Rich style embedding for intrinsic plagiarism detection

Abstract

Stylometry plays an important role in the intrinsic plagiarism detection, where the goal is to identify potential plagiarism by analyzing a document involving undeclared changes in writing style. The purpose of this paper is to study the interaction between syntactic structures, attention mechanism, and contextualized word embeddings, as well as their effectiveness on plagiarism detection. Accordingly, we propose a new style embedding that combines syntactic trees and the pre-trained Multi-Task Deep Neural Network (MT-DNN). Additionally, we use attention mechanisms to sum the embeddings, thereby experimenting with both a Bidirectional Long Short-Term Memory (BiLSTM) and a Convolutional Neural Network (CNN) maxpooling for sentences encoding. Our model is evaluated on two sub-task; style change detection and style breach detection, and compared with two baseline detectors based on classic stylometric features.