An artificial intelligence expert system based on knowledge management system and its construction method

Abstract

The present invention relates to an artificial intelligence expert system based on a knowledge management system and a construction method thereof, wherein the system comprises: a knowledge acquisition module, a knowledge parsing module, an input module, an evaluation module and a display module. The system of the present invention is based on a constructed composite heterogeneous graph neural network and a heterogeneous behavior network dual network, and constructs a dual attention mechanism based on an attention mechanism of a composite heterogeneous graph neural network and a heterogeneous behavior attention mechanism, which realizes the attention mechanism for knowledge operation behavior, and further strengthens the user participation through relearning, and forms an online expert based on the user's knowledge operation behavior, providing a further efficient effect for the imparting of knowledge. The analysis results of the attention mechanism of the dual network unique to the present invention can reflect the user's mastery and application of knowledge in the knowledge field from the side, and strengthen the guidance of knowledge acquisition according to the user's interest in or depth of research in a specific field of knowledge.

Description

Technical Field

[0001] The present invention relates to an artificial intelligence expert system and a construction method thereof, in particular to an artificial intelligence expert system based on a knowledge management system and a construction method thereof, and belongs to the field of knowledge intelligent systems. Background Art

[0002] The existing technology combines user behavior data (API) with modeling analysis to obtain a knowledge graph. However, this knowledge graph is only a knowledge context organization and intelligent perception of knowledge. It cannot analyze the collective behavior of users with the same interests, hobbies and learning goals, and provide specific knowledge to relevant users, such as not only graphs, but also other related knowledge materials. This puts higher requirements on the construction of expert systems.

[0003] The existing analysis of APIs has not been refined to the classification of corresponding behaviors. In real-world knowledge-related operations, there are queries, downloads, uploads, Q&A, discussions, and combinations of these behaviors. They represent different levels of knowledge requirements and the degree of activity in knowledge in different fields, which indirectly reflects the user's mastery and application of knowledge in the knowledge field. API cannot classify the user's interest in or depth of research in a specific field of knowledge based on the specific behavior type, nor can it rate the user's performance in knowledge. This allows users to become online experts with the knowledge mastery and research characteristics of high-knowledge users, rather than just letting the system act as a machine expert. Moreover, the existing technology does not provide specific modeling, and technicians cannot obtain specific methods for analyzing and processing API data, but only describe them from the perspective of building a framework.

[0004] Heterogeneous graph network is a data configuration for the degree of association between different objects. Through deep neural network analysis of data configuration, we obtain a mechanism called heterogeneous graph neural network attention mechanism (HGNNAM). Its advantage is that the data configuration itself is abstract, but it can be visualized, giving the distribution, relationship, and degree of association between objects, thereby sorting out the association between related knowledge. Summary of the invention

[0005] Taking into account the above-mentioned shortcomings of the existing API, the present invention provides solutions to the following problems: first, the classification of knowledge behaviors; second, the data configuration problem of knowledge behaviors themselves; third, the specific analysis of knowledge behaviors and the configuration of knowledge data presented to users; and fourth, the re-learning of users' knowledge behaviors to form online experts.

[0006] In order to implement the above solution, the present invention provides an artificial intelligence expert system based on a knowledge management system, comprising:

[0007] The knowledge acquisition module is used to acquire knowledge from the knowledge generation source and record the spatial location and generation time of the knowledge generation. The knowledge includes knowledge related to the equipment related to the knowledge, specifically including the name of the equipment, the knowledge generation source and / or user to which it belongs, the purpose and technical principle related to the equipment, as well as scientific and technological resources, social and humanities resources, specifically including documents, pictures, videos, and audio.

[0008] The knowledge parsing module is connected to the knowledge acquisition module and is used to construct a bi-network (bi-N) of a complex heterogeneous graph neural network (CHGNN) and a heterogeneous behavior network (HBN) and to construct a bi-attention mechanism (bi-Attention Mechanism) based on a complex heterogeneous graph neural network (CHGNN) and a heterogeneous behavior attention mechanism (HBAM) by using the knowledge and spatial locations input by the received knowledge generation source. Mechanism, bi-AM), and after the construction is completed, the knowledge context is retrieved again based on the new input (in most cases, the new input is a search operation, that is, the relevant knowledge input is used to search for any relevant knowledge information of interest, or there are other purposes of use, learning, and organization) and the recorded knowledge operation behavior, and the analysis results of the dual attention mechanism are given based on the knowledge context and knowledge operation behavior, and bi-N and bi-AM are updated within a predetermined time period.

[0009] The input module is connected to the computer network and is used to receive the newly input knowledge of the user, and input it into the knowledge analysis module through the computer network. The knowledge is clustered in a preset knowledge field with a field label. When the user newly inputs knowledge, it is necessary to select a target field for the newly input knowledge and select a field label for the knowledge. At the same time, the knowledge operation behavior is recorded according to the user's new input behavior analysis, including at least one of query, download, upload, question and answer, and discussion, and a combination thereof.

[0010] The evaluation module is used to conduct a comprehensive analysis of the retrieved knowledge context and the analysis results of the dual attention mechanism, and give a final evaluation that is suitable for the user's true search meaning, as well as re-learning to form an online expert system, online knowledge transfer based on the online expert system, and,

[0011] The display module is used to display the analysis results of the knowledge context, the dual attention mechanism, and the comprehensive analysis results and to operate the visualized bi-N.

[0012] The knowledge generation source periodically inputs the knowledge generated according to the confidentiality classification into the knowledge acquisition module, and transmits it to the knowledge analysis module for the update.

[0013] Optionally, the knowledge generation sources include: administrative departments, enterprises, institutions, companies, school research institutions, social groups and organizations, and individuals.

[0014] It is easy to understand that the equipment related to knowledge refers to any key equipment under the knowledge or equipment obtained from the knowledge generation source. For example, various instruments and equipment in the field of analytical chemistry, various equipment in the field of physics, such as wind tunnels in fluid mechanics, such as electron microscopes in the field of microscopy, etc. The purpose includes the knowledge field in which the equipment is used and the function of the equipment.

[0015] Optionally, the knowledge acquisition module parses the IP address or obtains the spatial location based on the name and address of the input knowledge generation source. The generation time is obtained by recording the time of receiving the knowledge through the knowledge acquisition module. The knowledge is reviewed manually or by computer to decide whether to adopt the knowledge and record the spatial location and generation time. The operation includes dragging the visualized bi-N, clicking on it to pop up a dialog box or search bar to search and / or view, modify, delete and / or restore the knowledge.

[0016] The method of constructing a dual network of a composite heterogeneous graph neural network and a heterogeneous behavior network and utilizing the knowledge and spatial locations input by the received knowledge generation source to construct a dual attention mechanism based on the attention mechanism of the composite heterogeneous graph neural network and the heterogeneous behavior attention mechanism includes:

[0017] S1 establishes a geographical map, marks the spatial locations on the geographical map as nodes of the heterogeneous graph neural network, constructs a heterogeneous graph network G=(V, E) based on real spatial locations, and forms a heterogeneous graph network neural network of real spatial locations, where V is a node set, which can not be dragged and modified once recorded, but can be deleted and restored, and E is a spatial edge set. The spatial edge represents the spatial relationship between different nodes.

[0018] That is, with the help of geographic maps, the spatial location of equipment, knowledge, and knowledge generation sources are defined, thus laying the foundation for spatial distribution and providing a big data foundation for the research on knowledge creation, dissemination, learning, and use. However, the existing heterogeneous graph neural network is only a logical abstract node, and does not correspond to the actual geographical location of the node. Through the real geographical location, that is, the correspondence of the spatial location, it is possible to carry out collaborative association from the trinity of space, time, and degree of association, so that the mathematical model is clear and the principle is clear. The source of knowledge generation can be either the user or the user, so the system increases the degree of participation of various knowledge subjects in system updates and improvements.

