A Research On Entity Relation Extraction Model And Performance Improvement

Free texts contain a large quantity of unstructured and important information,which is hard to use directly.Entity relation extraction is a kind of technique that aims to transform information in unstructured text into structured information.In recent years,the combination of machine learning and entity relation extraction is highly focused among researchers.Supervision learning requires a large amount of human-labeld data which is expensive to obtain,and weakly-supervised learning is suffered from wrong labeled problem,along with overfitting problem comes with machine learning method.To address these problems,the main improvement work can be listed as follows:To address the wrong label problem,we design an end-to-end model based on Squeeze-and-Excitation network.Traditional deep learning method usually use shallow network as sentence encoder and can not model the noise in weakly-supervised learning.Our proposed model can enhance the expressive ability of network,alongwith soft-label method,eliminate the noise in distant supervision,and further improve the performance of model.To address the problem of imbalance problem and hard sample problem in distant supervision,we improve the scoring function by substituting cross entropy with focal loss.Focal loss can reshape the loss function to down-weight easy examples and thus focus training on hard negatives and further improve the performance of model.To overcome the flaws of max pooling like losing position information in convolution networks that we apply in Natrual Language Processing task,we design and realize double pooling.We also design several regularization methods,such as random depth algorithm,random word dropout and adversarial training to overcome overfitting problem and improve the performance of model.