AN ATTENTION-GUIDED MULTIMODAL DEBERTA FRAMEWORK FOR POLITICAL FAKE NEWS DETECTION WITH CONTEXTUAL METADATA FUSION AND STATE-OF-THE-ART PERFORMANCE ON THE LIAR DATASET
Volume: 15 - Issue: 01 - Date: 01-01-2026
Approved ISSN: 2278-1412
Published Id: IJAECESTU480 | Page No.: 101-105
Author: Jitendra Malviya
Co- Author: Nitya Khare
Abstract:-The rapid dissemination of fake news through digital platforms has emerged as a
critical societal challenge, particularly in the political domain where misinformation can influence
public opinion and democratic decision-making. Although transformer-based language models
have significantly improved text classification performance, most existing fake news detection
approaches rely solely on textual content, ignoring crucial contextual and source-related
information. This paper proposes an attention-guided multimodal fake news detection framework
built upon the DeBERTa-v3-base transformer, augmented with structured contextual metadata.
The proposed model jointly learns semantic representations of political statements and auxiliary
metadata features, including speaker credibility, subject category, and historical truthfulness
indicators, using an attention-based fusion mechanism. The framework is evaluated on the widely
used LIAR benchmark dataset containing 12,836 manually verified political statements.
Experimental results demonstrate that the proposed approach achieves an accuracy of 65.11%,
outperforming the recent state-of-the-art baseline by 9.32%. Comprehensive evaluation using
confusion matrix analysis, class-wise performance metrics, and convergence analysis confirms the
robustness and generalization capability of the model. Furthermore, an enterprise-ready
implementation is developed to support real-time and batch fake news detection. The results
validate that integrating contextual metadata with advanced transformer architectures
significantly enhances fake news detection performance in politically sensitive environments.
Key Words:-Fake News Detection, DeBERTa, Multimodal Learning, LIAR Dataset, Political Misinformation, Attention Mechanism
Area:-Engineering
Download Paper:
Preview This Article
