Method, system and device for constructing criminal investigation knowledge graph based on multi-agent cooperation

Abstract

The present application relates to the technical field of knowledge graph, in particular to a criminal investigation knowledge graph construction method, system and device based on multi-agent cooperation, aiming at improving the quality of knowledge graph. The criminal investigation knowledge graph construction system based on multi-agent cooperation comprises a planning agent, an analysis agent and a verification agent. The planning agent is used to generate a first search parameter and a planning prompt word according to user demand, and obtain first task data from a data platform according to the first search parameter. The analysis agent is used to generate a structured expression of a target knowledge graph according to the planning prompt word, and initialize the target knowledge graph according to the first task data. The verification agent is used to check whether there is a conflict in the elements of the target knowledge graph, and if there is, select the element with the highest authenticity probability. The present application not only improves the efficiency of constructing the knowledge graph, but also improves the quality of the knowledge graph.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knowledge graph, and particularly relates to a criminal investigation knowledge graph construction method, system and device based on multi-agent cooperation. BACKGROUND

[0002] As a key intelligent technology, the knowledge graph applied to criminal investigation is gradually becoming the core force to improve the efficiency of investigation. The knowledge graph is a structured semantic knowledge base, which stores and displays the complex relationship between entities in a graphical way, and provides a comprehensive and dynamically updated intelligence analysis platform for the field of criminal investigation.

[0003] In the technical field of automatically constructing a criminal investigation knowledge graph, there are currently three models: a universal information extraction (UIE) model, a large language model (LLM) and a single agent model. However, the quality of the knowledge graph constructed by these existing technologies when applied to the field of criminal investigation is not satisfactory. SUMMARY

[0004] In order to solve the above problems in the prior art, the present application proposes a criminal investigation knowledge graph construction method, system and device based on multi-agent cooperation, which improves the quality of the constructed knowledge graph.

[0005] In a first aspect of the present application, a criminal investigation knowledge graph construction system based on multi-agent cooperation is proposed, which comprises:

[0006] A planning agent is configured to generate a first search parameter and a planning prompt word according to user demand, and to obtain first task data from a data platform according to the first search parameter;

[0007] An analysis agent is configured to generate a structured expression of a target knowledge graph according to the planning prompt word, and to initialize the target knowledge graph according to the first task data;

[0008] A verification agent is configured to check whether there is a conflict in the elements of the target knowledge graph, and if there is, to select the element with the highest probability of authenticity.

[0009] Preferably, the system further comprises:

[0010] An extension agent is configured to search whether there is an extensible entity on the leaf node in the target knowledge graph, and if there is, to generate an extension demand feedback to the planning agent;

[0011] The planning agent is further configured to generate a second retrieval parameter and an extension prompt word according to the extension requirement, and obtain second task data from the data platform according to the second retrieval parameter;

[0012] The analysis agent is further configured to extend the target knowledge graph according to the extension prompt word and the second task data.

[0013] The verification agent is further configured to check whether there is a conflict in the elements of the extended target knowledge graph, and if there is, select the element with the highest authenticity probability.

[0014] Preferably, the system further comprises:

[0015] The first training module is configured to train the planning agent, the analysis agent, the verification agent and the extension agent.

[0016] Preferably, the first training module comprises:

[0017] The pre-training unit is configured to preheat train the planning agent, the analysis agent, the verification agent and the extension agent respectively;

[0018] The collaborative training unit is configured to collaboratively train the planning agent, the analysis agent, the verification agent and the extension agent that have been preheated trained by using a MADDPG algorithm until the quality score r1 of the generated knowledge graph reaches a maximum value.

[0019] Preferably, the pre-training unit comprises:

[0020] The first sub-unit is configured to generate a knowledge graph based on the planning agent and the analysis agent while fixing the analysis agent, train the planning agent by using an RLHF process and a PPO algorithm until the quality score r2 of the generated knowledge graph reaches a maximum value.

