Industrial knowledge base dynamic construction method and system based on multi-source heterogeneous data

Abstract

The invention provides an industry knowledge base dynamic construction method and system based on multi-source heterogeneous data, and relates to the technical field of data processing.The method comprises the steps that target industry multi-source heterogeneous data, user historical behavior data, a natural language question text and a processing text are collected to generate a query data stream; mapping the user question into a query semantic vector, converting the user interaction behavior into a behavior feature vector, and generating a joint query vector; industry knowledge units are automatically extracted from multi-source heterogeneous data and external knowledge sources, and a dynamic knowledge topological graph is constructed; establishing a dynamic knowledge routing table, taking the joint query vector as a routing request, matching a target knowledge node, recording a query path and caching an association rule; and counting and querying path node transition probability to identify a hotspot path, mining a deep association rule, and reversely adjusting a knowledge topological graph structure, so that user intentions can be accurately matched, an industry knowledge base can be dynamically optimized, and knowledge retrieval and association efficiency can be improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for dynamically constructing an industry knowledge base based on multi-source heterogeneous data. Background Technology

[0002] The dynamic construction method of industry knowledge bases is a core technology for intelligent information services and knowledge management, capable of integrating industry data and supporting intelligent querying and decision-making. This technology has broad application prospects in scenarios such as intelligent customer service, professional consulting, and enterprise knowledge management.

[0003] Currently, existing technologies often employ a multi-source heterogeneous data industry knowledge base construction scheme based on retrieval enhancement. This involves first representing the text as a vector, and then performing similarity retrieval to complete knowledge matching. Such schemes pre-build static knowledge structures and perform knowledge extraction and query matching according to fixed rules.

[0004] However, this solution struggles to fully understand the user's true intent, suffers from low knowledge matching accuracy, and cannot adaptively optimize the knowledge structure based on query habits. It also struggles to balance query response speed with the depth of knowledge association, leading to knowledge lag and matching failures with long-term use. Therefore, existing technologies suffer from weak knowledge base adaptability and poor query matching performance. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for dynamically constructing an industry knowledge base based on multi-source heterogeneous data, so as to solve the problems of weak self-adaptation capability and poor query matching effect of the existing technology.

[0006] To address the aforementioned technical problems, firstly, this application provides a method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data, comprising: Collect multi-source heterogeneous data from target industry scenarios. The multi-source heterogeneous data includes user historical behavior data and natural language expression texts submitted by users to the intelligent agent. Process the natural language expression texts to generate query data streams. The system maps user questions in the query data stream into query semantic vectors, and converts user interaction behaviors extracted from user historical behavior data into behavioral feature vectors to generate joint query vectors that can represent user intent. Automated extraction of industry knowledge units from multi-source heterogeneous data and associated external knowledge sources to construct a dynamic knowledge topology graph; A dynamic knowledge routing table is constructed, using the joint query vector as the routing request. The target knowledge node with the highest semantic consistency with the current query is calculated and matched in the dynamic knowledge topology graph. The nodes traversed by the query are recorded as the query path, and the association rules on the query path are dynamically stored in the cache. By statistically analyzing the transition probabilities between nodes based on accumulated query paths, hotspot paths are identified. Deep association rules are then mined from these hotspot paths, and the nodes and connections in the dynamic knowledge topology graph are adjusted in reverse according to these association rules, thereby achieving adaptive construction and optimization of the industry knowledge base.

[0007] Optionally, based on the cumulative query path statistics, the transition probability between nodes is used to identify hotspot paths. Deep association rules are then mined from these hotspot paths, and the nodes and connections in the dynamic knowledge topology graph are adjusted in reverse according to these association rules to achieve adaptive construction and optimization of the industry knowledge base. This includes: Collect accumulated historical query paths, calculate the transition frequency between adjacent nodes in the query path, deduce the transition probability based on the transition frequency, and filter path segments with a transition probability higher than a threshold as hot paths. Analyze the co-occurrence patterns and order relationships of nodes on hotspot paths, and extract deep association rules from the co-occurrence patterns and order relationships; By applying deep association rules, nodes can be added or deleted in the dynamic knowledge topology graph, and the connection relationships between nodes can be updated.

