Emergency response knowledge graph reasoning method and system based on dynamic prediction and completion

Abstract

The invention discloses an emergency response mapping knowledge domain reasoning method and system based on dynamic prediction and completion. The method comprises the following steps: summarizing a rule set from an emergency response mapping knowledge domain by a rule-based AnyBURL (Uniform Resource Locator) method; in combination with rule quality and rule length information provided by a rule set, proposing a short-path relationship of an emergency target in a dynamic prediction function guide model attention atlas; performing dynamic prediction according to the current state of the intelligent agent, and constructing a hidden candidate entity set; designing a completion algorithm, and generating an additional action space by using the hidden candidate entity set to realize dynamic completion of the action space of the intelligent agent; according to the method, the deficiency of key relations can be complemented, the efficiency of emergency response knowledge graph reasoning is remarkably improved when underground engineering risks in the fields of hydropower, thermal power and the like occur, and powerful support is provided for rapid and accurate decision making under emergency situations.

Description

Technical Field

[0001] The present invention relates to an emergency response knowledge graph reasoning method and system based on dynamic prediction and completion, belonging to the technical field of knowledge engineering, and is used for emergency response and decision support in sudden situations. Background Art

[0002] In emergency response, knowledge graphs can help integrate and manage a large amount of complex emergency-related information. They can effectively represent and connect multi-dimensional information elements such as emergency events, resources, locations, time, equipment, personnel, and their relationships. However, since emergency events are often unpredictable and information is not fully obtained, the constructed emergency response knowledge graphs are usually sparse and lack complete entity relationship connections. When traditional methods perform reasoning on such graphs, they may face problems such as broken reasoning paths and missing information, which will affect the speed and accuracy of emergency response.

[0003] Rule-based reasoning methods rely on rule bases constructed manually or by expert experience, and reason about knowledge graphs through these predefined rules. This type of method has strong interpretability, and each reasoning step is supported by clear rules, but the rule coverage is not comprehensive, and it is difficult to cope with complex and changing emergency scenarios. Embedding-based reasoning methods map entities and relationships to low-dimensional vector spaces for reasoning. This type of method can capture the latent semantic features of entities and relationships that do not have significant features in explicit rules, but due to the sparsity of emergency knowledge graphs, insufficient training data affects model performance and lacks interpretability.

[0004] The path reasoning model based on reinforcement learning simulates the intelligent agent's path search in the knowledge graph environment, and gives the intelligent agent corresponding rewards according to the action and state transition results, thereby adjusting its behavior strategy and optimizing the reasoning path. However, since the emergency response knowledge graph is usually sparse and the connectivity of the graph is poor, the intelligent agent may not be able to find a reasonable path in the limited information, resulting in a broken reasoning path or an incorrect emergency response decision. Summary of the invention

[0005] Purpose of the invention: In view of the shortcomings of the prior art in the emergency response knowledge graph, the present invention provides an emergency response knowledge graph reasoning method and system based on dynamic prediction and completion. The information on the emergency response knowledge graph is summarized to construct a rule set, and the rule length is introduced to realize the dynamic prediction of the relationship, shorten the average reasoning step length, and help decision makers make quick decisions within a limited time. A dynamic completion algorithm is designed, and the dynamic prediction results are used to realize the dynamic completion of the action space state of the intelligent agent, thereby effectively alleviating the problem of missing paths in the emergency response knowledge graph. This method significantly improves the practical value of the emergency response knowledge graph and provides more reliable intelligent support for emergency management decisions.

[0006] Technical solution: A knowledge graph reasoning method for emergency response based on dynamic prediction and completion, including the following steps:

[0007] (1) Constructing a rule set: The Anytime Bottom-Up Rule Learning (AnyBURL) method is introduced to extract rules from the sparse emergency response knowledge graph and construct a rule set Q.

[0008] (2) Design of dynamic prediction function: Since emergency response often requires making quick decisions within a limited time, the limitation of reasoning step length also becomes a key factor. In order to improve the efficiency of path reasoning and shorten the average reasoning step length, the rule length is additionally introduced into the dynamic prediction calculation method, and a new dynamic prediction function is proposed.

[0009] (3) Obtaining the hidden candidate entity set: Combine the constructed rule set Q and the dynamic prediction function to calculate the probability vector set p of the entity as the correct target entity (arranged in descending order of probability), and select the entity set e with the highest ranking in the probability vector set p p Form a hidden candidate entity set to complete the dynamic prediction of the current state of the agent.