[0021] The second sub-unit is configured to generate a knowledge graph based on the planning agent and the analysis agent while fixing the planning agent, train the analysis agent by using an RLHF process and a PPO algorithm until the quality score r3 of the generated knowledge graph reaches a maximum value.

[0022] The third sub-unit is configured to train the verification agent based on a graph training set with conflicting elements until the recognition rate of the conflicting elements reaches a maximum value.

[0023] The fourth sub-unit is configured to train the extension agent based on a graph training set with expandable entities until the recognition rate of the expandable entities reaches a maximum value.

[0024] Preferably, the system further comprises:

[0025] a reward model, configured to calculate the quality score r1, the quality score r2 or the quality score r3 of the knowledge graph during the training according to the following formula:

[0026] r G =w P P G +w R R G +w r r s

[0027] wherein,

[0028] r G denotes the quality score r1, the quality score r2 or the quality score r3;

[0029] P G denotes precision; R G denotes recall; r s denotes subjective score; w P , w R and w r each denote a weight, which is calculated by a multivariate gradient regression model.

[0030] Preferably, the system further comprises:

[0031] the subjective score model, configured to calculate the subjective score;

[0032] a second training module, configured to train the subjective score model;

[0033] the second training module comprises:

[0034] an initialization unit, configured to set random parameters for the planning agent, the analysis agent, the verification agent and the extension agent respectively;

[0035] a graph construction unit, configured to construct K graph samples using the planning agent, the analysis agent, the verification agent and the extension agent; K is a preset value;

[0036] a sample generation unit, configured to sort K graph samples according to the scores of the case handling personnel, and generate training samples according to the sorting;

[0037] a training unit, configured to train the subjective score model based on the training samples and a sorting loss function;

[0038] wherein,

[0039] Each of the training samples is a triple (x, y1, y2), x represents input data for generating a knowledge graph, y1 and y2 represent the graph with a higher score and a lower score obtained by a case-handling personnel respectively;

[0040] The ranking loss function is:

[0041]

[0042] D represents A data set composed of the training samples from the data set D;

[0043] represents taking expectation on the training samples from the data set D;

[0044] r θ (x, y1) represents a predicted score of the model when inputting (x, y1);

[0045] r θ (x, y2) represents a predicted score of the model when inputting (x, y2);

[0046] σ represents a sigmoid function.

[0047] Preferably, the system further comprises:

[0048] a reality evaluation model for evaluating the reality probability of a graph element;

[0049] a third training module for training the reality evaluation model;

[0050] The third training module comprises:

[0051] a sample generation unit for collecting authoritative open-source knowledge graph data as positive samples, and randomly replacing a certain attribute of the positive samples with a false description to generate negative samples;

[0052] a training set construction unit for labeling the reality probability of the positive samples and the negative samples, and then constructing a training set;

[0053] a training unit for selecting a graph element z from the training set, recording all element vectors connected to the element within N hops as characterizing the authority of the data source of the element as a weight vector and taking as the input of the reality evaluation model, and training by using a back propagation method.

[0054] In a second aspect of the present application, a criminal investigation knowledge graph construction method based on multi-agent collaboration is provided, and the method comprises:

[0055] The planning agent generates a first retrieval parameter and a planning prompt word according to the user demand, and obtains first task data from a data platform according to the first retrieval parameter;

[0056] The analysis agent generates a structured expression of a target knowledge graph according to the planning prompt word, and initializes the target knowledge graph according to the first task data;

[0057] The verification agent checks whether there is a conflict in the elements in the target knowledge graph, and if there is, selects an element with the highest authenticity probability.

[0058] In a third aspect, the application provides a computer readable storage device, which stores a computer program capable of being loaded and executed by a processor to perform the method described above.

[0059] The application has the following beneficial effects:

[0060] The criminal investigation knowledge graph construction system based on multi-agent collaboration has the planning agent, the analysis agent and the verification agent to undertake the planning, construction and verification functions respectively, and can automatically verify authenticity when there is a conflict in data.

