RAGAT-Mind: A Multi-Granular Modeling Approach for Rumor Detection Based on MindSpore

Zhenkai Qin 1,2,3 1 School of Information Technology 2 Network Security Research Center 3 Big Data and Policing Technology Laboratory Guangxi Police College, Nanning, China [email protected] Guifang Yang School of Information Technology Guangxi Police College 530028, China yangguifang@gxjcxy. · 2025

Read RAGAT-Mind: A Multi-Granular Modeling Approach for Rumor Detection Based on MindSpore by Zhenkai Qin 1,2,3 1 School of Information Technology 2 Network Security Research Center 3 Big Data and Policing Technology Laboratory Guangxi Police College, Nanning, China [email protected] Guifang Yang School of Information Technology Guangxi Police College 530028, China yangguifang@gxjcxy. (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-17574.txt.

Introduction

Zhenkai Qin and co-authors present RAGAT-Mind, a model for spotting rumors in Chinese social posts, in a 2025 paper. False claims spread fast online, so automatic detection matters. The model is built on the MindSpore learning framework. It reads each post at several scales at once, from local phrases to the whole text.

Four parts divide the work. One extracts local meaning, one tracks word order in both directions, one focuses attention across the full post, and one maps how words co-occur as a graph. Tests on a Weibo rumor data set gave near-perfect accuracy scores. The authors stress that the model also generalizes well and can show its reasoning. Platforms gain a candidate guard against viral falsehoods.

Catalog record

CREATORZhenkai Qin 1,2,3 1 School of Information Technology 2 Network Security Research Center 3 Big Data and Policing Technology Laboratory Guangxi Police College, Nanning, China [email protected] Guifang Yang School of Information Technology Guangxi Police College 530028, China yangguifang@gxjcxy.
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE42 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-17574

