Research On Joint Extraction Method Of Entity Relationship Based On Deep Learning

Named entity recognition and relational extraction are two very important fundamental tasks in the field of text mining.The traditional research method is to use the pipeline model to process the task in two steps,that is,first to identify the entity and then extract the relationship.Although the pipeline model can design the whole system flexibly,there are two problems in the process: 1)It separates the intrinsic link between the two tasks of named entity recognition and relationship extraction,and discards useful information between some subtasks.2)Error propagation,that is,when the upstream named entity recognition task has an error,it will be transmitted to the relationship extraction task without difference and without feedback,which has an impact on the accuracy performance of the entire system.In order to overcome the above problems of the pipeline model,some researchers have developed a joint model of named entity recognition and relationship extraction.Traditional joint models are generally heavily dependent on feature engineering.In addition,many models often only examine the entity pairs extracted by the relationship,rather than directly modeling the entire sentence,so that it is impossible to infer multiple pairs of entities and relationships in the same sentence.The previous joint extraction work is mostly based on the representation of word embedding.Since the traditional word embedding cannot solve the problem of the representation of each meaning item in the polysemy,the training effect of the model is inevitably affected.In order to solve the problem of word embedding,a text representation based on pre-training has been proposed in recent years.Among them,the BERT model is the most outstanding.The model fully combines the advantages of previous work,and proposes a bidirectional encoder representation based on Transformer.Effect.In this paper,a pre-trained BERT model is used as our text representation,and a new joint framework for named entity recognition and relation extraction is proposed.Previous models relied on features,and model accuracy was affected by the extracted features.Our model can automatically learn features without being limited by external conditions.Our model regards named entity recognition as sequence labeling problem,and uses CRF layer to obtain global optimal output sequence.For relation extraction task,through sigmoid layer,it also predicts whether there is relationship between two entities and the relationship type between them.And can also identify pairs of relationships that a sentence may have.In the training process,the sum of the loss functions of the two tasks of named entity recognition and relationship extraction is optimized as the final loss function,and the joint model of named entity recognition and relationship extraction is realized.Finally,the results of the sequential pattern were used to correct the results.The experimental results show that our model is better than the neural network model that automatically extracts features before,and the performance is still good compared with the feature-based neural network model.