Multimodal sentiment exploration in disaster crowdfunding campaigns: Analyzing descriptions, images, and public comments

Crowdfunding increasingly complements institutional disaster relief, yet the emotional signals surrounding these campaigns are distributed across multiple channels: the written pitch, the featured image, and the surrounding public conversation. Building on advances in multimodal analysis from crisis informatics, this paper examines sentiment distributions and cross-modality consistency in disaster crowdfunding. We assemble a corpus of 319 campaigns from GoFundMe and 7,420 associated Facebook comments. Sentiment is classified for (i) campaign descriptions, (ii) campaign images, and (iii) user-generated comments (UGC). Descriptions and images skew negative—as expected in crisis appeals—at approximately 57% and 62% respectively, while UGC is comparatively less negative at around 56%, with higher shares of neutral and positive sentiment. Inter-modality agreement was found to be very low, with Cohen’s κ ranging from 0.068 to 0.159, indicating that official campaign messaging and community responses often diverge. We discuss the implications of this divergence for multimodal fusion strategies and campaign design, arguing that such variation can itself provide valuable signals for predictive modeling and practical decision-making.