Methodology
Authors and posts
The atlas maps 164,291 Twitter authors discussing AI between 30 October 2022 and 27 February 2023. The source cohort begins with 363,618 authors retweeted at least once. Its largest connected component contains 202,710 authors. The atlas includes authors with at least three usable posts after text cleaning and matching analysis outputs.
Network and positions
Retweets connect authors in a directed network. The atlas cohort has 782,974 connections. Leiden identifies communities in the retweet network. Positions follow the network structure. The 3D layout covers its largest connected component of 161,035 authors; all 164,291 appear in 2D. Community labels summarise prevalent content within each group.
Topics
Each author’s cleaned posts form one document, including original posts, replies, quotations and retweets. The default LDA model uses twelve topics and a TF–IDF word representation, chosen after comparing candidate models. Each author has a mixture of topic weights. The topic lens colours authors by their highest-weight topic. An eight-topic model is also available.
Sentiment and emotion
CardiffNLP Twitter models assign sentiment and emotion probabilities to post text. Scores are averaged across each author’s posts, including retweeted material. The lenses display these author-level averages.
Nearest and characteristic authors
Closest retweet connections rank directly connected authors by total retweets in both directions within the atlas cohort. Characteristic authors are closest to the selected group’s mean mixture. The identified version includes illustrative post excerpts and searches author IDs, community labels, topics and those excerpts. The anonymous version searches community labels and topics.
Associations
Pairwise associations compare community and leading-topic assignments using bias-corrected Cramér’s V. Detailed comparisons show Pearson residuals, calculated from all row and column totals.