[0061] The extension agent is used to detect extendable entities and feed back to the planning agent, which provides an effective way for dynamic expansion of the target knowledge graph, and makes the target knowledge graph information more and more perfect, thereby providing a good decision basis for criminal investigation personnel.

[0062] In the first training module, the multi-agents are preheated and trained respectively, and then are collaboratively trained, so that the training cost is reduced, and the quality of the knowledge graph is further improved. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 is a main structure schematic diagram of the criminal investigation knowledge graph construction system based on multi-agent collaboration of the application;

[0064] Figure 2 is a schematic diagram of the target knowledge graph generated in the embodiment of the application;

[0065] Figure 3 is a main structure schematic diagram of the criminal investigation knowledge graph construction system based on multi-agent collaboration of the application. DETAILED DESCRIPTION

[0066] The preferred embodiments of the present application will be described below with reference to the accompanying drawings. Those skilled in the art will understand that these embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0067] To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0068] It should be noted that, in the description of the present application, the terms "first", "second" are only for the convenience of description, and do not indicate or imply the relative importance of the devices, elements or parameters, and therefore cannot be understood as a limitation of the present application. In addition, the term "and / or" in the present application is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper, if not specially stated, generally represents a "or" relationship between the associated objects before and after.

[0069] The construction of criminal investigation knowledge graph mainly depends on several key elements:

[0070] (1) Entity recognition: In the construction of knowledge graph, entity refers to those independent objects that can be clearly identified and distinguished, including personnel, cases, articles, scenes, files, etc. Each entity has its unique attributes, such as the name of personnel, the number of cases, the description of articles, etc. These attributes provide accurate identity for entities.

[0071] (2) Relationship modeling: The relationship between entities constitutes the skeleton of the knowledge graph, which reveals the mutual influence and connection between entities. For example, the relationship between personnel and cases may cover suspects, victims, witnesses, etc. The relationship between cases and scenes may involve the location of the crime, the time, etc.

[0072] (3) Attribute extraction: Each entity is equipped with a series of attributes, which provide detailed description for the entity. Taking the case entity as an example, it may contain the specific time, place, suspect, victim information, etc. of the case.

[0073] (4) Data fusion: Knowledge graph aims to reduce information redundancy and improve the consistency and accuracy of data by integrating data from different sources, such as transcript data, police data, scene investigation, social media, police records, etc.

[0074] (5) Dynamic updating: Given that criminal intelligence reconnaissance is a constantly changing process, the knowledge graph must have the ability to update in real time to reflect the latest intelligence and data. This includes the entry of new cases, the discovery of new evidence, the identification of new suspects, etc., to ensure that the intelligence analysis platform is always up to date.

[0075] (6) Quality assessment: A strict quality assessment system is the guarantee of the credibility of the knowledge graph, which needs to be evaluated from multiple aspects to ensure the accuracy and authenticity of the graph elements and their actual contribution to case investigation.

[0076] Multi-Agent Deep Deterministic Policy Gradient (MADDPG) is a reinforcement learning algorithm used in multi-agent environments, which is an extension of the Deep Deterministic Policy Gradient (DDPG) algorithm in multi-agent scenarios. MADDPG is mainly used to solve the problem of cooperation and competition in multi-agent environments, especially when the interaction between agents can be very complex.

[0077] MADDPG algorithm is based on Actor-Critic structure, where each agent has its own Actor and Critic networks. Actor is used to output action policy, while Critic is used to evaluate the value of actions. The uniqueness of this algorithm is that the Critic network is global, that is, it not only depends on the state and action of a single agent, but also uses the state and action of all agents.

[0078] In the embodiments of the present application, planning agents, analysis agents, verification agents and extension agents are used to perform the construction and updating of the criminal investigation knowledge graph, each agent is a LLM-Based Agent, and each agent includes an Actor network and a Critic network.

[0079] Figure 1 is the main component schematic diagram of the first embodiment of the criminal investigation knowledge graph construction system based on multi-agent collaboration of the present application. As shown in Figure 1 the system 100 of the present embodiment includes a planning agent 101, an analysis agent 102 and a verification agent 103.