[0008] Optionally, the co-occurrence patterns and order relationships of nodes on hotspot paths are analyzed, and deep association rules are extracted from the co-occurrence patterns and order relationships, including: Traverse the hotspot path, extract multiple consecutive nodes on the hotspot path, and combine the node sequence obtained from each extraction as a node combination to capture the co-occurrence pattern and order relationship of the nodes. Count the number of times each node combination appears in all historical query paths, and calculate the co-occurrence frequency of each node combination based on the number of occurrences; All node combinations are filtered based on a preset frequency threshold, and node combinations whose co-occurrence frequency exceeds the frequency threshold are retained as valid node combinations. For each valid combination of nodes, a deep association rule is generated from the predecessor node to the successor node based on the order of the nodes in the sequence.

[0009] Optionally, a dynamic knowledge routing table is constructed, using the joint query vector as the routing request. The target knowledge node with the highest semantic consistency with the current query is calculated and matched in the dynamic knowledge topology graph. The nodes traversed by the query are recorded as the query path, and the association rules on the query path are dynamically stored in the cache, including: Build and initialize a dynamic knowledge routing table, and input the joint query vector into the dynamic knowledge routing table; In the dynamic knowledge topology graph, starting from the joint query vector, the shortest path optimization algorithm is applied to traverse the nodes and connections of the dynamic knowledge topology graph, calculate the semantic distance between the joint query vector and each node, update the path weight based on the semantic distance, and iteratively find the optimal path to each node. Select the node with the smallest semantic distance as the target knowledge node, and backtrack the optimal path from the starting point to the target knowledge node, recording the node sequence on the optimal path as the query path; Extract the connection relationships between adjacent nodes on the query path and write these connection relationships as association rules into the cache.

[0010] Optionally, in the dynamic knowledge topology graph, starting from the joint query vector, a shortest path optimization algorithm is applied to traverse the nodes and connections of the dynamic knowledge topology graph, and the semantic distance between the joint query vector and each node is calculated. The path weights are updated based on the semantic distance, and the optimal path to each node is iteratively found, including: Set the joint query vector as the starting point for path calculation, and initialize a record table to store the cumulative distance between each node and the starting point; Starting from the starting point, in the dynamic knowledge topology graph, visit the unvisited nodes that are directly connected to the currently visited nodes in sequence; When visiting each unvisited node, calculate the semantic distance between the joint query vector and that node, and add the semantic distance to the cumulative distance to the corresponding predecessor node to obtain the cumulative path distance to the corresponding node. Compare the cumulative path distances of different paths to the same node, update the value of the corresponding node in the record table to the minimum cumulative path distance, and record the corresponding predecessor node to update the path weight. The process of repeatedly accessing, calculating, comparing, and updating continues until the optimal cumulative path distance of all nodes in the record table no longer changes, thus obtaining the optimal path from the starting point to each node.

[0011] Optionally, user questions in the query data stream are uniformly mapped to query semantic vectors, and user interaction behaviors extracted from user historical behavior data are converted into behavioral feature vectors to generate a joint query vector that can represent user intent, including: Based on the query data stream, extract the text string from the user's question, and use the text vector tool to convert the text string into a first numerical array as the query semantic vector; From the user's historical behavior data, we analyze the log records of user interaction behavior, extract the operation type and time series from the log records, construct a behavior pattern representation based on the operation type and time series, and use a vectorization tool to convert the behavior pattern representation into a second numerical array as a behavior feature vector. The query semantic vector and behavioral feature vector are fused together, and a joint query vector that can represent the user's intent is generated by array concatenation and weight adjustment.

[0012] Optionally, industry knowledge units are automatically extracted from multi-source heterogeneous data and associated external knowledge sources to construct a dynamic knowledge topology graph, including: By traversing multi-source heterogeneous data, core terms and descriptive fields in the data content are detected and packaged into industry knowledge units. Connect with related external knowledge sources, extract relevant information blocks from the external knowledge sources, and decompose the relevant information blocks as supplementary knowledge units; Using industry knowledge units and supplementary knowledge units as graph nodes, connecting lines are added based on the contextual relationships between industry knowledge units, and the semantic relevance between graph nodes is used as the initial edge weights to construct a dynamic knowledge topology graph.

