Joint named entity recognition and relation extraction is an essential natural language processing task that aims to identify entities and extract the corresponding relations in an end-to-end manner. At present, compared with the named entity recognition task, the relation extraction task performs poorly on complex text. To solve this problem, we proposed a novel joint model named extracting Entity-Relations viaImproved Graph Attention networks (ERIGAT), which enhances the ability of the relation extraction task. In our proposed model, we introduced the graph attention network to extract entities and relations after graph embedding based on constructing symmetry relations. To mitigate the over-smoothing problem of graph convolutional networks, inspired by matrix factorization, we improved the graph attention network by designing a new multi-head attention mechanism and sharing attention parameters. To enhance the model robustness, we adopted the adversarial training to generate adversarial samples for training by adding tiny perturbations. Comparing with typical baseline models, we comprehensively evaluated our model by conducting experiments on an open domain dataset (CoNLL04) and a medical domain dataset (ADE). The experimental results demonstrate the effectiveness of ERIGAT in extracting entity and relation information.

Object detection is an essential computer vision task that aims to detect target objects from an image. The traditional models are insufficient to generate a high-quality anchor box. To solve the problem, we propose a novel joint model called guided anchoring Region proposal networks and Cascading Grid Region Convolutional Neural Networks (RCGrid R-CNN), enhancing the ability of object detection. Our proposed model design is a joint object detection algorithm containing an anchor-based and an anchor-free branch in parallel and symmetry. In the anchor-based, we use nine-point spatial information fusion to obtain better anchor box location and introduce the shape prediction method of Guided Anchoring Region Proposal Networks (GA-RPN) to enhance the accuracy of the predicted anchor box. In the anchor-free branch, we introduce the Feature Selective Anchor-Free module (FSAF) to reduce the overlapping anchor boxes to obtain a more accurate anchor box. Furthermore, inspired by cascading theory, we cascade the new-designed detectors to improve the ability of object detection by setting a gradually increasing Intersection over Union (IoU) threshold. Compared with typical baseline models, we comprehensively evaluated our model by conducting experiments on two open datasets: Pascal VOC2007 and COCO2017. The experimental results demonstrate the effectiveness of RCGrid R-CNN in producing a high-quality anchor box.

Drug entity and adverse effect relation extraction is a critical task that aims to recognize drug entity and extract adverse effect relation from the unstructured medical text. Many works have used statistical learning approach, traditional or new deep learning approach to solve similar pharmacovigilance problem. Recent works tended to employ the graph convolutional network to enhance the ability of drug and adverse effect information extraction. However, the injective problem generated by a single aggregator and the weak generalization of the summation aggregator cause the graph neural network lack the ability to extract sufficient relation information. To solve these problems, we propose a new end-to-end model named recognizing Drug entities and extracting Adverse effect relations via the Principal Neighbourhood Aggregation network (DAPNA). Moreover, we compared the DAPNA with baseline models on the Adverse Drug Effect (ADE) dataset using multiple metrics. The experimental results demonstrate the method proposed in this paper achieves state-of-the-art results and can be applied to other drug and adverse effect information extraction tasks.

Deep learning has been successfully applied to computer vision, speech recognition and other domains. In image processing, the CNN model has been relatively mature in the training of labeled data. There are labeled data that need to be manually labeled in supervised learning, but now there are a large number of data without labels or with a small number of labels that need to be processed. We propose an image classification algorithm named OCFC based on self-supervised learning without manual labeling. After image preprocessing, features are extracted by three-layer Convolution Restricted Boltzmann Machine. Then the extracted feature clusters are labeled with pseudo-labels by Fuzzy C-means algorithm. Finally, the CNN model is used to classify and predict other image categories. The self-supervised learning model can be arbitrarily transferred to a shallow model or a deep model. The experimental results show that this method can effectively avoid the complexity of manually extracting features, and the accuracy on the STL-10 dataset reaches 82.7%.