WuRain-Senti: a dual-dimensional (relevance x emotion) Weibo corpus of Wuhan rainstorm discourse, 2010-2025

Abstract

Annotation labels, frozen data splits, per-item model predictions, evaluation metrics and recomputation code for the WuRain-Senti benchmark: 10,942 Weibo posts about Wuhan rainstorm events between 2010 and 2025, each labelled on two decoupled dimensions, topical relevance (R) and emotion (E, five classes). Every accuracy, macro-F1, confusion cell, McNemar p-value, bootstrap interval and Cohen's kappa reported in the accompanying paper can be recomputed from the files in this record alone, using code/recompute_metrics.py, in a few seconds and without a GPU, model weights or the original post texts. Contents. Labels for the 10,942-post emotion corpus with the frozen 80:10:10 split; per-item predictions of all benchmark configurations on the 1,095-post test set, including three random seeds; class probabilities of the proposed configuration; the relevance-gate cascade output; affect-eligibility gate predictions with both operating points on its 4,230-post test partition; the two annotators' independent labels on a stratified 400-post re-annotation; year-by-category counts for the full 42,310-post candidate pool; 18 frozen metric files; and the annotation codebook in English. Not included. Original post texts, user identifiers, usernames, profile fields, URLs, exact timestamps, model checkpoints and the collection software. Posts are represented only by a salted one-way hash; the salt is not distributed. Temporal information is truncated to year and month. The de-identified texts are available from the corresponding author on reasonable request, for non-commercial research use and subject to the platform's terms of service. Data are released under CC BY 4.0; the recomputation code under the MIT licence. Companion records for the other studies in this project reuse the same salted post_id, so they can be joined post by post with this record.