Abstract: In order to solve the problem of lag in the actual application of the existing microblog rumor detection algorithms and the insufficient feature extraction ability of traditional deep learning model, this paper proposes an early detection model of microblog rumors based on RoBERTa-BiSRU++-AT. The original text is used as the input, which does not contain forwarding and comment or other relevant information. The RoBERTa pre-trained model assigns dynamic and contextual semantics to words, which sloves the polysemy and improves the semantic representation ability of words. Furthermore, the bidirectional simple recurrent unit with built-in self-attention extracts the deep semantic features, so self-attention can learn the dependencies between words, which can provide comprehensive high-dimensional features. The soft attention mechanism is introduced to calculate the weight of words on classification result, which can solve the problem that different features have the same influence on classification result, and the model can focus on key features. The classification probability is calculated by Softmax layer, and the label corresponding to the maximum probability is taken as the result. Comparative experiment was carried out on the Chinese microblog rumors dataset. The experimental results prove that the F1 index of RoBERTa-BiSRU++-AT is 98.16%, higher than other microblog rumor detection algorithms, which proves that the proposed model has better early recognition ability for microblog text rumor.