An enterprise knowledge graph automatic construction and intelligent retrieval method

Abstract

The present application provides an enterprise knowledge graph automatic construction and intelligent retrieval method, the method comprises the following steps: collecting multi-source data from heterogeneous enterprise information systems, performing data cleaning and standardization processing; and based on the comprehensive scoring mechanism of field similarity and behavior semantic vector, merging the entities from different systems to generate a standard entity set with unique identification; based on the standard entity set, combining the scoring mechanism of task-type relationship and collaborative relationship, extracting semantic relationship from behavior records to construct enterprise knowledge graph structure; extracting representative semantic paths from the knowledge graph structure, screening high-quality paths through a path scoring model, and organizing them into a structured path index set; receiving a natural language query sentence, encoding the natural language query sentence into a semantic vector, matching it with the path index set, and executing a query in the graph and returning the result in combination with the permission control mechanism.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of data analysis, and particularly relates to an enterprise knowledge graph automatic construction and intelligent retrieval method. BACKGROUND

[0002] With the acceleration of enterprise digitization process, internal information systems are increasingly complex, involving conference systems, operation and maintenance log systems, work order transfer platforms, and personnel master data systems, and other types of heterogeneous environments. These systems each store key information such as personnel, tasks, and behavior records, but in actual business, the same object often appears with different identifiers, for example, the same employee may be recorded as a speaker's name in the conference system, as a login account in the log, and as a work number or email in the work order platform. This multi-source heterogeneity and identity fragmentation phenomenon leads to serious fragmentation and redundancy problems in the structured expression of enterprise knowledge. In addition, the task behavior within an enterprise often has event-driven and chain characteristics, and the completion of a business often involves multiple roles, multiple links, and multi-hop interactions. Traditional knowledge organization methods based only on static entities and single relationships cannot fully reflect the business flow logic, resulting in a lack of dynamicity and semantic integrity in the knowledge graph. Even in some existing systems that attempt to introduce relationship extraction or knowledge indexing, they mostly remain at the static template matching stage and cannot effectively deal with the problems of irregular semantic chains, serious path redundancy, and complex permission isolation in enterprise scenarios, resulting in insufficient accuracy of the final retrieval results, low response efficiency, and difficulty in meeting the requirements of actual business use. Therefore, there is an urgent need for a complete method that can realize the unification of multi-source entities, dynamic relationship modeling, semantic path organization, and natural language available queries to solve the problems of entity fragmentation, single relationship, path redundancy, and uncontrollable retrieval in existing technologies. SUMMARY

[0003] The purpose of the present application is to design an enterprise knowledge graph automatic construction and intelligent retrieval method that can realize the unified modeling of entities in a multi-source heterogeneous environment and extract task behavior relationships to construct a complete enterprise knowledge graph structure, and then form a path organization method highly aligned with natural language queries through a semantic path indexing mechanism, and finally accurately map the user's query statement to the graph path and return the results that meet the permission and semantic constraints.

[0004] To achieve the above purpose, the present application provides an enterprise knowledge graph automatic construction and intelligent retrieval method, which comprises:

[0005] Multi-source data is collected from heterogeneous enterprise information systems, and data cleaning and standardization processing is performed; the multi-source data includes conference record data, operation and maintenance log data, and personnel master data; and based on a comprehensive scoring mechanism of field similarity and behavior semantic vector, entities from different systems are merged to generate a standard entity set with unique identifiers;

[0006] Based on the standard entity set, the semantic relationship is extracted from the behavior record in combination with the scoring mechanism of task type relationship and collaborative type relationship, and the enterprise knowledge graph is constructed;

[0007] Representative semantic paths are extracted from the enterprise knowledge graph, high-quality paths are screened through a path scoring model, and a structured path index set is organized;

[0008] A natural language query statement is received, the natural language query statement is encoded into a semantic vector, the semantic vector is matched with the path index set, and the query is executed in the enterprise knowledge graph in combination with a permission control mechanism and the result is returned.

