유의어 가중치를 적용한 논문 표절 유사도 측정 방법

Abstract

With the advancement of the Internet, various information can be easily found without the restrictions of time and place. However, there are cases where such information is abused in plagiarizing others’ theses. Plagiarism mostly occurs when using the content of other people’s paper without citation. Some of the common forms of plagiarisms are plagiarizing text, ideas, rewriting other people’s work, and interweaving various sources together in work. CopyKiller, the free service commonly used in plagiarism has standard processing criteria of judging plagiarism when six word-phrases are the same. However, this criterion seems insufficient since there can be rewordings.

The result of the earlier study showed that one document had an algorithm word A often appeared in the introduction and related studies but another word B which is a modification from the algorithm A, appeared in the main body. Since the two words are measured to be similar from the test, TF-IDF is modified to give weight by the paragraph of the introduction, related studies, main body, and conclusion, to improve the accuracy of similarity test on the words alike but have different meanings.

Hence, this paper proposes a method of measuring similarity by finding synonyms between the words and applying weight by the paragraph for the improvement of accuracy in document similarity checks. After preprocessing the data, vector values of the word are extracted by using Doc2Vec and word with weights by paragraph is clustered by DBSCAN(Density-based spatial clustering of applications with noise). The main word in each cluster is set as the representative word, and other words in the same cluster are set as synonyms. The distance value between the representative word and a synonym is used as synonym weight. Since the number is too low to be applied as weight with the distance value of the word alone, an L-R table is created via L-R syntactic parsing. The noun score is calculated using the score for the possible noun and is applied to the synonym weight.

To evaluate the similarity between the documents, similarity criteria by sentence are found first and the sentences having the similarity of 0.7 and higher are identified as the plagiarized sentence. The performance of the method suggested by this paper will be verified through a comparative evaluation with the existing studies.