Japanese Author Attribution Using BERT Finetuning with Stylometric Features

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F26%3AYZ8DATAH" target="_blank" >RIV/00216208:11320/26:YZ8DATAH - isvavai.cz</a>

Výsledek na webu

<a href="http://dx.doi.org/10.1007/978-981-96-5123-8_20" target="_blank" >http://dx.doi.org/10.1007/978-981-96-5123-8_20</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-981-96-5123-8_20" target="_blank" >10.1007/978-981-96-5123-8_20</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Japanese Author Attribution Using BERT Finetuning with Stylometric Features

Popis výsledku v původním jazyce

This study investigates author attribution (AA) in Japanese texts through fine-tuning the pre-trained BERT model “cl-tohoku/bert-large-japanese-v2” with Japanese-specific stylometric features. Experiments explored combinations of these features with classifiers such as LR, SVM, and RF, across varying author counts from 5 to 75. Focusing solely on native Japanese compositions, the study utilized the “Composition Bilingual Database” from the National Institute for Japanese Language and Linguistics to maintain linguistic consistency. The BERT model combined with LR achieved the highest accuracy of 96.3% for 5 authors, demonstrating deep learning’s potential in Japanese AA. However, high-dimensional stylistic features introduced noise when integrated, highlighting challenges in feature alignment. Future work will explore advanced non-linear models like XGBoost, LightGBM, and CatBoost for improved feature integration, and low-resource classification methods such as prototypical networks to enhance performance without extensive dataset expansion. Additionally, further testing of alternative Japanese pre-trained language models will be conducted to capture linguistic nuances more effectively. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

Název v anglickém jazyce

Japanese Author Attribution Using BERT Finetuning with Stylometric Features

Popis výsledku anglicky

This study investigates author attribution (AA) in Japanese texts through fine-tuning the pre-trained BERT model “cl-tohoku/bert-large-japanese-v2” with Japanese-specific stylometric features. Experiments explored combinations of these features with classifiers such as LR, SVM, and RF, across varying author counts from 5 to 75. Focusing solely on native Japanese compositions, the study utilized the “Composition Bilingual Database” from the National Institute for Japanese Language and Linguistics to maintain linguistic consistency. The BERT model combined with LR achieved the highest accuracy of 96.3% for 5 authors, demonstrating deep learning’s potential in Japanese AA. However, high-dimensional stylistic features introduced noise when integrated, highlighting challenges in feature alignment. Future work will explore advanced non-linear models like XGBoost, LightGBM, and CatBoost for improved feature integration, and low-resource classification methods such as prototypical networks to enhance performance without extensive dataset expansion. Additionally, further testing of alternative Japanese pre-trained language models will be conducted to capture linguistic nuances more effectively. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2025.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

—

Rok uplatnění

2025

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název statě ve sborníku

Commun. Comput. Info. Sci.

ISBN

978-981-96-5122-1

ISSN

—

e-ISSN

—

Počet stran výsledku

15

Strana od-do

293-307

Název nakladatele

Springer Science and Business Media Deutschland GmbH

Místo vydání

—

Místo konání akce

Beijing

Datum konání akce

1. 1. 2026

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—