AI-Based Tension Prediction and Human-Agent Interaction Analysis in U.S. Crisis Response Simulators

Abstract

Predicting when a conversation is about to boil over during a crisis is a tough task for any support system. Emotions tend to spike fast, sometimes in just a few back-and-forth exchanges. Because of that, spotting signs of distress early on is a big deal for anyone trying to help. This research looks at a hybrid machine learning setup designed to see these tension spikes coming. It uses a tool called the Composite Tension Index (CTI) to measure emotional weight by looking at how language, timing, and interaction patterns shift over time. To build this index, the framework pulls in various signals like the mood of the words used, how urgent someone sounds, and how much the emotional tone swings. It also tracks things like how long it takes for someone to reply or who is dominating the talk. After gathering this data, the study tests a few different machine learning models, mostly ensemble methods and regularized regressions, to see if they can guess if the next part of the chat will get more heated. The results show that this system actually picks up on these rising patterns pretty well. It turns out that looking at the conversation as a moving, changing process works way better than just analyzing a single message in a vacuum. The data also suggests that the way people interact matters. When responses are reflective and the dialogue stays balanced, tension usually stays low. On the other hand, long silences or bossy tones make things much more unpredictable. Essentially, these escalations work like a tipping point in any complex system. This framework acts as a digital early warning, catching those rising signals before things get out of hand, which helps human responders know exactly when to step in.