[0019] S2 constructs a knowledge heterogeneous graph network H = (K, F) based on V, where K is a set of knowledge nodes in different knowledge fields and F is a set of knowledge edges. Knowledge edges represent the association between knowledge in different knowledge fields or in the same knowledge field, and constructs a device heterogeneous graph network I = (U, W) based on V, where U is a set of device nodes representing a single device or a set of devices, and W is a set of device edges. Device edges represent the association between devices or sets of devices belonging to the same knowledge field or similar knowledge fields. In this way, any new input from any user can be associated with the relevant device, thereby gaining a deeper understanding of the application of knowledge.

[0020] S3 constructs a heterogeneous behavior network B={X,Y} based on V, where X is a node representing the knowledge operation behavior, and Y is a behavior edge set between knowledge operation behaviors. The behavior edge set represents the association between various knowledge operation behaviors.

[0021] S4 constructs the meta-path I in G = (V, E), H = (K, F), I = (U, W), and B = {X, Y}: Metapath II: Meta Path III: Meta Path III: Traversing Metapath V: And define the neighbors of the meta-path, that is, for any node N∈V,K,U,X, all nodes connected to N by the meta-path are neighbors, where A p ∈V,B q ∈K,J d ∈U,Q f ∈X,T z ∈V,K,U,X,p,q,d,f,z∈N,N is a natural number, and p∈[1,m+1],q∈[1,n+1],d∈[1,w+1],f∈[1,v+1],z∈[1,t+1],R1,R2,…,R m is the spatial relationship of the nodes, each of which is a double-arrow vector representing the adjacent nodes in the meta-path I on the geographic map; C1, C2, ..., Cn is the relationship between knowledge, each of which is a double arrow representing the adjacent nodes in the meta-path II in the knowledge field of scientific and technological resources and social and humanities resources (indicating that the knowledge fields are mutually related), M1, M2, …, M w is the association between devices (for example, the various types of instruments and equipment required to complete a certain experiment or solve a certain type of problem in a knowledge field). Each of them represents a double arrow in the knowledge field related to the device in the meta-path III (indicating that the knowledge fields to which the devices belong are related to each other), P1, P2, ..., P w Each of the two-way arrows in the meta-path IV represents the relationship between the adjacent nodes of the knowledge operation behaviors. This can describe the reversible traceability between the operations and understand the process of user knowledge acquisition.

[0022] For the traversal metapath V neighbors, define the traversal neighbors. Select the focus node T1∈V, or T1∈K, or T1∈U, or T1∈X from G=(V,E), H=(K,F), I=(U,W), B={X,Y}, and t∈N. Let r∈N, r∈[2,t+1] and the range of r is less than or equal to t and greater than 0. Let T r ∈T2,…,T t+1 ,T r ∈V, then in, is the complement of the range of r, D1, D2, ..., D t To focus on the traversal relationship between the node and the traversal neighbor and between the traversal neighbors (local traversal and remote traversal), it is expressed as T r and The double arrows between form a crossing edge, where and The value range of s∈r, then =s, then the traversal metapath V is called the full traversal metapath, and the non-full traversal metapath is called the partial traversal metapath. That is, the nodes in the full traversal metapath are distributed in all four heterogeneous graph networks; the nodes in the partial traversal metapath are in any two of the four heterogeneous graph networks at most, thus forming a composite heterogeneous graph neural network composed of G=(V,E), H=(K,F), I=(U,W) and bi-N with B={X,Y}. The traversal path inside the pure composite heterogeneous graph neural network only describes the association between spatial locations, knowledge, and equipment. The traversal path between the composite heterogeneous graph neural network and B={X,Y} describes the association between the knowledge operation behaviors of the two networks. For example, in a certain region and in a certain knowledge field, the level of users' knowledge needs can be seen through the full traversal metapath, such as just viewing, understanding, or citing, or downloading and learning, or uploading and imparting knowledge, or participating in questions and answers and even discussions. This can determine whether the user can be selected as an object of re-learning to become an online expert.

[0023] It should be emphasized that for equipment I = (U, W), its neighbors can be divided according to the property rights and usage attributes of the equipment and the knowledge generation source, and the relationship between the partners. For the traversal path containing I = (U, W), it reflects the relationship between the equipment itself and the knowledge and spatial location.

[0024] As for the knowledge operation B={X,Y}, it can be independent of the composite heterogeneous graph neural network, and the study of the relationship between the operation behaviors can also be a study of the relationship between the composite heterogeneous graph neural network, so as to conduct a holistic study on the spatial location, knowledge, equipment, and knowledge operation behavior, and pay attention to the operation behavior of users in a certain geographical area on the knowledge field that they are interested in, reflecting the users' knowledge needs, knowledge mastery, application, and research.

[0025] S5 first establishes attention mechanisms in G, H, I, and B according to meta-path I, meta-path II, meta-path III, and meta-path IV respectively. It first builds the node attention mechanism, then the semantic attention mechanism, and finally the subgraph attention mechanism. The node attention mechanism, semantic attention mechanism, and subgraph attention mechanism built based on B are called heterogeneous behavior attention mechanisms.

[0026] Secondly, based on the traversing meta-path V, a traversing attention mechanism is established between G, H, I, and B. First, one of G=(V,E), H=(K,F), I=(U,W), and B={X,Y} is selected to select the focus node. According to the traversing meta-path, the node attention mechanism is also first constructed, then the semantic attention mechanism is constructed, and finally the subgraph attention mechanism is constructed. Then, the nodes of each of the remaining two of G=(V,E), H=(K,F), I=(U,W), and B={X,Y} are selected in turn as the focus nodes. Again, according to the traversing meta-path, the node attention mechanism is first constructed, and then the semantic attention mechanism is constructed. Finally, the bi-AM is obtained, wherein the number of nodes of meta-path I, meta-path II, meta-path III, and traversing meta-path IV are all adjustable between 3 and 20. Among them, the attention mechanism purely constructed based on the composite heterogeneous graph neural network in the bi-AM is CHGNNBAM, which together with HBAM constitutes the bi-AM. The bi-AM includes the attention mechanism of the node traversing path belonging to X.

[0027] Optionally, the node attention mechanism construction method is as follows:

[0028] S5-1 For the node pair (i, j) connected by meta-path I, meta-path II, meta-path III, meta-path IV, and traversal path V, for L∈meta-path I, meta-path II, meta-path III, meta-path IV, and traversal path V, five types of node attention are constructed respectively. Att Node∈G (·), Att Node∈H (·), Att Node∈I (·), Att Node∈B (·), Att Node∈III (·) are five types of node attention deep neural networks, g i ,g j ∈V,h i ,h j ∈K,u i ,u j ∈U,x i ,x j ∈X,T III ,t j ∈V,K,U,X are the representations of nodes i and j when L∈ meta-path I, meta-path II, meta-path III, meta-path IV, and L∈ crosses path IV, respectively. III For node i, select the representation of the focus node T1;

[0029] S5-2 calculates the attention of all neighbors j of node i based on meta-path L, and the following formula (2) obtains the unified expression of the normalized coefficients of the attention of the four types of nodes in formula (1): || is the concatenation operator, is the node attention vector of meta-path L. The superscript T indicates transposition, N iL is the number of neighbor nodes of node i, LeakyReLU[·] is the activation function, L i and L j They are the unified expression of node i and j’s attention in four types of nodes, and the unified representation,

[0030] S5-3 Unified representation of features of computing node i based on four types of meta-paths L Finally, the attention mechanism is repeated K times to obtain Indicates that the attention mechanism is repeated K times for splicing;

[0031] The semantic attention mechanism is constructed as follows:

[0032] S5-4 Calculate the weights of the five types of meta-paths in

[0033] || represents the modulus of the corresponding node set. In the fifth formula, K, V, U, and X are determined according to the specific node set to which the selected attention node belongs (if the node belongs to any two or more of the four, it includes the modulus of the union of the nodes). The q with subscripts and the w and b with subscripts are the attention vector, weight, and bias representation of the corresponding semantics, respectively.