[0013] Secondly, this application provides a dynamic construction system for industry knowledge bases based on multi-source heterogeneous data, including: The data acquisition module is used to collect multi-source heterogeneous data in the target industry scenario. The multi-source heterogeneous data includes user historical behavior data and natural language expression text submitted by users to the intelligent agent. The natural language expression text is processed to generate query data stream. The mapping module is used to uniformly map user questions in the query data stream into query semantic vectors, and at the same time convert user interaction behaviors extracted from user historical behavior data into behavioral feature vectors to generate joint query vectors that can represent user intent. The module is used to automatically extract industry knowledge units from multi-source heterogeneous data and associated external knowledge sources to build a dynamic knowledge topology graph. The matching module is used to build a dynamic knowledge routing table. It uses the joint query vector as the routing request, calculates and matches the target knowledge node with the highest semantic consistency with the current query in the dynamic knowledge topology graph, records the nodes passed through by the query as the query path, and dynamically stores the association rules on the query path into the cache. The identification module is used to identify hot paths based on the cumulative query path statistics of the transition probability between nodes, mine deep association rules from the hot paths, and adjust the nodes and connection relationships of the dynamic knowledge topology graph in reverse according to the association rules to achieve adaptive construction and optimization of the industry knowledge base.

[0014] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute computer programs to implement the steps of the method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data as described in the first aspect above.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can implement the steps of the method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data as described in the first aspect above.

[0016] The method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data provided in this application can provide a data foundation for intent understanding and knowledge query by collecting multi-source heterogeneous data of the target industry and processing natural language text to generate query data streams. By converting user questions into query semantic vectors and interactive behaviors into behavioral feature vectors and generating joint query vectors, it can accurately represent the user's true intent. By automatically extracting industry knowledge units to construct a dynamic knowledge topology graph, it can form a structured and scalable industry knowledge organization form. By constructing a dynamic knowledge routing table to match target knowledge nodes, recording query paths and caching association rules, it can improve knowledge matching accuracy and query response speed. By statistically analyzing query path transition probabilities to identify hot paths, mining deep association rules and adjusting the knowledge topology graph, it can achieve adaptive optimization of the knowledge base.

[0017] Furthermore, by collecting historical query paths and calculating node transfer frequency and probability, hotspot paths with high transfer probabilities are identified. The co-occurrence patterns and order relationships of nodes on these hotspot paths are analyzed to extract deep association rules. Based on these rules, nodes are added, deleted, and their connections are updated in the dynamic knowledge topology graph. This step accurately identifies high-frequency query logic, effectively uncovers implicit knowledge associations, and continuously iterates and optimizes the knowledge topology structure, making the industry knowledge base more aligned with actual query patterns and improving the stability and usability of knowledge services. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data, provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data, provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an industry knowledge base dynamic construction system based on multi-source heterogeneous data, provided in an embodiment of this application. Detailed Implementation

[0020] Existing industry knowledge base construction technologies mostly adopt static knowledge structures and fixed rule-based retrieval and matching methods, which make it difficult to accurately understand the user's true intent, resulting in low knowledge matching accuracy. At the same time, they cannot adaptively update and optimize the knowledge structure according to the user's actual query habits. Long-term use is prone to problems such as knowledge lag, low query efficiency, and poor matching effect, making it difficult to meet the needs of efficient, accurate, and dynamic knowledge services in industry scenarios.

[0021] To address the aforementioned issues, this application proposes a dynamic construction method for industry knowledge bases based on multi-source heterogeneous data. This method generates joint query vectors by fusing user natural language queries and historical behavior data to accurately represent user intent. It then achieves efficient knowledge matching and path caching based on a dynamic knowledge topology graph and a dynamic knowledge routing table. Furthermore, it iteratively optimizes the knowledge structure by statistically analyzing query hotspots and mining association rules. This solution effectively improves the accuracy of user intent understanding and knowledge matching precision, enabling adaptive updates and dynamic optimization of the knowledge base. It fundamentally solves the technical problems of weak knowledge base adaptability and poor query matching performance in existing technologies.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The core of this application is to provide a method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data. A flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes: S101. Collect multi-source heterogeneous data in the target industry scenario. The multi-source heterogeneous data includes user historical behavior data and natural language expression text submitted by users to the intelligent agent. Process the natural language expression text to generate query data stream.

