Integrated ensemble of BERT- and feature-based models for authorship attribution in Japanese literary works

Traditional authorship attribution (AA) research has primarily relied on statistical analysis and classification based on stylistic features extracted from textual data. Although pre-trained language models like BERT have gained prominence in text classification tasks, their effectiveness in small-sample AA scenarios remains insufficiently explored. A critical unresolved challenge is developing methodologies that effectively integrate BERT with conventional feature-based approaches to advance AA research.

This study aims to substantially enhance performance in small-sample AA tasks through the strategic combination of traditional feature-based methods and contemporary BERT-based approaches. Furthermore, we conduct a comprehensive comparative analysis of the accuracy of BERT models and conventional classifiers while systematically evaluating how individual model characteristics interact within this combination to influence overall classification effectiveness.

We propose a novel integrated ensemble methodology that combines BERT-based models with feature-based classifiers, benchmarked against conventional ensemble techniques. Experimental validation is conducted using two literary corpora, each consisting of works from 10 distinct authors. The ensemble framework incorporates five BERT variants, three feature types, and two classifier architectures to systematically evaluate model effectiveness.

BERT demonstrated effectiveness in small-sample authorship attribution tasks, surpassing traditional feature-based methods. Both BERT-based and feature-based ensembles outperformed their standalone counterparts, with the integrated ensemble method achieving even higher scores. Notably, the integrated ensemble significantly outperformed the best individual model on Corpus B-which was not included in the pre-training data- improving the F1 score from 0.823 to 0.96. It achieved the highest score among all evaluated approaches, including standalone models and conventional ensemble techniques, with a statistically significant margin (p < 0.012, Cohen's d = 4.939), underscoring the robustness of the result. The pre-training data used in BERT had a significant impact on task performance, emphasizing the need for careful model selection based not only on accuracy but also on model diversity. These findings highlight the importance of pre-training data and model diversity in optimizing language models for ensemble learning, offering valuable insights for authorship attribution research and the broader development of artificial general intelligence systems.