[0034] S5-5 For the subgraphs corresponding to K, V, U, and X, the attention of node i based on each meta-path L is calculated, and the following formula (5) obtains the unified expression of the normalization coefficients of the four types of semantic attention: Where W L is the unified expression of formula (4),

[0035] S5-6 Constructing four types of meta-path attention β L∈子图 =Att L∈子图 (Z L∈子图 ) (6), It means that for any path L∈ subgraph, the meta-path attention is represented by Z under the corresponding pathL∈子图 Deep neural network Att via semantic attention L∈子图 (·) is learned, then the features of the semantic attention mechanism are uniformly expressed as Z 子图 =∑ L∈子图 β L∈子图 ·Z L∈子图 (7);

[0036] The subgraph attention mechanism is constructed as follows:

[0037] S5-7 is the same as the node attention mechanism and the semantic attention mechanism. For the knowledge domain to which the knowledge input by the knowledge generation source belongs and the subgraphs corresponding to the knowledge operation behaviors (that is, the subgraph set composed of the nodes corresponding to each domain and operation behavior, the subgraph set can be a subgraph composed of non-traversal paths or a subgraph composed of traversal paths), constructing the subgraph attention mechanism involves the following corresponding formulas (8)-(11):

[0038] is the subgraph association degree, is a deep neural network for subgraph attention,

[0039] W 子图 is the subgraph weight, with superscripts and subscripts w 子图 、b 子图 are the corresponding sub-graph attention vector, weight, and bias representation, respectively.

[0040]

[0041] Unified representation for subgraph attention mechanism.

[0042] It can be seen that for the subgraph When the equal sign is obtained, it corresponds to the attention mechanism based on the whole image. When it corresponds to a local attention mechanism, including CHGNNBAM and heterogeneous behavior attention, it can construct meta-paths and sub-graphs through five types of meta-paths, so as to obtain the importance and relevance of predictions in local sub-graphs through corresponding attention mechanisms, so as to know the importance or relevance of each node in neighbors, meta-paths, and sub-graphs, that is, the status. In this way, the source of knowledge, knowledge itself, equipment, and knowledge operation behavior are fully included in a bi-N, providing a big data foundation for the construction of online expert systems.

[0043] It should be noted that ZL∈子图 The solution is obtained through the corresponding degree of association and attention mechanism, and the optimized network parameters are obtained through training through the following steps.

[0044] Substitute Z into Att Node∈G (·), Att Node∈H (·), Att Node∈I (·), Att Node∈B (·), Att Node∈III (·), Att L∈子图 (·), Any one or a combination of them, i.e., multi-layer sensing mechanism, (R is a real number) to calculate the cross entropy loss L = -∑ l∈o Y l ln(C·Z l ), where l represents the number of the domain tag, o is the domain tag set, and Y l is the domain label numbered l, C is the node classifier parameter, By training each attention deep neural network until the cross entropy loss stabilizes and minimizes, we get bi-AM, where

[0045] Since Z belongs to the subgraph attention mechanism representation, it contains the components of the semantic (sixth item) and node (first five items) attention mechanisms, or in other words, it comes from the semantic and node attention mechanisms. Therefore, there may be a certain degree of deviation when different attention mechanisms are substituted. Therefore, there is an adjustment of the combination coefficient. When only one of the attention mechanisms is substituted, although it does not hinder training, it will deviate from the attention mechanism trained by the representation of a better corresponding level. Therefore, different combinations are selected according to different accuracy requirements, that is, the adjustment of the combination coefficient. Generally speaking Each item of the combined attention mechanism represents the prediction result of the degree of synergy correlation of the attention mechanism at the corresponding level. The combination coefficient value is different depending on the specific meta-path and sub-graph selection. For example, when the meta-path does not include the traversal path, γ = 0, and when the meta-path selection only includes the meta-path I in G, β = γ = δ = ε = 0.

[0046] The training methods include:

[0047] S5-8 The knowledge generation source inputs the knowledge generated according to the confidentiality classification into the input module regularly through the computer network, and the input module transmits it to the knowledge acquisition module to record the spatial location, generation time, and knowledge operation behavior of the knowledge generation; S5-9 The knowledge parsing module obtains the trained attention mechanisms of the five types of meta-paths in G, H, I, B, and between G, H, I, and B through step S5, receives the input knowledge and the generated spatial location, divides it into a training set and a verification set, and the ratio of the two is 9-3:4-1, processes the training set to form a node representation, forms a domain label according to the knowledge domain input by the knowledge generation source, and defines the number of meta-path nodes, obtains all corresponding meta-paths of the knowledge corresponding nodes, and predicts the corresponding normalization coefficients according to formulas (1)-(3). and According to formulas (4)-(7), the corresponding normalization coefficient β is obtained L∈子图 and Z 子图 , and according to formulas (8)-(11), the corresponding subgraph association degree, that is, the normalization coefficient and Z, and the cross entropy loss values ​​are calculated respectively,

[0048] S5-10 back-propagates the cross entropy loss value to optimize the parameters of each deep neural network, uses the verification set to verify the accuracy, and loops through steps S5-7-S5-9 to optimize the network parameters and verify the accuracy.

[0049] The knowledge context is retrieved based on the newly input knowledge and the recorded knowledge operation behavior, and the analysis result of the dual attention mechanism is given based on the knowledge context and the knowledge operation behavior. The specific method is:

[0050] S6 inputs knowledge into the trained bi-AM to obtain the predicted knowledge domain and related knowledge operation behaviors.

[0051] In the predicted knowledge domain, S7 draws the knowledge context according to the corresponding spatial location and generation time of the knowledge, and retrieves the drawn knowledge context.

[0052] S8: marking nodes and / or subgraphs and / or knowledge operation behaviors corresponding to the new input knowledge in the drawn knowledge context.

[0053] That is to say, if the corresponding knowledge domain is mostly viewed, the knowledge viewing function is marked. For example, if the predicted knowledge operation behaviors are highly correlated with viewing, downloading, uploading, etc., including high correlation between nodes and high correlation between meta-paths formed by a series of nodes, then these operations and related paths are marked, so as to show the level of knowledge demand in this knowledge domain according to the marking.

[0054] Among them, the predicted knowledge domain described in S6 is specifically based on the knowledge domain to which the domain label that may be involved is represented by the largest normalization coefficient under each domain label. The drawing method in S7 is to select all nodes and / or subgraphs belonging to the predicted knowledge domain in the geographic map, H, and I, and in the predicted knowledge domain, sort them according to the chronological order of knowledge generation to form a historical context his. S8 is marked in the geographic map and the visualized H, I, and B.

[0055] Optionally, a knowledge context corresponding to at least one other field among the fields whose normalization coefficients under each field label are not the largest is selected for drawing, and the visualization is to display the nodes belonging to V on the corresponding geographical map.

[0056] Optionally, the comprehensive analysis includes statistically analyzing at least one of the following types of distributions: a first distribution on the geographic map, namely, a G distribution, in the predicted knowledge field corresponding to the knowledge and the at least one other field;

[0057] The second distribution in the visualized H, the H distribution,

[0058] The third distribution in the visualized I, the I distribution,

[0059] The fourth distribution in the visualized B, namely the B distribution,

[0060] The crossing distribution between G, H, I, and B, namely the GHIB distribution, and

[0061] According to the user's knowledge generation source, the knowledge field involved in the knowledge generation source is displayed, corresponding to at least one of the G distribution, H distribution, I distribution, B distribution, and GHIB distribution on the geographic map, and,

[0062] Statistically analyzing the historical context his in each selected distribution, that is, the historical context his of knowledge in the nodes belonging to V, K, and U in each selected distribution, the final evaluation adapted to the true meaning of the user's search includes comparing the predicted knowledge field with the field label corresponding to the at least one other field and the field label of the user input knowledge, and taking the comparison result as the final evaluation.

