Deep Learning Based Chinese Entity Relation Extraction Research

With the continuous development of Internet technology and the continuous emergence of massive data,how to extract usefully structured information from massive unstructured data has become a research hotspot at this stage,and relation extraction has emerged as the times require.As an upstream technology for providing basic data,relation extraction has important application value in many downstream fields,such as knowledge graph,semantic understanding,recommendation retrieval,machine translation,and intelligent question answering.In recent years,deep learning models have become the state-of-the-art method for relation extraction,and existing work has achieved considerable results,but there are still problems such as entity nesting,relation overlap,and exposure bias,which seriously affect the performance of relation extraction models.From the perspective of solving the above problems,this paper proposes two entity-relationship joint extraction models.The main work includes the following aspects.1.Aiming at the entity nesting problem existing in the existing entity relation extraction methods,different from the original idea of relation extraction based on token,this paper adopts the idea of span for relation extraction,and designs and uses it.Sliding windows and three mapping strategies recombine and rearrange the token sequences into span sequences.2.Aiming at the problems of exposure bias and relationship overlap existing in existing entity relationship extraction methods,a joint entity relationship extraction method(Span based Multi Head Selection,SMHS)based on span multi-head selection is proposed,which converts entity relationship extraction into span-level multi-head relationship selection question.First,the span semantic vector is constructed by means of span marker and span embedding,and the original token sequence is converted into span sequence by combining the proposed span mapping strategy,and then LSTM and multi-head self-attention mechanism are used to extract span features.The selection mechanism performs segment-level relation decoding and introduces span classification task to assist training,and decodes relation triples in a single step.3.Aiming at the large time complexity and slow reasoning speed of SMHS,an entity relation extraction model(Span-Labeling Based Model,SLM)based on span-labeling is proposed,which transforms the entity relation extraction problem into the span-labeling problem.First,convert the token vector into span semantic vector,and combine the span mapping strategy to convert the original token sequence into span sequence,then use GRU and multi-head self-attention mechanism to extract span features,and finally use carefully designed span relationship.The labels are classified by relation labels,and the relation triples are decoded in a single step.4.Based on the experiments on the authoritative Chinese relation extraction dataset Du IE2.0,the labeling form of the dataset is re-modified.To verify the performance of the model,the current mainstream relation extraction model was selected for comparative experiments;in order to verify the effectiveness of the proposed module,an ablation experiment was carried out;in order to explore the influence of model parameters on the model,an experiment of influencing factors was carried out.Experiments show that the two models proposed in this paper have achieved better results than the current mainstream extraction methods;the proposed modules can indeed improve the performance of the model;the potential impact of related parameters on the model is determined,and the effectiveness and efficiency of the model are verified.superiority.Comparing the two models,the accuracy of SMHS is higher than that of SLM,but SLM has advantages in terms of time and space complexity and reasoning speed.