The lack of effective domain adaptation methods

The gap

Evidence profile

Stated in the future work and cells future research and cells research gap sections of the source papers, classified as general, spanning 4 journals.

Supporting evidence — 5 representative gaps

• Sentiment analysis of social media for enhancing disaster response strategies (2026) · Frontiers in Public Health · doi

  In this work, we aimed to address the limitations existing sentiment analysis systems in the context of disaster response, where social media plays a crucial role in real-time situational awareness. Recognizing the challenges posed by the nuanced and often ambiguous nature of disaster-related discourse, we proposed framework comprising SentEMBNet—a sentimenta novel enhanced multi-branch neural a polarity-aligned curriculum optimization strategy. SentEMBNet integrates lexicon-guided embeddings, graph-based syntactic modeling, and transformer abstractions to capture sentiment signals at both local and discourse levels. PACO further enhances training efficacy by aligning model learning with semantic clarity and polarity variance.

  generalstated in future workevidence 5/5
  Keywords: sentiment disaster discourse sentembnet polarity paco aimed address limitations existing systems context response social media
• Cross-disaster Domain Adaptation Using Co-training Variants (2026) · Proceedings of the International ISCRAM Conference · doi

  This study evaluated domain-adaptive and semi-supervised approaches for cross-disaster tweet classification under limited supervision. Across multiple source–target pairs and label regimes, DeCoTa-based variants consistently outperformed Qwen3-14B zero-shot, supervised learning, self-training, and UDA, particularly at moderate-to-high label counts. While all methods benefited from additional supervision, structured co-training mechanisms enabled more effective cross-domain transfer and revealed important performance–calibration tradeoffs. In particular, cross-view training provided the most balanced improvement, achieving strong predictive performance while maintaining comparatively stable calibration. Our findings have direct implications for the deployment of automated social media classification systems in crisis response. In practice, models trained on historical disasters should not be applied to new events without adaptation, as cross-disaster domain shift can significantly degrade performance. Instead, practitioners should prioritize rapid WiP Paper – Social Media & Crisis Communication: narratives, signals, and sentiments Proceedings of the 23rd ISCRAM Conference – The Hague, the Netherlands June 2026 Caroline Rizza, Apoorva Chauhan, Amy Matser, Joyce Kox, Willem Treurniet and Jeroen Wolbers eds. Khushboo et al. Cross-disaster Domain Adaptation annotation of a small, representative subset of target data early in an event, as even a limited number of labeled samples can substantially improve model effectiveness when used with structured semi-supervised approaches. Unlabeled data, while abundant, must be leveraged carefully using mechanisms that control pseudo-label noise and uncertainty, rather than naive self-training. Additionally, model confidence scores should be interpreted cautiously, as overconfidence under domain shift may lead to misleading prioritization of information. These observations suggest that adaptive, human-in-the-loop systems—where models are continuously updated with new data and used to assist rather than replace analysts—are best suited for reliable deployment in dynamic crisis environments. Future work will focus on leveraging the observed strengths of individual DeCoTa variants to design a unified, more adaptive algorithm. Our results reveal complementary behaviors across variants: weighted-confidence achieves strong peak performance, low-confidence improves calibration stability, cross-view provides balanced robustness, and LLM-initialized training benefits early-label regimes. A promising direction is to integrate these mechanisms within a dynamically controlled framework that adjusts confidence weighting, agreement constraints, and pseudo-label selection based on model readiness and uncertainty signals during training.

  generalstated in future workevidence 5/5
  Keywords: cross domain label training performance adaptive confidence supervised disaster variants mechanisms calibration crisis adaptation early
• Cascaded Attention-Based Multimodal Framework for Robust Disaster Tweet Classification (2026) · IDRiM Journal · doi

