Research On Encoding And Decoding Technology For Joint Extraction Of Entities And Relations

In the era of big data,information is mainly stored and transmitted through text as a carrier.People process unstructured text data through information extraction technology to obtain structured or semi-structured data,and obtain information that users are really interested in.Named entity recognition and relationship extraction are the basis of information extraction technology,which is of great significance for advanced applications in the field of natural language processing.Traditional methods use pipelined processing of entity and relation extraction tasks,which have the disadvantages of error propagation and redundant information,and have poor performance.There are cross-dependence between entities and relationships,so joint extraction of entities and relationships is necessary.This thesis studies the encoding and decoding technology for joint extraction of entities and relationships.The neural network structure is used to automatically extract the semantic features of the input text.It does not rely on external natural language processing tools to avoid external tool errors.The proposed model includes four modules: word embedding layer,encoding layer,relation extraction module and named entity recognition module.The word embedding layer uses word2 vec word embedding of words and character-level word embedding;the encoding layer uses a bidirectional long short term memory network,which can make full use of the contextual semantic features of the text and effectively relieve the problem of long-distance dependence.The word embedding layer and the encoder layer belong to the shared bottom layer,and the output information is also used as the input of the named entity recognition decoder and the relation extraction decoder.The two decoder modules share the text features and encoder parameters,and can constrain each other.The relation extraction decoder uses a convolutional neural network,which can extract multiple relationship instances.Add an attention mechanism to the input layer of CNN,and generate context information for each word.The named entity module adopts softmax decoder and conditional random field decoder respectively,and the experimental results show that the result of conditional random field is better.In the end,the Bi LSTM-ATTCNN-CRF model is adopted.Compared with the previous joint extraction method,the model proposed in this topic has improved in F1 value.