Target intention recognition method, system and electronic equipment

Abstract

The present invention provides a target intent recognition method, system, and electronic device, comprising: constructing a spatiotemporal knowledge graph containing target and intent scenarios based on historical intents between different types of targets; performing node-level feature extraction on targets at a time granularity and a spatial granularity based on the spatiotemporal knowledge graph to obtain feature information of paths that carry relatively more information about the target intent to be recognized; performing relationship-level feature extraction on the targets based on the spatiotemporal knowledge graph to obtain the target intent tendency vector based on the inherent characteristics of the spatiotemporal knowledge graph; inputting the feature information and the intent tendency vector into a target intent recognition fusion model to obtain the intent of the target to be recognized; the model is used to fuse the feature information and the intent tendency vector at a time granularity and a spatial granularity to obtain a single-dimensional intent recognition result, and to fuse the intent recognition results of different dimensions to obtain a target intent recognition result. The present invention improves the reliability of target intent recognition results.

Description

Technical Field

[0001] The present invention belongs to the field of target intention recognition, and more specifically, relates to a target intention recognition method, system and electronic equipment. Background Art

[0002] Research on target intention recognition began relatively early abroad, and significant progress has been made in both theoretical frameworks and system development. Typical and notable research achievements in target intention recognition in China include: Yao Chunyan of the National University of Defense Technology's situation estimation temporal reasoning method, which proposed inference methods such as fuzzy time, maximum a posteriori probability, and hypothesis testing statistics; Dr. Wu Ji of Xidian University's planning recognition situation estimation model, which utilizes a two-level planning recognition model and path planning recognition for intention estimation; and the plan recognition model developed by Kang Yaohong et al. of Xidian University. These system models partially implement certain functions of intention estimation. From knowledge-based systems to logical template matching and multi-agent plan recognition methods, researchers have gradually realized that the characterization of various information, including factors such as the target's environment, directly affects the final results of situation prediction and intention estimation.

[0003] In summary, the existing methods ignore the ambiguity and randomness of the target's environment and the subjectivity of the aerial target operator, and the target intention recognition results obtained are unreliable. Summary of the Invention

[0004] In view of the defects of the prior art, the purpose of the present invention is to provide a target intention recognition method, system and electronic equipment, aiming to solve the problem of poor reliability of the existing target intention recognition method.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for identifying target intentions, comprising the following steps:

[0006] A spatiotemporal knowledge graph containing targets and intent scenarios is constructed based on the historical intents between targets of different types; the nodes in the spatiotemporal knowledge graph are entities corresponding to targets, and edges represent the existence of intents between two targets. A target has at least one node in the spatiotemporal knowledge graph, and a node corresponds to the encoding of a target at a time granularity and a spatial granularity. Each target has encodings at multiple time granularities and / or multiple spatial granularities, and each target belongs to a target type.

[0007] Based on the spatiotemporal knowledge graph, node-level features of the target to be identified at a time granularity and a space granularity are extracted to obtain feature information of a path that carries relatively more information about the intention of the target to be identified;

[0008] Extracting relational features of the target to be identified based on the spatiotemporal knowledge graph to obtain an intention tendency vector of the target to be identified according to the inherent characteristics of the spatiotemporal knowledge graph;

[0009] The feature information and intention tendency vector are input into a pre-trained target intention recognition fusion model to obtain the intention of the target to be identified; the target intention recognition fusion model is used to fuse the feature information and intention tendency vector at a time granularity and a spatial granularity to obtain the intention recognition result of the target to be identified at a time granularity and a spatial granularity, and fuse the intention recognition results of the target to be identified at different combinations of different time granularities and different spatial granularities to obtain intention recognition fusion results of different dimensions, and then determine the intention of the target to be identified based on the intention recognition fusion results of different dimensions.

[0010] In a possible implementation, the node-level feature extraction is specifically as follows:

[0011] In the spatiotemporal knowledge graph, a directed graph is constructed starting from the target to be identified, and all paths between the target to be identified and the opposite target that do not contain loops are obtained; wherein the intention of the target to be identified is the intention of the target to be identified relative to the opposite target;

[0012] Selecting multiple paths from all paths whose path lengths are less than a preset value and whose path characteristic values ​​are relatively high in descending order; the multiple paths are paths that carry a relatively large amount of information about the target intention to be identified;

[0013] Inputting the multiple paths into a long short-term memory network (LSTM) respectively to obtain a feature vector for each path;

[0014] The feature information is obtained by concatenating the feature vectors of each path.