  The rapid dissemination of disaster-related information on social media, including text, images, and metadata, offers unprecedented opportunities for real-time crisis response. This paper presents a novel cascaded attention-based multimodal framework that integrates textual and visual modalities to classify tweets as informative or non-informative. By leveraging Bi- LSTM for text processing, VGG-16 for image analysis, and a combination of self- and cross- attention mechanisms, the model dynamically captures both intra- and inter-modal relationships, thereby overcoming the limitations of traditional fixed-weight fusion strategies. Evaluated on the CrisisMMD dataset, the proposed framework achieves an accuracy of 90.81%, precision of 84.29%, recall of 88.78%, and F1-score of 85.63%, surpassing unimodal baselines and state-of-the-art multimodal models MCAN and DSACA by significant margins. These results demonstrate its ability to accurately identify critical disaster-related information, enhancing situational awareness for humanitarian agencies. This framework has helped in improving disaster response and management. It will depict informative content related to reports of casualties, infrastructure damage, or urgent needs. Humanitarian agencies can prioritise their action based on the informative tweets. By enabling rapid identification of actionable information, the system can support timely resource allocation and targeted relief efforts, particularly in fast-evolving disaster scenarios. By doing so, it supports faster resource allocation and targeted relief in fast-evolving disaster scenarios. Future research will extend the framework to incorporate audio and video modalities, further enriching multimodal analysis. Additionally, exploring transfer learning and domain adaptation will address challenges in data-scarce disaster scenarios, ensuring robust performance for emerging crises. By advancing attention-based multimodal learning, this work paves the way for more effective disaster response and resource allocation, ultimately saving lives. 22 IDRiM (2026) 16 (1) DOI10.5595/001c.162527 DATA AVAILABILITY ISSN: 2185-8322 The data used to support the findings of this study are available from the corresponding author upon request. FUNDING This work was carried out without funding from any agency in the public, commercial, or not- for-profit sectors. DECLARATION OF GENERATIVE AI AND AI-ASSISTED TECHNOLOGIES During the preparation of this work, the author(s) used ChatGPT, developed by OpenAI, to enhance the readability and clarity of the text (e.g., improving grammar, sentence flow, and language polishing). After using this tool/service, the author(s) reviewed and edited the content generated by the tool as necessary and take(s) full responsibility for the content of the publication. REFERENCES Acerbo, F. S., & Rossi, C. (2017). Filtering informative tweets during emergencies: A machine learning approach. In Proceedings of the 1st Context Workshop on ICT Tools for Emergency Networks and Disaster Relief. Agarwal, M., Leekha, M., Sawhney, R., & Shah, R. R. (2020). Crisis-DIAS: Multimodal damage analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, 34, 346–353. https://doi.org/10.1609/aaai.v34i01.5369 Asif, A., Khatoon, S., Hasan, M. M., Alshamari, M. A., Abdou, S., Elsayed, K. M., & Rashwan, M. (2021). Automatic analysis of social media images to identify disaster type and infer 83. appropriate https://doi.org/10.1186/s40537-021-00464-4 Big Data, emergency response.

  generalstated in future workevidence 5/5
  Keywords: disaster multimodal informative response framework related information text attention based tweets content resource allocation relief
• DisasterVQA: A Visual Question Answering Benchmark Dataset for Disaster Scenes (2026) · Proceedings of the International AAAI Conference on Web and Social Media · doi

  Future research should focus on developing more robust and operationally meaningful vision-language models for disaster response. - Future research should explore the use of multimodal data, including text, images, and videos, to improve the accuracy of damage assessments. - Future research should investigate the use of transfer learning and few-shot learning to improve the performance of vision-language models in disaster response.

  generalstated in cells future researchevidence 5/5
  Keywords: future research focus developing robust operationally meaningful vision-language
• Cross-disaster Domain Adaptation Using Co-training Variants (2026) · Proceedings of the International ISCRAM Conference · doi

  The lack of effective domain adaptation methods for cross-disaster tweet classification. - The need for a temporally grounded benchmark for cross-disaster classification systems. - The gap in understanding the impact of pseudo-label utilization strategies on cross-event generalization and calibration.

  generalstated in cells research gapevidence 5/5
  Keywords: lack effective domain adaptation methods cross-disaster tweet classification

Questions about this gap