AI detects user deaths from tweets with 90% accuracy
New BERT-based model spots post-mortem tweets with 90% accuracy...
Researchers developed machine learning classifiers that automatically detect when social networking site users have died, based on posts associated with their profiles. Using a new dataset built from Wikidata and Twitter, they trained traditional and deep learning models—including BERT, which outperformed all others—to distinguish post-mortem content from pre-mortem content. The study found clear linguistic differences, including higher negative sentiment, more sadness, anger, surprise, and fear, and more frequent use of words, personal pronouns, verbs, family words, religious words, death words, and swear words in post-mortem tweets, while pre-mortem tweets showed more neutral sentiment, happy expressions, impersonal pronouns, and informal words.
- BERT model achieves 90% accuracy in detecting user deaths from tweets using Wikidata and Twitter data
- Post-mortem tweets show higher negative sentiment, death-related words, and analytical thinking compared to pre-mortem tweets
- The system outperforms traditional ML models (Random Forest, SVM) and deep learning models (BiLSTM, CNN)
Why It Matters
This could automate death notifications for social platforms, enabling faster memorialization and reducing manual moderation workload.