Abstract: Ethical issues and potential risks associated with artificial intelligence have attracted increasing attention in recent years. As one of the most common forms of human communication, dialogue introduces important ethical concerns when combined with artificial intelligence systems. To address this issue, dialogue ethics detection has become an important research task. In this study, a Chinese multi-turn dialogue dataset, EthicDialog, is constructed from multiple sources, including film subtitles and existing dialogue corpora, and its statistical characteristics are systematically analyzed. Furthermore, a multi-task learning framework for dialogue ethics detection is proposed, which jointly performs ethical attribute shift detection and dialogue-level ethics classification. By introducing dedicated masking strategies into Transformer layers, the model effectively captures global contextual information, local semantic information, and speaker-related information, thereby enabling a more comprehensive understanding of semantic context. Experimental results demonstrate the effectiveness of the constructed Ethicdialog dataset and verify that the proposed model consistently outperforms baseline methods.