[0015] In a possible implementation, the path characteristic value is determined by the following steps:

[0016] Determine the relationship between each node on the path and the two adjacent nodes;

[0017] Starting from the head of the path, each additional node generates a subpath, and multiple subpaths are obtained until the tail node is reached;

[0018] Determine the number of relationships for each node in the path in the spatiotemporal knowledge graph; the number of relationships is the total number of intents corresponding to all single edges with the node as the head node, where the total number of intents is obtained by adding the number of edges corresponding to each intent, and the number of edges for each intent divided by the number of relationships is the weight of each intent in the number of relationships for the node;

[0019] The weight of the intention type corresponding to the edge between the previous node and the next node in the path in the number of relationships of the previous node is used as the intention weight of the sub-path with the next node as the tail node compared to the previous node;

[0020] Determine the additional nodes of the next sub-path compared to the previous sub-path, multiply the path characteristic value of the previous sub-path by the intention weight of the added node to obtain the path characteristic value of the next sub-path, and solve the path characteristic value of each sub-path in turn to obtain the path characteristic value of the path from the head node to the tail node; the path characteristic value of the sub-path corresponding to the head node in the path is a preset value.

[0021] In a possible implementation, the relationship-level feature extraction is specifically as follows:

[0022] Determine the nodes corresponding to the target to be identified and the opposite target in the spatiotemporal knowledge graph;

[0023] The tendency value of the target to be identified under each intention is determined by the tendency function based on translation embedding and the intention between the corresponding node of the target to be identified and the corresponding node of the opposite target, and the tendency values ​​under all intentions are spliced ​​to obtain the intention tendency vector of the target to be identified.

[0024] It should be noted that the inherent characteristics of the knowledge graph can be understood as follows: if there is a certain relationship between A and B, that is, there is an edge, then the following equation will be satisfied: A's projection vector + relationship representation vector = B's projection vector. If there is no relationship between A and B, the above equation is not satisfied.

[0025] In addition, precisely because A and B have the above equation relationship when they are associated, the corresponding knowledge graph can be constructed based on the above equation.

[0026] In one possible implementation, the tendency value for each intention is determined by the following steps:

[0027] Convert the encoding of the target node h to the corresponding hyperplane projection vector h ⊥ , convert the encoding of the opposite target node t into the projection vector t corresponding to the hyperplane ⊥ ;

[0028] Convert the encoding of the intention r into the representation vector d corresponding to the hyperplane r , then the tendency value of the entity pair (h, t) for intention r is:

[0029] In a possible implementation, the target intent recognition fusion model includes: a feature fusion module and an intent multi-classifier;

[0030] The feature fusion module is used to fuse the feature information and intention tendency vector at a time granularity and a space granularity to obtain the intention recognition result of the target to be identified at a time granularity and a space granularity;

[0031] The intention multi-classifier is used to classify the intention of the target to be identified based on the fusion results of intention recognition in different dimensions, determine the probability of each preset intention, and take the preset intention with the largest probability value as the intention of the target to be identified.

[0032] In a second aspect, the present invention provides a target intention recognition system, comprising:

[0033] A knowledge graph construction unit is configured to construct a spatiotemporal knowledge graph containing targets and intent scenarios based on historical intents between targets of different types; nodes in the spatiotemporal knowledge graph are entities corresponding to targets, edges represent the existence of intents between two targets, a target has at least one node in the spatiotemporal knowledge graph, a node corresponds to the encoding of a target at a time granularity and a spatial granularity, each target has encodings at multiple time granularities and / or multiple spatial granularities, and each target belongs to a target type;

[0034] a feature extraction unit configured to perform node-level feature extraction on a target to be identified at a time granularity and a spatial granularity based on the spatiotemporal knowledge graph to obtain feature information of a path that carries relatively more information about the target's intention; and to perform relationship-level feature extraction on the target to be identified based on the spatiotemporal knowledge graph to obtain an intention tendency vector of the target to be identified based on inherent features of the spatiotemporal knowledge graph;

