Optimizing Text Classification: A Comparative Analysis of Pre-Trained Models and Custom Neural Network Architectures on the AuTextification Dataset

One of the biggest challenges in the rapidly developing field of natural language processing is telling machine-generated text from human-generated material. In this work, a comparative analysis based on the AuTextification dataset—a heterogeneous mix of texts—is presented. Two primary strategies were used: optimizing pre-trained models and utilizing embeddings within a unique neural network design. Models like DistilBERT Base, BERT Base Uncased, and RoBERTa (both Base and Large variants) were included in the first approach; the RoBERTa Base model achieved a maximum accuracy of 52%. The second method produced a binary output by using a specially designed bi-LSTM network architecture with three dense layers. Interestingly, 70% accuracy was obtained when embeddings were made using the T5-3B model; however, 79% accuracy was achieved when an embedding layer that was directly implemented in Keras was used. The potential for improving text categorization tasks through embedding strategy optimization is highlighted by this study.