Pemodelan Derajat Severitas Dangerous Speech pada Teks Daring Berdasarkan KOmbinasi Fitur Konten dan Struktur

Findawati, Yulian (2026) Pemodelan Derajat Severitas Dangerous Speech pada Teks Daring Berdasarkan KOmbinasi Fitur Konten dan Struktur. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7025211026_Doctoral.pdf Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Dangerous speech atau ujaran berbahaya di media sosial—terutama konten yang menghasut kekerasan terhadap kelompok tertentu—telah menjadi ancaman serius karena perannya dalam memicu konflik dunia nyata dan polarisasi masyarakat. Meskipun penelitian klasifikasi ujaran berbahaya telah banyak dilakukan, sebagian besar studi terdahulu berfokus pada konten tekstual dan fitur struktural tanpa mempertimbangkan dimensi kontekstual serta substansi pesan secara mendalam. Banyak penelitian mengabaikan apakah sebuah konten ujaran mengandung aspek sosiopolitik maupun narasi spesifik seperti dehumanisasi, accusation in a mirror (tuduhan cermin), serangan terhadap kelompok rentan (wanita/anak-anak), loyalitas terhadap grup, serta ancaman terhadap integritas kelompok, yang merupakan indikator kunci dari tingkat bahaya suatu ujaran. Munculnya dua atau lebih aspek secara bersamaan dalam satu tweet menunjukkan bahwa konten ujaran mengandung aspek multilabel. Untuk mengatasi kesenjangan tersebut, penelitian ini mengusulkan pemodelan derajat severitas berupa klasifikasi multiclass dangerous speech yang mengintegrasikan fitur multimodal teks, aspek dangerous speech berbasis multilabel, dan fitur struktural berupa metrik keterlibatan (seperti jumlah like dan retweet) serta tingkat pengaruh pengguna berbasis graf dengan target netral, hate speech dan dangerous speech. Evaluasi dilakukan secara komprehensif dengan menguji berbagai kombinasi fitur, meliputi fitur teks, aspek, struktur pengaruh pengguna, dan metrik keterlibatan untuk menentukan pengaruh masing-masing komponen terhadap performa model. Penelitian ini menggunakan dataset sebanyak 3.254 tweet dari periode 2019–2022, masa puncak polarisasi politik pasca-Pemilu di Indonesia. Tiga model dievaluasi: TF-DangerML (tradisional), BERT-DangerML (hibrida), dan Multi-DangerBERT (Pendekatan Multimodal Neural Network). Hasil ujicoba menunjukkan bahwa integrasi fitur berbasis aspek secara signifikan meningkatkan performa klasifikasi multiclass. Model TF-DangerML dengan XGBoost mencapai akurasi tertinggi sebesar 96,3% dan macro F1-score 95,9% melalui kombinasi fitur teks dan aspek. Di sisi lain, Multi-DangerBERT memberikan performa yang lebih stabil dan seimbang di seluruh kelas dengan skor ROC-AUC 0,983.Temuan ini menegaskan bahwa pendekatan multimodal yang menggabungkan analisis konten dan struktural sangat krusial untuk mengidentifikasi ujaran berbahaya secara lebih akurat.
=============================================================================================================
Dangerous speech on social media—particularly content inciting violence against specific groups—has become a serious threat due to its role in fueling real-world conflict and societal polarization. Although automatic classification research has been widely conducted, most prior studies focus on textual content and simple structural features without deeply considering contextual dimensions and message substance. Many studies overlook whether an utterance contains sociopolitical aspects or specific narratives such as dehumanization, accusation in a mirror, attacks on vulnerable groups (women/children), group loyalty, and threats to group integrity, which are key indicators of the danger level of a speech. The co-occurrence of two or more aspects within a single tweet indicates that the speech content contains multilabel aspects. To address this gap, this study proposes modeling severity levels through multiclass Dangerous Speech classification that integrates multimodal text features, multilabel-based Dangerous Speech aspects, and structural features—namely engagement metrics (such as likes and retweets) and graph-based user influence levels—targeting Neutral, Hate Speech, and Dangerous Speech classes. Evaluation was conducted comprehensively by testing various feature combinations, including text, aspects, user influence structure, and engagement metrics, to determine the contribution of each component to model performance. This study utilizes a dataset of 3,254 tweets from the 2019–2022 period, the peak of post-election political polarization in Indonesia. Three models were evaluated: TF-DangerML (traditional), BERT-DangerML (hybrid), and Multi-DangerBERT (a multimodal neural network approach). The experimental results show that the integration of aspect-based features significantly improves multiclass classification performance. The TF-DangerML model with XGBoost achieved the highest accuracy of 96.3% and a macro F1-score of 95.9% by combining text and aspect features. Conversely, Multi-DangerBERT provided more stable and balanced performance across all classes with a macro ROC-AUC score of 0.983. These findings confirm that a multimodal approach combining content and structural analysis is crucial for identifying dangerous speech more accurately.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	dangerous speech, derajat severitas, multiclass, multilabel, fitur konten, fitur struktural, multimodal,severity level, , content features, structural features.
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20001-(S3) PhD Thesis
Depositing User:	Yulian Findawati Findawati
Date Deposited:	22 Jan 2026 23:55
Last Modified:	22 Jan 2026 23:55
URI:	http://repository.its.ac.id/id/eprint/130059

Actions (login required)

View Item