Deteksi Misinformasi Berita yang Tidak Sesuai Konteks Menggunakan Evidence Semantic Clustering Large Language Model

Budhiantono, Cinta Aushafa Putri (2026) Deteksi Misinformasi Berita yang Tidak Sesuai Konteks Menggunakan Evidence Semantic Clustering Large Language Model. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221002-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Berita hoaks masih marak menyebarkan misinformasi yang dapat menyebabkan berbagai dampak, mulai dari kesalahpahaman publik hingga konflik antarnegara. Salah satu jenis hoaks yang paling umum namun sulit dideteksi adalah hoaks *out-of-context*, yaitu berita yang menggunakan gambar atau informasi asli dalam konteks yang berbeda sehingga menyesatkan. Penelitian Tugas Akhir ini bertujuan mendeteksi berita hoaks *out-of-context* dari artikel yang memuat gambar dan teks. Penelitian ini mengembangkan model **Evidence Semantic Clustering Large Language Model (ESC-LLM)** untuk mendeteksi hoaks *out-of-context* berbasis model Sniffer. Model Sniffer masih bergantung pada *evidence* berupa citra asli sehingga efektivitas pendeteksiannya terbatas. ESC-LLM terdiri atas dua modul utama, yaitu (i) modul *internal checking* yang memanfaatkan model I-JEPA dan Q-Former untuk membentuk representasi fitur semantik antara citra dan *caption* berita secara efisien, serta (ii) modul *external checking* yang menambahkan *Semantic Clustering* untuk menyediakan *evidence* berupa teks berita asli. Penelitian ini mengumpulkan data berita politik di Indonesia dari Agustus 2020 hingga Desember 2025. Hasil eksperimen menunjukkan bahwa model ESC-LLM menghasilkan performa yang lebih baik dibandingkan metode *baseline* pada sebagian besar metrik evaluasi.
================================================================================================================================
Hoax news continues to spread misinformation, leading to consequences ranging from public misunderstanding to conflicts between countries. One of the most common yet challenging types of hoax is **out-of-context** news, in which authentic images or information are presented in misleading contexts. This final project aims to detect out-of-context hoax news from articles containing both images and text. To achieve this objective, an **Evidence Semantic Clustering Large Language Model (ESC-LLM)** is proposed based on the Sniffer model. The original Sniffer model relies primarily on evidence extracted from original images, which limits its effectiveness in detecting out-of-context hoaxes. The proposed ESC-LLM consists of two main modules: (i) an *internal checking* module based on I-JEPA and Q-Former to efficiently construct semantic feature representations between news images and captions, and (ii) an *external checking* module that incorporates *Semantic Clustering* to provide supporting evidence from original news texts. This study collected Indonesian political news published between August 2020 and December 2025. Experimental results demonstrate that the proposed ESC-LLM outperformed the baseline method across most evaluation metrics.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Berita Palsu, Deteksi Misinformasi Out-of-Context, Multimodal Large Language Model, Semantic Clustering ======================================================================================================================================================= Hoax News, Out-of-Context Misinformation, Multimodal Large Language Model, Semantic Clustering
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Cinta Aushafa Putri Budhiantono
Date Deposited:	04 Aug 2026 10:22
Last Modified:	04 Aug 2026 10:22
URI:	http://repository.its.ac.id/id/eprint/143457

Actions (login required)

View Item