A Dynamic Parameter Enhanced Network for distant supervised relation extraction

Yanjie Gou, Yinjie Lei, Lingqiao Liu, Pingping Zhang, Xi Peng

Abstract: Highlights • We utilize the class connection with entity types to boost Distant Supervised Relation Extraction. • We propose a dynamic parameter enhanced network to improve the prediction accuracy. • We propose a relation-aware attention over entity types to select the discriminative entity types. • Extensive experiments show that our method gives new state-of-the-art performance. Abstract Distant Supervised Relation Extraction (DSRE) is usually formulated as a problem about classifying a bag of sentences that contains two query entities into the predefined relation classes. Most existing methods consider those relation classes as distinct semantic categories while ignoring their potential connections to query entities. In this paper, we propose to leverage this connection to improve the relation extraction accuracy. Our key ideas are twofold: (1) For sentences belonging to the same relation class, the keywords to express the relation can vary according to the input query entities, i.e., style shift. To account for this style shift, the model can adjust its parameters in accordance with entity types. (2) Some relation classes are semantically similar, and the entity types appear in one relation may also appear in others. Therefore, it can be trained across different relation classes and further enhance those classes with few samples, i.e., long-tail relations. To unify these two arguments, we developed a novel Dynamic Parameter Enhanced Network (DPEN) for Relation Extraction, which introduces a parameter generator that can dynamically generates the network parameters according to the input query entity types and relation classes. By using this mechanism, the network can simultaneously handle the style shift problem and enhance the prediction accuracy for long-tail relations. Through extensive experiments, our method which is built on the top of the non-BERT-based or BERT-based models, can achieve superior performance over the state-of-the-art methods.