"MGR-DEL"

"The dataset file utilized for the MGR-DEL model is structured around two primary dataset files, encompassing both inference outputs and common sense analysis. It is composed of three JSON files, which correspond to the training, testing, and validation splits. Alongside these JSON files, an image folder is included to store associated visual data. Each JSON entry contains a series of well-defined fields that facilitate multimodal rumor detection and analysis.Specifically, the 'content'\u200b field holds the textual information or news body. The 'label'\u200b indicates the ground-truth classification, such as real or fake. The 'time'\u200b and 'source_id'\u200b provide temporal and source identifier metadata. The 'image_description'\u200b gives a descriptive caption of the accompanying image, while 'image_text_relevance'\u200b quantifies the alignment between the image and the text.Crucially, the 'td_rationale'\u200b and 'cs_rationale'\u200b fields capture reasoning for the text-driven and common sense inferences, respectively. The 'td_pred'\u200b and 'cs_pred'\u200b show the predicted labels from each analysis branch, and 'td_acc'\u200b and 'cs_acc'\u200b reflect the correctness of these predictions relative to the ground truth. The 'split'\u200b specifies whether the entry belongs to the train, test, or validation set. The 'image_path'\u200b links to the corresponding image file in the image folder, and the 'tag'\u200b may include additional categorical or keyword labels. This comprehensive structure supports multi-granularity reasoning by integrating visual, textual, temporal, and logical features for enhanced rumor veracity assessment."