[0009] Further, the step of merging entities from different systems includes: constructing a matching pair between entity identifiers in different systems; calculating the structural similarity of each pair of entities in the name, number, and mailbox static fields; extracting the last three operation behavior texts of each entity, generating behavior semantic vectors through word-level attention and sentence-level aggregation network coding, and calculating their similarity; the structural similarity and behavior semantic similarity are weighted and summed to obtain a comprehensive matching score; if the score exceeds a preset threshold, the entities are merged and a unified number and source identifier are assigned.

[0010] Further, the generation method of behavior semantic vector is: using a word-level attention network to encode the behavior text, outputting a vector with a dimension of 128; then compressing through a sentence-level aggregation layer to generate a semantic vector with a dimension of 64; finally, the similarity of the behavior semantic vectors of two entities is calculated by cosine similarity.

[0011] Further, the scoring of the task type relationship specifically includes: the co-occurrence frequency normalized value of the entity in the original data; the time distance inverse of the behavior record between entities; the appearance order direction confidence in the behavior text; the role category difference penalty term between entities.

[0012] Further, the scoring of the collaborative type relationship specifically includes:

[0013] Co-occurrence frequency, time overlap ratio of joint speech in meeting records, and current shortest path distance between two entities in the graph; through the product of the co-occurrence frequency and the time overlap ratio, divided by one plus the product of the distance penalty term and the shortest path distance, the collaborative edge retention value is obtained, and whether to retain the edge is determined through a threshold.

[0014] Further, the path scoring model specifically includes:

[0015] semantic weight of each edge in the path; normalized redundancy of the number of occurrences of each edge in the graph structure; semantic directionality weight of the edge; business recurrence frequency of the path in historical real task data; path length; and sparse regularization term for penalizing high-frequency paths.

[0016] Further, each index record in the path index set comprises:

[0017] an ordered sequence composed of semantic labels of each edge in the path; a uniform number of the start entity of the path; a uniform number of the end entity of the path; an integrated score value of the path.

[0018] Further, the encoding mode of the semantic vector is:

[0019] the first layer attention network is used to extract context dependency between query keywords; the second layer attention network is used to perform semantic combination at the sentence level; and a 64-dimensional vector is output to represent the overall semantic intention of the query.

[0020] Further, the matching process with the path index set is to calculate a path matching score, and the calculation elements of the path matching score include:

[0021] cosine similarity between the query semantic vector and the path structure vector; score of the path itself; offset value of the actual length of the path and the average optimal path length of the system; and offset penalty coefficient for penalizing excessively long paths.

[0022] Further, the specific implementation of the permission control mechanism is:

[0023] a permission identification mask is set for each edge in the knowledge graph; during query execution, it is judged whether the two entity nodes connected by the edge are visible to the current user according to the access permission of the current user; only when both nodes are visible, the edge is considered as an effective edge to participate in query matching.

[0024] The beneficial technical effects of the present application are at least in the following aspects:

[0025] To solve the above problems, the present application provides an enterprise knowledge graph automatic construction and intelligent retrieval method, which realizes the merging of the same entity in different systems by introducing a comprehensive scoring mechanism based on field similarity and behavior semantic vector, generates a unique and traceable entity set; then designs a relationship scoring method combining occurrence frequency, time distance and role difference and a collaboration relationship density control mechanism to construct a graph structure of enterprise task and collaboration dual-type relationship; on this basis, a path scoring model is proposed, which fuses edge semantic weight, redundancy penalty and business recurrence frequency, compresses the multi-hop task chain into a representative semantic path index, and forms a structured index set that can be directly called for query; finally, through the joint matching mechanism of natural language query semantic vector and path index structure, the path length offset penalty and permission access mask are introduced to realize the closed loop from natural language intention to graph query execution. Through the above steps, the present application can significantly improve the unity of enterprise knowledge, the accuracy of business relationship expression, the efficiency of path retrieval and the controllability of query results, and construct an automatic construction and intelligent retrieval system of knowledge graph that conforms to the actual business semantics of enterprises and supports the use of natural language. BRIEF DESCRIPTION OF DRAWINGS

[0026] The present application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For ordinary skilled in the art, other drawings can be obtained without creative labor on the basis of the following drawings.