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

RAGAT-Mind: A Multi-Granular Modeling Approach for Rumor Detection Based on MindSpore Zhenkai Qin 1,2,3 1 School of Information Technology 2 Network Security Research Center 3 Big Data and Policing Technology Laboratory Guangxi Police College, Nanning, China [email protected] Guifang Yang School of Information Technology Guangxi Police College 530028, China [email protected] Dongze Wu School of Information Technology Guangxi Police College 530028, China [email protected] Abstract As false information continues to proliferate across social media platforms, effective rumor detection has emerged as a pressing challenge in natural language processing. This paper proposes RAGAT-Mind , a multi-granular modeling approach for Chinese rumor detection, built upon the MindSpore deep learning framework. The model integrates TextCNN for local semantic extraction, bidirectional GRU for sequential context learning, Multi-Head Self-Attention for global dependency focusing, and Bidirectional Graph Convolutional Networks (BiGCN) for structural representation of word co-occurrence graphs. Experiments on the Weibo1-Rumor dataset demonstrate that RAGAT-Mind achieves superior classification performance, attaining 99.2% accuracy and a macro-F1 score of 0.9919. The results validate the effectiveness of combining hierarchical linguistic features with graph-based semantic structures. Furthermore, the model exhibits strong generalization and interpretability, highlighting its practical value for real-world rumor detection applications. K eywords Rumor Detection; Multi-Granular Modeling; Graph Attention; TextCNN; BiGCN; MindSpore; Chinese Social Media 1 Introduction With the rapid development of social media, platforms such as Weibo, Twitter, and Facebook have become vital channels for information dissemination. These platforms offer users convenient content publishing tools and powerful diffusion mechanisms, enabling information to spread rapidly within a short period of time [ 1 ] . However, this highly efficient dissemination also provides fertile ground for the spread of misinformation, rumors, and erroneous public opinion. In scenarios such as emergencies, social hotspots, and public crises, the wide circulation of false content can easily mislead public perception and may even pose serious threats to social stability, policy implementation, and national security [ 2 , 3 ] . In response to this issue, rumor detection has garnered increasing attention from both academia and industry. Traditional rumor detection methods are primarily based on handcrafted feature engineering, which involves extracting textual content, user behavior, and propagation graph structure, and then applying classifiers such as Support Vector Machines (SVM) and Random Forest to identify rumors [ 4 ] . Although these methods achieved promising results in the early stages, they heavily rely on domain expertise, exhibit poor generalizability, and struggle with the informal and diverse linguistic expressions prevalent in social media. Consequently, their robustness and adaptability are limited [ 5 ] . With the widespread adoption of deep learning in natural language processing, an increasing number of studies have explored end-to-end rumor detection using neural networks. Convolutional Neural Networks (CNNs) excel at capturing local features and are well-suited for modeling n-gram patterns in text; Recurrent Neural Networks (RNNs), along with their variants such as Gated Recurrent Units (GRUs), are capable of capturing contextual dependencies, thereby enhancing the model’s understanding of textual sequences [ 6 , 7 ] . In recent years, the introduction of the Transformer architecture and attention mechanisms—especially Multi-Head Self-Attention—has significantly improved performance in long-text and complex semantic modeling tasks, and has been widely applied in sentiment analysis, event detection, and related areas [ 8 ] . However, social media texts exhibit not only sequential characteristics but also explicit structural relationships, such as word co-occurrence and semantic association. Solely relying on sequence modeling often fails to capture the nonlinear structural dependencies between words. To address this limitation, Graph Neural Networks (GNNs) have been introduced into textual modeling tasks in recent years, enabling structure-aware representation learning. In particular, Bidirectional Graph Convolutional Networks (BiGCNs) leverage both forward and backward propagation to extract deep structural relationships from bidirectional semantic graphs. They have shown great potential in tasks such as text classification, sentiment analysis, and information extraction [ 9 ] . MindSpore is an all-scenario AI framework independently developed by Huawei. It supports hybrid execution of dynamic and static computational graphs, operator fusion optimization, and automatic parallelism, thereby enabling efficient training and deployment of deep neural networks. In Chinese natural language processing tasks, MindSpore provides flexible graph computation capabilities, operator support for heterogeneous structural modules, and efficient multi-module scheduling mechanisms. These features make it particularly suitable for complex modeling tasks that integrate convolutional, recurrent, attention-based, and graph-based neural architectures. Building upon these foundations, this paper proposes a multi-granular modeling framework for Chinese rumor detection, named RAGAT-Mind (Rumor-aware Graph Attention and TextCNN model based on MindSpore), which achieves deep fusion between semantic and structural representation learning. The model integrates four key architectural modules: TextCNN for local semantic feature extraction, GRU for modeling temporal dependencies, Multi-Head Attention (MHA) for focusing on salient information, and a Bidirectional Graph Convolutional Network (BiGCN) for structural representation via word co-occurrence graphs. Through a parallel dual-path design, the model demonstrates strong expressiveness and robust discriminative capacity in handling Chinese social media rumors characterized by high linguistic ambiguity, fragmented structure, and frequent semantic discontinuities. 2 Related Work With the rapid proliferation of social media platforms, the dissemination of information has become more efficient, but the spread of misinformation and online rumors has also intensified. As a result, automated rumor detection has emerged as a critical task in the field of natural language processing. Existing studies in this area primarily fall into three methodological categories: traditional machine learning approaches, deep semantic modeling methods, and graph-based neural modeling techniques. Early research focused on handcrafted feature extraction combined with conventional classifiers. For instance, Rizzo et al. [ 10 ] extracted statistical features from Weibo posts and integrated temporal propagation cues for early rumor detection. Shi et al. [ 11 ] applied transfer learning techniques to improve cross-domain adaptability among various Chinese datasets. Zhang et al. [ 12 ] combined user behavioral graphs with textual features to estimate the credibility of information sources. Although these approaches provided valuable insights, they exhibited limited semantic expressiveness, inadequate structural modeling capabilities, and poor robustness to informal and ambiguous linguistic expressions prevalent in social media contexts. With the advancement of deep learning, the field has shifted toward end-to-end representation learning. Convolutional Neural Networks (CNNs) have been widely adopted for capturing local n-gram semantics [ 13 ] , and Gated Recurrent Units (GRUs) have shown effectiveness in modeling temporal dependencies [ 14 ] . Multi-Head Self-Attention (MHA), in particular, has improved the ability to focus on semantically salient regions [ 15 ] . Gao et al. [ 16 ] introduced a sentiment-aware BERT framework to enhance the classification of fake news. However, most of these methods ignore the structural information embedded in text, which is essential for understanding the underlying topology of word associations. To compensate for this deficiency, Graph Neural Networks (GNNs) have been introduced into rumor detection tasks to capture word co-occurrence and topological semantics. Zichao et al. [ 17 ] proposed TextGCN to learn structure-aware representations by constructing heterogeneous word-document graphs. Yao et al. [ 7 ] improved this framework by incorporating edge weighting and position encoding. Bian et al. [ 9 ] developed a Bidirectional Graph Convolutional Network (BiGCN) to model bidirectional semantic propagation. In the context of Chinese rumor detection, Vu et al. [ 18 ] proposed a multimodal fusion model combining textual features with user propagation paths via GCNs. Despite

Read the complete text