[0035] The target intention recognition unit is used to input the feature information and intention tendency vector into a pre-trained target intention recognition fusion model to obtain the intention of the target to be identified; the target intention recognition fusion model is used to fuse the feature information and intention tendency vector at a time granularity and a spatial granularity to obtain the intention recognition result of the target to be identified at a time granularity and a spatial granularity, and to fuse the intention recognition results of the target to be identified at different combinations of different time granularities and different spatial granularities to obtain intention recognition fusion results of different dimensions, and then determine the intention of the target to be identified based on the intention recognition fusion results of different dimensions.

[0036] In one possible implementation, the feature extraction unit constructs a directed graph in the spatiotemporal knowledge graph with the target to be identified as the starting point, and obtains all paths between the target to be identified and the opposite target that do not contain loops; wherein the intention of the target to be identified is the intention of the target to be identified relative to the opposite target; selects multiple paths from all paths whose path lengths are less than a preset value and whose path feature values ​​are relatively high in descending order; the multiple paths are paths that carry a relatively large amount of information about the intention of the target to be identified; the multiple paths are respectively input into the long short-term memory network LSTM to obtain a feature vector for each path; and the feature vectors of each path are spliced ​​to obtain the feature information.

[0037] In one possible embodiment, the feature extraction unit determines the relationship between each node on the path and the two adjacent nodes; starting from the head of the path, each additional node obtains a sub-path, and multiple sub-paths are obtained until the tail node is increased; the number of relationships of each node in the path in the spatiotemporal knowledge graph is determined; the number of relationships is the total number of intentions corresponding to all single edges with the node as the head node, wherein the total number of intentions is obtained by adding the number of edges corresponding to each intention, and the number of edges for each intention divided by the number of relationships is the weight of each intention in the number of relationships of the node; the weight of the intention type corresponding to the edge between the previous node and the next node in the path in the number of relationships of the previous node is used as the intention weight of the sub-path of the next node compared with the previous node as the tail node; and the added node of the next sub-path is determined compared with the previous sub-path, the path feature value of the previous sub-path is multiplied by the intention weight of the added node to obtain the path feature value of the next sub-path, and the path feature value of each sub-path is solved in turn to obtain the path feature value of the path composed of the head node to the tail node; the path feature value of the sub-path corresponding to the head node in the path is a preset value.

[0038] In a third aspect, the present invention provides an electronic device comprising: at least one memory for storing programs; and at least one processor for executing the programs stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation of the first aspect.

[0039] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program. When the computer program runs on a processor, the processor executes the method described in the first aspect or any possible implementation of the first aspect.

[0040] In a fifth aspect, the present invention provides a computer program product, which, when executed on a processor, enables the processor to execute the method described in the first aspect or any possible implementation of the first aspect.

[0041] In general, the above technical solutions conceived by the present invention have the following beneficial effects compared with the prior art:

[0042] The present invention provides a target intention recognition method, system and electronic device. Based on spatiotemporal coding, a spatiotemporal knowledge graph is constructed, and an uncertain target is represented by the arrangement and combination of determined spatiotemporal nodes, so as to take into account the ambiguity and randomness of the target's environment and the subjectivity of the aerial target operator, simplifying the complexity of target intention recognition and strengthening the control of the target's environmental situation. The present invention proposes an experience-based path selection method, which makes the feature path selected in the path sorting algorithm more biased, thereby making the calculation result more in line with reality and improving the accuracy of target intention recognition. The present invention proposes a target intention recognition method, which uses a feature extraction module in the sub-algorithm to extract multi-level and multi-angle feature information, and finally inputs it into the feature fusion module. Compared with existing methods, the technical solution proposed by the present invention considers more comprehensive angles and extracts more comprehensive features, and the target intention recognition result is more accurate, which improves the reliability of the target intention recognition result and strengthens the analysis and prediction of the target's environmental situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of a target intention recognition method provided by an embodiment of the present invention;