[0027] Figure 1 A flow chart of an enterprise knowledge graph automatic construction and intelligent retrieval method of the present application. DETAILED DESCRIPTION

[0028] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference signs represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are only used to explain the present application, and cannot be understood as a limitation on the present application.

[0029] In one or more embodiments, as shown in Figure 1 An enterprise knowledge graph automatic construction and intelligent retrieval method is disclosed, which comprises the following:

[0030] S1: Collect multi-source data from heterogeneous enterprise information systems, perform data cleaning and standardization processing; the multi-source data includes meeting record data, operation and maintenance log data and personnel master data; and based on a comprehensive scoring mechanism of field similarity and behavior semantic vector, the entities from different systems are merged to generate a standard entity set with unique identification;

[0031] Specifically, in an enterprise environment where multiple systems operate in parallel, the same employee, the same task, or the same organizational unit is often represented differently in different business systems. For example, in a meeting system, employee Zhang San might appear as "Speaker: Zhang San"; in a log system, it might be recorded as "User ID: U3856"; and in an operations and maintenance system, it might be identified by the operator's email address or login account. If these entities are used directly for graph construction without differentiation, it can easily lead to fragmented nodes in the graph structure and duplicate modeling of the same entity, affecting the accurate representation and retrieval of knowledge. Therefore, this step aims to merge synonymous entities appearing in different systems through field alignment and behavioral semantic modeling, generating a unified set of standard entities to provide accurate and unique node inputs for subsequent graph structure construction.

[0032] This step involves input from three main data sources. First, meeting minutes data, from the structured interface of the meeting system, includes fields such as name, speaking tag, and organization. Second, log data, from the structured log collection tool of the operations and maintenance system, includes fields such as operator account, operation type, and login terminal. Third, personnel file data from the master data system, includes fields such as employee number, unified account, organizational level, and on-duty status. After being collected uniformly through a pre-configured interface, this data enters the data cleaning process, completing field unification, encoding standardization, and character format normalization. Subsequently, based on each original entity representation, the system constructs all possible entity matching pairs and calculates their similarity in turn.

[0033] To determine whether two entities can be merged into the same actual object, the following scoring function is designed:

[0034] ;

[0035] in: For the first The entity and the first A comprehensive matching score between entities; For the first The and the first The structural similarity of entities on static fields (such as name, ID, email), with a value range of [value range missing]. ; For the first The and the first The similarity of entities in their behavioral semantic representations, with a value range of . The specific calculation is shown in the following formula; This is the weighted coefficient between structural field similarity and behavioral semantic similarity, ranging from... The value is usually taken as to , is preset by the system.

[0036] The calculation formula of the behavior semantic similarity is as follows:

[0037]

[0038] Wherein: is the context semantic vector of the entity , which is obtained by encoding the latest three operation behavior texts (such as conference speeches, operation description, and work order submission instructions) of the entity through a word-level attention network (the first layer, the vector dimension is ) and a sentence-level aggregation layer (the second layer, the vector dimension is ); is the context semantic vector of the entity , and the obtaining manner is the same as ; the numerator is the dot product of the two semantic vectors, and the denominator is the product of the respective lengths, the result is the cosine similarity, the value range is , but after vector normalization processing, the actual output is limited to the interval .

[0039] When (for example, the threshold ), it is considered that the entity and the entity should be merged into the same object. When merging, the number in the personnel master data system is preferentially used as the unified identity number, and if not, the unique identification code generation algorithm is used to automatically generate the number. The new entity formed after merging is denoted as , and its fields include: unified number; original identification set (used to track the source of merging); department attribution path; source system identification; latest behavior density level, etc.