[0063] Relearning and forming an online expert system specifically includes the following steps:

[0064] S9 retrieves the knowledge operation behaviors of all users and knowledge generation sources, and according to the statistical results, takes the meta-path users and knowledge generation sources whose number of nodes in meta-path IV in distribution B exceeds 3 as online expert re-learning objects;

[0065] S10, according to the GHIB distribution, retrieves the GHIB distribution corresponding to the object, and analyzes at least one of the G distribution, H distribution, I distribution, and B distribution therein;

[0066] S11 establishes a convolutional neural network CNN, forms a distribution image for at least one of G distribution, H distribution, I distribution, B distribution and / or GHIB distribution, divides the distribution image into a training set and a validation set, and the ratio of the two is 5-3:1-2. The convolutional neural network is trained using the training set, and the accuracy is verified by the validation set. The network parameters are optimized by calculating the loss function to obtain a trained model CNN, and the distribution image is identified as an expert type as a recommended expert to form an online expert system.

[0067] That is, when the user completes the new input, the system will form various distribution statistics and final evaluation according to the above algorithm steps. Based on this, the user adjusts the knowledge domain and completes the acquisition of the distribution image, and finally inputs it into the trained model CNN in the expert system to obtain the recommended expert.

[0068] Optionally, the method of forming the distribution image is to form at least one of the visualized G distribution, H distribution, I distribution, B distribution, and GHIB distribution into a G distribution map, an H distribution map, an I distribution map, and a B distribution map, respectively, and rearrange them in two dimensions according to the distribution method of the G distribution map.

[0069] It is understandable that by selecting experts based on the distribution of knowledge, equipment, and knowledge operation behaviors based on geographical distribution, it reflects the geographical distribution of experts in the knowledge field and the distribution of knowledge points in the field of expertise, thereby providing experts with more local, knowledge field, and related equipment characteristics. All of this information is reflected in the distribution image. At the same time, the distribution image can also further help with the acquisition of knowledge through social connections in a better and more efficient way.

[0070] Optionally, the expert system includes a trained model CNN and a display module for displaying at least one of the G distribution, H distribution, I distribution, B distribution, and GHIB distribution, and has an operation visualization distribution to perform the operation on the knowledge.

[0071] Optionally, the input module also has a network connection module connected to a computer network, for realizing the connection between the knowledge operation behavior and the Internet, that is, the knowledge operation behavior is realized with the help of a computer network, realizing the knowledge operation behavior on the domain.

[0072] Online knowledge transfer based on online expert system specifically includes the following steps:

[0073] S12: Based on the new input of the user, the online expert system obtains at least one recommended expert (that is, the first several recommended experts predicted by CNN);

[0074] S13 The user selects at least one recommended expert and acquires the knowledge through the display module and the network connection module by performing the operation and / or knowledge operation behavior.

[0075] Another object of the present invention is to provide a method for constructing the above artificial intelligence expert system based on heterogeneous behaviors, comprising:

[0076] The first step is to construct the modules in the artificial intelligence expert system based on heterogeneous behaviors, realize the electrical connection between the modules, and connect the system to the computer network;

[0077] In the second step, the input from knowledge generation sources and users is continuously collected, and the system completes the construction of bi-N, bi-AM, the generation function test of G distribution, H distribution, I distribution, B distribution and GHIB distribution, as well as the drawing of knowledge context and the training of re-learning CNN model.

[0078] Beneficial Effects

[0079] Based on the constructed composite heterogeneous graph neural network and heterogeneous behavior network dual network, the attention mechanism based on the composite heterogeneous graph neural network and the dual attention mechanism of the heterogeneous behavior attention mechanism are constructed. While realizing the attention mechanism for knowledge operation behavior, the user's participation is further strengthened through re-learning, so that the knowledge data accumulated by the user's knowledge operation behavior is recognized by the system, so that through the system's re-learning, the user becomes an online expert, which provides a further efficient effect for the imparting of knowledge. In addition, the analysis results of the dual network's attention mechanism not only give the correlation and distribution between knowledge, equipment, and knowledge generation sources, but also, due to the coupling of knowledge operation behavior, it can reflect the user's mastery and application of knowledge in the knowledge field from the side, and strengthen the guidance of knowledge acquisition based on the user's interest in specific field knowledge or the depth of research. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 FIG. 1 shows a configuration diagram of an artificial intelligence expert system according to an embodiment of the present invention.

[0081] Figure 2 A visualization of the subgraphs G = (V, E) and I = (U, W), where the north direction is given.

[0082] Figure 3 G=(V,E), I=(U,W), H=(K,F), B={X,Y} sub-graph visualization diagram, where the north direction is given,

[0083] Figure 4 a- Figure 4 e is a schematic diagram of the corresponding node i and its neighbors for inputting new knowledge in the five-category meta-path III. Figure 4 f is a schematic diagram of j2 and its neighbors when j2 is the corresponding node for inputting new knowledge.

[0084] Figure 5 A simplified flow chart of methods S1-S4 for constructing bi-AM.

[0085] Figure 6 A simplified diagram of the node attention mechanism construction method.

[0086] Figure 7 A simplified diagram of the construction process of the semantic attention mechanism and the subgraph attention mechanism.

[0087] Figure 8 A flowchart of a specific method for retrieving a knowledge context based on newly input knowledge and providing an analysis result of an attention mechanism based on the knowledge context.

[0088] Fig. 9 Flowchart of methods for various statistical distributions, historical contexts, and final evaluations adapted to the true meaning of the user's search.

[0089] Fig.10 Study again and form a simple flow chart of the online expert system method.

[0090] Fig.11 Schematic diagram of the two-dimensional distribution image. DETAILED DESCRIPTION

[0091] Figure 1 As shown, an artificial intelligence expert system configuration based on a knowledge management system is given, including a knowledge acquisition module with a gray background and black rounded corners, which is used to acquire knowledge from the knowledge generation source and record the spatial location (indicated by the blue arrow) and generation time (indicated by the red arrow) of the knowledge generation. The knowledge includes the name of the device, the knowledge generation source and user to which it belongs, the purpose and technical principle related to the device, as well as scientific and technological resources, social and humanities resources, including documents, pictures, videos, and audio.

[0092] It also includes a knowledge acquisition module represented by a black rounded box with a gray background, which is used to construct a dual network (bi-N) of a composite heterogeneous graph neural network (CHGNN) and a heterogeneous behavior network (HBN) (indicated by the black arrow) and utilize the knowledge (indicated by the purple arrow) and spatial location (indicated by the blue single-line arrow) input by the received knowledge generation source to construct a dual attention mechanism (bi-AM) based on the attention mechanism of the composite heterogeneous graph neural network (CHGNNBAM) and the heterogeneous behavior attention mechanism (HBAM). After the construction is completed, the knowledge context is retrieved again based on the new input (indicated by the purple one-way arrow) knowledge and the recorded knowledge operation behavior (indicated by the lake blue single-line arrow), and the analysis results of the dual attention mechanism are given based on the knowledge context and knowledge operation behavior (indicated by the red arrow), and the bi-N and bi-AM are updated within a predetermined time period (not shown in the figure).

[0093] The input module represented by the black gray box with rounded corners includes a network connection module (not shown in the figure) connected to the computer network (indicated by the lake blue line segment), which is used to realize the connection between the knowledge operation behavior and the Internet, and receive new knowledge input by the user (indicated by the purple one-way arrow), and input it into the knowledge parsing module through the computer network (indicated by the lake blue arrow).

[0094] The evaluation module represented by the black gray box with rounded corners is used to conduct a comprehensive analysis of the analysis results of the retrieved knowledge context and attention mechanism, and give a final evaluation that is suitable for the true meaning of the user's search (indicated by the black arrow). The display module represented by the black gray box with rounded corners is used to display the analysis results of the knowledge context and attention mechanism (indicated by the black line arrow) and the results of the comprehensive analysis and final evaluation (indicated by the black arrow), and to display and operate the visualized CHGNN. At the same time, the knowledge operation behavior is recorded according to the user's new input behavior analysis, including query, download, upload, question and answer, and discussion.

[0095] The evaluation module represented by the black gray box with rounded corners is used to conduct a comprehensive analysis of the retrieved knowledge context and the analysis results of the dual attention mechanism, and give a final evaluation that is suitable for the user's true search meaning, as well as re-learning to form an online expert system, online knowledge transfer based on the online expert system, and,

[0096] The display module is represented by a black gray-bottom rounded box, which is used to display the analysis results of the knowledge context, the dual attention mechanism, and the results of the comprehensive analysis and to operate the visualized bi-N.

