A Multi-level Analysis And Reasoning Model Based On Improved Probabilistic Soft Logic

The data corresponding to a large number of problems in the field of artificial intelligence are characterized by complex structures and large scales,from social networks to knowledge maps to images,video and natural language processing.Therefore,we need an effective response.The uncertainty of the problem,and the ability to express complex structural information with a simple description,the emergence of probabilistic soft logic(PSL)largely meets this requirement.It integrates first-order logic and probabilistic models,allowing users to model and reason about structured data using logic rules like natural language.Currently,PSL has been applied to many fields such as collective classification,link prediction,ontology alignment,personalized medicine,social network modeling,knowledge mapping and text sentiment analysis.However,in practice,PSL mainly has two dilemmas: 1)As the number of rules and the complexity of rules increased,the reasoning performance of PSL will drop sharply;2)PSL needs to give a lot of common sense and domain knowledge as human beings.Prerequisites for the establishment of rules,which are often very expensive to acquire and incorrect information contained therein may reduce the correctness of reasoning.In order to alleviate the above dilemmas,this paper starts from the rules of the rules and the construction of the rules themselves to build the model.The main work of this paper includes three aspects.: 1)Use the multi-level reasoning method to adjust the running structure of PSL rules,and all the rules simultaneously The way of running changes to hierarchical operation,improving the reasoning performance of the model.2)Using the decision tree algorithm to build the rule automatic mining module of the PSL model,so that the PSL has the ability to process classification or prediction problems.3)Further improve the PSL by introducing AMIE + The algorithm enables PSL to exploit the rules of complex knowledge map data and use it for new knowledge discovery.Experiments on datasets from several different fields shows that the proposed method by this is feasible and has better performance.