Disasters on Social Media from CrowdFlower (Figure Eight): 10,876 tweets that the publisher says were culled with searches for words such as 'ablaze', 'quarantine' and 'pandemonium', each judged by contributors on whether it refers to a disaster event (as opposed to a joke, a film review or something non-disastrous). Judgments: 6,187 Not Relevant, 4,673 Relevant, 16 Can't Decide. Each row has the crowd job's unit ID, the judgment and its confidence, the search keyword (221 distinct), the author's free-text location, the tweet text, and two number columns named tweetid and userid. The package also holds two files from the folder named kaggle-nlp-getting-started as their own tables: train.csv (7,613 rows with a 0/1 target) and test.csv (3,263 rows without one).
First GOP debate sentiment analysis: 13,871 tweets about the Republican presidential primary debate in Ohio in August 2015, as labeled by Figure Eight (CrowdFlower) contributors, who judged whether each tweet was relevant, which candidate it mentioned, which subject it concerned and its sentiment. The publisher says the non-relevant messages were removed, and all 13,871 rows are marked relevant. Sentiment: 8,493 Negative, 3,142 Neutral, 2,236 Positive. Candidate: 11 values plus 96 empty. Each row has the four labels with their confidence values, the author's screen name, the tweet text, retweet count, time, free-text location and time zone; 21 tweets carry coordinates. The package also holds the copy of the data from the folder named kaggle-v2 (Sentiment.csv, 13,871 rows) as its own table.
2015 New Year's resolutions from CrowdFlower: 5,011 tweets, a Twitter sentiment analysis of users' 2015 New Year's resolutions that, per the publisher, contains demographic and geographical data of the users and resolution categorizations. Each tweet has a resolution topic (115 distinct values) and one of 10 resolution categories (1,781 Personal Growth, 1,010 Humor, 840 Health & Fitness, 467 Recreation & Leisure, 351 Family/Friends/Relationships, 176 Finance, 126 Career, 89 Education/Training, 87 Time Management/Organization, 84 Philanthropic). Each row also has the tweet text, the author's screen name and a gender value (2,518 male, 2,493 female), the retweet count, time, free-text location, state, region and time zone; 227 tweets carry coordinates.
Twitter US Airline Sentiment, a CrowdFlower job on the problems of each major U.S. airline: 14,640 tweets about 6 airlines (United 3,822, US Airways 2,913, American 2,759, Southwest 2,420, Delta 2,222, Virgin America 504); the publisher says the tweets were scraped in February 2015. Contributors first classified each tweet as positive, negative or neutral (9,178 negative, 3,099 neutral, 2,363 positive) and then gave a reason for the negative ones (9,178 tweets have a reason, in 10 categories). Each row has the crowd job's unit ID, the sentiment and the reason with the confidence of each, the airline, the author's screen name, the tweet text, its retweet count and time, and the author's free-text location and time zone; 1,019 tweets carry coordinates. The package also holds the copy of the tweets from the folder named kaggle-v2 (Tweets.csv, 14,640 rows) as its own table.
20,050 Twitter profiles from October 2015, each with one random tweet, released by CrowdFlower (Data for Everyone): crowd workers looked at each profile and judged whether the account belongs to a woman (6,700), a man (6,194) or a brand (5,942), with a confidence score; 1,117 could not be told. Each row has the tweet, the profile description, profile colours and counts. The tweet IDs in the original file were saved in scientific notation, so the tweets cannot be fetched again. Useful for author profiling, gender inference and telling people from brands. Licensed CC BY 4.0 by Figure Eight; it contains screen names and profile descriptions.
Weather sentiment from CrowdFlower (Figure Eight): 1,000 weather-related tweets, each graded by 20 contributors for the emotion the author expresses specifically about the weather (271 Negative, 261 Neutral / author is just sharing information, 235 Tweet not related to weather condition, 231 Positive, 2 I can't tell), plus the follow-up job in which 10 contributors judged whether the crowd's grade was correct, for the same 1,000 tweets (answers: 992 Yes, 5 No, 3 I can't tell). Each tweet has an ID, its text, the emotion and the confidence upstream computed for it; the evaluation file's judgments are in their own table.
Twitter (X)