[0044] Figure 2 This is an overall flow chart of the intent recognition method provided by an embodiment of the present invention;

[0045] Figure 3 is a schematic diagram of a path selection algorithm provided by an embodiment of the present invention;

[0046] Figure 4 This is a block diagram of a single-level intent recognition model provided by an embodiment of the present invention;

[0047] Figure 5 This is a block diagram of the intent recognition fusion model provided by an embodiment of the present invention;

[0048] Figure 6 This is a diagram of the target intent recognition system architecture provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0050] The term "and / or" as used herein describes an association between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. The symbol " / " as used herein indicates that the related objects are in an "or" relationship, for example, A / B means either A or B.

[0051] In the embodiments of the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present invention should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0052] In the description of the embodiments of the present invention, unless otherwise specified, “plurality” means two or more.

[0053] First, the technical terms involved in the embodiments of the present invention are introduced.

[0054] (1) Knowledge Graph

[0055] A knowledge graph is a graph-based data structure consisting of nodes and edges. Each node represents an "entity" and each edge is a "relationship" between entities. A knowledge graph is essentially a semantic network.

[0056] (2) Entity

[0057] Entities can refer to things in the real world, such as people, places, companies, telephones, animals, etc.; relationships are used to express a certain connection between different entities.

[0058] Next, the technical solutions provided in the embodiments of the present invention are introduced.

[0059] Target intent recognition is a crucial component of assessing the target's environment. Rapidly and accurately identifying the target's intentions allows us to promptly follow up on them and respond appropriately. Intelligent intent recognition methods, such as deep learning and knowledge graph reasoning, have made some progress. However, issues such as incomplete datasets and inaccurate characterization of the target's environment severely hinder the application of intent recognition technology in various fields.

[0060] Figure 1 is a flow chart of a target intention recognition method provided by an embodiment of the present invention; Figure 1 As shown, the following steps are included:

[0061] S101: Construct a spatiotemporal knowledge graph containing targets and intent scenarios based on historical intents between targets of different types; nodes in the spatiotemporal knowledge graph are entities corresponding to targets, edges represent the existence of intents between two targets, a target has at least one node in the spatiotemporal knowledge graph, a node corresponds to an encoding of a target at a time granularity and a spatial granularity, each target has encodings at multiple time granularities and / or multiple spatial granularities, and each target belongs to a target type;

[0062] S102, performing node-level feature extraction on the target to be identified at a time granularity and a space granularity based on the spatiotemporal knowledge graph to obtain feature information of a path that carries relatively more information about the target's intention;

[0063] S103, performing relational feature extraction on the target to be identified based on the spatiotemporal knowledge graph, so as to obtain an intention tendency vector of the target to be identified according to the inherent characteristics of the spatiotemporal knowledge graph;

[0064] S104, input the feature information and intention tendency vector into a pre-trained target intention recognition fusion model to obtain the intention of the target to be identified; the target intention recognition fusion model is used to fuse the feature information and intention tendency vector at a time granularity and a spatial granularity to obtain the intention recognition result of the target to be identified at a time granularity and a spatial granularity, and fuse the intention recognition results of the target to be identified at different combinations of different time granularities and different spatial granularities to obtain intention recognition fusion results of different dimensions, and then determine the intention of the target to be identified based on the intention recognition fusion results of different dimensions.

[0065] Figure 2 : is a general flow chart of a specific intent recognition method provided by an embodiment of the present invention; illustratively, it includes the following steps:

[0066] S1. Construct sample data: Based on a specific entity coding method, map the entities in the target environment into entity coding vectors composed of spatial coding, temporal coding, and type coding.

[0067] S2. Split sample code: Split the entity code of the sample training set and perform the operations of steps S3 and S4 respectively.

[0068] Specifically, the above splitting refers to splitting the same target at different time granularities and different spatial granularities and performing feature extraction on each of them.

[0069] S3. Node-level feature extraction: Using the path sorting algorithm, the connection paths between nodes are used as features to determine the possible relationships between entity pairs.

[0070] S4. Relationship-level feature extraction: Determine the most likely relationship between entity pairs through a tendency function based on translation embedding.

