The Tone of Awareness: Topic, Sentiment, and Toxicity Maps During Mental Health Month on TikTok Dataset

Indexed Metrics Dataset for TikTok Mental Health Month Analysis This dataset accompanies the manuscript: The Tone of Awareness: Topic, Sentiment, and Toxicity Maps During Mental Health Month on TikTok Authors: Henrique Ferraz de Arruda, Andreia Sofia Teixeira, Pranay Gundala Reddy, Anindya Mondal, Kleber Andrade Oliveira, and Filipi Nascimento Silva. Affiliations include the Institute for Biocomputation and Physics of Complex Systems (BIFI), University of Zaragoza; ARAID Foundation; BRAN Lab, Network Science Institute, Northeastern University London; Kent Medway Medical School; LASIGE, Faculdade de Ciencias da Universidade de Lisboa; Observatory on Social Media, Indiana University; Social Dynamics Research Lab, University of Limerick; and CSSI, Kellogg School of Management, Northwestern University. Dataset Summary The release contains two CSV files with derived topic, sentiment, and toxicity metrics for TikTok videos and comments collected for a Mental Health Awareness Month study. The files are designed to preserve the analytical relationship between videos and comments without redistributing TikTok identifiers, usernames, text, captions, hashtags, URLs, or other direct content fields. The key relational field is video_index: videos.csv contains one row per exported video and assigns each video a release-local integer video_index. comments.csv contains one row per exported comment and uses video_index to point to the corresponding row in videos.csv. Comment rows do not include a direct topic column. To analyze comments by topic, join comments.csv to videos.csv on video_index and use the video's topic. The integer indices are generated only for this release and are not TikTok IDs. Files File Rows Columns Description videos.csv 11,736 15 Video-level topic, UTC date, Detoxify scores, and XLM-T sentiment scores. comments.csv 40,896 15 Comment-level UTC date, Detoxify scores, XLM-T sentiment scores, and video_index link to videos.csv. The exported video dates range from 2023-04-15 to 2024-06-14. The exported comment dates range from 2023-04-15 to 2025-02-10, because comments attached to videos in the study window may have been created later. The combined export contains 277 distinct dates after privacy filtering. Topic Labels The video topic labels in videos.csv are: Anxiety Bipolar Borderline Depression Duet Grief Mental Health Mental Health Month Positivity Psychisch Self Care Suicide Prevention Sunset Venting 1 Venting 2 Topic labels come from the paper-locked topic model variant used for the manuscript. Columns videos.csv Column Type Description topic string Human-readable topic label assigned to the video in the paper-locked topic model. video_index integer Release-local video identifier. Use this to join comments to videos. date string UTC calendar date of video creation, formatted as YYYY-MM-DD. detoxify_toxicity float Detoxify multilingual model score for toxicity. detoxify_severe_toxicity float Detoxify multilingual model score for severe toxicity. detoxify_obscene float Detoxify multilingual model score for obscenity. detoxify_identity_attack float Detoxify multilingual model score for identity attack. detoxify_insult float Detoxify multilingual model score for insult. detoxify_threat float Detoxify multilingual model score for threat. detoxify_sexual_explicit float Detoxify multilingual model score for sexually explicit content. xlm_sentiment string Highest-probability XLM-T sentiment label: xlm_negative, xlm_neutral, or xlm_positive. xlm_polarity float xlm_positive - xlm_negative. Higher values indicate more positive sentiment under the model. xlm_negative float XLM-T probability-like score for negative sentiment. xlm_neutral float XLM-T probability-like score for neutral sentiment. xlm_positive float XLM-T probability-like score for positive sentiment. comments.csv Column Type Description video_index integer Release-local link to videos.csv. comment_index integer Release-local comment identifier. date string UTC calendar date of comment creation, formatted as YYYY-MM-DD. detoxify_toxicity float Detoxify multilingual model score for toxicity. detoxify_severe_toxicity float