[0080] The planning agent 101 is used to generate the first retrieval parameter and the planning prompt word according to the user demand, and to obtain the first task data from the data platform according to the first retrieval parameter.

[0081] The planning prompt word includes steps and precautions for constructing the knowledge graph, and is used to guide the analysis agent 102 to construct the knowledge graph. The data platform stores various data that may be used in case handling, such as a smart criminal investigation platform. The first task data is data required for creating a target knowledge graph according to user demand. The first task data can include case data, personnel data, recording data, and police situation data. In some embodiments, the planning agent 101 will also find an interface for obtaining the first task data according to the interface document of the data platform 300, and then obtain the first task data required for constructing the knowledge graph from the data platform according to the interface and the first search parameter.

[0082] The analysis agent 102 is configured to generate a structured expression of the target knowledge graph 400 according to the planning prompt word, and initialize the target knowledge graph 400 according to the first task data.

[0083] The step of initializing the target knowledge graph according to the first task data specifically includes extracting entities, relationships, and attributes from the first task data, and then filling them into the target knowledge graph.

[0084] The verification agent 103 is configured to check whether elements in the target knowledge graph 400 conflict, and if so, select the element with the highest authenticity probability.

[0085] Specifically, the verification agent 103 checks whether the entities have attributes with the same or similar meanings, checks whether these attributes conflict, sorts all conflict points and data sources of the conflict points, calculates authenticity probabilities of the conflict points, and updates the graph.

[0086] For example, the user demand is to construct an organizational relationship knowledge graph of B company according to the recording information of capturing suspects in a fraud case in A city.

[0087] The first search parameter generated by the planning agent 101 includes an involved organization information search parameter (company ID), an involved personnel information search parameter (company ID and case ID), a case information search parameter (case ID), an involved personnel information search parameter, an involved personnel recording information search parameter, and the like.

[0088] The prompt planning words generated by the planning agent 101 include: (1) entity recognition: identifying entities such as companies, individuals, cases, etc. from the first task data, and assigning a unique identifier to each entity; (2) relationship extraction: determining the relationship between entities, such as the relationship between employees and companies, the relationship between suspects and cases, etc.; (3) attribute assignment: assigning appropriate attributes to each entity, such as company name, employee position, case type, etc.; (4) graph construction: using a graph database or other suitable tools to construct a knowledge graph based on entities and relationships; (5) graph optimization: optimizing the graph, including merging duplicate entities, adjusting relationship direction and type, etc. to improve the readability and usability of the graph; (6) graph verification: verifying the accuracy and completeness of the graph to ensure that all key information is correctly represented.

[0089] The target knowledge graph constructed and initialized by the analysis agent 102 according to the planning prompt words is shown in FIG. 4. Figure 2 In this example, the verification agent 103 finds that the victim attribute information has a conflict, i.e. the number of victims stated in the Zhang San record is 60, which is inconsistent with the number of 130 victims recorded in the case investigation data. After calculating the authenticity probability of the two attribute values respectively, it is considered that the value of 130 is more authentic, and therefore the number of victims in the target knowledge graph is set to 130.

[0090] Figure 3 FIG. 3 is a schematic diagram of the main components of the second embodiment of the criminal investigation knowledge graph construction system based on multi-agent collaboration of the present application. As shown in FIG. 3, the system of this embodiment includes a planning agent 201, an analysis agent 202, a verification agent 203, and an extension agent 204. Figure 3

[0091] The planning agent 201 is used to generate first search parameters and planning prompt words according to user requirements, and to obtain first task data from the data platform according to the first search parameters, and is also used to generate second search parameters and extension prompt words according to the extension requirements sent by the extension agent 204, and to obtain second task data from the data platform 300 according to the second search parameters. The extension prompt words include steps and precautions for extending the knowledge graph, and are used to guide the analysis agent 202 to extend the target knowledge graph.

