A Fakeddit (multimodal fake news benchmark) dataset of 965,647 Reddit submissions with text, images and metadata, collected by Nakamura et al. via the Reddit API (2020). Each record holds the post title, text, image URL, comment count, score, timestamp and 2-, 3- and 6-way veracity labels (6-way: true, satire/parody, misleading content, imposter content, false connection, manipulated content). Suitable for fake news detection, multimodal learning and misinformation classification.
A LIAR (fact-checking benchmark) dataset of 12,789 labeled claim statements from Politifact, collected by Wang via Politifact fact-checking (2017). Each record contains claim text, 6-class veracity label, speaker metadata (party, state, job title), credit history counts (barely_true/false/half_true/mostly_true/pants_fire), and context. The dataset is notable for its benchmark dataset for fake news detection with fine-grained 6-class labels, rich speaker metadata, and credit history. Intended for research into fake news detection, fact-checking, political discourse analysis, and misinformation research.
A Meme Descriptions (Cheezburger dataset) dataset of 252,284 meme descriptions and 117 images, collected by Wang and Wen via Cheezburger (2015). Each record contains meme image and text descriptions, popularity scores. The dataset is notable for its multimodal dataset for meme popularity prediction and generation. Intended for research into multimodal learning, meme generation, popularity prediction, and computational humor.
A #petpeeve (annoying behaviors) dataset of 3,375 tweets labeled with annoying behavior categories, collected by Wang and Yang via Twitter API (2015). Each record contains tweet text and annoying behavior category label. The dataset is notable for its dataset for automatic categorization of annoying behaviors using lexical and frame-semantic embeddings. Intended for research into social media analysis, text classification, and behavioral categorization.
Yelp coffee-shop reviews for computational branding analytics (Wang, Lin & Kominek, EMNLP 2013): 17,894 reviews in train (10,495), dev (3,617) and test (3,782) splits. Each row holds the brand (Starbucks, Dunkin', other), satisfaction (negative, positive, neutral), the reviewer's gender (missing as '?' in some rows), US region, and the review text. There are no review, business or date fields. Research use only, per the README.
Reddit
Twitter (X)
Yelp