MULTILINGUAL, MULTIMODAL AND EXPLAINABLE APPROACHES FOR AUTOMATED FACT-CHECKING PROBLEM

Date

2025-1-10

Author

Çekinel, Recep Fırat

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

62
views

41
downloads

Cite This

Automated fact-checking methods primarily rely on content-based approaches, utilizing deep neural networks to extract sophisticated features from text for prediction. However, the inherently black-box nature of these models makes their decision-making processes challenging to interpret. Another challenge for automated fact-checking models is their dependence on language-specific data, with limited multilingual datasets available for training. Moreover, the multimodal nature of fake posts—including text, images, and speech—presents an additional challenge. This thesis addresses automated fact-checking research, aiming to predict the veracity of claims while extending contributions to explainable solutions for fact-checking and sarcasm detection. We propose explainable models through multi-task learning and causal inference, evaluate cross-lingual transfer learning for low-resource languages, and examine how recent VLMs utilize text and image information for fact-checking. Our multi-task learning approach involves a T5-based encoder-decoder model trained for text summarization and veracity prediction, with generated summaries serving as explanations for predicted veracity labels. Moreover, a Turkish fact-checking dataset is released and experiments are conducted using transfer learning and machine translation to address data scarcity. In multimodality, we investigate VLMs' effectiveness in representing text and image information, finding that while multimodal embeddings generally enhance performance, discrete text-only and image-only models often outperform them. Lastly, we apply causal inference to text analysis, examining how sarcastic linguistic features and punctuation impact text popularity and leveraging clustering and topic modeling to uncover latent information on irony and popularity.

Subject Keywords

fact-checking, explainability, cross-lingual learning, multimodality, causal inference

URI

https://hdl.handle.net/11511/113463

Collections

Graduate School of Natural and Applied Sciences, Thesis