基于深度学习的翻译洗稿抄袭检测算法

Deep Learning-Based Translation Laundering Plagiarism Detection Algorithm

摘要: 为应对多媒体技术和互联网快速发展带来的多样化和新型化洗稿抄袭问题，本文提出了一种基于深度学习的翻译洗稿的抄袭检测算法，该算法通过融合多轮翻译后的特征来增强翻译文本的特征，从而得到高质量的文本表示，并利用对比学习架构拉近原文本在语义向量空间中与翻译文本的距离，同时保持其与负样本的距离。此外，本文通过改进的对比损失函数增强模型检测洗稿文本的能力。最后利用所构建的多元组翻译洗稿数据集来进行训练和验证，使之达到检测翻译洗稿抄袭的能力。实验结果表明，本文所提出的算法产生了质量更高的文本表示，从而在翻译洗稿抄袭检测任务上优于先前的方法，Spearman相关系数的结果也证明了所构建模型的优越性。

Abstract: In response to the increasingly diverse and novel plagiarism issues arising from the rapid development of multimedia technologies and the Internet, this paper introduces a deep learning-based plagiarism detection algorithm for translation laundering. This algorithm enhances the characteristics of translated texts by fusing features from multiple translation rounds, thereby achieving high-quality text representations. It utilizes a contrastive learning framework to narrow the distance between the original text and the spun text within the semantic vector space, while maintaining separation from negative samples. Additionally, the model’s ability to detect spun texts is bolstered by an improved contrastive loss function. The algorithm was trained and validated on a specially constructed multiset translation laundering dataset, to effectively detect plagiarism via translation laundering. Experimental results show that the proposed algorithm produces higher quality text representations and surpasses previous methods in detecting translation laundering plagiarism. The effectiveness of the constructed model is further affirmed by the Spearman correlation coefficient results.