[0010] (4) Dynamic completion: In the event of sudden disasters, the emergency response knowledge graph may not have sufficient emergency response paths. Therefore, a dynamic completion strategy based on prediction information is proposed. During the reasoning process, the action space of the intelligent agent is divided into two different states, and the action space is dynamically completed or replaced respectively.

[0011] Furthermore, the specific steps of constructing a rule set based on the AnyBURL method in step (1) are as follows:

[0012] (1.1) Horn rules are extracted from the graph based on path sampling in a bottom-up manner. The rules are expressed as follows:

[0013] H re (X,Y)←b1(X,A1)∧…∧b n (A n ,Y)

[0014] The obtained Horn rule is represented by H, H re (X,Y) represents the extracted relationship rules, X and Y are the starting and target entities respectively, re is the relationship between entities, A1 and A n Auxiliary entities representing intermediate connections, b n (…) indicates the association between entities. ∧ indicates the logical “and” operation. The extracted rule is true only when all the preconditions are true at the same time.

[0015] (1.2) AnyBURL constructs a rule set based on the existing path relationships in the graph, and evaluates each rule in the rule set based on the inference results in pre-training. The support of rule r indicates how many triples in a given knowledge graph the rule can match in the inference task, and the formula is as follows:

[0016]

[0017] in, Indicates the number of correct triples that satisfy the rule. Because each rule selected in the pre-training inference prediction is probabilistic, the confidence of the rule is evaluated and determined by calculating the ratio of correct facts of the rule prediction head node. The confidence of rule r is expressed as follows:

[0018]

[0019] in, Indicates the number of all triples matched by the rule.

[0020] Furthermore, the specific steps of designing the dynamic prediction function in step (2) are as follows: whether an entity node is a high-quality hidden candidate entity is not just whether it constitutes a highly credible rule. In order to respond to emergencies in a limited time, the limitation of the reasoning step length also becomes a key factor. Therefore, it is necessary to consider whether the entity can reach the correct entity through a shorter reasoning path. First, the evaluation index provided by the rule set Q generated in step (1) is calculated, and then normalized. In order to improve the efficiency of path reasoning and shorten the average reasoning step length, the rule length is additionally introduced in the dynamic prediction calculation method to evaluate the entity e. i The probability vector p i The calculation formula is as follows:

[0021]

[0022] Among them, p i Represents the value of the i-th dimension of vector p, representing entity e i is the probability of the correct tail entity, for the candidate tail entity e i The rules of i The confidence of the rule set Q is defined as conf(r i ), support is defined as support(r i ). l(r) represents the length of rule r.

[0023] Furthermore, the specific steps of obtaining the hidden candidate entity set in step (3) are as follows:

[0024] (3.1) Using the rule set Q generated in step (1) and the dynamic prediction function designed in step (2), calculate the probability vector set p of all entities as tail entities (correct target entities). Since the dynamic prediction function takes the rule length into consideration when calculating the prediction probability, shorter rules are prioritized in the probability vector set. This design encourages and constrains the model to focus more on rules with short paths during dynamic prediction. Short path rules usually have better generalization capabilities because they can quickly link to key entities (such as spillways, pumping stations, and emergency equipment in dam emergency response), ensuring that the system can make accurate decisions in the shortest time.

[0025] (3.2) Select the top-ranked entity set e in the vector set p p As a hidden candidate entity set, dynamic prediction of the current state of the agent is completed.

[0026] Furthermore, the specific steps of performing dynamic completion in step (4) are as follows:

[0027] (4.1) The threshold parameter M is introduced as the classification threshold of the agent action space in the knowledge reasoning task, which indicates the minimum action space on which the agent can complete reasoning. In addition, in order to improve the quality of the relationship in the action space, the agent can reach the target entity in fewer steps. The optimization parameter k is introduced, which indicates that the relationship in the original action space is optimized using the first k hidden candidate entities in the dynamic prediction results.

[0028] (4.2) If at time t, the action space is less than M, it means that the agent has few relationships to choose from, that is, there are not enough paths in the emergency response knowledge graph to complete the reasoning. To complete the action space, replace N in the probability vector set p calculated in step (3) add hidden candidate entities and their corresponding relations to generate additional action space Add to the current action space to increase the action space size to M. Additional action space size N add The formula is as follows:

[0029] N add =MN

[0030] Where N represents the size of the current action space.

