Context-Aware Detection of Content Evidencing Illegal or Predatory Fintech Lending Practices on Indonesian Digital Platforms Using Transformer-Based Language Models
Abstract
Illegal and predatory fintech lending practices generate digital content that may be difficult to identify using keyword-based screening because harmful practices can be expressed through formal, colloquial, and context-dependent language. This study benchmarks transformer-based language models for detecting content evidencing illegal or predatory fintech lending practices and examines whether Indonesian-specific pretraining provides an advantage over multilingual pretraining. A corpus of 10,000 Indonesian texts was collected from news, complaint platforms, regulatory reports, forums, consumer advocacy sources, and social media. The texts were annotated using six predefined indicators of harmful lending practices and normalised using a domain-specific slang dictionary and Sastrawi stemming. Multilingual BERT, IndoBERT, and IndoBERTweet were fine-tuned using weighted cross-entropy and compared with a TF-IDF logistic regression baseline. Multilingual BERT achieved the highest macro F1-score (0.9616), followed by IndoBERT (0.9582) and IndoBERTweet (0.9449), compared with 0.8974 for the baseline. Multilingual BERT missed only 9 of 413 harmful-content instances, reducing missed cases by 77.5% relative to the baseline. Indonesian-specific pretraining showed no clear advantage on the predominantly formal corpus, while stemming may reduce its benefit. The findings support transformer-based screening as a research benchmark for regulatory content monitoring, while human review remains necessary for enforcement decisions.
Related
- A lightweight hybrid CNN–vision transformer framework for image-based Android malware detection
- Transformer-aided guard: A hybrid spatio-temporal framework for confidence-aware intrusion detection in cyber-physical systems
- Benchmarking Embedding Models for POJK Regulatory Retrieval in Indonesian Insurance Documents
- Comparative phytochemical profiling and bioactivity evaluation of three Indonesian pomegranate (Punica granatum) peel varieties: Antioxidant, antibacterial, and in vitro antiproliferative activities
- One-Layer Transformer Provably Learns Multiclass One-Nearest Neighbor in Context
BibTeX
@article{suriani2026contextaware,
title = {{Context-Aware Detection of Content Evidencing Illegal or Predatory Fintech Lending Practices on Indonesian Digital Platforms Using Transformer-Based Language Models}},
author = {Uci Suriani and Netti Herawati and Ilsa Palingga Ninditama},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1903},
url = {https://doi.org/10.63158/journalisi.v8i5.1903},
note = {CC BY},
}
RIS
TY - JOUR
TI - Context-Aware Detection of Content Evidencing Illegal or Predatory Fintech Lending Practices on Indonesian Digital Platforms Using Transformer-Based Language Models
AU - Uci Suriani
AU - Netti Herawati
AU - Ilsa Palingga Ninditama
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1903
UR - https://doi.org/10.63158/journalisi.v8i5.1903
AB - Illegal and predatory fintech lending practices generate digital content that may be difficult to identify using keyword-based screening because harmful practices can be expressed through formal, colloquial, and context-dependent language. This study benchmarks transformer-based language models for detecting content evidencing illegal or predatory fintech lending practices and examines whether Indonesian-specific pretraining provides an advantage over multilingual pretraining. A corpus of 10,000 Indonesian texts was collected from news, complaint platforms, regulatory reports, forums, consumer advocacy sources, and social media. The texts were annotated using six predefined indicators of harmful lending practices and normalised using a domain-specific slang dictionary and Sastrawi stemming. Multilingual BERT, IndoBERT, and IndoBERTweet were fine-tuned using weighted cross-entropy and compared with a TF-IDF logistic regression baseline. Multilingual BERT achieved the highest macro F1-score (0.9616), followed by IndoBERT (0.9582) and IndoBERTweet (0.9449), compared with 0.8974 for the baseline. Multilingual BERT missed only 9 of 413 harmful-content instances, reducing missed cases by 77.5% relative to the baseline. Indonesian-specific pretraining showed no clear advantage on the predominantly formal corpus, while stemming may reduce its benefit. The findings support transformer-based screening as a research benchmark for regulatory content monitoring, while human review remains necessary for enforcement decisions.
SN - 2656-5935
SN - 2656-4882
ER -