| Unstructured text contains rich named entity and relation information.How to quickly and accurately extract this semantic information from text is a basic task in the field of natural language processing,which can provide data support for downstream tasks such as knowledge graph construction and information retrieval.The joint entity relation extraction aims to extract the entity relation triples contained in the text at the same time.Because it better solves the problem of error propagation and ignoring task dependency in the pipeline framework,it has been widely concerned by researchers.The current research on joint entity relation extraction mainly has the following challenges:(1)Long-distance entity relations are difficult to identify.Due to problems such as the vanishing gradient,the relation between long-distance entities is difficult to obtain.(2)The directionality of the relation is difficult to obtain.When extracting the relation,the first entity and the second entity should be assigned correctly.(3)The overlapping relation is difficult to extract.Because there is reasoning between overlapping relations,how to use reasoning information is the key to solving the problem of overlapping relation extraction.In response to the above challenges,this paper proposes two joint entity relation extraction models based on graph convolutional neural networks.The main contents are as follows:(1)Aiming at the problem of long-distance entity relation recognition and relation directionality acquisition,a joint entity relation extraction model based on pruning dependency tree graph convolution is proposed.The model consists of two parts: the sequence layer and the graph convolutional neural network layer.The sequence layer is used to extract entities while the graph convolutional neural network layer is applied to extract relations.In order to make the model have the ability to acquire the directionality of the relation,the model successively concatenates two entities and sentence semantic features.In order to solve the long-distance entity problem,the model adopts a path-centric pruning strategy to prune the syntactic dependency tree.Finally,the validity of the model is verified by experiments on the NYT dataset.(2)Aiming at the problem of overlapping relations extraction,a joint entity relation extraction model based on relation-aware graph convolution is proposed.The model divides joint extraction into two subtasks: detecting entity spans and identifying entity relation types at the same time.In order to obtain the reasoning information between the overlapping relation triples,this paper proposes a relation-aware attention mechanism to obtain the relation representations between the two entity spans,and construct a fully connected graph according to the entity spans and the relation representations.In addition,in order to use adjacent node features and edge information at the same time when encoding node representations,this paper improves the original graph convolutional neural network.Finally,the experimental results performed on the NYT and Web NLG datasets illustrate the effectiveness of the proposed model. |