[0071] S5. Single-level feature fusion: The features extracted from the above two levels are fused through a neural network to obtain single-dimensional intent feature information.

[0072] S6. Overall feature fusion: The intention feature information of multiple dimensions is fused through the intention recognition fusion model to obtain a more accurate enemy target intention recognition result.

[0073] This paper proposes a method for identifying target intent, analyzing and determining the intent relationships between targets at both the entity and relationship levels. At the entity level, an empirically based path selection method is proposed based on the data characteristics of the target's environment, improving the path sorting algorithm. At the relationship level, an embedding-based propensity function is proposed using the embedding principle. Finally, a feature fusion module combines the judgment results from these two levels to obtain more comprehensive and accurate intent identification results.

[0074] It is understood that those skilled in the art can apply the above-mentioned target intent recognition method to various fields according to actual needs, such as teaching, maritime communication, drone confrontation, AI, etc. When the method is applied to different fields, technicians in the corresponding technical fields can preset the classification of specific intents according to actual needs so as to use the target intent recognition method provided by the present invention to identify the corresponding intent. The present invention does not impose any restrictions on the types of specific intents or the total number of categories.

[0075] In this example, the relationships in the knowledge graph are intents, including platform-to-platform intent, platform-to-group intent, group-to-platform intent, and group-to-group intent. Intents, like platform models, are encoded in decimal form, starting at 0 and incrementing by 1, in the order of their first appearance.

[0076] In this example, the full set of intents can be: attack, reconnaissance, defense, support, and unknown. The full set of targets refers to the set of all targets that appear in the sample training set.

[0077] S1. Construct sample data: According to the following entity coding method, the target entity at sea is mapped into an entity coding vector composed of space coding, time coding and type coding. The simulation system generates the simulation scene at sea and the real intention information. Figure 3 Tuple batches must be no less than 10,000.

[0078] S2. Split sample code: Split the entity code of the sample training set into separate training sets and perform the operations of steps S3 and S4 respectively.

[0079] S3. Node-level feature extraction: Using the path sorting algorithm, the connection paths between nodes are used as features to determine the possible relationships between entity pairs.

[0080] like Figure 3 As shown, the algorithm flow is as follows:

[0081] 1. Select an enemy target h at sea:

[0082] 2. Construct a directed graph with h as the starting point;

[0083] 3. Start breadth-first search from the starting point;

[0084] 4. List all the paths starting from the starting point to avoid loops;

[0085] 5. If there are more than K paths between the starting point and a target, select K paths by limiting the path length and calculating the path feature value;

[0086] 6. Select another target at sea:

[0087] 7. Call the path characteristic value calculation function to calculate the path characteristic values ​​of all paths between target h and target t;

[0088] 8. For the path sequence S(h,t) = {P1(h,t), P2(h,t), ..., Ps(h,t)} starting from h and reaching t, sort the path characteristics from high to low and take the top K paths;

[0089] 9. Input the K paths into the long short-term memory network (LSTM) to obtain the final output vector h t , K h t Linked together and input into the target intention recognition fusion model;

[0090] In one embodiment, the calculation process of the path characteristic value calculation function is as follows:

[0091] enter:

[0092] 1. Spatiotemporal knowledge graph of maritime targets; 2. Head target h; 3. Tail target t;

[0093] Output:

[0094] The characteristic values ​​of all paths between target h and target t.

[0095] Sub-function flow:

[0096] 1) Let the relationship R between target h and target t be R(h), where R(h) = {e:R(h,e)}, that is, R(h) is the node on the path from h to t.

[0097] 2) Assume that the |R(h)| relations with the selected target h as the head node contain m different relation types, r1, r2, ..., r m are the number of m relationships, r1+r2+…+r m =|R(h)|;

[0098] 3) When there is a connecting path between target h and target t, that is, there are relations R1(h,e1), R2(e1,e2), ..., R n+1 (e n ,t), and record this connection path as P(h,t)=R1R2…R n+1 ;

[0099] 4) Eigenvalue F of path P(h,t) P Defined as:

[0100]

[0101] Where P'=R1R2…R n , r j Target e n R in the relationship with the head node n+1 The number of types.