[0040] The final output is a set of de-duplicated, structurally consistent standard entity sets:

[0041] ;

[0042] Wherein, each represents a merged entity node, has a unique number and complete attribute structure, and will be directly used as a graph node in subsequent graph structure construction.

[0043] This step solves the inconsistency of entities between multiple systems in an enterprise by constructing a unified identity recognition mechanism combining structural field matching and behavior semantic analysis. This scheme not only relies on traditional static fields such as name, number, and email for comparison, but also extracts the behavior semantic features of entities in actual business processes as a reference basis for entity merging, thereby significantly improving the matching accuracy and robustness. Through the behavior-driven entity merging method, a unified, unique, and context-rich entity node set is established.​

[0044] S2: Based on the standard entity set, the semantic relationship is extracted from the behavior record by combining the scoring mechanism of task type relationship and collaboration type relationship, and the enterprise knowledge graph is constructed;

[0045] Specifically, this step will be based on the set, combined with the actual enterprise running scene, task, behavior, operation and other behavior records, to automatically build a structured graph relationship reflecting the real business semantics of the enterprise.

[0046] Unlike general text relationship extraction tasks, this step is specially designed for the characteristics of enterprise internal knowledge flow: first, the behavior data of the enterprise has the characteristics of "event-driven", "incomplete structure", "unstable role", etc. For example, the speaker role in the meeting record is not clear, the responsibility attribution in the work order process is ambiguous, and the task operation in the log has no subject. Secondly, the behavior relationship in the enterprise scene is not only one type, but also "task type relationship" and "collaboration type relationship" coexist. Therefore, this step needs to capture the two types of relationships from semantics and model them separately in the graph structure.

[0047] The system first models the task type semantic relationship as a triple form , where represents the directional relationship of "responsible", "execute", "be assigned", etc. To ensure the credibility of the triple, combined with the structure matching, behavior time density, and role semantic consistency, the following triple scoring function is designed:

[0048] ;

[0049] Where: represents the comprehensive credibility score of the relationship between entities ; represents the frequency normalization result of and appearing together in the original data; represents the time distance inverse of the behavior records of the two, used to judge whether the behavior has actual interaction possibility; represents the directional confidence of the appearance order structure of the two in the original behavior text, for example, "Zhang San reports to Li Si" where Zhang San appears first, the directionality is higher; represents the role category number (such as department number, post type) to which the entity belongs, which is a role difference penalty term, used to suppress unreasonable cross-level edge construction; the coefficient is a weighting parameter, satisfying , is a role difference penalty factor, usually not greater than ​.

[0050] The last term of the above scoring function is the role difference item: traditional graph construction often only focuses on the degree of behavioral association, ignoring the influence of phenomena such as "cross-level authorization" and "misplaced responsibility" in the actual operation of the enterprise on the credibility of the relationship. Through the role number function The introduction of a penalty mechanism can significantly improve the accuracy of graph structure in enterprise internal process modeling.

[0051] For collaborative relationships (such as "joint processing", "joint discussion in meetings", and "joint operation of a system"), the system uses undirected edges to model and introduces a relationship density control function to ensure that an overly dense and redundant graph is not formed. For all candidate collaborative relationship triplets , if belongs to the collaborative relationship set, the following judgment value is calculated:

[0052] ;

[0053] Where: represents the structural retention value of the collaborative edge ; is consistent with the previous formula, representing the co-occurrence frequency; represents the time overlap ratio (taking values ) of the two in the meeting records; represents the current shortest path distance between the two in the graph structure; is the distance penalty term, which avoids the presence of a large number of redundant collaborative edges in the short path. When (the system set threshold), this undirected edge is retained in the graph, otherwise it is filtered to avoid excessive connection of the graph.

[0054] The system unifies all task-type and collaborative-type semantic edges that meet the scoring conditions into an edge set , and the final graph structure is:

[0055] ;

[0056] Where, is the standard entity set output in step one, is the task-type and collaborative-type semantic edge set selected according to the scoring function in this step. Each edge is accompanied by a relationship label (such as "execution", "collaboration", and "responsibility chain") and a scoring value or to support subsequent path analysis and semantic queries.

