A new AI framework called EMPATH tracks how emotions shift, turn by turn, inside crisis text-counseling conversations.
Researchers built EMPATH to analyze mental health chats at three levels - individual turn labels, the probability of emotions transitioning from one state to another, and broader conversation-level patterns. They tested it on text-based crisis conversations with self-identified Black texters discussing grief. The analysis found persistent negative affect throughout the chats, a gradual drift toward hope rather than sudden shifts, clearly different emotional roles for texters versus volunteers, and recovery paths that varied widely from one conversation to the next.
Most computational work on crisis conversations still scores each message with a single static emotion label, missing how a conversation actually moves over time. EMPATH's three-tier approach - turn, transition, and archetype - offers a more realistic map of how support unfolds, which could help crisis lines spot when a conversation is stalling or drifting the wrong way, and give researchers a sharper tool for comparing counseling styles.
It is a research paper, not a deployed tool, and grief is one narrow slice of what crisis lines handle. Whether any hotline has the staff or systems to act on turn-by-turn emotion data is a separate question from whether the data is interesting.