[0031] In this case, the original action space is combined with the additional action space generated by dynamic prediction Merge to get a new action space The formula is as follows:

[0032]

[0033] In the knowledge reasoning task based on reinforcement learning, the environment usually refers to the entire knowledge graph, and the state of the tth time step is defined as a tuple (e t , e s , r q ), e t represents the entity accessed in the emergency response knowledge graph at step t, e t is the source entity, e s The target entity, r q is the query relation. The action space set A at time t t By entity t All connected relationships are composed of: A t ={(r,e)|(e t ,r,e)∈G}, e represents the source entity e in the knowledge graph t And all the connection relationships r it has correspond to the connected entities.

[0034] (4.3) If at time t, the action space is greater than or equal to M, it means that the current action space of the agent is relatively rich, and the emergency response knowledge graph needs short path relationships to achieve action space optimization. To this end, the k hidden candidate entities and their corresponding relationships in the probability vector set p are used to construct an additional action space The size of the additional action space N add The definition is as follows:

[0035] N add =k

[0036] In order to improve the quality of the action space without blindly expanding the size of the original action space, the merged action space is sorted according to the dynamic prediction score in the probability vector set p, and the first N actions are taken as the new action space, thereby keeping the size of the original action space unchanged. The new action space is expressed as follows:

[0037]

[0038] An emergency response knowledge graph reasoning system based on dynamic prediction and completion, including the following modules:

[0039] (1) Rule construction module: The rule set Q is constructed from the sparse emergency response knowledge graph in a bottom-up manner through the rule-based AnyBURL method;

[0040] (2) Dynamic prediction module: Since emergency response often requires making quick decisions within a limited time, whether an entity node is a high-quality hidden candidate entity is not only determined by whether it constitutes a highly credible rule, but also by the limitation of the inference step length. In order to improve the efficiency of path inference and shorten the average inference step length, the rule length is additionally introduced into the dynamic prediction calculation method, and a new dynamic prediction function is proposed.

[0041] (3) Hidden candidate entity set acquisition module: Combine the constructed rule set Q and the dynamic prediction function to calculate the probability vector set p, and select the entity set e with the highest ranking in the vector set p. p Form a hidden candidate entity set to complete the dynamic prediction of the current state of the agent;

[0042] (4) Dynamic completion module: In the event of sudden disasters, there may not be enough emergency response paths (relationships) in the emergency response knowledge graph. Therefore, a dynamic completion strategy based on prediction information is proposed. During the reasoning process, the action space of the intelligent agent is divided into two different states, and the action space is dynamically completed or replaced respectively.

[0043] The implementation process of the system and method is the same and will not be repeated here.

[0044] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the emergency response knowledge graph reasoning method based on dynamic prediction and completion as described above is implemented.

[0045] A computer-readable storage medium stores a computer program for executing the emergency response knowledge graph reasoning method based on dynamic prediction and completion as described above.

[0046] Beneficial effects: The present invention introduces considerations of dynamic prediction functions and rule lengths, and can give priority to short-path rules, thereby significantly improving reasoning efficiency and ensuring faster decision-making in emergency situations. In the emergency response knowledge graph, some emergency measures or equipment information may be missing due to insufficient data, resulting in a sparse graph. The present invention dynamically predicts the potential relationships of the current entity through dynamic completion and replacement strategies, thereby generating additional action spaces, alleviating the problem of missing paths on sparse graphs, and adjusting the decision space during the reasoning process, avoiding fixed path searches, and allowing the system to flexibly adjust decisions based on the current emergency status and information already available in the graph. Dynamic prediction and dynamic completion alleviate the problems of sparsity, lengthy paths, and missing information in the emergency response knowledge graph, improving the efficiency, accuracy, and flexibility of the emergency response knowledge graph reasoning system, and providing strong support for responding to sudden disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0048] Figure 2 This is an overall framework diagram of emergency response knowledge graph reasoning based on dynamic prediction and completion in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0050] Step (1) Construct a rule set: Use the rule-based AnyBURL method to construct a rule set Q from a sparse emergency response knowledge graph (taking the risk emergency response knowledge graph of underground hydropower and thermal power projects as an example).

