Use adjacent turns to classify replies without inventing causes
When a conversation-level classifier leaves a large residual, decompose it into reply states by using only the immediately preceding human-support turn within a bounded window. Keep explicit problem statements at higher priority, and report contextual states—such as a response to…