[0024] The multi-source heterogeneous data includes user historical behavior data and natural language expressions submitted by users to the intelligent agent. The query data stream is a processed, uniformly formatted collection of user query data.

[0025] In one specific implementation, this step collects historical user behavior data from the target industry's business systems and interaction logs, while simultaneously collecting natural language text input by the user during interactions with the intelligent agent. The natural language text is then subjected to noise filtering, formatting, and sentence segmentation to form a standardized query data stream that can be directly used for subsequent calculations.

[0026] For example, in the smart after-sales consultation scenario of the home appliance industry, historical user behavior data such as user consultation records, equipment repair records, and product browsing and operation records are collected from the after-sales management system and smart customer service interaction logs of home appliance companies. Simultaneously, natural language text entered by users through the smart customer service window is captured in real time, such as "How to fix a refrigerator that isn't cooling?" or "What to do about abnormal noises during the washing machine's spin cycle?". This natural language text undergoes noise filtering, removing invalid characters, interjections, and meaningless short sentences. It is then standardized into a standard format of questions with interrogative intonation, breaking down long sentences into concise and easy-to-understand core query statements. Finally, a query data stream containing standardized entries such as "Refrigerator not cooling" and "Washing machine spin cycle noise" is generated.

[0027] S102. Map user questions in the query data stream into query semantic vectors, and convert user interaction behaviors extracted from user historical behavior data into behavioral feature vectors to generate joint query vectors that can represent user intent.

[0028] Among them, the query semantic vector is a numerical array formed after processing the user's question text, which is used to quantify the core meaning of the user's query; the behavioral feature vector is a numerical array formed after transforming the user's interaction behavior, which is used to quantify the user's operating habits and behavioral preferences; the joint query vector is a comprehensive numerical array obtained by integrating the query semantic vector and the behavioral feature vector, which can more comprehensively and accurately reflect the user's true query intent.

[0029] This step is designed in conjunction with industry-specific intelligent after-sales consultation scenarios. Its core purpose is to solve the problem that it is difficult to accurately determine the user's true needs based on a single user's problem or single behavioral data, so that subsequent knowledge matching can better meet the user's actual needs.

[0030] Optionally, step S102 may specifically include the following steps: S1021. Based on the query data stream, extract the text string from the user's question, and convert the text string into a first numerical array using the text vector tool, which serves as the query semantic vector.

[0031] Among them, the text-to-quantity tool is a tool that can convert natural language text into a computable numerical array, achieving the correspondence between text and numerical values ​​without complicated operations; the first numerical array is the query semantic vector, where each value in the array corresponds to a different semantic feature in the text, and the closer the values ​​are, the more similar the semantic meanings are.

[0032] S1022. From the user's historical behavior data, analyze the log records of user interaction behavior, extract the operation type and time series from the log records, construct a behavior pattern representation based on the operation type and time series, and use a vectorization tool to convert the behavior pattern representation into a second numerical array as a behavior feature vector.

[0033] Among them, the log records detailed data on the interaction process between users and the smart customer service and after-sales system for home appliances, including information such as operation behavior and operation time; operation type refers to the user's specific interactive actions, such as consultation, repair, and browsing products; behavior pattern is a description that reflects the user's behavioral habits after organizing the user's operation type and time sequence; the second numerical array is the behavior feature vector, which is used to quantify the user's behavioral patterns.

[0034] S1023. The query semantic vector and behavioral feature vector are fused together, and a joint query vector that can represent the user's intent is generated by array concatenation and weight adjustment.

[0035] In this embodiment, user question text is extracted and converted into query semantic vector in S1021, user historical behavior logs are parsed and converted into behavior feature vector in S1022, and the two vectors are concatenated and their weights are adjusted in S1023 to finally generate a joint query vector, thus forming a complete user intent representation process.