[0057] S3: Extract representative semantic paths from the enterprise knowledge graph, filter high-quality paths through a path scoring model, and organize them into a structured path index set;

[0058] Specifically, the goal of this step is to extract representative, semantically-mappable path sets from the graph structure , and organize them into an index structure that can be directly used by natural language queries. Unlike general graphs, enterprise business graphs have two characteristics:

[0059] Irregularity of paths: business processes often change in depth and structure due to factors such as organizational structure, event dynamics, and task assignment;

[0060] Strong dependence of semantic chain: some query intentions (such as "Did Zhang San participate in the handling of customer A's problem?") must rely on the "task-task-customer" chain, otherwise they cannot be matched.

[0061] Therefore, this step does not enumerate all paths from the graph, but selectively extracts high-quality semantic chains from the graph , and organizes them into a queryable index set .

[0062] The system first performs multi-hop path scanning in the graph . For each pair of entities , within a range of no more than (usually set to ), all path sets are enumerated . Each path represents a directed, relationally continuous node sequence.

[0063] To evaluate whether these paths have preservation value, the system designs a path scoring function that combines "structural stability" and "semantic consistency". This function considers the semantic weight of each edge in the path, the redundancy of the path, and the recurrence frequency of the path in historical business processes, and is calculated as follows:

[0064] ;

[0065] Where: represents the score value of path ; is the semantic weight of edge , derived from the scoring function in step two; is the normalized value of the number of occurrences of edge in the graph structure, indicating its redundancy, the higher the value, the more common but the lower the information content; is the edge semantic directionality weight (for example, "report" is more causal than "notify"); is the path length; represents the number of times path is repeated in historical real task data by business records; is the path sparsity regular term coefficient, which controls the punishment degree of high frequency path, usually taking .

[0066] The formula integrates edge-level semantic expression, redundancy penalty and historical reproducibility three mechanisms. Especially through term avoids "path drift" (i.e. common connection leads to low quality path flooding), and combines to enhance the priority reservation of real business chain.

[0067] The system includes all paths with scores higher than the set threshold as semantic path index. Each path will be converted into a semantic template, that is, a sequence of ordered combination of a group of relationship labels:

[0068] ;

[0069] Among them, represents the semantic label represented by the th edge in the path, such as "report", "review", "deploy", "archive", etc.

[0070] The system organizes each path template with its start and end entity number as a structured path index, forming the final index set:

[0071] ;

[0072] The index set supports quick path matching through "semantic path → start and end entity → reachability → score", which is the necessary interface for the natural language analysis system to map user intent to structured path in the next step.

[0073] The output result is stored in the intermediate service component, supporting high frequency path priority caching, low frequency path regular refreshing mechanism, and achieving the balance of query efficiency and system load.

[0074] S4: receiving a natural language query statement, encoding the natural language query statement into a semantic vector, matching with the path index set, combining the permission control mechanism to execute the query in the enterprise knowledge graph and return the result.

[0075] Specifically, the task of this step is to convert the user's natural language query request into path matching and entity retrieval behavior based on the graph structure, realizing the complete closed loop from language intent to structured knowledge return.

[0076] The user proposes a natural language query ("Did Zhang San participate in the problem handling of customer A?"), the system first encodes the query into a semantic vector. A double-layer attention mechanism network is used, the first layer to extract the context-dependent relationship between keywords, and the second layer to combine the semantics at the sentence level, finally outputting a query vector representing the overall structural semantics of the query intent.

[0077] The system then matches with the structural vectors of all paths in the path index set , where each is the weighted average of all semantic label embedding vectors in the path . The path matching score is calculated using the following formula:

[0078] ;

[0079] where: represents the matching score of the natural language query with the path ; represents the query sentence semantic vector; represents the path structure vector; is the path score from the path index structure; is the path length offset value, equal to the actual length of the path minus the average optimal path length; is the offset penalty coefficient, which controls the degree of punishment for structurally deep paths, usually taking .

