The present invention relates to the technical field of artificial intelligence, and more particularly, to an entity relationship joint extraction method based on reinforcement learning. Firstly, theunstructured text, segmented words, and training word vectors for entity relation extraction can be obtained, and are input in LSTM by taking words as a unit, because the same entity in a sentence mayappear in different forms in different locations, and we do not know where the entities really useful for relationship extraction, so we can use reinforcement learning method to select these entities; after the entity selection is completed, if there is a consecutive one, we need to merge it into one entity. Finally, after removing the redundancy, if two entities are picked out, the word vectorsof these two entities and the sentence vectors of the final output of LSTM are stitched together, and the relations are classified by a fully connected neural network, otherwise the sentence is considered to be noisy.