantic space. The Earth Mover’s Distance (EMD), instead of exhaustive matching, is employed to retrieve relevant documents, which enables us to markedly shrink the searching domain. Two PD algorithms are designed and implemented to efficiently flag the suspected plagiarized document sources. We conduct extensive experimental verifications including document retrieval, PD, the study of the effects of parameters, and the empirical study of the system response. The results corroborate that the proposed approach is accurate and computationally efficient for performing PD.

Introduction

The Internet has, undoubtedly, become an indispensable component of our daily life ranging from restaurant booking to technology research. The online fashion is, however, posing a severe challenge to textual intellectual property because the Internet and computer technology have made disseminating knowledge across the world facile. People can search, copy, save, and reuse online sources in ease. The most flagrant instance of plagiarism is to copy a document from another source without any kind of modifications. But this type of plagiarism is easy to be identified using the plagiarism detection (PD) system. Less obvious examples occur when people integrate an existing work into their work. They attempt to bypass the detection system by conducting substitution of words or sentences within an already existing document, or pasting some phrases from an outside source into a new document. Cut-and-paste PD, at present, has become a growing concern in education system. One of the difficulties of efficiently detecting plagiarism is to search the source with speedy query response because people may copy from one of millions of documents in the Internet, where each document usually involves thousands of words.

Existing techniques for anti-plagiarism include fingerprinting, a method developed specifically for detecting co-derivatives, and ranking, a method developed for document retrieval. Hoad and Zobel [1] investigated the performance of these techniques and demonstrated that the ranking method is superior to the fingerprinting method. Chow et al. [2] also reported promising results by using the ranking approach. Following this line, this paper presents a coarse-to-fine framework to detect plagiarism using multilevel matching (MLM). The proposed approach delivers a number of desirable features that include generality, robustness, and efficiency. Concretely, these features can be described as follows:

• The generality refers to the multilevel-structured document representation and its encoding features. We use document–paragraph–sentence structure to form a coarse-to-fine representation of each document. In document and paragraph level, principal component analysis (PCA), a traditional dimensionality reduction tool, is used to capture the hidden latent semantic topics. Instead of PCA, any other latent semantic analysis or dimensionality reduction techniques can be incorporated into this scheme.

• The proposed system is robust due to its use of signature matching. The signature in document and paragraph level is constructed by involving the length and the histograms of terms in each component. Each sentence is featured by using the index number of each term that indicates the presence of the corresponding term in vocabulary. In this signature encoding, we do not consider the sequence of terms in a sentence, which is reasonable because plagiarists strive to substitute words in each sentence or reorganize the sentence structure so as to bypass the PD system.

• Document modeling and its applications are notoriously computational intensive due to their involvement of at least thousands of words. Our proposed system is based on depth matching by using a coarse-to-fine strategy to filter out the unpromising searching domain. This pruning capability enables us to bring large computational efficiency. Therefore, the proposed approach can be used for a large dataset and practical online applications.

The main contributions of this paper are threefold. First, we propose a multilevel-structured document representation together with encoding features. Second, we investigate MLM approaches, i.e. histogram based MLM (MLMH) and signature based MLM (MLMS), for relevant document retrieval (DR). Third, two detection algorithms are implemented by setting appropriate thresholds such that undesirable paths are pruned in advance during multilevel matching process.

The remaining sections of this paper are organized as follows. A brief overview of document modeling and its applications are presented in Section 2. The relationship of PD versus document categorization (or classification, and or clustering) and DR are also discussed, respectively. Section 3 introduces a multilevel-structured document representation together with the document segmentation, dimensionality reduction, and feature encoding. Document segmentation is done using HTML tags. In Section 4, we discuss various document retrieval approaches based on histograms and signatures. Two detection algorithms are implemented in Section 5. We conduct extensive experimental verifications in Section 6. Section 7 lists the discussion based on observed results and proposes the system framework from a practical viewpoint. Finally, Section 8 ends the paper with conclusion and future work propositions.