[0036] As a specific example, firstly, the text string of the user's question is extracted from the query data stream via S1021. Word2Vec is selected as the text-to-vector tool. This tool learns the semantic features of a large amount of industry-related text through pre-training, and can directly map natural language text into a fixed-dimensional numerical array without additional complex training, thereby obtaining the query semantic vector, providing semantic-level quantitative data for subsequent vector fusion. The choice of text-to-vector tool can be determined according to actual needs. In addition to Word2Vec, other commonly used text-to-vector tools can also be used; this application does not impose any limitations on this.

[0037] For example, user question text can be extracted from the query data stream. Word2Vec can be used as the text-to-vector tool. The extracted natural language question text is input into this tool and converted into a fixed-dimensional query semantic vector, such as [0.21, 0.35, 0.18, 0.42, 0.31]. This achieves accurate quantification of user question semantics and provides a foundation for subsequent vector fusion. The above example is only one example of this application. In practical applications, other text-to-vector tools can be selected according to requirements, and this application does not limit them.

[0038] Secondly, in step S1022, user interaction logs are extracted from historical user behavior data. A log parsing algorithm is used to extract operation types and time series from the logs. Based on the frequency of operation types and the chronological order of the time series, a behavioral pattern representation reflecting user habits is constructed. Then, using the Word2Vec vectorization tool from step S1021, the behavioral pattern representation is converted into behavioral feature vectors, corresponding to the previously obtained query semantic vectors, providing quantitative data at the behavioral level for subsequent vector fusion. The log parsing algorithm uses a conventional algorithm known in the field; the appropriate algorithm type can be selected based on the specific needs of the scenario, which will not be elaborated upon in this application.

[0039] For example, by analyzing a user's interaction logs over the past 30 days, the operation types and time series are extracted using a log parsing algorithm. A user behavior pattern representation is constructed by combining operation frequency and chronological order. Then, using the Word2Vec tool from S1021, the behavior pattern representation is converted into a behavior feature vector with the same dimension as the query semantic vector, for example, [0.38, 0.29, 0.45, 0.17, 0.33]. This complements the query semantic vector, improving the quantitative data of user intent. The above example is merely one example of this application; the behavior pattern construction method can be adjusted in practical applications, and this application does not limit it.

[0040] Finally, in step S1023, the query semantic vector obtained in S1021 and the behavioral feature vector obtained in S1022 are fused. First, an array concatenation algorithm is used to combine the two vectors sequentially into a longer numerical array. Then, based on the actual scenario requirements, a weight allocation algorithm is used to assign different weights to the two vectors, with the vector more closely related to the user intent having a higher weight. The joint query vector is obtained through weighted calculation, achieving a comprehensive representation of the user intent. The weighted calculation can use a weighted fusion formula, such as a linear weighted fusion formula, which will not be elaborated on in this application.

[0041] For example, the query semantic vector and behavioral feature vector are first concatenated sequentially into a longer numerical array. Then, based on the actual scenario requirements, a weight of 0.6 is assigned to the query semantic vector and a weight of 0.4 to the behavioral feature vector. Using a weighted fusion method known in the art, the resulting joint query vector is [0.28, 0.33, 0.29, 0.32, 0.32, 0.15, 0.12, 0.18, 0.07, 0.13]. This vector integrates the semantics of the user's question and historical behavioral features, accurately representing the user's true intent. The above example is merely one illustration of this application; the weight allocation can be adjusted in practical applications, and this application does not limit this.

[0042] In another specific implementation, the text-to-vector tool can use the BERT model, which learns relevant text features of the home appliance industry through pre-training. After inputting the user's question text, it directly outputs the corresponding query semantic vector. The vectorization process of the behavioral feature vector can be optimized in combination with the frequency of user operations, so that the vector is more in line with the user's long-term behavioral habits.

[0043] This application addresses the challenges of accurately capturing users' true intentions through single text queries and the inability of behavioral data to effectively assist in intent judgment by combining user question semantics with historical behavioral features, thus achieving precise representation of user intent. Compared to traditional single-vector representation methods, it adapts to the user interaction characteristics of specific industries, improving the accuracy and reliability of intent recognition and effectively avoiding inefficient subsequent services caused by misunderstandings of intent.

[0044] S103. Automatically extract industry knowledge units from multi-source heterogeneous data and associated external knowledge sources to construct a dynamic knowledge topology graph.

