Identifying artificial intelligence-generated content using the DistilBERT transformer and NLP techniques

Abstract

Natural language processing (NLP) has evolved significantly with the emergence of large language models (LLMs), leading to the rapid growth of artificial intelligence-generated content (AIGC). This expansion raises critical challenges in ensuring content authenticity and preventing the spread of misinformation and plagiarism. The identification of AIGC is an active research area and is significant for maintaining the authenticity and credibility of digital content in academic and professional environments. In this research study, we focus on identifying AIGC based on state-of-the-art deep learning-based transformers and by exploring deep features of textual content. The proposed DistilBERT transformer is an advanced and lightweight form of bidirectional encoder representations from transformers (BERT), a well-known transformer of many LLMs that utilizes a distilled transformer architecture with self-attention mechanisms that dynamically weigh textual elements based on contextual relevance, effectively capturing linguistic patterns. Additionally, this study explores both traditional machine learning with textual features and deep learning models integrated with word embeddings such as GloVe and Fast Text. Experimental analysis demonstrated that the proposed DistilBERT-based model achieved a superior predictive accuracy of 98%, outperforming traditional deep learning models, such as long short-term memory (LSTM) with GloVe embeddings, which achieved 93% accuracy. Furthermore, qualitative assessments validate the model's effectiveness in confidently classifying diverse textual samples, reinforcing its practical reliability.

Topics & Concepts