Detoxify multilingual model score for severe toxicity. detoxify_obscene float Detoxify multilingual model score for obscenity. detoxify_identity_attack float Detoxify multilingual model score for identity attack. detoxify_insult float Detoxify multilingual model score for insult. detoxify_threat float Detoxify multilingual model score for threat. detoxify_sexual_explicit float Detoxify multilingual model score for sexually explicit content. xlm_sentiment string Highest-probability XLM-T sentiment label: xlm_negative, xlm_neutral, or xlm_positive. xlm_polarity float xlm_positive - xlm_negative. Higher values indicate more positive sentiment under the model. xlm_negative float XLM-T probability-like score for negative sentiment. xlm_neutral float XLM-T probability-like score for neutral sentiment. xlm_positive float XLM-T probability-like score for positive sentiment. Methods Videos were collected with the TikTok Research API for mental-health-related keywords and hashtags during the Mental Health Awareness Month study window, April 15 through June 15, for 2023 and 2024. Comments were collected for the eligible videos. The released files were generated from the paper-locked variant: Variant: paper_locked_multilingual_detoxify_mpnet Toxicity model: Detoxify multilingual Sentiment model: cardiffnlp/twitter-xlm-roberta-base-sentiment Video embedding model for topic modeling: sentence-transformers/all-mpnet-base-v2 Topic modeling: BERTopic with UMAP and HDBSCAN, using the manuscript settings Topic mode: paper-locked labels used in the manuscript The CSV export step keeps only: the video-comment relationship through video_index; the video topic label; UTC dates; Detoxify toxicity scores; XLM-T sentiment labels and scores. It removes TikTok video IDs, TikTok comment IDs, usernames, raw text, processed text, captions, hashtags, URLs, audio transcripts, and other direct content fields. Privacy Filtering To reduce identification risk from sparse dates, the export excludes calendar dates with fewer than 10 total rows across the combined video and comment export. The filtering is iterative: Count total video and comment rows per UTC date. Remove rows on dates with fewer than 10 total rows. Remove comments whose parent video is no longer present. Recompute date counts and repeat until all remaining dates meet the threshold. After filtering, the minimum combined number of exported rows on any remaining date is 10. This release is still derived from public social media behavior and mental-health-related content. It should be used only for aggregate analysis. Do not attempt to reidentify users, videos, comments, or accounts. Example Join import pandas as pd videos = pd.read_csv("videos.csv") comments = pd.read_csv("comments.csv") comments_with_topic = comments.merge( videos[["video_index", "topic"]], on="video_index", how="left", validate="many_to_one", ) Reproducibility The export can be regenerated from the local, non-redistributable paper-locked variant files with: python Scripts/08_export_zenodo_indexed_metrics_dataset.py \ --variant-dir old/Variants/paper_locked_multilingual_detoxify_mpnet \ --output-dir zenodo/indexed_metrics_dataset \ --min-entries-per-date 10 This command does not recollect TikTok data and does not rerun the scoring or topic-modeling stages. It only exports the shareable indexed metrics CSVs from existing local feather files. Integrity SHA-256 checksums for this export: 71971193022496de60fabd4d1758538f5d7d5c9bf51e52f258bc1675b944f5e9 comments.csv e938ec74706b29a1a3a120fa3fa3dbc0c96b8c1547d9aeaecbb2a312db8f9f04 videos.csv Citation When using this dataset, cite both the Zenodo record and the accompanying manuscript: Henrique Ferraz de Arruda, Andreia Sofia Teixeira, Pranay Gundala Reddy, Anindya Mondal, Kleber Andrade Oliveira, and Filipi Nascimento Silva. The Tone of Awareness: Topic, Sentiment, and Toxicity Maps During Mental Health Month on TikTok. Add the final Zenodo DOI and manuscript DOI once available. License Set the license on the Zenodo record according to the authors' data-sharing decision and the applicable TikTok Research API, institutional, and publication requirements. This README does not grant rights beyond the license selected for the deposited Zenodo record.