Research On Entity Relation Extraction With Deep Learning

Entity relation extraction aims to extract structured relation triples from unstructured text data,which is the basis of constructing large-scale knowledge graphs.The extractive model in the entity relation extraction task can be divided into the pipeline and joint models.The joint model can be subdivided into the model based on shared parameters and the model based on joint decoding.The pipeline and model based on shared parameters are multi-stage decoding models,which suffer from exposure bias and error accumulation.The joint decoding model avoids exposure bias and error accumulation from the structure through a single-stage decoding process,but the model’s decoding structure will become complex.Relation overlap is a critical problem that needs to be dealt with in the entity relation extraction task.Current extractive models generally use stacked labeling layers to deal with this problem.However,the calculation of many labeling layers is actually redundant,which makes the labeling matrix sparser and affects the model’s extraction performance.This paper mainly studies optimizing the decoding structure of the multi-stage extractive model to reduce the impact of error accumulation and sparse matrix on the model’s performance.The research content is as follows:The current multi-stage decoding model uses a relatively simple way to deal with the feature fusion between the stages,which will increase the influence of error accumulation and affect the model’s extraction performance.This paper proposes an entity relation extraction model with dual relation prediction and feature fusion.The model encodes the text by BERT and uses a two-stage dual relation prediction structure to decode the triples.Two relation prediction corresponds to two entity alignment processes,which can effectively filter out redundant entity pairs.The feature fusion structure of the gated linear unit and conditional layer normalization can effectively enrich the information contained in the entity pair fusion features to reduce the influence of error accumulation on the extraction performance.Experimental results on two general data sets show that this model has better extraction performance than the baseline model.The current extractive model generally uses stacked labeling layers to deal with the problem of relation overlap,which increases the labeling matrix’s sparsity and affects the model’s final fitting effect.This paper proposes an entity relation extraction model with chain decoding structure that shares the BERT output.The model divides the decoding process of triples into five stages to avoid the increase of the sparsity of the labeling matrix.The chain decoding structure reduces the effect of error accumulation by filtering the information layer by layer,and conditional layer normalization is used to ensure the effect of the information filtering.The gated attention unit and gated linear unit are used to enhance the fitting performance of the model,and the head-tail separation and the relation correction module are used for multiple verifications of the relationship set.The model achieves better extraction results on two general data sets and proves the feasibility of the multi-stage chain decoding structure.In order to show that the proposed model can be applied in practice,the relational triplet labeling system is designed according to the entity relation extraction model with chain decoding structure.The relational triplet labeling system is divided into two modes: single sentence and batch.The purpose of this system is to help users the labeling process of entity relation extraction related corpus.In this paper,the design and implementation of the system are introduced in detail,and the practical effects of corresponding functions are shown.