Abstract: Extracting entity relationship information from text big data quickly and accurately is very important to build knowledge maps. The existing main methods for remote supervised relationship extraction often ignore the type information and syntactic information of entity pairs. In this work, a bi-directional long short-term memory (BiLSTM) model combined with an attention mechanism layer of words around entities is utilized as the first module of sentence encoding. Then, an entity type embedding module is added to the model to enrich sentence encoding information. Finally, a semantic dependency parsing module is also included to the model. Thus, the three modules form a relation extractor. In addition, most of distant supervised relationship extraction models are designed to reduce noises in packets and sentences, they ignore the impacts of noise labels on model performances. Focused on noise reduction of labels, this work designs a label learner, which can learn soft labels of sentences on the basis of reinforcement learning so as to modify noisy labels. A novel relationship extraction framework for text entities based on deep reinforcement learning is built from our designed relationship extractor and label learner. The experiment results for a self-constructed dataset and two public datasets, ACE2005 and Chinese-Literature-NER-RE-Dataset show that our proposed method outperforms several state-of-the-art models in precision and recall rate.