[0051] (1.1) Horn rules are extracted from the graph based on path sampling in a bottom-up manner. The rules are expressed as follows:

[0052] H re (X,Y)←b1(X,A1)∧...∧b n (A n ,Y)

[0053] The obtained Horn rule is represented by H, X, Y, A n Represent different entities respectively. The entity association in the rule is represented by b n (...) indicates that. For the constructed Horn rule H re (e i ,e j ), which is equivalent to the fact triple in the emergency response knowledge graph (e i ,re,e j ). Some of the rules included in the constructed rule set Q are shown in Table 1;

[0054] Table 1 Some rules included in the rule set Q

[0055]

[0056] (1.2) AnyBURL constructs a rule set based on the existing path relationships in the graph, and evaluates each rule in the rule set based on the inference results in pre-training. The support of rule r indicates how many triples in a given knowledge graph the rule can match in the inference task, and the formula is as follows:

[0057]

[0058] in, Indicates the number of correct triples that satisfy the rule. Because each rule selected in the pre-training inference prediction is probabilistic, the confidence of the rule is evaluated and determined by calculating the ratio of correct facts of the rule prediction head node. The confidence of rule r is expressed as follows:

[0059]

[0060] in, Indicates the number of all triples matched by the rule.

[0061] Step (2) Design a dynamic prediction function: In the emergency response knowledge graph, since emergency response often requires making quick decisions within a limited time, the limitation of the reasoning step length also becomes a key factor. To this end, it is necessary to ensure that the entity can reach the final emergency response entity through a shorter reasoning path (such as activating the early warning system, reinforcing the surrounding rock support structure, restricting the working area, activating the backup ventilation system, strengthening gas monitoring and extraction, suspending construction and evacuating personnel, etc.). Figure 2 As shown in Figure 1, the evaluation index provided by the rule set Q generated in step (1) is first calculated and then normalized. In order to improve the efficiency of path reasoning and shorten the average reasoning step length, the rule length is additionally introduced in the dynamic prediction calculation method to calculate the entity e. i The probability vector p i The calculation formula is as follows:

[0062]

[0063] Among them, p i Represents the value of the i-th dimension of vector p, representing entity e i is the probability of the correct tail entity, for the candidate tail entity e i The rules of i The confidence of the rule set Q is defined as conf(r i ), support is defined as support(r i ). l(r) represents the length of rule r.

[0064] Step (3) Obtain the hidden candidate entity set: Combine the constructed rule set Q and the dynamic prediction function to calculate the probability vector set p, and select the entity set e with the highest ranking in the vector set p. p The hidden candidate entity set is composed of Figure 2 shown.

[0065] (3.1) Using the rule set Q generated in step (1) and the dynamic prediction function designed in step (2), calculate the probability vector set p of all entities as the tail entity (correct target entity). Assuming that the current gas concentration exceeds the safety value, according to the rule set Q, the system calculates the probability of the corresponding entity being the tail entity, as shown in Table 2:

[0066] Table 2 The probability of each entity being the correct target entity

[0067]

[0068]

[0069] (3.2) Select the entity set e that ranks top in the vector set p p As hidden candidate entity sets, dynamic prediction of the current state of the agent is completed. The current instance selects entities with higher probability in Table 2 as hidden candidate entity sets, namely "backup ventilation system" and "gas monitoring and extraction enhancement".

[0070] Step (4) Dynamic completion: In order to cope with the situation where the sudden disaster map may not have enough emergency response paths, a dynamic completion strategy based on prediction information is proposed. During the reasoning process, the action space of the intelligent agent is divided into two different states, and the action space is dynamically completed or replaced respectively, such as Figure 2 As shown;

[0071] (4.1) The threshold parameter M = 3 is introduced as the classification threshold of the agent action space in the knowledge reasoning task, indicating the minimum action space on which the agent can complete reasoning. In addition, in order to improve the quality of the relationship in the action space, the agent can reach the target entity in fewer steps. The optimization parameter k = 2 is introduced, indicating that the relationship in the original action space is optimized using the first two hidden candidate entities in the dynamic prediction results;

[0072] (4.2) Assume that at time t, the entity being visited is “gas concentration detector”, and the corresponding associated action space N = 2, which is less than M, indicating that the agent has fewer relationships to choose from and is insufficient to complete the reasoning. To complete the action space, replace N in the probability vector set p calculated in step (3) with add =1(N add =MN) Hide candidate entities and their corresponding relations to generate additional action space Add the current action space to increase the size of the action space to M.

