Aimed at the problem that the demands of the modulation recognition algorithm based on deep learning for the labeled samples are too heavy, a multi task training strategy based on meta learning is adopted. This strategy obtains a ability of cross task signal recognition through a large number of different task training networks to make the network quickly adapt to new signal categories with a small number of samples. In order to extract the features from signal samples more comprehensively, a hybrid parallel feature extraction network is designed, completing the recognition task by measuring the distance between sample feature vectors. And a joint loss function is introduced to take both inter class and intra class distance into account, making the samples realize comparison more efficiently after feature extraction. The experimental results show that the algorithm can achieve a highest recognition accuracy of 88.43% when there are only five labeled signal samples.