[0045] Among them, industry knowledge units are independent knowledge modules containing core information extracted from multi-source heterogeneous data, and are the basic units that constitute the industry knowledge base; external knowledge sources are external information sources related to the target industry, used to supplement and improve the knowledge content; dynamic knowledge topology graph is a structured knowledge presentation form with knowledge units as nodes and knowledge associations as connections, which can dynamically adjust the node and connection relationships according to subsequent needs.

[0046] The purpose of this step is to achieve automated extraction and structured organization of industry knowledge, solve the problems of knowledge being scattered and difficult to associate efficiently, and build a structured knowledge system.

[0047] Optionally, step S103 may specifically include the following steps: S1031. By traversing multi-source heterogeneous data, detect the core terms and description fields in the data content, and package the core terms and description fields into industry knowledge units.

[0048] Among them, core terms are keywords or phrases from multi-source heterogeneous data that can represent core industry information and are the core identifiers of knowledge units; description fields are related information used to explain and supplement core terms, making the content of knowledge units more complete. S1032, connect and associate external knowledge sources, capture relevant information blocks from external knowledge sources, and decompose relevant information blocks as supplementary knowledge units.

[0049] Among them, information blocks are information fragments with independent meaning from external knowledge sources; supplementary knowledge units are auxiliary knowledge modules extracted from external knowledge sources to improve industry knowledge units, and together with industry knowledge units, they constitute a complete knowledge system.

[0050] S1033. Using industry knowledge units and supplementary knowledge units as graph nodes, add connecting lines based on the contextual relationships between industry knowledge units, and use the semantic relevance between graph nodes as the initial edge weights to construct a dynamic knowledge topology graph.

[0051] In this context, a graph node is the basic unit that carries knowledge units in a dynamic knowledge topology graph, with each node corresponding to one knowledge unit; connecting lines are used to represent the relationships between different knowledge units; semantic relevance is an indicator that measures the semantic similarity or closeness of two knowledge units; edge weight is a quantitative representation of semantic relevance, used to reflect the strength of the relationship between knowledge units; and a dynamic knowledge topology graph is a structured knowledge graph that can adjust nodes, connecting lines, and edge weights according to subsequent rules.

[0052] In this embodiment, core industry knowledge units are first extracted from multi-source heterogeneous data, then knowledge gaps are supplemented by external knowledge sources, and finally the two types of knowledge units are structured and organized into a dynamic knowledge topology graph, forming a complete automated extraction and structured construction process for industry knowledge.

[0053] First, by traversing multi-source heterogeneous data through S1031, the core terms and descriptive fields in the data are identified using a keyword detection algorithm. The two are then integrated and packaged to form industry knowledge units, which serve as the core nodes for constructing the knowledge topology graph.

[0054] The keyword detection algorithm adopts a conventional algorithm known in the field. Its core is to accurately detect core terms and descriptive fields by identifying words that appear frequently in the data and are representative of the industry, as well as their corresponding descriptive information. The appropriate algorithm type can be selected according to the actual scenario requirements, which will not be elaborated on in this application.

[0055] For example, in the home appliance industry, core terms and their corresponding descriptive fields are extracted from multi-source heterogeneous data. Keyword detection algorithms identify core terms such as refrigerator refrigeration and washing machine spin-drying, which are then combined with their corresponding fault causes and solutions to form industry knowledge units. The above example is merely one illustration of this application; in practical applications, the detection method for core terms can be adjusted according to requirements, and this application does not impose any limitations on it.

[0056] Secondly, by establishing a connection with the associated external knowledge source through S1032, web crawling technology is used to capture relevant information blocks from the external knowledge source, and the information blocks are decomposed to obtain supplementary knowledge units, thus filling the knowledge gaps in multi-source heterogeneous data.

[0057] Among them, the web crawler technology adopts conventional technology known in the field. The core is to traverse and associate external knowledge sources through automated programs to crawl effective information fragments related to the industry. The crawling strategy can be adjusted according to the type of external knowledge source to improve the targeting and efficiency of information crawling. This application will not elaborate on this further.

