A Population-based Plagiarism Detection using DistilBERT-Generated Word Embedding

Plagiarism is the unacknowledged use of another person’s language, information, or writing without crediting the source. This manuscript presents an innovative method for detecting plagiarism utilizing attention mechanism-based LSTM and the DistilBERT model, enhanced by an enriched differential evolution (DE) algorithm for pre-training and a focal loss function for training. DistilBERT reduces BERT’s size by 40% while maintaining 97% of its language comprehension abilities and being 60% quicker. Current algorithms utilize positive-negative pairs to train a two-class classifier that detects plagiarism. A positive pair consists of a source sentence and a suspicious sentence, while a negative pair comprises two dissimilar sentences. Negative pairs typically outnumber positive pairs, leading to imbalanced classification and significantly lower system performance. To combat this, a training method based on a focal loss (FL) is suggested, which carefully learns minority class examples. Another addressed issue is the training phase, which typically uses gradient-based methods like back-propagation for the learning process. As a result, the training phase has limitations, such as initialization sensitivity. A new DE algorithm is proposed to initiate the back-propagation process by employing a mutation operator based on clustering. A successful cluster for the current DE population is found, and a fresh updating approach is used to produce potential solutions. The proposed method is assessed using three datasets: SNLI, MSRP, and SemEval2014. The model attains excellent results that outperform other deep models, conventional, and population-based models. Ablation studies excluding the proposed DE and focal loss from the model confirm the independent positive incremental impact of these components on model performance.

@article{JING2023,
  title     = {A Population-based Plagiarism Detection using DistilBERT-Generated Word Embedding},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {8},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Yuqin JING and Ying LIU},
  doi       = {10.14569/IJACSA.2023.0140868},
  url       = {https://doi.org/10.14569/IJACSA.2023.0140868}
}