Research And Implementation Of Cross-domain Relation Extraction Based On Adversarial Network

With the development of computer network technology and the popularization of smart devices,people’s enthusiasm for communicating via the Internet has increased,and the information on the Internet has become more complex and diverse.If truly useful information can be extracted quickly and accurately from a large number of unstructured electronic texts,it will help people to understand and make better use of these information resources.In the face of the above challenges,information extraction techniques have emerged,and the relational extraction studied in this paper is the core task of information extraction.By modeling text information,relationship extraction can automatically extract entity pairs and their semantic relationships and complete information semantic mining.The resulting ternary group in the form of<entity,entity,relationship>can be used to build a knowledge spectrum,an intelligent question and answer and an intelligent search system,from which users can quickly obtain the knowledge they need and avoid the tedious and time-consuming work of filtering and browsing.In this paper,the cross-domain relation extraction problem is studied based on the adversarial network approach.Unlike the extraction of noun entity pairs in generalized relational extraction,the main target of this paper is to extract adjective noun pairs in text messages,which can better help human beings to understand the opinion messages conveyed by texts.First,a cross-domain relation extraction framework that incorporates weakly supervised labels in an adversarial neural network is proposed.The method takes full advantage of the graph representation of the graphical model to analyze the structural relationships between different word pairs and between packages and mentions,which in turn generates weakly supervised labels on the target domain.Based on the weakly supervised labels obtained above,a new objective function is designed to complete the domain adaptive model training in the adversarial network framework.Afterwards,to address the problems of low probabilistic graphical model computation efficiency and slow neural network training speed,in order to ensure that the model is more adaptive in business scenarios with high demands on immediacy,this paper proposes an adversarial network training framework that combines sample and feature migration.The framework performs the adversarial domain adaptive training after removing domain-related samples,which can effectively improve the model training efficiency.The effects of the above two algorithmic models are experimentally verified on Amazon’s five major product review datasets.Finally,the cross-domain relation extraction system is designed and implemented based on the above two algorithms,and the system covers the processes of data processing,word pair mention generation,extraction algorithm,and model prediction to automate the extraction of target relationships.