[0058] For example, in the home appliance industry, connections can be established with external knowledge sources such as professional industry databases and official technical documents. Web crawling technology can be used to extract information blocks related to products like refrigerators and washing machines. These extracted information blocks can then be broken down and filtered to extract effective information such as product repair specifications and common troubleshooting techniques as supplementary knowledge units, complementing the industry knowledge units obtained earlier. The above example is merely one illustration of this application; in practical applications, the information extraction and decomposition methods can be adjusted according to needs, and this application does not impose any limitations on this.

[0059] Finally, industry knowledge units and supplementary knowledge units are used as graph nodes through S1033. Connecting lines are added based on the contextual association between each knowledge unit. The semantic similarity calculation method known in the field is used to obtain the semantic association degree between nodes, which is used as the initial edge weight. Finally, a dynamic knowledge topology graph is constructed, forming a complete knowledge structured organization process.

[0060] The semantic similarity calculation can employ well-known methods in the field, such as the cosine similarity formula. The core is to determine the edge weight by quantifying the degree of semantic association between two knowledge units. The appropriate calculation method can be selected based on actual needs, which will not be elaborated on in this application.

[0061] For example, in the home appliance industry scenario, industry knowledge units such as refrigerator refrigeration fault handling and washing machine spin-drying noise repair, as well as supplementary knowledge units such as product repair specifications, are used as graph nodes. Connecting lines are added based on the contextual relationships between these knowledge units. The semantic similarity between nodes is calculated using a known semantic similarity method, and this degree of semantic association is used as the initial edge weights to construct a dynamic knowledge topology graph, thus achieving the structured organization of knowledge in the home appliance industry. The above example is merely one illustration of this application; the setting of edge weights can be adjusted in practical applications, and this application does not limit this.

[0062] In another specific implementation, the detection of core terms and description fields can be optimized using natural language processing technology to improve the accuracy of detection; the crawling of external knowledge sources can be combined with interface calling methods to improve the efficiency and security of information crawling; the calculation of semantic relevance can be combined with industry characteristics to adjust parameters so that the edge weights are more in line with the industry knowledge relevance rules.

[0063] This application solves the problems of low efficiency, large knowledge gaps, and chaotic knowledge organization caused by manual knowledge extraction by automatically extracting industry knowledge units and supplementing them with external knowledge, thus realizing the automated and structured organization of industry knowledge. Compared with the traditional method of manually constructing knowledge graphs, it improves the efficiency and accuracy of knowledge extraction. The constructed dynamic knowledge topology graph can clearly present knowledge relationships, effectively avoiding the problem of inefficient knowledge retrieval caused by scattered knowledge and unclear relationships.

[0064] S104. Construct a dynamic knowledge routing table, using the joint query vector as the routing request, calculate and match the target knowledge node with the highest semantic consistency with the current query in the dynamic knowledge topology graph, record the nodes traversed by the query as the query path, and dynamically store the association rules on the query path into the cache.

[0065] The dynamic knowledge routing table is a structured table used to manage knowledge query paths and record node relationships, and can be dynamically updated according to query requirements; the routing request is the instruction that triggers knowledge query and matching, which is acted by the joint query vector; the target knowledge node is the knowledge unit node in the dynamic knowledge topology graph that has the highest semantic consistency with the current query; the query path is the sequence of all nodes traversed from the query starting point to the target knowledge node; the association rule is the connection logic and association relationship between adjacent nodes on the query path; and the cache is a storage area used to temporarily store association rules, which can enable fast retrieval for subsequent queries.

[0066] The purpose of this step is to achieve accurate matching between query semantics and knowledge nodes, optimize query efficiency, enable efficient reuse of association rules, and solve the problems of inefficient knowledge matching and cumbersome association rule calling.

[0067] Optionally, step S104 may specifically include the following steps: S1041. Construct and initialize the dynamic knowledge routing table, and input the joint query vector into the dynamic knowledge routing table.

[0068] Among them, building a dynamic knowledge routing table refers to creating a structured table to record the paths, relationships, and weights of knowledge nodes.

