Abstract: In this study, we explore the application of DNN algorithm for detection of fake-news, focusing on the role of attention mechanisms in improving performance. Four Algorithms were developed and assessed CNN, Bi-LSTM, Attention Convolutional Neural Network (ACNN), & Attention Bidirectional LSTM (ABiLSTM) to investigate their ability to accurately identified faked news by appropriate capturing context and semantic information in text. The LIAR database was applied to comprehensively assess the performing of such algorithm across training, validation, and test sets. Our results show that deep learning technique could improved the ability of deep models to focus on key components of the text, thereby improving such abilities to distinguish fake from true news. Among the models, the two attention-based methods, ACNN and ABiLSTM, demonstrated higher test accuracy of 0.56, reflecting a slight improvement over their non-attention counterparts. Furthermore, these models maintained a desirable balance between precision and recall, which underscores their robustness and ability to perform well across different evaluation criteria.
Additionally, the F1-scores of the attention models were notably higher. Specifically, the ACNN and ABi-LSTM technique achieved F1-scores of 0.748 and 0.77 on the test set, outperforming the non-attention variants (CNN and BiLSTM). On the validation set, the F1-scores were 0.66 and 0.67, further validating their improved ability to extract and leverage important context-dependent features in the text. Among the two, ABiLSTM performed slightly better, suggesting that combining bidirectional LSTM with an attention mechanism is particularly effective for detecting fakes-news.
Overall, this studies highlights the potential of attention mechanisms to enhance deep neural network models by focusing on the most informative components of text. The results underscore the importance of integrating attention into deep architectures to achieve greater robustness, accuracy, and generalization in detecting fakes news tasks.
Keywords: Bi-LSTM, Hybrid Model, Fake News
|
DOI:
10.17148/IJARCCE.2025.14681
