Research On Techniques Of Relation Extraction Based On Mixed Supervision

Relation extraction is an important topic in the field of natural language processing and knowledge graph building.The knowledge graph consists essentially of(entity-relation-entity)triples.In the process of building a knowledge graph,the most important task is to solve the knowledge acquisition.In addition to acquiring knowledge from structured information,a large amount of relevant knowledge is hidden in unstructured text.Therefore,how to extract structured knowledge from unstructured text,especially the relationship between entities,is urgently needed to be solved at this stage.This paper studies the problem of the extraction of domain-restricted sentence-level relations extraction based on deep learning.Methods to solve the problem of the domain-restricted sentence-level relationship extraction methods based on deep learning are mainly divided into two categories:full-supervised relationship extraction based on manual annotated data,and distant supervised relationship extraction based on existing knowledge graphs and unlabeled data.Despite the relationship between full-supervised relationship extraction and distant supervised extraction has achieved good results,but each has its own problems.First of all,there is a problem that the total supervision has insufficient data volume and the manual labeling data cost is too high.Then distant supervision has the problem that the generated data is inaccurate and needs to be noise-reduced.This paper first proposes a distant supervised data generator model based on Bi-LSTM+ATT,and designs the knowledge graph relationship vector,and uses the knowledge graph relationship vector as an additional feature inputted to model to assist data generation.In this paper,the APCNN+D distant supervision model is reconstructed using the Bi-LSTM+ATT model.The experiment proves that the reconstruction result is highly consistent with the performance of the original model,and the reconstruction is basically realized.The fourth chapter of this paper proposes the concept of mixed supervision,which combines full supervision and distant supervision.Furthermore,a mixed supervision model mix GAN combining manual annotation data and knowledge graph is proposed.In the specific implementation of the model,we innovatively propose to use the generated adversarial network,the generator is used to generate distant supervised annotation data,the discriminator is used to identify other people's annotation data and machine generated data as much as possible,and the generator improves the quality of the generated data.It is possible to confuse the discriminator to achieve adversarial training and ultimately get a high quality distant supervised data generator.This paper uses the Bi-LSTM+ATT model mentioned in Chapter 3 as a generator,and designs a discriminator model for generating adversarial training.Experiments on Freebase knowledge graph and NTY text data prove that the generated adversarial model can significantly improve the performance of the Bi-LSTM+ATT model.