[0080] After the path score is calculated, the system selects the top path index with the highest matching score and enters the graph query phase. The system traverses the graph structure , finds all start and end node pairs that meet the path structure , and constructs entity chains for result return.

[0081] To ensure the permission compliance of query results in enterprise multi-organization scenarios, the system introduces a graph edge access control mechanism in the query execution phase. Each graph edge will be accompanied by a permission identification mask, indicating whether the edge is open to the current query user. The judgment formula is as follows:

[0082] ;

[0083] where: represents the effectiveness of the edge in the current session; represents whether the original edge of the graph exists (1 or 0); indicates whether two entity nodes can be accessed by the current user (permission control identifier).

[0084] The mechanism has strong adaptability in enterprise tenant isolation, departmental data restriction scenarios, and ensures that the results are reasonable and compliant.

[0085] Specifically, when the user queries whether "Zhang San participated in the decision-making process of customer A's problem", the system determines that the semantic intention is the "person-task-decision-customer" path through semantic coding. The system finds the structure template in the path index, the starting point is the entity number of Zhang San , the end point is the entity number of customer A , the path score is as high as , and the matching score is ranked first, third and third. The system enters the graph traversal stage, retrieves all path chains that meet the template, and finally returns as the hit customer entity, and the complete path structure is attached as the response result.

[0086] The embodiment of the application also provides an enterprise knowledge graph automatic construction and intelligent retrieval device, comprising a processor, a memory and a computer program stored in the memory and configured to be executed by the processor, wherein the processor implements the steps in the above-mentioned enterprise knowledge graph automatic construction and intelligent retrieval method embodiment when executing the computer program, such as steps S1-S4 described in Figure 1 ; or, the processor implements the functions of each module in each system embodiment described above when executing the computer program.

[0087] Illustratively, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the application. The one or more modules can be a series of computer program instruction segments that can complete a specific function, which are used to describe the execution process of the computer program in the enterprise knowledge graph automatic construction and intelligent retrieval device.

[0088] The enterprise knowledge graph automatic construction and intelligent retrieval device can be a desktop computer, a notebook, a palm computer and a cloud server, etc. The enterprise knowledge graph automatic construction and intelligent retrieval device can include, but is not limited to, a processor, a memory. Those skilled in the art can understand that the enterprise knowledge graph automatic construction and intelligent retrieval device can also include input / output devices, network access devices, buses, etc.

[0089] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASAC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or can also be any conventional processor, etc. The processor is the control center of the enterprise knowledge graph automatic construction and intelligent retrieval device, and connects various parts of the enterprise knowledge graph automatic construction and intelligent retrieval device through various interfaces and lines.

[0090] The memory can be used to store the computer programs and / or modules, and the processor realizes various functions of the enterprise knowledge graph automatic construction and intelligent retrieval device by running or executing the computer programs and / or modules stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required by a function, etc.; and the data storage area can store data created according to the running of the air conditioner controller, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0091] The module integrated with the enterprise knowledge graph automatic construction and intelligent retrieval device can be stored in a computer readable storage medium if it is realized in the form of a software function unit and sold or used as an independent product. Based on this understanding, all or part of the processes in the above-mentioned embodiment methods can also be completed by a computer program instructing related hardware. The computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of the above-mentioned various method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0092] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned various method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0093] The above is the preferred embodiment of the present application. It should be noted that those of ordinary skill in the art can make several improvements and refinements without departing from the principles of the present application. These improvements and refinements are also considered within the scope of protection of the present application.