[0073] In this case, the original action space is combined with the additional action space generated by dynamic prediction Merge to get a new action space The formula is as follows:

[0074]

[0075] Among them, the action space set A at time t t By entity t All connected relationships are composed of: A t ={(r,e)|(e t ,r,e)∈G}. The original action space and the completed action space are merged to obtain the expanded action space: {(trigger, alarm system), (trigger, evacuate personnel), (start, local ventilator)};

[0076] (4.3) Assume that at time t, the action space N = 4 of the currently visited entity "vent" is greater than M, which means that the current action space of the agent is relatively rich, and the model needs short path relationships to achieve action space optimization. To this end, the k hidden candidate entities and their corresponding relationships in the probability vector set p are used to construct an additional action space The size of the additional action space N add The definition is as follows:

[0077] N add =k

[0078] In order to improve the quality of the action space without blindly expanding the size of the original action space, the merged action space is sorted according to the dynamic prediction score in the probability vector set p. The first four actions are used as the new action space, so as to keep the size of the original action space unchanged. In this example, the new action space is: {(trigger, alarm system), (trigger, personnel evacuation), (activate, local ventilator), (notify, ventilation system adjust air volume)}.

[0079] The emergency response knowledge graph reasoning system based on dynamic prediction and completion includes the following modules:

[0080] (1) Rule construction module: Using the rule-based AnyBURL method, starting from the sparse emergency response knowledge graph, a bottom-up strategy is adopted to construct the rule set Q;

[0081] (2) Dynamic prediction module: Given that emergency response requires making decisions in a short time, when evaluating whether an entity node is a high-quality hidden candidate entity, not only whether it can form a credible rule but also the efficiency of the reasoning step should be considered. In order to improve the path reasoning speed and shorten the reasoning time, the rule length factor is introduced, and a new dynamic prediction function is proposed to optimize the path reasoning process.

[0082] (3) Hidden candidate entity set acquisition module: Based on the constructed rule set Q and dynamic prediction function, the probability vector set p of the entity is calculated, and the top-ranked entity set e is selected from it. p Form a hidden candidate entity set to complete the dynamic prediction of the current state of the agent;

[0083] (4) Dynamic completion module: When responding to sudden disasters, there may be a lack of sufficient emergency paths in the emergency response knowledge graph. To solve this problem, a dynamic completion strategy based on prediction information is proposed. During the reasoning process, the action space of the agent is divided into two states, and the action space is completed or replaced respectively to ensure the integrity and effectiveness of the reasoning process.

[0084] Obviously, those skilled in the art should understand that the various steps of the emergency response knowledge graph reasoning method based on dynamic prediction and completion or the various modules of the emergency response knowledge graph reasoning system based on dynamic prediction and completion of the above-mentioned embodiments of the present invention can be implemented by a general-purpose computing device, which can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented with executable program codes of computing devices, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules respectively, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A knowledge graph reasoning method for emergency response based on dynamic prediction and completion, characterized in that: The following steps are involved: (1) Constructing a rule set: Introducing a bottom-up rule learning method at any time to extract rules from the sparse emergency response knowledge graph and construct a rule set Q; (2) Design of dynamic prediction function: In the dynamic prediction calculation method, the rule length is additionally introduced to optimize the path reasoning and a new dynamic prediction function is proposed; (3) Obtaining the hidden candidate entity set: Combine the constructed rule set Q and the dynamic prediction function to calculate the probability vector set p of the correct target entity. The probability vector set p is sorted in descending order of probability, and the entity set e ranked in the top m in the probability vector set p is selected. p Form a hidden candidate entity set to complete the dynamic prediction of the current state of the agent; (4) Perform dynamic completion: A dynamic completion strategy based on prediction information is proposed. During the reasoning process, the action space of the intelligent agent is divided into two different states, and dynamic completion or replacement of the action space is performed respectively.

2. The emergency response knowledge graph reasoning method based on dynamic prediction and completion according to claim 1 is characterized in that: The specific steps of constructing a rule set based on the AnyBURL method in step (1) are as follows: (1.1) Horn rules are extracted from the graph based on path sampling in a bottom-up manner. The rules are expressed as follows: H re (X,Y)←b1(X,A1)∧…∧b n (A n ,Y) The obtained Horn rule is represented by H, H re (X,Y) represents the extracted relationship rules, X and Y are the starting and target entities respectively, re is the relationship between entities, A1 and A n Auxiliary entities representing intermediate connections, b n (…) indicates the association between entities; ∧ indicates the logical "and" operation, and the extracted rule is true only when all the preconditions are true at the same time; (1.2) AnyBURL constructs a rule set based on the existing path relationships in the graph and evaluates each rule in the rule set based on the inference results in pre-training. The support of rule r indicates how many triples in a given knowledge graph the rule can match in the inference task. The formula is as follows: in, Represents the number of correct triples that satisfy the rule; the confidence of the rule is evaluated and determined by calculating the ratio of correct facts of the rule prediction head node; the confidence of rule r is expressed as follows: in, Indicates the number of all triples matched by the rule.