[0102] 5) When P(h,t) is an empty path, F P Defined as:

[0103]

[0104] That is, the above formula sets the path characteristic value of the sub-path including only the head node to 1. Those skilled in the art can set it to other specific values ​​according to actual needs.

[0105] S4. Relationship-level feature extraction: Determine the most likely relationship between entity pairs through a tendency function based on translation embedding.

[0106] 1. Select an enemy target h;

[0107] 2. Select another target t;

[0108] 3. Convert the encoding of the target h and t into projection vectors on the corresponding hyperplane. If there is a certain relationship r between the target h and the target t, the word vectors of the three should satisfy h ⊥ +d r ≈t ⊥ ;

[0109] 4. Take the representation vector d of the intention r on the corresponding hyperplane r Calculation tendency:

[0110]

[0111] 5. For all types of intent, perform steps 3 and 4;

[0112] 6. Combine the tendencies of target h and target t for all intentions into a tendency vector, i.e., E(h,t) = [E(h,r1,t),E(h,r2,t),…,E(h,rn,t)], and input E(h,t) into the target intention recognition fusion model;

[0113] 7. For all targets except h, perform steps 2-6;

[0114] 8. Select the next enemy target and repeat steps 1-7;

[0115] S5. Feature-level fusion: see Figure 4 As shown in the figure, the features extracted from the above two levels are fused through a neural network to obtain a single-dimensional intent recognition result.

[0116] The features obtained by the above two algorithms are combined together through the feature fusion module to obtain the comprehensive feature vector F.

[0117]

[0118] S6. Recognition result fusion: see Figure 5 As shown in the figure, the intent recognition fusion model fuses the intent recognition results of multiple dimensions to obtain a more accurate enemy target intent recognition result. The full connection is used to fuse multiple intent recognition classification results to obtain an even more accurate intent recognition result.

[0119] Figure 6 is a diagram of the target intention recognition system architecture provided by an embodiment of the present invention; Figure 6 Shown, including:

[0120] A knowledge graph construction unit 610 is configured to construct a spatiotemporal knowledge graph containing targets and intent scenarios based on historical intents between targets of different types. The nodes in the spatiotemporal knowledge graph are entities corresponding to targets, and edges represent the existence of intents between two targets. A target has at least one node in the spatiotemporal knowledge graph, and each node corresponds to the encoding of a target at a time granularity and a spatial granularity. Each target has encodings at multiple time granularities and / or multiple spatial granularities, and each target belongs to a target type.

[0121] A feature extraction unit 620 is configured to perform node-level feature extraction on the target to be identified at a time granularity and a spatial granularity based on the spatiotemporal knowledge graph to obtain feature information of a path that carries relatively more information about the intent of the target to be identified; and to perform relationship-level feature extraction on the target to be identified based on the spatiotemporal knowledge graph to obtain an intent tendency vector of the target to be identified based on the inherent characteristics of the spatiotemporal knowledge graph;

[0122] The target intention recognition unit 630 is used to input the feature information and intention tendency vector into a pre-trained target intention recognition fusion model to obtain the intention of the target to be identified; the target intention recognition fusion model is used to fuse the feature information and intention tendency vector at a time granularity and a spatial granularity to obtain the intention recognition result of the target to be identified at a time granularity and a spatial granularity, and to fuse the intention recognition results of the target to be identified at different combinations of different time granularities and different spatial granularities to obtain intention recognition fusion results of different dimensions, and then determine the intention of the target to be identified based on the intention recognition fusion results of different dimensions.

[0123] It is understandable that the detailed functional implementation of each of the above units can be found in the description of the aforementioned method embodiment, and will not be repeated here.

[0124] It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program unit in the device are similar to those described in the above-mentioned method. The working process of the device can refer to the corresponding process in the above-mentioned method and will not be repeated here.

[0125] Based on the methods described in the above embodiments, an embodiment of the present invention provides an electronic device that may include: at least one memory for storing programs and at least one processor for executing the programs stored in the memory. When the programs stored in the memory are executed, the processor is configured to execute the methods described in the above embodiments.

[0126] Based on the method in the above embodiment, an embodiment of the present invention provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method in the above embodiment.

