Research And Implementation Of Unsupervised Domain Adaptation Technology In Relation Extraction

Entity and relation extraction are the key steps of knowledge graph construction and information extraction.It mainly extracts some kind of connection between two or more entities in order to obtain triples(entity 1,relationship,entity 2).The domain adaptation problem refers to how the models learned in one domain can be effectively applied to another domain.In real life,we are unlikely to obtain data in all fields,and it is impossible to label data for all fields.Therefore,the unsupervised domain adaptation problem is particularly important in relation extraction.The research problem of this paper is to use the labeled data in the source domain and the unlabeled data in the target domain to improve the relation extraction of the target domain.It is mainly divided into the following aspects.1.In order to avoid introducing private domain features and more effectively extract shared domain features,a cross-view adaption network is proposed.The network uses cross-view training in the target domain.The labeled data in the source domain and the unlabeled data in the target domain are fed into the shared feature extractor together to learn the shared feature representation.These features are then used to generate relation predictions.In addition to these full-view data,some restricted view data are also constructed,which will lose some context information on the target domain,such as entity words.These restricted view data are also fed into a shared feature extractor to generate predictions.The predictive distribution of full-view data will act as a "teacher" to enable different restricted view data to learn the same predictions.By fitting the predicted distribution,the model can learn some contextual information that does not depend on domain-specific features(such as entity words).Experiments on the ACE2005 dataset show that the model improves the F1 value by up to 2.1%compared to the existing work,and achieves the best current results.2.Aiming at the problem that the existing models cannot capture non-sequential features in different domains well,an end-to-end graph adaptation network is proposed to align non-local features between domains.In the tripartite graph constructed,edges exist only between domain private words and shared words,source domain words and target domain words are indirectly connected through shared words as medium,and word co-occurrence information is used as the fixed weight of edges.However,the word co-occurrence information has a strong dependence on the corpus,and it will inevitably introduce some noise,so it also introduces the dynamic weight of the graph.The dynamic weights of domain private words are caculated by attention mechanism.The fixed weight and dynamic weight are added as the final weight of the edge.Domain private words are aligned by a graph convolutional neural network,and then the vector representation of these words is fed into a shared feature extractor for feature extraction,which can effectively avoid the introduction of domain specific information.Experiments on the ACE2005 dataset show that the graph adaptation network can effectively align the domain specific features of non-sequential,and pay more attention to the alignment of some words with strong domain correlation.The experimental results show that the F1 value is increased by up to 2.7%,which also illustrates the effectiveness of the model.3.Based on the above innovations,an unsupervised cross-domain relationship extraction module was designed and developed to visually show the effects of the model,furthermore make it easy to integrate into the knowledge extraction system or other downstream tasks.