3. The emergency response knowledge graph reasoning method based on dynamic prediction and completion according to claim 1 is characterized in that: The specific steps of designing the dynamic prediction function in step (2) are as follows: first, the evaluation index provided by the rule set Q generated in step (1) is calculated, and then normalized. The rule length is additionally introduced into the dynamic prediction calculation method. i The probability vector p i The calculation formula is as follows: Among them, p i Represents the value of the i-th dimension of vector p, representing entity e i is the probability of the correct tail entity, for the candidate tail entity e i The rules of i The confidence of the rule set Q is defined as conf(r i ), support is defined as support(r i ), l(r) represents the length of rule r.

4. The emergency response knowledge graph reasoning method based on dynamic prediction and completion according to claim 1 is characterized in that: The specific steps of obtaining the hidden candidate entity set in step (3) are as follows: (3.1) Using the rule set Q generated in step (1) and the dynamic prediction function designed in step (2), calculate the probability vector set p of all entities as tail entities; (3.2) Select the top-ranked entity set e in the vector set p p As a hidden candidate entity set, dynamic prediction of the current state of the agent is completed.

5. The emergency response knowledge graph reasoning method based on dynamic prediction and completion according to claim 1 is characterized in that: The specific steps of performing dynamic completion in step (4) are as follows: (4.1) The threshold parameter M is introduced as the classification threshold of the agent action space in the knowledge reasoning task; the optimization parameter k is introduced, which means that the relationship of the original action space is optimized using the first k hidden candidate entities in the dynamic prediction results; (4.2) If at time t, the action space is less than M, it means that there are not enough paths in the emergency response knowledge graph. To complete the action space, replace N in the probability vector set p calculated in step (3) with add hidden candidate entities and their corresponding relations to generate additional action space Add to the current action space to increase the action space size to M; additional action space N add The formula for size is as follows: N add =M-N Among them, N represents the size of the current action space; In this case, the original action space is combined with the additional action space generated by dynamic prediction Merge to get a new action space The formula is as follows: Here, S represents the state space in the reasoning process, and the state of the tth time step is defined as a tuple S t =(e t , e s , r q ), e t represents the entity accessed in the emergency response knowledge graph at step t, e s is the source entity, r q is the query relation; the action space set A at time t t By entity t All connected relationships are composed of: A t ={(r,e)|(e t ,r,e)∈G}; (4.3) If at time t, the action space is greater than or equal to M, the emergency response knowledge graph needs a short path relationship to achieve action space optimization; use the k hidden candidate entities and their corresponding relationships in the probability vector set p to construct an additional action space The size of the additional action space N add The definition is as follows: N add =k The merged action space is sorted according to the dynamic prediction scores in the probability vector set p, and the first N actions are taken as the new action space; the new action space is expressed as follows:

6. An emergency response knowledge graph reasoning system based on dynamic prediction and completion, characterized in that: Includes the following modules: (1) Rule construction module: The rule set Q is constructed from the sparse emergency response knowledge graph in a bottom-up manner through the rule-based AnyBURL method; (2) Dynamic prediction module: Introducing additional rule length into the dynamic prediction calculation method and proposing a new dynamic prediction function; (3) Hidden candidate entity set acquisition module: Combine the constructed rule set Q and the dynamic prediction function to calculate the probability vector set p, and select the entity set e with the highest ranking in the vector set p. p Form a hidden candidate entity set to complete the dynamic prediction of the current state of the agent; (4) Dynamic completion module: Based on the dynamic completion strategy of prediction information, the action space of the intelligent agent is divided into two different states during the reasoning process, and the action space is dynamically completed or replaced respectively.

7. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the emergency response knowledge graph reasoning method based on dynamic prediction and completion are implemented as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program for executing the emergency response knowledge graph reasoning method based on dynamic prediction and completion as described in any one of claims 1-5.

Images