Go to ACL ARR 2025 July homepageASIDE: Adaptive Self-Learning for Misinformation Detection via Dual Supervision
Abstract: Detecting misinformation on high-volume social media platforms such as X and Facebook is challenging, exacerbated by the cost and inconsistency of human annotation. To address this, we propose a novel adaptive self-learning framework, ASIDE, that leverages active learning with limited labels and a dual teacher-student approach. Our framework uniquely employs a teacher text model for content and a hyper-teacher to capture conversational structure. During active learning, ASIDE dynamically transitions its data acquisition strategy based on its confidence, prioritizing uncertain but influential samples using a hybrid method and an uncertainty threshold. Simultaneously, it employs a dynamic sampling strategy to adapt to performance changes. Extensive experiments demonstrate that ASIDE significantly outperforms state-of-the-art methods, including active learning, graph-based, and unsupervised approaches, in misinformation detection.
Paper Type: Long
Research Area: Efficient/Low-Resource Methods for NLP
Research Area Keywords: Active Learning, Misinformation Detection
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency
Languages Studied: Arabic
Reassignment Request Area Chair: This is not a resubmission
Reassignment Request Reviewers: This is not a resubmission
A1 Limitations Section: This paper has a limitations section.
A2 Potential Risks: N/A
A2 Elaboration: Our work does not involve personal data, or sensitive populations. Therefore, risk discussion was not applicable in this context.
B Use Or Create Scientific Artifacts: Yes
B1 Cite Creators Of Artifacts: Yes
B1 Elaboration: We cited the original dataset creators' model backbones (GPT-2, BERT) in the main text and reference section.
B2 Discuss The License For Artifacts: Yes
B2 Elaboration: All used datasets are publicly available for research purposes, and model were used under their respective open-source licenses (e.g., HuggingFace ).
B3 Artifact Use Consistent With Intended Use: Yes
B3 Elaboration: All datasets and models are publicly available were used strictly for academic research and analysis, consistent with their intended use. No distribution or modification beyond experimental usage was performed.
B4 Data Contains Personally Identifying Info Or Offensive Content: No
B4 Elaboration: We used publicly available datasets that have been anonymized or filtered by their creators. Our preprocessing pipeline does not use any usernames, mentions, or user metadata.
B5 Documentation Of Artifacts: No
B5 Elaboration: We used existing benchmark datasets (Twitter15, Twitter16, PHEME) which are already well-documented in prior work. We did not create new datasets or artifacts requiring further documentation.
B6 Statistics For Data: Yes
B6 Elaboration: We reported statistics on rumor vs non-rumor class distribution and dataset splits for each benchmark (Twitter15, Twitter16, and PHEME) in Section 5 and Table 1.
C Computational Experiments: Yes
C1 Model Size And Budget: No
C1 Elaboration: We used publicly available pretrained models (GPT-2, BERT) without modification. We did not benchmark or analyze model size or compute cost explicitly, as our focus was on methodological contributions.
C2 Experimental Setup And Hyperparameters: Yes
C2 Elaboration: We report full training and hyperparameter settings in Section 9.1 (Appendix A.1), including learning rate, dropout, number of layers, sampling thresholds, and acquisition parameters.
C3 Descriptive Statistics: Yes
C3 Elaboration: We report mean performance (accuracy, precision, recall, F1) for each model and dataset in Tables 2–7. In ablation and warm-start analysis, we include switch counts and confidence-related variation.
C4 Parameters For Packages: Yes
C4 Elaboration: We used HuggingFace Transformers and PyTorch for model implementations. Configurations such as dropout, learning rate, and optimizer are described in Appendix A.1.
D Human Subjects Including Annotators: No
D1 Instructions Given To Participants: N/A
D2 Recruitment And Payment: N/A
D3 Data Consent: N/A
D3 Elaboration: All datasets used (Twitter15, Twitter16, PHEME) are publicly released and widely used in prior work. We did not collect any new human data.
D4 Ethics Review Board Approval: N/A
D4 Elaboration: No new human data was collected in this study. We used existing publicly available datasets that have been approved and widely used in previous research.
D5 Characteristics Of Annotators: N/A
E Ai Assistants In Research Or Writing: No
E1 Information About Use Of Ai Assistants: No
E1 Elaboration: We did not use AI assistants such as ChatGPT or Copilot during any stage of research, writing, coding, or analysis. All work was completed manually by the authors.
Author Submission Checklist: yes
Submission Number: 241
Loading