Plagiarism of academic texts among college students is common, and there is an urgent need for an efficient detection method to strengthen the management of academic texts among college students. This paper proposes a Bidirectional Encoder Representations from Transformers-Bidirectional Long Short-Term Memory (BERT-BiLSTM)-attention matching model based on deep learning to determine whether sentence pairs match, thereby detecting plagiarism. Word vectors were obtained through BERT, and deep features were extracted from texts based on BiLSTM-attention. The Softmax function outputted the results. Experiments found that the BERT-BiLSTM-attention model achieved a F1 value of 87.56% for the Microsoft Research Paraphrase Corpus (MSRP) set, outperforming methods such as term frequency-inverse document frequency, and a Plagdet of 82.38% for the PAN12 set, thus outperforming other methods in the literature being compared. The model achieved an F1 value of 91.4% for real academic texts. The results demonstrate the excellent performance of the BERT-BiLSTM-attention model in detecting text plagiarism behavior, demonstrating its practical applicability.
Shen Jiang (Wed,) studied this question.