Speech Act Modeling Improves Conversation Derailment Forecasting

Using speech act information as an auxiliary learning signal improves the prediction of conversational hostility in low-data and cross-domain settings. This approach reduces lexical noise and improves the generalizability of moderation models.