A Research And Technology Implementation To Linkage Revealing With Label Self-correcting For Property Crime Incidents

Public security departments’ transcript text and online record data are essential intelligence sources for solving property crimes.Currently,the information links contained in these data rely heavily on the experience of senior police officers to establish and reduce the many complex information of the case into a whole and form a vein before it can be used for detection.However,the processing of these underlying big data has been consuming the limited senior police resources of the grassroots police stations,and the high consumption and low efficiency is a serious challenge to the public security authorities’ "fast and accurate" disposal principle.By discovering the composition of property crimes from semi-structured or even unstructured data and constructing the discovery into a computerized system,the low-level consumption of the police force can be transformed into efficient analysis by machines,which can discover the linkage between property crimes in depth and form the case chain,providing a basis for case consolidation,reducing the intensity of manual analysis of individual cases and improving the efficiency of solving them.In this way,the deep discovery of the connection between property crimes becomes a research issue.First,this paper reviews the current state of affairs and identifies the following shortcomings in the discovery of property crime incident linkages: data representation relies on manual analysis and does not fully recognize the semantically hidden nature of crime incident texts,which results in non-domain and non-topical semantics（noisy semantics）from distant supervision and the challenges of its recognition and disambiguation;the challenges of incident linkage representation and machine cognition;and the challenges of non-deterministic local features and machine cognition brought by discrete semantics and semantic weak structure.Secondly,this paper proposes to study the characteristics of entity relations and their representations of property crime incidents from a new perspective of observing and characterizing incidents from entity relations.We also develop a Label self-correction property crime incident linkages discovery（LEOPOLD）model.Third,the open ambiguity between valid and noisy samples in semantics is investigated and found,for which a new strategy of label self-correction sample screening is proposed based on PCNN,based on which a mathematical model of label self-correction extraction of entity relations is established,and a new method of label self-correction extraction of relations is constructed from it.Fourth,we propose a new strategy of "graph coupling" representation of incident linkages,and construct a deep discovery method and sub-algorithm of incident linkages based on graph neural network.Fifthly,we construct two new key algorithms,"Label self-correction extraction of relations" and "Deep discovery of incident linkage",and construct these two key algorithms.Based on these two key algorithms,we construct and design the LEOPOLD framework algorithm,analyze and give the control complexity of the above important algorithms,and provide a theoretical upper limit for the effectiveness analysis of LEOPOLD technology.Sixth,design and implement the prototype system of LEOPOLD technology.The functional and performance effectiveness of the prototype system is tested based on test cases.There are three novel works in this paper.To address the challenges of machine characterization methods for property crime incidents,we study the entities and their relations involved in property crime incidents and entity relation extraction,and give entity relations characterization of property crime incidents.To address the challenge of ambiguity recognition in entity relation extraction,we study the characteristics of noisy samples,automatic sample screening methods and label self-correction methods,and give the sample screening and label correction methods based on matching degree Z-score,and give the label self-correction extraction technique of entity relation,referred to as REASON technique.The graph representation and graph classification methods of incident linkages are studied,and the incident graph construction method and the graph neural network-based incident linkage deep discovery technique are given.We also construct a prototype system for LEOPOLD.The results of theoretical analysis and practical test show that the proposed method of LEOPOLD does not require manually extracted crime incident features as a prerequisite,and the Recall can reach 99.20±0.81%,which is 31.70% higher than the baseline method;and the accuracy can still reach 99.23±0.35%,which is close to the baseline method,and the proposed REASON on the "NYT＿10" dataset are better than the baseline method.It shows that the incident linkage deep discovery method can reduce the labor intensity and miss fewer criminal incidents,while maintaining the overall accuracy close to that of the baseline method.