Fine-tuning Pretrained Language Models for Multigenerator, Multidomain Machine Generated Text Detection

Machine-generated texts have become increasingly sophisticated due to the rise of large language models (LLMs), making it challenging to distinguish them from human-written content and raising ethical concerns. This has led to a surge in research on detecting machine-generated texts; however, most studies focus on just one or two popular generators and are limited to a few domains. To address these challenges, we developed a comprehensive detection system that identifies both the origin (human or machine) and the specific LLM responsible for generating the text. Our investigation utilized various machine learning (ML) algorithms (SVC, SGD, MNB, LR, XGBoost, and LightGBM) along with features like BoW and TF-IDF. Additionally, we fine-tuned pretrained language models (PLMs) such as BERT, Longformer, DeBERTa-v3, RoBERTa-base, and RoBERTa-large. Among these methods, RoBERTa-base achieved the highest accuracy, with 0.89 for distinguishing between human and machine-generated texts (Task A) and 0.86 for identifying the specific generator (Task B), highlighting the effectiveness of advanced transformer models in addressing the challenges of machine-generated text detection.