[0092] The analysis agent 202 is used to generate a structured representation of the target knowledge graph 400 according to the planning prompt words, and to initialize the target knowledge graph 400 according to the first task data, and is also used to extend the target knowledge graph 400 according to the extension prompt words and the second task data.

[0093] ​The verification agent 203 is used to check whether there is any conflict between the elements in the target knowledge graph. If so, the element with the highest probability of authenticity is selected.

[0094] The line expansion agent 204 is used to retrieve whether there is an expandable entity on the leaf node in the target knowledge graph. If so, it generates a line expansion demand and feeds it back to the planning agent 201.

[0095] For example, the line expansion agent 204 detects Figure 2 The leaf node in the target knowledge graph shown in the figure has an attribute that is the ID card information of "Zhang San". Then, information such as family members and criminal records related to the ID card information can be obtained through the data platform. At this time, the entity "Zhang San" is considered to be expandable, and the expansion demand is generated and fed back to the planning agent 201. The planning agent 201 generates a second search parameter and expansion prompt word based on the expansion demand, and obtains the second task data (Zhang San's family members, criminal records, etc.) based on the second search parameter. The analysis agent 202 expands the target knowledge graph based on the expansion prompt word, thereby providing more detailed information for the case handlers.

[0096] In an optional embodiment, the knowledge graph construction system of the present invention may also include: a first training module, a reward model, a subjective scoring model, a second training module, an authenticity evaluation model and a third training module.

[0097] In this embodiment, the first training module is used to train the planning agent, the analysis agent, the verification agent, and the line expansion agent. The first training module may include: a pre-training unit and a collaborative training unit.

[0098] The pre-training unit is used to pre-train the planning agent, the analysis agent, the verification agent, and the line expansion agent. The pre-training unit may include: a first sub-unit, a second sub-unit, a third sub-unit, and a fourth sub-unit.

[0099] Among them, the first sub-unit is configured as follows: when the analysis agent is fixed, a knowledge graph is generated based on the planning agent and the analysis agent, and the planning agent is trained using the RLHF process and the PPO algorithm until the quality score r2 of the generated knowledge graph reaches the maximum value; the second sub-unit is configured as follows: when the planning agent is fixed, a knowledge graph is generated based on the planning agent and the analysis agent, and the analysis agent is trained using the RLHF process and the PPO algorithm until the quality score r3 of the generated knowledge graph reaches the maximum value; the third sub-unit is configured as follows: based on the graph training set with conflicting elements, the verification agent is trained until the recognition rate of the conflicting elements reaches the maximum value; the fourth sub-unit is configured as follows: based on the graph training set with extensible entities, the line expansion agent is trained until the recognition rate of the extensible entities reaches the maximum value.

[0100] The collaborative training unit is used to collaboratively train the planning agent, analysis agent, verification agent and line expansion agent after preheating training using the MADDPG algorithm until the quality score r1 of the generated knowledge graph reaches the maximum value.

[0101] In this embodiment, the reward model is used to calculate the quality score r1, quality score r2 or quality score r3 of the knowledge graph during the training process according to the method shown in the following formula (1):

[0102] r G = w P P G +w R R G +w r r s (1)

[0103] Among them, r G represents the quality score r1, quality score r2 or quality score r3; P G represents the precision rate; R G represents the recall rate; r s Indicates subjective rating; w P 、w R and w r Both represent weights, which are calculated by the multivariate gradient regression model.

[0104] The precision rate is the ratio of the intersection of the knowledge graph and the baseline (i.e., the reference standard or known correct knowledge graph) to the knowledge graph, P G The higher it is, the more accurate the information in the constructed knowledge graph is (consistent with the baseline). Recall is the ratio of the intersection of the knowledge graph and the baseline to the baseline, R G The higher it is, the more the constructed knowledge graph successfully covers more information in the baseline.

[0105] In this embodiment, the subjective rating model adopts the ResNet network structure to calculate the subjective rating r in formula (1) s .

[0106] In this embodiment, the second training module is used to train the subjective scoring model, and the second training module includes: an initialization unit, a graph construction unit, a sample generation unit and a training unit.

