Reducing Misunderstanding in Human-AI Communication: A Speech Act Model for Better Language Interaction

Abstract

Human-AI talk often fails when a system reads sentence form but misses user action. A user may ask, hint, soften, refuse, or seek repair. Many systems answer the words on the screen, not the act behind them. This paper studies that gap and offers a speech act model for better language interaction. The study uses secondary analysis of public resources and published experiments. The paper reviews classic speech act theory and recent benchmarks of pragmatic ability. The public materials include DailyDialog, MultiWOZ, the Stanford Politeness Corpus, the Switchboard Dialogue Act Corpus, PUB, DialogBench, and INDIR-IT. The review shows four repeated failure points. Systems struggle with indirect requests, politeness, dialogue act balance, and repair after confusion. Based on these findings, the paper proposes a four-stage model: act detection, context reading, face-risk check, and repair-based response planning. The model treats misunderstanding as a mismatch between user intention and system reply. It uses clarification when force is unclear. It also adjusts tone to the social setting. It links classic pragmatics with open benchmark practice. The paper argues that better human-AI talk needs pragmatic design, not only fluent text. The model offers a clear path for future testing, system tuning, and linguistics-based evaluation of intelligent language systems.