Geographic Entity Relationship Extraction Model Based on Piecewise Convolution of Residual Network

Nowadays, geographic entity relationship extraction systems generally rely on artificial feature extraction. These features either require complex and complete data sets, or cannot describe deep features such as semantics. And data sets that can be used for geographic relationship extraction are scarce. To tackle these problems, this paper uses distant supervision to map existing knowledge bases into rich unstructured data which contributes to a large amount of training data. In training, this paper uses the deep residual network to extract more abstract and deeper features. Then the piecewise max pooling and selective attention mechanisms are used to further improve the accuracy of the model. Finally, the experimental results show that the deeper network and the piecewise max pooling significantly improve the extraction results.