Plagiarism detection in documents requires appropriate definition of document similarity and efficient computation of the similarity. This paper evaluates the validity of using vector representation of words for defining a document similarity in terms of the processing time and the accuracy in plagiarism detection. This paper proposes a plagiarism detection algorithm based on the score vector weighted by vector representation of words. The score vector between two documents represents the number of matches between corresponding words for every possible gap of the starting positions of the documents. The vector and its weighted version can be computed efficiently using convolutions. In this paper, two types of vector representation of words, that is, randomly generated vectors and a distributed representation generated by a neural network-based method from training data, are evaluated with the proposed algorithm. The experimental results show that using the weighted score vector instead of the normal one for the algorithm can reduce the processing time with a slight decrease of the accuracy, and that randomly generated vector representation is more suitable for the algorithm than the distributed representation in the sense of a tradeoff between the processing time and the accuracy.
