Research And Implementation Of Relation Extraction Model Based On Interactive Graphical Model

With the rapid development of Internet economy,more and more data are generated by users on the network,and most of them are text data.Huge amounts of data contain a lot of valuable information,therefore an important challenge is how to mine these imformation from complex text data.Among many text mining tasks,relation extraction is one of the current research hotspots.It not only has important theoretical value,but also has extensive applications in bioinformatics,e-commerce,social network,information retrieval and other scenarios.In the current research,discriminative models based on deep learning have achieved good performance.This kind of model requires a large amount of labeled data,however it costs a lot to obtain high-quality labels.Though crowdsourcing and semi-supervised learning methods can reduce the labeling workload,the discriminant results in the training process are not accurate enough due to the lack of label quality and quantity.To mitigate the impact of insufficient label quality and quantity,related data dependencies on domain and task have been taken into consideration.On the way to this purpose,the main challenges are how to mine and describe the multiple dependencies between data through a combination of qualitative and quantitative methods,discover model problems by using the dependencies,and obtain better performance through further debugging.In view of the above challenges,in order to help users define and fine-tune data dependencies,an interactive graphical model construction method is proposed.First of all,it is oriented to the task of relation extraction,fully mining multiple dependencies between text data,and mapping them to the structural characteristics of MentionGraph graphical model.By selecting appropriate dependencies through interactive debugging and gradually building a graphical model,the issues of noise labels and insufficient labels can be effectively alleviated.In order to support good user interaction during debugging,this paper proposes effective interaction principles,and designs debugging semantics and operators.Secondly,based on the above methods and principles,an interactive model construction and debugging system is designed to help users observe the model operation and explore the graphical model space through a good user-system interaction interface.Finally,the debugging process is not efficient enough due to the large amount of calculation of the graphical model,three accelerated optimization methods are proposed to solve the problem,which effectively improve the convergence speed of the graphical model and then optimize the user experience.The experimental results on the relation extraction task show that the proposed interactive graphical model constructed by the system can be generalized from a small training set to a larger corpus in another domain,therefore it reflects the ability to learn the dependencies of cross-domain corpora.Moreover,it shows that the proposed interactive graphical model is superior to the comparison methods on both crowdsourcing tasks and semi-supervised learning tasks.