Indonesian Hoax News Classification with Multilingual Transformer Model and BERTopic

Abstract

Technology and information growth enable internet users to play a role in disseminating information, including hoax news. One way that to avoid hoax news is to look for sources of information, but valid news is not always perceived as 'true' by individuals because human judgments can lead to bias. Several studies on automatic hoax news classification have been carried out using various deep learning approaches such as the pre-trained multilingual transformer model. This study focuses on classifying Indonesian hoax news using the pre-trained transformer multilingual model (XLM-R and mBERT) combined with a BERTopic model as a topic distribution model. The result shows that the proposed method outperforms the baseline model in classifying fake news in the low-resource language (Indonesian) with accuracy, precision, recall, and F1 results of 0.9051, 0.9515, 0.8233, and 0.8828 respectively. Povzetek: Raziskava se ukvarja z identifikacijo lažnih novic v Indoneziji s pomočjo modelov XLM-R and mBERT.

Simple Hack for Transformers Against Heavy Long-Text Classification on a Time- and Memory-Limited GPU Service

“…There are limited studies discussing Transformer models over Indonesian long-text classification. All of them use news bodies [20]- [24]. Hutama and Suhartono [20] study binary fake news detection using 1,100 news articles.…”

“…All of them use news bodies [20]- [24]. Hutama and Suhartono [20] study binary fake news detection using 1,100 news articles. Fawaid et al [21] study the same, using the same data, with an additional set containing 1,116 news articles.…”

“…Hutama et al [20], Fawaid et al [21], and Juarto et al [22] remove stopwords from the text and keep punctuation. Unfortunately, they neither report the comparison between the text with and without stopwords nor mention other studies to confirm that the method results better.…”

“…Hutama et al [20] and Fawaid et al [21] use multilingual models that we avoid when monolingual models are available. Multilingual models use shared vocabulary and the tokenizers produce more tokens [26], [27].…”

A Highly Accurate Internet-Based Fake Information Detection Tool for Indonesian Twitter

“…In more recent study, Hutama and Suhartono used the pre-trained transformer multilingual model (XLM-R and mBERT) in conjunction with a BERTopic model as a topic distribution model to categorize Indonesian fake news. They obtained an accuracy value of 90.51% [25].…”