Research On Vertical Domain Entity Relation Analysis Method Based On Deep Learning

Relation extraction is the core task in natural language processing(NLP)and can provide structural knowledge,semantic understanding and relational reasoning for downstream tasks of NLP,such as question answering,dialogue system,and knowledge graph.Such as the dialogue system,relation extraction converts semi-structured or unstructured data into structured data,which is crucial for the natural language understanding of the dialogue text and the construction of the knowledge base.Previous relation extraction models rely heavily on large-scale supervised data,they do not have generalizations and robustness,and difficult to provide algorithmic technical support for industrial application.In the relation extraction task,we combine machine learning algorithms and feature fusion technology of deep learning to extract structured knowledge from plain text and apply entity relation extraction to the actual application scenario.The main works of this paper are as follows.1)Distantly supervised relation extraction.In order to obtain large-scale labeled training data,the distant supervised methods that automatically label data are widely used in relation extraction tasks.In this paper,in order to solve the problem of data loss and deformation in deep network propagation,we propose a compensation mechanism in the deep residual neural networks.We use the attention mechanism to extract sentences features,which alleviates the influence of noise data on model performance.To address deep networks to be impressionable to overfitting and highly sensitive to noise,we introduce adversarial learning to train the model.In the distant supervised methods,the experimental results show that the proposed model achieves significant and consistent improvements on distantly supervised datasets(NYT).2)Few-Shot relation classification.Although distant supervised methods show great success in relation classification,they also bring the problem of data noise.Inspired by people learning new knowledge from only a few samples,we focus on predicting formerly unseen classes with a few labeled data.In this paper,we propose a heterogeneous graph neural network for few-shot relation classification,which contains sentence nodes and entity nodes.We build the heterogeneous graph based on the message passing between entity nodes and sentence nodes in the graph,which can capture rich neighborhood information of the graph.In the few-shot learning relation classification,experimental results have demonstrated that our model outperforms the state-of-the-art baseline models on the FewRel dataset.3)Document-Level relation extraction.Most of the existing relation extraction models are based on intra-sentence relations for single entity pairs,which is difficult to extend to practical applications.We propose a document-level multi-entity relation extraction model,which can collect the co-occurrence information in the document to refer to the disambiguation,and use the attention mechanism to make full use of the context information contained in the document.In the document-level relation extraction,we use the attention mechanism and the co-reference information to give the model reasoning ability,and the performance evaluation on the DocRED document-level relationship extraction dataset exceeds the most advanced experimental results in the previous work.The experiment proves that the method proposed in this paper is real and effective in the relation extraction task,and has certain value in scientific research and practical application.Combining deep learning to improve the overall performance of the relational extraction model is very important for realizing the practical application of vertical domain entity relation extraction.At the end of the paper,we review the current works and look forward to future works.