Abstract: Distant supervision has a good effect in relation extraction tasks. Meanwhile, most methods use multi-instance learning to reduce the impact of training data been wrong labelled in distant supervision. However, the effect of multi-instance learning depends on the sentence feature vector extracted by the neural network. At present, most methods for extracting sentence features only pay attention to the structural features of sentences, while ignoring semantic features. As a result, structural feature sentences and semantic feature sentences cannot occupy the same proportion in multi-instance learning, which further influences the precision of the model. To alleviate this issue, we propose a BiLSTM-CNN-Attention model (BLCANN) based on semantic dependency graph to extract sentence features. In this model, we extract the shortest dependency path between the two entities from the semantic dependency graph as the input to the model. The shortest path combines the structural and semantic features of the sentence, which contributes to distinguishing between positive and negative examples in multi-instance learning. Experimental results show that our model is adept in extracting structural features and semantic features. Our model has increased the precision of the relationship extraction on Top100 by 10 percent compared to the baseline [9].