[0097] The knowledge generation source inputs the knowledge generated according to the classification of confidentiality into the knowledge acquisition module regularly (1-3 years), and transmits the knowledge analysis module for the update ( Figure 1 not shown).

[0098] Next, we will Figure 1 The three types of functions represented by the red, blue, and green dashed rectangles are described in detail. Figure 2-Figure 4 As shown, the method of constructing CHGNN and constructing CHGNNBAM based on the knowledge and spatial location input by the heterogeneous graph neural network to receive the knowledge generation source includes:

[0099] like Figure 2 As shown, S1 establishes a geographical map, marks the spatial locations AH, J, K, M on the geographical map as nodes of a heterogeneous graph neural network, and constructs a heterogeneous graph network G = (V, E) based on real spatial locations, and Figure 2 The figure shows an example of a subgraph consisting of nodes AD and M in G = (V, E), forming a heterogeneous graph network neural network of a real space location, where V is a node set and E is a spatial edge set.

[0100] At the same time Figure 3 G=(V,E) constructed from the geographical map of V, and then step S2 is performed to construct a knowledge heterogeneous graph network H=(K,F) based on V. Figure 2 The corresponding exemplary display part H = (K, F) subgraph, K is the knowledge node set of different knowledge fields, and F is the knowledge edge set. These H = (K, F) subgraphs are composed of Figure 3 As shown, nodes A and B are used as an example (for other Figure 2 The corresponding two black nodes i'∈K and their surroundings are represented by concentric lake blue and grass green dotted circles of different radii, with different degrees of proximity to them. These neighbors of different distances represent knowledge clusters of different degrees of proximity in different fields (based on multi-level subject classification and its branches, or patent IPC or CPC classification as the division standard), specifically, knowledge clusters represented by scientific and technological resources and social humanities resources. Thus, the H=(K,F) subgraph is visualized, and the full graph H=(K,F) can be visualized in the same way. It can be seen that the abstract knowledge heterogeneous graph structure is visualized as concentric circles.

[0101] like Figure 2 and 3 As shown, for the visualization of the device heterogeneous graph network I = (U, W), concentric circles of different radii are also used to represent it, and they are represented near nodes AH, J, K, M. It can be displayed in the display module and the user can drag and click the nodes on it with the mouse to pop up a dialog box for search, viewing, and modification. The H = (K, F) graph can also perform these operations, but once the nodes AH, J, K, M are recorded, they do not need to be dragged or modified. Figure 4As shown in a-4d, the neighbors j1-j5 of i with different distances in I=(U,W) are represented by multiple concentric circles of different colors (black, purple) with 5 red nodes i as the center. The neighbors j1, j2, j3 (black) on the black dotted circle are in the same field and are the closest neighbors to i, and the neighbors j4, j5 (purple) on the purple dotted circle are in the same field and are the second closest neighbors to i. Similarly, Figure 3 Although not shown in the figure, the different lake blue and grass green dashed circles in the H = (K, F) subgraph have multiple other neighbors of i' (if they exist on the meta-path II).

[0102] like Figure 4 As shown in e, there are other lesser neighbors on other concentric circles, as well as other neighbors on black and purple dotted circles (as long as they are on the set meta-path III), which are omitted by three black and red dots respectively. Figure 4 f, if the new knowledge input by the user belongs to j2, then i, j1, and j3 are the nearest neighbor, the next nearest neighbor, and the next nearest neighbor respectively. Figure 3 In the H=(K,F) subgraph in , the input new knowledge can also belong to other nodes other than i', then the visualization effect of the lake blue and grass green dotted circles will change, as shown in Figure 4 f and Figure 4 a- Figure 4 Changes between e. Figure 2 and Figure 3 In the G=(V,E) subgraph, near nodes AD and M, they are Figure 4 A subgraph of I(U,W) represented by the five-class meta-path composed of the five i nodes and their neighbors in a-4e ( Figure 4 The meta-path edges are not shown in e), that is, Figure 4 The double-sided double arrows in a-4d represent the metapath III (the same below): as well as The order of the single arrows of each node corresponds to the defined meta-path III: In the node, for example Here, i corresponds to J1, and so on for the corresponding relationships of nodes in other defined meta-path types.

[0103] like Figure 3As shown, after constructing G=(V,E), H=(K,F), I=(U,W), step S3 is performed to construct a heterogeneous behavior network B={X,Y} based on V, where X is the node representing the knowledge operation behavior and Y is the behavior edge set between the knowledge operation behaviors. The behavior edge set represents the association between various types of knowledge operation behaviors. In the figure, an example of visualization B={X,Y} is given next to the local visualization H=(K,F) above nodes A and B. On the former node, there is only a meta-path IV with a high correlation of view-download, that is, The viewing status is equivalent to Q1, while downloading is equivalent to Q2. P1 represents the knowledge operation behavior edge between the two, that is, the reversible double-arrow relationship between viewing and downloading (not shown in the figure, but only represented by the dark blue vertical line simplified from the meta-path unidirectional arrow). The query-download-upload-question-answer of the latter node is analyzed in the same way, which will not be explained here.

[0104] It can be seen that the level of activity in this knowledge field in B is significantly higher than that in A, and the probability of being selected as an expert is also high. Figure 3 The acquired knowledge corresponding to the local H = (K, F) is recommended to be at the geographical location represented by B. This is the guiding advantage of knowledge acquisition achieved through dual networks.

[0105] In the following S4, five types of meta-paths are defined based on G = (V, E), H = (K, F), I = (U, W), and B = {X, Y}. Figure 2 As shown, the meta-path I consisting of nodes AD and M is given: The foregoing Figure 3 There are two types of meta-paths II (starting from i'), Figure 4 a-4e five-category metapath III, and Figure 3 There are two types of meta-paths IV defined in: view-download, query-download-upload-question-answer (here a horizontal line is used to simply represent a one-way arrow); and four types of traversal meta-paths IV: (two purple lines), Query and Download (two lake blue), and define the neighbors of each type of meta-path. Among the four types of traversal paths, the two purple ones are local traversal paths, and the two lake blue ones are full traversal paths. The focus nodes are B∈V, A∈V, and i∈U. Traversal Path They are all local crossing paths, and the lake blue For local crossing, the lake blue To travel to a different place.

[0106] For different locations of equipment, for example, for different locations B and A, the visualization local H = (K, F), that is, Figure 3The visualization local H=(K,F) represented in the figure shows that different devices are used to conduct certain experiments in the same knowledge field. For example, the principles are different. As a result, different research directions are formed in the two places. The local visualization B={X,Y} reflects the different high-correlation paths displayed by a new user input. This reflects that the research direction of the former B is more popular. This is another manifestation of the guiding advantage of knowledge acquisition achieved through dual networks.

[0107] Therefore, in summary, the dual attention mechanism bi-AM built on the dual network bi-N gives the attention analysis results from the perspective of the source of knowledge generation, knowledge itself, equipment, and knowledge operation behavior as a whole, and comprehensively analyzes the user's interest in knowledge and level of cognition reflected behind the user's new input, as well as targeted expert recommendations.

[0108] Then proceed to S5. First, establish the attention mechanism in G, H, I, and B according to meta-path I, meta-path II, meta-path III, and meta-path IV respectively. First, build the node attention mechanism, then build the semantic attention mechanism, and finally build the subgraph attention mechanism. The node attention mechanism, semantic attention mechanism, and subgraph attention mechanism built based on B are called heterogeneous behavior attention mechanisms.

