A Study On Joint Extraction Of Textual Entity And Relation With Deep Reinforcement Learning

As the main link of knowledge extraction in natural language processing,joint extraction of textual entity and relation is a hot research topic.However,the traditional pipeline based joint extraction of entity and relation method cannot achieve good results,and the existing joint extraction methods usually need complex feature engineering.This paper studies the joint extraction of textual entity and relation based on deep reinforcement learning,and proposes the pre-training method of joint extraction of entity and relation based on parameter transfer learning and the joint extraction of entity and relation model based on depth policy gradient.The main research work and innovation contents are as follows:1)Research on the pre-training method of joint extraction of entity and relation based on parameter transfer learning.Firstly,this paper uses the transfer learning algorithm based on shared parameters to realize the pre-training of the joint extraction of entity and relation,and uses two kinds of shared parameter pre training techniques shared word vector parameters and shared LSTM joint extractor parameters.In this method,the latest ALBERT word vector combined with entity location information is embedded to construct the shared word vector parameters;part of the data is extracted to construct the LSTM model,and the pre-trained weight matrix is shared in the LSTM joint extractor with the same parameters.These two shared parameter technologies are combined to extract entity relationship.2)Joint extraction of textual entity and relation based on deep policy gradient.Based on the pre-training method of joint extraction of entity and relation based on parameter transfer learning,this paper proposes a joint extraction of entity and relation model based on deep policy gradient.In this paper,the output value of the joint extractor of LSTM after the pre-training method is taken as the action value of the model.Secondly,it initializes the target network trainer in reinforcement learning module,and disrupts the bag order;calculates the reward value of each statement orstate through reinforcement learning algorithm based on the policy gradient,and defines the target function by the policy gradient algorithm to maximize the total reward;defines the optimization function by the policy gradient algorithm to update the parameters of the model until the end of training rounds.Finally,we train the model to optimize the super parameters,set the parameter values until the model converges,and generate the final joint extraction scheme of entity relations.In this paper,the above method and model are tested on the public data set and medical text data set.The experimental results show that the joint extraction of textual entity and relation method based on deep reinforcement learning is effective,and it can more closely apply the information between entities and relations,better predict the entities and relations of complex text,and effectively improve the efficiency of joint extraction of entity and relation.