Claims

1. A method for automatic construction and intelligent retrieval of enterprise knowledge graphs, characterized in that, The method includes: Data from multiple sources is collected from heterogeneous enterprise information systems, and then cleaned and standardized. The multi-source data includes meeting minutes, operation and maintenance logs, and personnel master data. Based on a comprehensive scoring mechanism of field similarity and behavioral semantic vectors, entities from different systems are merged to generate a set of standard entities with unique identifiers. Based on the aforementioned standard entity set, and combined with scoring mechanisms for task-oriented and collaborative relationships, semantic relationships are extracted from behavioral records to construct an enterprise knowledge graph. The scoring of task-oriented relationships specifically includes: normalized co-occurrence frequency of entities in the original data; the reciprocal of the temporal distance between behavioral records between entities; confidence in the directional order of phrases appearing in the behavioral text; and penalties for differences in role categories between entities. The scoring of the collaborative relationship specifically includes: Co-occurrence frequency, time overlap ratio of joint speeches in meeting minutes, and current shortest path distance between two entities in the graph; by multiplying the co-occurrence frequency and time overlap ratio, dividing by one, and adding the product of distance penalty term and shortest path distance, the cooperative edge retention value is obtained, and a threshold is used to determine whether to retain the edge; Representative semantic paths are extracted from the enterprise knowledge graph, high-quality paths are selected through a path scoring model, and organized into a structured set of path indexes. The system receives natural language query statements, encodes them into semantic vectors, matches them with the path index set, executes the query in the enterprise knowledge graph in conjunction with the access control mechanism, and returns the results.

2. The method for automatic construction and intelligent retrieval of enterprise knowledge graphs according to claim 1, characterized in that, The steps for merging entities from different systems include: constructing matching pairs between entity identifiers in different systems; calculating the structural similarity of each pair of entities in the static fields of name, ID, and email; extracting the three most recent action texts of each entity, generating action semantic vectors through word-level attention and sentence-level aggregation network encoding, and calculating their similarity; weighted summing of structural similarity and action semantic similarity to obtain a comprehensive matching score; if the score exceeds a preset threshold, the entities are merged and assigned a unified ID and source identifier.

3. The method for automatic construction and intelligent retrieval of enterprise knowledge graphs according to claim 1, characterized in that, The behavioral semantic vector is generated as follows: the behavioral text is encoded using a word-level attention network to output a vector with a dimension of 128; then it is compressed through a sentence-level aggregation layer to generate a semantic vector with a dimension of 64; finally, the similarity between the behavioral semantic vectors of the two entities is calculated using cosine similarity.

4. The method for automatic construction and intelligent retrieval of enterprise knowledge graphs according to claim 1, characterized in that, The path scoring model specifically includes: The semantic weights of each edge in the path; the normalized redundancy of the number of times each edge appears in the graph structure; the semantic directional weights of the edges; the frequency of business recurrence of the path in historical real task data; the path length; and the sparse regularization term used to penalize high-frequency paths.

5. The method for automatic construction and intelligent retrieval of enterprise knowledge graphs according to claim 1, characterized in that, Each index record in the path index set includes: An ordered sequence of semantic labels for each edge in the path; a unique identifier for the starting entity of the path; a unique identifier for the ending entity of the path; and a comprehensive score for the path.

6. The method for automatic construction and intelligent retrieval of enterprise knowledge graphs according to claim 1, characterized in that, The encoding method of the semantic vector is as follows: The first-layer attention network is used to extract the contextual dependencies between query keywords; the second-layer attention network is used to perform sentence-level semantic combination; and a 64-dimensional vector is output to represent the overall semantic intent of the query.

7. The method for automatic construction and intelligent retrieval of enterprise knowledge graphs according to claim 1, characterized in that, The matching process with the path index set involves calculating a path matching score, the calculation elements of which include: The query calculates the cosine similarity between the semantic vector and the path structure vector; the path's own score; the offset between the actual path length and the system's average optimal path length; and the offset penalty coefficient used to penalize excessively long paths.

8. The method for automatic construction and intelligent retrieval of enterprise knowledge graphs according to claim 1, characterized in that, The specific implementation of the access control mechanism is as follows: Set an access control mask for each edge in the enterprise knowledge graph; during query execution, determine whether the two entity nodes connected to the edge are visible to the current user based on the current user's access permissions; the edge is considered a valid edge for query matching only when both edge nodes are visible.

Images