Research on Semantic Matching of Thesis Themes Based on Comparative Learning and Adversarial Training

At the doctoral level in the field of computer science, the domain of literature plagiarism detection encompasses numerous plagiarism detection algorithms, but relatively fewer have focused on methods specifically designed for topic-based plagiarism detection. To address this research gap, this paper introduces a text semantic matching algorithm based on contrastive learning and adversarial training, which we have named "Comparative Adversarial Training (CAT)." Specifically, during the fine-tuning process, we generate affirmative and negative sentences using paper titles and keywords and optimize the model's performance through contrastive learning. Simultaneously, we generate adversarial samples by perturbing the word embedding matrix of the model and then employ these adversarial samples for adversarial training. Through experiments conducted on our proprietary internal dataset, the fine-tuned BERT model exhibited an average performance improvement of 3.71% compared to the baseline BERT model. Similarly, the fine-tuned RoBERTa model demonstrated a performance enhancement of 3.18% relative to the RoBERTa baseline model. Based on these experimental results, our algorithm's effectiveness has been thoroughly validated.