In this work, we study on semi-supervised semantic parsing under a multi-task learning framework to alleviate limited performance caused by limited annotated data. Two novel strategies are proposed to leverage unlabeled natural language utterances. The first one takes entity predicate sequences as training targets to enhance representation learning. The second one extends Mean Teacher to seq2seq model and generates more target-side data to improve the generalizability of decoder network. Different from original Mean Teacher, our strategy produces hard targets for the student decoder and update the decoder weights instead of the whole model. Experiments demonstrate that our proposed methods significantly outperform the supervised baseline and achieve more impressive improvement than previous methods.