[0109] Secondly, based on the traversing meta-path V, a traversing attention mechanism is established between G, H, I, and B. First, one of G=(V,E), H=(K,F), I=(U,W), and B={X,Y} is selected to select the focus node. According to the traversing meta-path, the node attention mechanism is also first constructed, then the semantic attention mechanism is constructed, and finally the subgraph attention mechanism is constructed. Then, the nodes of each of the remaining two of G=(V,E), H=(K,F), I=(U,W), and B={X,Y} are selected in turn as the focus nodes. Again, according to the traversing meta-path, the node attention mechanism is first constructed, and then the semantic attention mechanism is constructed. Finally, the bi-AM is obtained, wherein the number of nodes of meta-path I, meta-path II, meta-path III, and traversing meta-path IV are all adjustable between 3 and 20. Among them, the attention mechanism purely constructed based on the composite heterogeneous graph neural network in the bi-AM is CHGNNBAM, which together with HBAM constitutes the bi-AM. The bi-AM includes the attention mechanism of the node traversing path belonging to X. Figure 5 An outline of the steps S1-S4 is given.

[0110] like Figure 6 As shown, Figure 1 The node attention mechanism construction method represented in the red dashed rectangle is as follows:

[0111] The node attention mechanism construction method is as follows:

[0112] S5-1 For the node pair (i, j) connected by meta-path I, meta-path II, meta-path III, meta-path IV, and traversal path V, for L∈meta-path I, meta-path II, meta-path III, meta-path IV, and traversal path V, five types of node attention are constructed respectively. Att Node∈G (·), Att Node∈H (·), Att Node∈I (·), Att Node∈B (·), Att Node∈III (·) are five types of node attention deep neural networks, g i ,g j ∈V,h i ,h j ∈K,u i ,u j ∈U,x i ,x j ∈X,T III ,t j ∈V,K,U,X are the representations of nodes i and j when L∈ meta-path I, meta-path II, meta-path III, meta-path IV, and L∈ crosses path IV, respectively. III For node i, select the representation of the focus node T1;

[0113] S5-2 calculates the attention of all neighbors j of node i based on meta-path L, and the following formula (2) obtains the unified expression of the normalized coefficients of the attention of the four types of nodes in formula (1): || is the concatenation operator, is the node attention vector of meta-path L. The superscript T indicates transposition, N iL is the number of neighbor nodes of node i, LeakyReLU[·] is the activation function, L i and L j They are the unified expression of node i and j’s attention in four types of nodes, and the unified representation,

[0114] S5-3 Unified representation of features of computing node i based on four types of meta-paths L Finally, the attention mechanism is repeated K times to obtain Indicates that the attention mechanism is repeated K times for splicing;

[0115] The semantic attention mechanism is constructed as follows:

[0116] S5-4 Calculate the weights of the five types of meta-paths in

[0117] || represents the modulus of the corresponding node set. In the fifth formula, K, V, U, and X are determined according to the specific node set to which the selected attention node belongs (if the node belongs to any two or more of the four, it includes the modulus of the union of the nodes). The q with subscripts and the w and b with subscripts are the attention vector, weight, and bias representation of the corresponding semantics, respectively.

[0118] S5-5 For the subgraphs corresponding to K, V, U, and X, the attention of node i based on each meta-path L is calculated, and the following formula (5) obtains the unified expression of the normalization coefficients of the four types of semantic attention: Where W L is the unified expression of formula (4),

[0119] S5-6 Constructing four types of meta-path attention β L∈子图 =Att L∈子图 (Z L∈子图 ) (6), It means that for any path L∈ subgraph, the meta-path attention is represented by Z under the corresponding path L∈子图 Deep neural network Att via semantic attention L∈子图 (·) is learned, then the features of the semantic attention mechanism are uniformly expressed as Z 子图 =∑ L∈子图 β L∈子图 ·Z L∈子图 (7);

[0120] The subgraph attention mechanism is constructed as follows:

[0121] S5-7 is the same as the node attention mechanism and the semantic attention mechanism. For the knowledge domain to which the knowledge input by the knowledge generation source belongs and the subgraphs corresponding to the knowledge operation behaviors (that is, the subgraph set composed of the nodes corresponding to each domain and operation behavior, the subgraph set can be a subgraph composed of non-traversal paths or a subgraph composed of traversal paths), constructing the subgraph attention mechanism involves the following corresponding formulas (8)-(11):

[0122] is the subgraph association degree, is a deep neural network for subgraph attention,

[0123] W 子图 is the subgraph weight, with superscripts and subscripts w 子图 、b 子图 are the corresponding sub-graph attention vector, weight, and bias representation, respectively.

[0124]

[0125] Unified representation for subgraph attention mechanism.

[0126] Substituting Z into the combined network, i.e. the multi-layer induction mechanism, The cross entropy loss is calculated in L = -∑ l∈o Y l ln(C·Z l ), where l represents the number of the domain tag, o is the domain tag set, and Y l is the domain label numbered l, C is the node classifier parameter, By training each attention deep neural network until the cross entropy loss stabilizes and minimizes, we get bi-AM, where

[0127] The training methods include:

[0128] S5-8 The knowledge generation source inputs the knowledge generated according to the confidentiality classification into the input module regularly through the computer network, and the input module transmits it to the knowledge acquisition module to record the spatial location, generation time, and knowledge operation behavior of the knowledge generation; S5-9 The knowledge parsing module obtains the trained attention mechanisms of the five types of meta-paths in G, H, I, B, and between G, H, I, and B through step S5, receives the input knowledge and the generated spatial location, divides it into a training set and a verification set, and the ratio of the two is 9-3:4-1, processes the training set to form a node representation, forms a domain label according to the knowledge domain input by the knowledge generation source, and defines the number of meta-path nodes, obtains all corresponding meta-paths of the knowledge corresponding nodes, and predicts the corresponding normalization coefficients according to formulas (1)-(3). and According to formulas (4)-(7), the corresponding normalization coefficient β is obtained L∈子图 and Z 子图 , and according to formulas (8)-(11), the corresponding subgraph association degree, that is, the normalization coefficient and Z, and the cross entropy loss values ​​are calculated respectively,

[0129] S5-10 back-propagates the cross entropy loss value to optimize the parameters of each deep neural network, uses the verification set to verify the accuracy, and loops through steps S5-7-S5-9 to optimize the network parameters and verify the accuracy.

[0130] Among them, Figure 8 As shown, Figure 1 The blue dotted rectangle box represents the retrieval of knowledge context based on the newly input knowledge and the recorded knowledge operation behavior, and the analysis result of the dual attention mechanism is given based on the knowledge context and knowledge operation behavior. The specific method is:

[0131] S6 inputs knowledge into the trained bi-AM to obtain the predicted knowledge domain and related knowledge operation behaviors.

[0132] In the predicted knowledge domain, S7 draws the knowledge context according to the corresponding spatial location and generation time of the knowledge, and retrieves the drawn knowledge context.

[0133] S8: marking nodes and / or subgraphs and / or knowledge operation behaviors corresponding to the new input knowledge in the drawn knowledge context.

[0134] The predicted knowledge domain described in S6 is specifically based on the knowledge domain to which the domain label that may be involved belongs, represented by the largest normalized coefficient under each domain label. The method of drawing in S7 is to select all nodes and subgraphs belonging to the predicted knowledge domain in the geographic map, H, and I, and in the predicted knowledge domain, according to Figure 1 The chronological order of knowledge generation forms the historical context. S8 indicates that Figure 3 Geographical maps and visualizations are marked in H, I, and B.

[0135] like Fig. 9 As shown, in the predicted knowledge domain and at least one other domain corresponding to the knowledge, the G distribution, H distribution, I distribution, B distribution, and GHIB distribution are statistically analyzed on the geographic map; then, according to the knowledge generation source to which the user belongs, the G distribution, H distribution, I distribution, B distribution, and GHIB distribution corresponding to the knowledge domain involved in the knowledge generation source on the geographic map are displayed. In addition, the historical context his in each type of distribution selected is statistically analyzed, and finally the domain labels corresponding to the predicted knowledge domain and the at least one other domain are compared with the domain labels of the user input knowledge, and the comparison result is used as the final evaluation.

[0136] Figure 1-Figure 3 The distribution of G in some of the geographical maps is given in Figure 2 and 3 Part of the H distribution, I distribution, B distribution, and GHIB distribution are given.

