Automated Approach for Dating English Text Using Transformer Neural Networks

The paper examines the existing methods of text dating using neural networks, highlighting their advantages and limitations. Text dating is a crucial task in fields such as history, archival studies, linguistics, and forensic science, as accurately determining the creation time of a document can help verify its authenticity, establish authorship, and detect forgeries. However, traditional methods based on stylometric or statistical approaches often lack accuracy, especially when dealing with large volumes of text data. This study proposes an approach for dating English-language texts using transformer neural networks. The model achieves an accuracy of 85 % within a 30-year range for texts written between the 15th and 20th centuries, outperforming existing models applied to English text. The core idea of the proposed automated approach is to utilize transfer learning to fine-tune a pre-trained transformer neural network, optimizing it for the classification of text fragments by decade. One key advantage of this approach is the use of transformer architecture, which, through the self-attention mechanism, effectively captures complex relationships within a text. Another significant benefit is the application of transfer learning, which reduces training time and computational resources compared to training a model from scratch. The approach was implemented in Python using the transformers libraries for training and testing the neural network, datasets for working with the dataset, and numpy for the calculations. Experimental results demonstrated high accuracy: 86 % within a 30-year range and 73 % within a 20-year range on the test dataset. For the 19th and 20th centuries, the model achieved an accuracy of 89% and 90%, respectively, while accuracy for earlier centuries was lower, averaging around 30%. The research also examines the possibility of identifying features that indicate a text's association with a specific period by extracting words with the highest attention scores. Future research will focus on improving the accuracy for underrepresented historical periods by expanding and refining the dataset. Further enhancements may be achieved by optimizing model hyperparameters and experimenting with alternative neural network architectures. Another direction for future research is to explore methods for identifying linguistic or stylistic features that mark texts as belonging to a certain historical period, in order to make the neural network's results more interpretable for the user. The proposed approach has potential applications in historical research, document authentication, plagiarism detection, literary studies, and forensic analysis.