Sexism detection remains as an extremely low-resource task for most of the languages including Chinese. To address this issue, we propose a zero-shot cross-lingual method to detect sexist speech in Chinese and perform qualitative and quantitative analyses on the data we employed. The proposed method aims to explicitly model the knowledge transfer process from rich-resource language to low-resource language using metric-based meta-learning. To overcome the semantic disparity between various languages caused by language-specific biases, a common label space of emotions expressed across languages is used to integrate universal emotion features into the meta-learning framework. Experiment results show that the proposed method improves over the state-ofthe-art zero-shot cross-lingual classification methods.