[0137] Regarding relearning, forming an online expert system specifically includes the following steps:

[0138] S9 retrieves the knowledge operation behaviors of all users and knowledge generation sources, and according to the statistical results, takes the meta-path users and knowledge generation sources whose number of nodes in meta-path IV in distribution B exceeds 3 as online expert re-learning objects;

[0139] S10 retrieves the GHIB distribution corresponding to the object according to the GHIB distribution, and analyzes the G distribution, H distribution, I distribution, and B distribution therein, such as Fig.10 As shown;

[0140] S11 Fig.10 As shown, a convolutional neural network CNN is established, and the GHIB distribution is two-dimensionalized to form a distribution image ( Fig.11 ), the distribution images are divided into training set and validation set with a ratio of 5:2. The training set is used to train the convolutional neural network, and the validation set is used to verify the accuracy. The network parameters are optimized by calculating the loss function to obtain the trained model CNN, which identifies the distribution images as expert types and serves as the recommended experts to form an online expert system.

[0141] Fig.11 An implementation method for two-dimensionalizing G distribution, H distribution, and I distribution is given, that is, nodes A and B remain unchanged according to the double-arrow vector in the geographic digital map, and I distribution, H distribution, and B distribution are arranged near nodes A and B to form a two-dimensional distribution image, which can actually be arranged according to the specified vector between the starting nodes of the meta-path.

Claims

1. An artificial intelligence expert system based on a knowledge management system, characterized in that: include: The knowledge acquisition module is used to acquire knowledge from the knowledge generation source and record the spatial location and generation time of the knowledge generation. The knowledge includes knowledge related to the equipment related to the knowledge, specifically including the name of the equipment, the knowledge generation source and / or user to which it belongs, the purpose and technical principle related to the equipment, as well as scientific and technological resources, social and humanities resources, specifically including documents, pictures, videos, and audio. The knowledge parsing module is connected to the knowledge acquisition module, and is used to construct a dual network bi-N of a composite heterogeneous graph neural network CHGNN and a heterogeneous behavior network HBN, and to construct a dual attention mechanism bi-AM based on the attention mechanism CHGNNBAM of the composite heterogeneous graph neural network and the heterogeneous behavior attention mechanism HBAM by using the knowledge and spatial location input by the received knowledge generation source. After the construction is completed, the knowledge context is retrieved again based on the newly input knowledge and the recorded knowledge operation behavior, and the analysis results of the dual attention mechanism are given based on the knowledge context and the knowledge operation behavior, and bi-N and bi-AM are updated within a predetermined time period. The input module is connected to the computer network and is used to receive the newly input knowledge of the user, and input it into the knowledge analysis module through the computer network. The knowledge is clustered in a preset knowledge field with a field label. When the user newly inputs knowledge, it is necessary to select a target field for the newly input knowledge and select a field label for the knowledge. At the same time, the knowledge operation behavior is recorded according to the analysis of the user's new input behavior, including at least one of query, download, upload, question and answer, and discussion, and a combination thereof. The input module also has a network connection module connected to the computer network, which is used to realize the connection between the knowledge operation behavior and the Internet. The evaluation module is used to conduct a comprehensive analysis of the retrieved knowledge context and the analysis results of the dual attention mechanism, and give a final evaluation that is suitable for the user's true search meaning, as well as re-learning to form an online expert system, online knowledge transfer based on the online expert system, and, The display module is used to display the analysis results of the knowledge context, the dual attention mechanism, and the comprehensive analysis results and to operate the visualized bi-N. The knowledge generation source periodically inputs the knowledge generated according to the confidentiality classification into the knowledge acquisition module, and transmits the knowledge analysis module for the update; The method of constructing a dual network of a composite heterogeneous graph neural network and a heterogeneous behavior network and utilizing the knowledge and spatial locations input by the received knowledge generation source to construct a dual attention mechanism based on the attention mechanism of the composite heterogeneous graph neural network and the heterogeneous behavior attention mechanism includes: S1 establishes a geographical map, marks the spatial locations on the geographical map as nodes of the heterogeneous graph neural network, and constructs a heterogeneous graph network based on real spatial locations , forming a heterogeneous graph network neural network of real space locations, where It is a collection of nodes. Once recorded, it cannot be dragged or modified, but can be deleted and restored. is the set of spatial edges; S2 is built on Knowledge Heterogeneous Graph Networks , is a collection of knowledge nodes in different knowledge fields. is a knowledge edge set, and constructs Device Heterogeneous Graph Networks on , is a collection of device nodes representing a single device or a collection of devices. A collection of device edges, so that any new user input can be associated with the relevant device, thus providing a deeper understanding of the application of knowledge; S3 is built on Heterogeneous Behavior Network , is a node representing the knowledge operation behavior, is the set of behavior edges between knowledge operation behaviors; S4 , , , Construct meta paths in : , meta path : , meta path : , meta path : , traverse the metapath : And define the neighbors of the meta-path, that is, for any node , then connect through the meta path All nodes of are neighbors, where , , , , , , is a natural number, and , is the spatial relationship of the nodes, each of which is a double-arrow vector representing an adjacent node in the meta-path I on the geographic map; is the relationship between knowledge, each of which represents a meta-path The double-arrow relationship between adjacent nodes in the knowledge fields of scientific and technological resources and social and humanities resources, is the association between devices, each of which represents a meta-path The double-arrow relationship between adjacent nodes in the device-related knowledge domain, is the association between each knowledge operation behavior, each of which represents a meta-path There is a double arrow relationship between adjacent nodes in the knowledge operation behavior. For traversal metapath Neighbors define crossing neighbors in , , , Select the focus node from one of the ,or ,or ,or ,and ,set up and The value range is less than or equal to Greater than 0, set ,but ,in, for The complement of the range of To focus on the traversal relationship between the node and the traversal neighbor and between the traversal neighbors, it is expressed as and The double arrows between form a crossing edge, where ,and , The value range of , for , at this time crossing the meta path The nodes in the full-traversal metapath are distributed in all four heterogeneous graph networks; the nodes in the local-traversal metapath are partially in any two of the four heterogeneous graph networks, thus forming , , Composite heterogeneous graph neural network and The traversal path inside the pure composite heterogeneous graph neural network only describes the relationship between spatial locations, knowledge, and devices. The crossing path between them describes the association relationship of knowledge operation behaviors between two networks; S5 First According to the meta path , meta path , meta path , meta path Establish an attention mechanism, first build a node attention mechanism, then build a semantic attention mechanism, and finally build a subgraph attention mechanism. The constructed node attention mechanism, semantic attention mechanism, and subgraph attention mechanism are called heterogeneous behavior attention mechanisms. Secondly, based on the traversal metapath ,exist Establish a cross-attention mechanism between , , , One of them selects the node to focus on, and builds the node attention mechanism first according to the meta-path, then the semantic attention mechanism, and finally the sub-graph attention mechanism, and then selects , , , The remaining two nodes in each of the two are taken as the focus nodes, and the node attention mechanism is constructed again according to the meta-path traversal, and then the semantic attention mechanism is constructed, and finally the bi-AM is obtained, in which the meta-path , meta path , meta path , traverse the metapath The number of nodes can be adjusted between 3 and 20. The attention mechanism in bi-AM is purely based on the composite heterogeneous graph neural network. CHGNNBAM and HBAM together constitute bi-AM. Bi-AM includes The attention mechanism of the node traversal path; The node attention mechanism construction method is as follows: S5-1 For the meta-path , meta path , meta path , meta path , Crossing Path Connected node pairs ,for Meta Path , meta path , meta path , meta path , Crossing Path , construct five types of node attention respectively ， ， ， ， （1）, , , , , They are five types of deep neural networks with node attention. , , , , They are Meta Path , meta path , meta path , meta path , Crossing Path Node at time The representation of which, For Node Select as focus node ' S5-2 Compute Node All meta-path based Neighbors The following formula (2) gives the unified expression of the normalized coefficients of the attention of the four types of nodes in formula (1): （2）， is the concatenation operator represented by, Metapath The node attention vector of represents transpose, For Node The number of neighbor nodes of is the activation function, , and Node The unified expression of attention in four types of nodes, as well as the unified representation, S5-3 Compute Node Based on four types of meta-paths The unified representation of features , and finally repeat the attention mechanism times, get （3）， Indicates that the attention mechanism is repeated Splicing is performed again; The semantic attention mechanism is constructed as follows: S5-4 Calculate the weights of the five types of meta-paths （4） in represents the mode of the corresponding node set, in formula (4) Determined according to the specific node set to which the selected focus node belongs, with superscripts and subscripts , with subscript , are the corresponding semantic attention vector, weight, and bias representation, respectively. S5-5 For In the corresponding subgraph, the calculation node All based on each meta path The following formula (5) gives the unified expression of the normalization coefficients of the four types of semantic attention: （5）， in is the unified expression of formula (4), S5-6 Constructing four types of meta-path attention (6), For any path The meta-path attention under the corresponding path is represented as Deep Neural Networks via Semantic Attention After learning, the characteristics of the semantic attention mechanism are uniformly expressed as （7）； The subgraph attention mechanism is constructed as follows: S5-7 is the same as the node attention mechanism and the semantic attention mechanism. For the knowledge domain to which the knowledge input by the knowledge generation source belongs and the subgraphs corresponding to the knowledge operation behaviors, constructing the subgraph attention mechanism involves the following corresponding formulas (8)-(11): （8）， is the subgraph association degree, is a deep neural network for subgraph attention, （9）， is the subgraph weight, with superscripts and subscripts , , are the corresponding sub-graph attention vector, weight, and bias representation, respectively. (10) (11) is a unified representation of the subgraph attention mechanism, use Substitution , , , , , , Any one or a combination of them, i.e., multi-layer sensing mechanism, , , , is a real number, and the cross entropy loss is calculated ,in represents the number of the field label, is a set of domain labels, For the number The field label, is the node classifier parameter, , by training each attention deep neural network until the cross entropy loss stabilizes and minimizes, we get bi-AM, where , The training methods include: S5-8 The knowledge generation source inputs the knowledge generated according to the confidentiality classification into the input module regularly through the computer network, and the input module transmits the knowledge to the knowledge acquisition module to record the spatial location, generation time, and knowledge operation behavior of the knowledge generation; S5-9 The knowledge parsing module obtains the In, and The trained attention mechanisms of the five types of meta-paths traversed between them receive the input knowledge and the generated spatial locations, and divide them into training set and validation set, with a ratio of 9-3:4-1. The training set is processed to form node representation, and the domain label is formed according to the knowledge domain input by the knowledge generation source, and the number of meta-path nodes is defined. All the corresponding meta-paths of the knowledge corresponding nodes are obtained, and the corresponding normalization coefficients are predicted according to formulas (1)-(3) and , and the corresponding normalization coefficients are obtained according to formulas (4)-(7) and , and according to formulas (8)-(11), the corresponding subgraph association degree, that is, the normalization coefficient and , and the cross entropy loss values ​​are calculated respectively, S5-10 back-propagates the cross entropy loss value to optimize the parameters of each deep neural network, uses the verification set to verify the accuracy, and loops through steps S5-7-S5-9 to optimize the network parameters and verify the accuracy.