[0107] The initialization unit is used to set random parameters for the planning agent, analysis agent, verification agent and line expansion agent respectively; the map construction unit is used to use the planning agent, analysis agent, verification agent and line expansion agent to construct K map samples, K is a preset value; the sample generation unit is used to sort the K map samples according to the case handlers' scores, and generate a sample based on the sorting. training samples; the training unit is used to train the subjective rating model based on the training samples and the ranking loss function.

[0108] Each training sample is a triple (x, y1, y2), where x represents the input data for generating the knowledge graph, and y and y2 represent the graphs that receive higher and lower scores from case handlers, respectively. The ranking loss function is shown in the following formula (2):

[0109]

[0110] Where D represents A dataset consisting of training samples; represents the expectation of the training samples from the dataset D; r θ (x,y1) represents the predicted score of the model when input (x,y1); r θ (x,y2) represents the predicted score of the model when input (x,y2); σ represents the sigmoid function.

[0111] In this embodiment, the authenticity assessment model adopts the ResNet network structure to evaluate the authenticity probability of the graph elements.

[0112] In this embodiment, the third training module is used to train the authenticity assessment model; the third training module includes: a sample generation unit, a training set construction unit and a training unit.

[0113] Among them, the sample generation unit is used to collect authoritative open source knowledge graph data as positive samples, and randomly replace a certain attribute of the positive sample with a false description to generate a negative sample; the training set construction unit is used to mark the authenticity probability of the positive and negative samples, and then construct the training set; the training unit is used to select the graph element z from the training set, and record all the element vectors connected within N hops of the element as Represent the authority of the data source of this element as a weight vector and will As the input of the authenticity assessment model, it is trained using the back-propagation method.

[0114] Based on the above system embodiment, the present invention also provides an embodiment of a method for constructing a criminal investigation knowledge graph based on multi-agent collaboration. The method of this embodiment mainly includes steps S10-S30:

[0115] Step S10: The planning agent generates a first search parameter and a planning prompt word according to user needs, and obtains first task data from the data platform according to the first search parameter.

[0116] Step S20, the analysis intelligent agent generates a structured expression of the target knowledge graph according to the planning prompt word, and initializes the target knowledge graph according to the first task data.

[0117] Step S30, the verification intelligent agent checks whether there is a conflict in the elements in the target knowledge graph, and if there is, selects the element with the highest authenticity probability.

[0118] Further, the present application also provides an embodiment of a computer readable storage device, and the storage device of the embodiment stores a computer program capable of being loaded and executed by a processor to perform the method as described above.

[0119] The computer readable storage device can include a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes.

[0120] Those skilled in the art should be aware that the method steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of both. In order to clearly illustrate the interchangeability of electronic hardware and software, the components and steps of each example have been described in the above description in general terms. Whether the functions are performed by electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0121] So far, the technical solution of the present application has been described in combination with the preferred embodiments shown in the drawings. However, those skilled in the art will readily understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after such changes or replacements will all fall within the protection scope of the present application.