[0127] Based on the method in the above embodiment, an embodiment of the present invention provides a computer program product. When the computer program product runs on a processor, the processor executes the method in the above embodiment.

[0128] It is understood that the processor in the embodiments of the present invention may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0129] The method steps in the embodiments of the present invention can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC.

[0130] In the above embodiments, all or part of the embodiments can be implemented using software, hardware, firmware, or any combination thereof. When implemented using software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted via the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0131] It should be understood that the various numerical numbers involved in the embodiments of the present invention are only used for the convenience of description and are not intended to limit the scope of the embodiments of the present invention.

[0132] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A target intention recognition method, characterized in that: The following steps are involved: A spatiotemporal knowledge graph containing targets and intent scenarios is constructed based on the historical intents between targets of different types; the nodes in the spatiotemporal knowledge graph are entities corresponding to targets, and edges represent the existence of intents between two targets. A target has at least one node in the spatiotemporal knowledge graph, and a node corresponds to the encoding of a target at a time granularity and a spatial granularity. Each target has encodings at multiple time granularities and / or multiple spatial granularities, and each target belongs to a target type. Based on the spatiotemporal knowledge graph, node-level features of the target to be identified at a time granularity and a space granularity are extracted to obtain feature information of a path that carries relatively more information about the intention of the target to be identified; Extracting relational features of the target to be identified based on the spatiotemporal knowledge graph to obtain an intention tendency vector of the target to be identified according to inherent features of the spatiotemporal knowledge graph; Inputting the feature information and the intention tendency vector into a pre-trained target intention recognition fusion model to obtain the intention of the target to be identified; The target intention recognition fusion model is used to fuse the feature information and intention tendency vector at a time granularity and a space granularity to obtain the intention recognition results of the target to be identified at a time granularity and a space granularity, and to fuse the intention recognition results of the target to be identified at different combinations of different time granularities and different space granularities to obtain intention recognition fusion results of different dimensions, and then determine the intention of the target to be identified based on the intention recognition fusion results of different dimensions; The node-level feature extraction is specifically as follows: In the spatiotemporal knowledge graph, a directed graph is constructed starting from the target to be identified, and all paths between the target to be identified and the opposite target that do not contain loops are obtained; wherein the intention of the target to be identified is the intention of the target to be identified relative to the opposite target; Selecting multiple paths from all paths whose path lengths are less than a preset value and whose path characteristic values ​​are relatively high in descending order; the multiple paths are paths that carry a relatively large amount of information about the target intention to be identified; Inputting the multiple paths into a long short-term memory network (LSTM) respectively to obtain a feature vector for each path; Concatenating the feature vectors of each path to obtain the feature information; The path characteristic value is determined by the following steps: Determine the relationship between each node on the path and the two adjacent nodes; Starting from the head of the path, each additional node generates a subpath, and multiple subpaths are obtained until the tail node is reached; Determine the number of relationships for each node in the path in the spatiotemporal knowledge graph; the number of relationships is the total number of intents corresponding to all single edges with the node as the head node, where the total number of intents is obtained by adding the number of edges corresponding to each intent, and the number of edges for each intent divided by the number of relationships is the weight of each intent in the number of relationships for the node; The weight of the intention type corresponding to the edge between the previous node and the next node in the path in the number of relationships of the previous node is used as the intention weight of the sub-path with the next node as the tail node compared to the previous node; Determine the additional nodes of the next sub-path compared to the previous sub-path, multiply the path characteristic value of the previous sub-path by the intention weight of the added node to obtain the path characteristic value of the next sub-path, and solve the path characteristic value of each sub-path in turn to obtain the path characteristic value of the path from the head node to the tail node; the path characteristic value of the sub-path corresponding to the head node in the path is a preset value.

2. The method according to claim 1, characterized in that The relationship level feature extraction is specifically as follows: Determine the nodes corresponding to the target to be identified and the opposite target in the spatiotemporal knowledge graph; The tendency value of the target to be identified under each intention is determined by the tendency function based on translation embedding and the intention between the corresponding node of the target to be identified and the corresponding node of the opposite target, and the tendency values ​​under all intentions are spliced ​​to obtain the intention tendency vector of the target to be identified.

