Distant Supervision Relation Extraction Combining Attention Mechanism and Ontology

Relational extraction extracts relationships from unstructured text and outputs them in a structured form. In order to improve the extraction accuracy and reduce the dependence on manual annotation, this paper proposes a distant supervision relationship extraction model based on attention mechanism and ontology, attention piecewise convolutional neural networks with ontology restriction (APCNNs+OR). The model is divided into feature engineering extraction module, classifier module and ontology restriction layer. In the classifier module, this paper introduces and improves the instance-level attention mechanism to learn the weight of each sentence in the data bag better, effectively reducing the noise interference introduced by the distant supervision hypothesis and the word information interference between the two entities in the sentence. In the ontology restriction layer, the extraction result is constrained by introducing the domain ontology, which improves the accuracy of the relationship extraction. The experiment results of SemMed and GoldStandard corpus show that the model can effectively reduce the noise interference of the wrong label and has better relation extraction performance than the existing models