Aiming at the problems of inaccurate identification and insufficient response ability of disaster help information, this paper proposes a text multi label classification model based on ALBERT and bidirectional GRU. Firstly, the text information is encoded by using the ALBERT preprocessing language model, and the dynamic word feature vector of the text is obtained. Then, it is sent to the bidirectional GRU neural network for training. According to different words, different weights are given, and the attention mechanism is used for decoding. The simulated annealing algorithm is used to solve the optimal threshold, and the micro average value is used as the evaluation function to determine the label category of samples. Compared with Logistic regression, Naive Bayes and Long-term and short-term memory neural network (LSTM), the results show that the multi label classification model has a higher accuracy rate of 95%, and the Hamming loss is only 0.05, which can better identify the disaster information and improve the rescue efficiency.