3. The method according to claim 2, characterized in that The propensity score for each intention is determined by the following steps: Convert the encoding of the target node h to the corresponding hyperplane projection vector h ⊥ , convert the encoding of the opposite target node t into the projection vector of the corresponding hyperplane t ⊥ ; Convert the encoding of the intention r into the representation vector corresponding to the hyperplane d r , then the tendency value of the entity pair (h, t) for intention r is: 。 4. The method according to any one of claims 1 to 3, characterized in that The target intention recognition fusion model includes: a feature fusion module and an intention multi-classifier; The feature fusion module is used to fuse the feature information and intention tendency vector at a time granularity and a space granularity to obtain the intention recognition result of the target to be identified at a time granularity and a space granularity; The intention multi-classifier is used to classify the intention of the target to be identified based on the fusion results of intention recognition in different dimensions, determine the probability of each preset intention, and take the preset intention with the largest probability value as the intention of the target to be identified.

5. A target intention recognition system, characterized in that: include: A knowledge graph construction unit is configured to construct a spatiotemporal knowledge graph containing targets and intent scenarios based on historical intents between targets of different types; nodes in the spatiotemporal knowledge graph are entities corresponding to targets, edges represent the existence of intents between two targets, a target has at least one node in the spatiotemporal knowledge graph, a node corresponds to the encoding of a target at a time granularity and a spatial granularity, each target has encodings at multiple time granularities and / or multiple spatial granularities, and each target belongs to a target type; a feature extraction unit configured to perform node-level feature extraction on a target to be identified at a time granularity and a spatial granularity based on the spatiotemporal knowledge graph to obtain feature information of a path that carries relatively more information about the target's intention; and to perform relationship-level feature extraction on the target to be identified based on the spatiotemporal knowledge graph to obtain an intention tendency vector of the target to be identified based on inherent features of the spatiotemporal knowledge graph; A target intention recognition unit is used to input the feature information and the intention tendency vector into a pre-trained target intention recognition fusion model to obtain the intention of the target to be recognized; The target intention recognition fusion model is used to fuse the feature information and intention tendency vector at a time granularity and a space granularity to obtain the intention recognition results of the target to be identified at a time granularity and a space granularity, and to fuse the intention recognition results of the target to be identified at different combinations of different time granularities and different space granularities to obtain intention recognition fusion results of different dimensions, and then determine the intention of the target to be identified based on the intention recognition fusion results of different dimensions; The feature extraction unit constructs a directed graph in the spatiotemporal knowledge graph with the target to be identified as the starting point, and obtains all paths between the target to be identified and the opposite target that do not contain loops; wherein the intention of the target to be identified is the intention of the target to be identified relative to the opposite target; selects multiple paths from all paths whose path lengths are less than a preset value and whose path feature values ​​are relatively high in descending order; the multiple paths are paths that carry a relatively large amount of information about the intention of the target to be identified; inputs the multiple paths into a long short-term memory network (LSTM) to obtain a feature vector for each path; and concatenates the feature vectors of each path to obtain the feature information; The feature extraction unit determines the relationship between each node on the path and the two adjacent nodes; starting from the head of the path, each additional node obtains a sub-path, and multiple sub-paths are obtained until the tail node is added; the number of relationships of each node in the path in the spatiotemporal knowledge graph is determined; the number of relationships is the total number of intentions corresponding to all single edges with the node as the head node, wherein the total number of intentions is obtained by adding the number of edges corresponding to each intention, and the number of edges for each intention divided by the number of relationships is the weight of each intention in the number of relationships of the node; the weight of the intention type corresponding to the edge between the previous node and the next node in the path in the number of relationships of the previous node is used as the intention weight of the sub-path of the next node compared to the previous node as the tail node; and the added node of the next sub-path is determined compared to the previous sub-path, the path feature value of the previous sub-path is multiplied by the intention weight of the added node to obtain the path feature value of the next sub-path, and the path feature value of each sub-path is solved in turn to obtain the path feature value of the path composed of the head node to the tail node; the path feature value of the sub-path corresponding to the head node in the path is a preset value.

6. An electronic device, characterized in that: include: at least one memory for storing a program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 4.

Images