2. The system according to claim 1, characterized in that The sources of knowledge generation include: administrative departments, enterprises, institutions, companies, school research institutions, social groups and organizations, and individuals.

3. The system according to claim 1 or 2, characterized in that: The knowledge acquisition module parses the IP address or obtains the spatial location based on the name and address of the input knowledge generation source. The generation time is obtained by recording the time of receiving the knowledge through the knowledge acquisition module. The knowledge is reviewed manually or by computer to decide whether to adopt the knowledge and record the spatial location and generation time. The operation includes dragging the visualized bi-N, clicking on it to pop up a dialog box or search bar to search and / or view, modify, delete and / or restore the knowledge.

4. The system according to claim 1, characterized in that in, The knowledge context is retrieved based on the newly input knowledge and the recorded knowledge operation behavior, and the analysis results of the dual attention mechanism are given based on the knowledge context and knowledge operation behavior. The specific method is: S6 inputs knowledge into the trained bi-AM to obtain the predicted knowledge domain and related knowledge operation behaviors. In the predicted knowledge domain, S7 draws the knowledge context according to the corresponding spatial location and generation time of the knowledge, and retrieves the drawn knowledge context. S8 marks the nodes and / or subgraphs and / or knowledge operation behaviors corresponding to the new input knowledge in the drawn knowledge context, wherein the predicted knowledge domain in S6 is specifically based on the knowledge domain to which the domain label that may be involved belongs, which has the largest normalization coefficient under each domain label. The method of drawing in S7 is to draw the knowledge domain on the geographical map, , Select all nodes and / or subgraphs belonging to the predicted knowledge domain, and sort them according to the chronological order of knowledge generation in the predicted knowledge domain to form a historical context , S8 is marked on the geographical map and visualization The bid is out; The comprehensive analysis includes statistically analyzing at least one of the following types of distributions: a first distribution on the geographic map in the predicted knowledge domain corresponding to the knowledge and the at least one other domain, namely distributed, In the visualization The second distribution in distributed, In the visualization The third distribution in distributed, In the visualization The fourth distribution in distributed, The crossing distribution between distribution, and According to the user's knowledge generation source, the corresponding knowledge field of the knowledge generation source is displayed on the geographical map. distributed, distributed, distributed, distributed, At least one of the distributions, and The historical context of various distributions selected for statistical analysis , that is, in the selected distributions, each belongs to The historical context of knowledge in the nodes The final evaluation adapted to the real meaning of the user's search includes comparing the predicted knowledge domain and the domain labels corresponding to the at least one other domain with the domain labels of the user input knowledge, and taking the comparison result as the final evaluation.

5. The system according to claim 4, characterized in that Select to draw the knowledge context corresponding to at least one other field among the fields with non-maximum normalization coefficients under each field label, and the visualization is in the field belonging to The nodes are displayed on the corresponding geographical map.

6. The system according to claim 4 or 5, characterized in that: Relearning and forming an online expert system specifically includes the following steps: S9 retrieves the knowledge operation behaviors of all users and knowledge generation sources, and converts them into Distribution of metapaths Meta-path users and knowledge generation sources with more than 3 nodes are used as online expert re-learning objects; S10 according to distribution, the object corresponding to Distributed call, and analyze the distributed, distributed, distributed, At least one of the distributions; S11 builds a convolutional neural network CNN. distributed, distributed, distributed, At least one of the distributions and / or The distribution forms a distribution image, which is divided into a training set and a validation set with a ratio of 5-3:1-2. The training set is used to train the convolutional neural network, and the validation set is used to verify the accuracy. The network parameters are optimized by calculating the loss function to obtain a trained model CNN. The distribution image is identified as an expert type and used as a recommended expert to form an online expert system. Online knowledge transfer based on online expert system specifically includes the following steps: S12: obtaining at least one recommended expert by the online expert system based on the new input of the user; S13 The user selects at least one recommended expert and acquires the knowledge through the display module and the network connection module by performing the operation and / or knowledge operation behavior.

7. The system according to claim 6, characterized in that The method of forming the distribution image is to visualize distributed, distributed, distributed, distributed, At least one of the distributions is formed Distribution map, Distribution map, Distribution map, distribution map, and divide it into The distribution pattern of the distribution map is rearranged in two dimensions; The expert system includes a trained model CNN and a display module for distributed, distributed, distributed, distributed, At least one of the distributions is displayed and has a distribution of operations visualized to perform the operations on the knowledge.

8. The method for constructing an artificial intelligence expert system according to any one of claims 1 to 7, characterized in that: include: The first step is to construct the modules in the artificial intelligence expert system, realize the electrical connection between the modules, and connect the system to the computer network; The second step is to continuously collect knowledge generation sources and user input, and the system completes the construction of bi-N and bi-AM. distributed, distributed, distributed, Distribution and Distribution generation function testing, as well as knowledge context drawing and relearning CNN model training.

Images