The pipeline model predicts object and relation with the subject, which leads to an inevitable accumulation of errors. In this paper, a multi-level labeling scheme is designed in order to remove the accumulation of errors, which is an end-to-end entity relation extraction model based on the joint annotation framework. Firstly, the word tokens are converted into vectors by a pre-trained model. Then, the four fully connected models are concatenated to obtain the feature vectors. Finally, these features are fed into a sigmoid classifier to predict the span of the subject and object corresponding to each relation. The experimental results show that our model has better results, and its F1 value is increased by 0.87, 2.11 and 0.58% on ADE, CoNLL04 and WebNLG datasets, respectively. Our study achieves state-of-the-art performance on ADE, CoNLL04 and WebNLG datasets.