Claims

1. A criminal investigation knowledge graph construction system based on multi-agent collaboration, characterized by: The system comprises: A planning agent, configured to generate a first search parameter and a planning prompt word according to user requirements, and obtain first task data from a data platform according to the first search parameter; an analysis agent, configured to generate a structured representation of a target knowledge graph according to the planning prompt words, and initialize the target knowledge graph according to the first task data; A verification agent is used to check whether there is any conflict between the elements in the target knowledge graph. If so, the element with the highest probability of authenticity is selected; The first training module is used to train the planning agent, the analysis agent, the verification agent, and the line expansion agent. The first training module includes: A pre-training unit, configured to perform pre-training on the planning agent, the analysis agent, the verification agent, and the line expansion agent respectively; A collaborative training unit is configured to collaboratively train the planning agent, the analysis agent, the verification agent, and the line expansion agent after preheating training using the MADDPG algorithm until the quality score r1 of the generated knowledge graph reaches a maximum value; The pre-training unit includes: The first subunit is configured to: generate a knowledge graph based on the planning agent and the analysis agent while the analysis agent is fixed, and train the planning agent using the RLHF process and the PPO algorithm until the quality score r2 of the generated knowledge graph reaches a maximum value; The second subunit is configured to: generate a knowledge graph based on the planning agent and the analysis agent while the planning agent is fixed, and train the analysis agent using the RLHF process and the PPO algorithm until the quality score r3 of the generated knowledge graph reaches a maximum value; The third subunit is configured to: train the verification agent based on the graph training set containing conflicting elements until the recognition rate of the conflicting elements reaches a maximum value; a fourth subunit configured to: train the line-expanding agent based on a graph training set containing expandable entities until the recognition rate of the expandable entities reaches a maximum value; A reward model is used to calculate the quality score r1, the quality score r2, or the quality score r3 of the knowledge graph according to the following formula during training: ； in, represents the quality score r1, the quality score r2 or the quality score r3; represents the precision rate; represents the recall rate; represents a subjective rating; 、 and Both represent weights, which are calculated by the multivariate gradient regression model; The subjective scoring model is used to calculate the subjective score; A second training module is used to train the subjective scoring model; The second training module includes: an initialization unit, configured to set random parameters for the planning agent, the analyzing agent, the verifying agent, and the line expansion agent, respectively; A graph construction unit, configured to construct K graph samples using the planning agent, the analyzing agent, the verifying agent, and the line expansion agent; K is a preset value; The sample generation unit is used to sort the K graph samples according to the scores of the case handlers, and generate a training samples; A training unit, configured to train the subjective rating model based on the training samples and the ranking loss function; in, Each of the training samples is a triple (x, , ), x represents the input data for generating the knowledge graph, and The graphs represent the graphs that received higher and lower scores from case handlers respectively; The ranking loss function is: D stands for A data set consisting of the training samples; represents the expectation of the training samples from the dataset D; (x, ) represents the input (x, ) when the model’s prediction score is (x, ) represents the input (x, ) when the model’s prediction score is σ represents the sigmoid function; an authenticity assessment model for assessing the authenticity probability of the graph elements; A third training module is used to train the authenticity assessment model; The third training module includes: The sample generation unit is used to collect authoritative open source knowledge graph data as positive samples, and randomly A certain attribute is replaced with a false description to generate a negative sample; A training set construction unit, configured to label the authenticity probabilities of the positive samples and the negative samples, thereby constructing a training set; The training unit is used to select a graph element z from the training set, and record all element vectors connected within N hops of the element as a ring , the authority of the data source of this element is represented as a weight vector and will As the input of the authenticity assessment model, back propagation method is used for training.

2. The criminal investigation knowledge graph construction system based on multi-agent collaboration according to claim 1 is characterized in that: The system further comprises: The line expansion agent is used to search whether there is an expandable entity on the leaf node of the target knowledge graph, and if so, generate a line expansion demand and feed it back to the planning agent; The planning agent is further configured to generate a second search parameter and a line expansion prompt word according to the line expansion requirement, and obtain second task data from the data platform according to the second search parameter; The analysis agent is further configured to expand the target knowledge graph according to the line expansion prompt word and the second task data; The verification agent is also used to check whether there are conflicts among the elements in the expanded target knowledge graph. If so, the element with the highest probability of authenticity is selected.

3. A method for constructing a criminal investigation knowledge graph based on multi-agent collaboration, characterized in that: The method comprises: The planning agent generates a first search parameter and a planning prompt word according to the user's needs, and obtains the first task data from the data platform according to the first search parameter; The analysis agent generates a structured expression of a target knowledge graph according to the planning prompt words, and initializes the target knowledge graph according to the first task data; The verification agent checks whether there are conflicts among the elements in the target knowledge graph, and if so, selects the element with the highest probability of authenticity.

4. A computer-readable storage device, characterized in that: A computer program is stored which can be loaded by a processor and execute the method according to claim 3 .

Images