Document Details
|
Document Type |
: |
Thesis |
|
Document Title |
: |
Arabic Rumor Detection Using Transfer Learning Based on Textual and Visual Features كشف الشائعات العربية باستخدام نقل المعرفة بالاعتماد على السمات النصية و المرئية |
|
Subject |
: |
Faculty of Computing and Information Technology |
|
Document Language |
: |
Arabic |
|
Abstract |
: |
Recently, the use of social media platforms has increased with ease of use and fast accessibility, making them a place of rumor proliferation owing to the lack of posting constraints and content authentication. Therefore, there is a need to leverage artificial intelligence techniques to detect rumors on social media platforms to prevent their adverse effects on society and individuals. Most existing studies that detect rumors in Arabic target the text of a tweet. Nevertheless, tweets contain different types of content, including text, images, videos, and URLs, and their visual features play an essential role in rumor diffusion.
This study proposes an Arabic rumor-detection model to detect rumors on Twitter using textual and visual image features through two types of multimodal fusion: early and late. Furthermore, it incorporated the transfer learning of recent Arabic versions of the pre-trained language model, bidirectional encoder representations from transformers (BERT), text-to-text-transfer-transformers (T5), and popular vision models VGG-19, ResNet50, and InceptionV3. Different experiments were conducted to select the best textual features among sixteen language models and visual feature extractors for building a multimodal model. MARBERTv2 was used as the textual feature extractor, whereas the ensemble of VGG-19 and ResNet50 was used as the visual feature extractor. Subsequently, the language and vision models of the single models were used as a baseline to compare the results with those of the multimodal models. The proposed models were evaluated using the publicly available dataset AraFacts, and regarding unbalanced classes, the multimedia dataset was collected and combined with AraFacts.
The experimental results of our multimodal model for early and late fusion were 85% and 84% f1-scores, respectively. Compared to the unimodal model, the MAR- BERTv2 and ensemble vision models achieve 90% and 79% f1-scores, respectively. The experimental results show that the multimodal models perform better than the vision-based model and cannot outperform the text-based model. Our findings demonstrate the effectiveness of textual features in rumor detection tasks compared with multimodal models.
Keywords: Arabic NLP, Artificial intelligence, Deep learning, Multimodal fusion, Rumor detection, Transfer learning. |
|
Supervisor |
: |
Dr. Amani Tariq Jamal |
|
Thesis Type |
: |
Master Thesis |
|
Publishing Year |
: |
1445 AH
2023 AD |
|
Co-Supervisor |
: |
Dr. Alaa Omar Khadidos |
|
Added Date |
: |
Friday, November 10, 2023 |
|
Researchers
| رشا مسلم البلوي | Albalawi, Rasha Musallam | Researcher | Master | |
|
Back To Researches Page
|