Fusing Stylometric and Embedding Systems to Estimate Authorship Likelihood Ratios in Japanese

A new study applies the likelihood ratio framework — the standard for logically and legally sound forensic evidence — to Japanese digital texts for the first time, fusing stylometric analysis with modern embedding-based systems to strengthen authorship comparisons [1]. The research, led by Shunichi Ishihara, used approximately 1,000-character excerpts from Japanese-language blogs to test whether combining traditional stylometric features with representations from pre-trained language models could improve forensic text comparison [1][2]. The likelihood ratio framework has been widely adopted across forensic sciences, but its application to authorship analysis had been limited to English-language texts until now [2]. The fused system maintained excellent calibration while increasing consistent-with-fact likelihood ratio magnitudes and decreasing contrary-to-fact magnitudes, improving overall discriminability [1]. The best-performing fusion achieved a log-likelihood-ratio cost of 0.32484 [1][2]. Authorship attribution in Japanese has historically relied on statistical classifiers such as Random Forests and Support Vector Machines, often tested on small groups of authors [3]. A separate study that fine-tuned a BERT model for Japanese author attribution reported 84 percent accuracy when identifying five authors and 82 percent accuracy across 80 authors using text data alone [3]. That work also found that adding stylistic features for a set of 25 authors reduced accuracy to 53 percent, underscoring the challenge of integrating feature-based and neural methods [3]. Other research has explored integrated ensembles of traditional feature-based and PLM-based methods on small-sample literary authorship tasks. One study reported that an integrated ensemble improved the F1 score by approximately 14 points over the best single model when the test corpus was not included in the pre-training data [4]. Ishihara’s earlier work with Satoru Tsuge, Mitsuyuki Inaba, and Wataru Zaitsu had already demonstrated the feasibility of deep-learning-based text representation for likelihood ratio estimation. In that 2022 study, documents of up to 100 words were mapped onto embedding vectors using RoBERTa, and scores from Cosine distance and probabilistic linear discriminant analysis were calibrated to likelihood ratios via logistic regression, yielding log-likelihood-ratio cost values of 0.55595 and 0.71591 respectively [5][6]. The new study extends that line of inquiry into Japanese and into fused multi-system architectures [1].