[0069] S1042. In the dynamic knowledge topology graph, starting from the joint query vector, the shortest path optimization algorithm is applied to traverse the nodes and connections of the dynamic knowledge topology graph, and the semantic distance between the joint query vector and each node is calculated. The path weight is updated according to the semantic distance, and the optimal path to each node is iteratively found.

[0070] Specifically, step S1042 may include the following process: setting the joint query vector as the starting point for path calculation, and initializing a record table storing the cumulative distance between each node and the starting point; starting from the starting point, in the dynamic knowledge topology graph, sequentially visiting unvisited nodes directly connected to the currently visited nodes; when visiting each unvisited node, calculating the semantic distance between the joint query vector and the node, adding the semantic distance to the cumulative distance to the corresponding predecessor node to obtain the cumulative path distance to the corresponding node; comparing the cumulative path distances of different paths to the same node, updating the value of the corresponding node in the record table to the minimum cumulative path distance, and recording the corresponding predecessor node to update the path weight; repeating the process of visiting, calculating, comparing, and updating until the optimal cumulative path distance of all nodes in the record table no longer changes, thus obtaining the optimal path from the starting point to each node.

[0071] In the above steps, the shortest path optimization algorithm is used to find the shortest and most efficient path from the starting point to each node; semantic distance is an indicator that measures the semantic consistency between the joint query vector and each knowledge node. The smaller the semantic distance, the higher the semantic consistency between the two; path weight is a quantitative representation of the quality of the query path, which is calculated from the semantic distance and can be dynamically updated; the optimal path is the path from the starting point to each node that has the minimum cumulative semantic distance and the highest semantic consistency.

[0072] A predecessor node is a node that is located before the current node and is directly connected to the current node in the query path; the cumulative distance record table is a table used to store the cumulative semantic distance between each node and the starting point, and it can be updated in real time; the cumulative path distance is the sum of the semantic distances on all paths from the starting point to the current node, and it is the core indicator for judging the quality of a path.

[0073] S1043. Select the node with the smallest semantic distance as the target knowledge node, and backtrack the optimal path from the starting point to the target knowledge node, and record the node sequence on the optimal path as the query path.

[0074] Among them, the target knowledge node is the node with the highest semantic consistency with the joint query vector in the dynamic knowledge topology graph, and the node sequence with the smallest semantic distance from the starting point refers to the ordered set formed by arranging all nodes on the optimal path in the order of access, i.e., the query path.

[0075] S1044. Extract the connection relationship between adjacent nodes on the query path and write the connection relationship as an association rule into the cache.

[0076] In this embodiment, a dynamic knowledge routing table is first constructed and initialized, and a joint query vector is input. Then, the knowledge topology graph is traversed through the shortest path optimization algorithm, the semantic distance is calculated, and the optimal path to each node is found. Next, the target knowledge node is selected and the query path is recorded. Finally, the association rules are extracted and written to the cache, forming a complete knowledge query, matching, and association rule storage process.

[0077] First, a structured dynamic knowledge routing table is constructed using S1041. The table is initialized by clearing initial redundant data and setting basic parameters. Then, the joint query vector obtained earlier is used as a routing request and input into the dynamic knowledge routing table to trigger the subsequent path calculation and node matching process.

[0078] For example, in the home appliance industry, a dynamic knowledge routing table is constructed to manage the query paths for home appliance knowledge. The table parameters are initialized to ensure that it can receive routing requests normally. The joint query vector representing the user's intent to inquire about the refrigerator not cooling is input into the routing table to complete the submission of the routing request. The above example is only one example of this application. In practical applications, the construction parameters of the routing table can be adjusted according to the requirements, and this application does not limit this.

[0079] Secondly, using the input joint query vector as the starting point, the shortest path optimization algorithm is applied through S1042 to traverse all nodes and connecting lines in the dynamic knowledge topology graph, calculate the semantic distance between the joint query vector and each knowledge node, update the path weight based on the semantic distance, and find the optimal path to each node through the iterative process.

[0080] The shortest path optimization algorithm employs a conventional algorithm known in the field. Its core principle is to accurately find the optimal path by traversing nodes, calculating cumulative path distances, and comparing path merits. The algorithm parameters can be adjusted according to the complexity of the knowledge topology graph, which will not be elaborated upon in this application. Semantic distance calculation can employ a method known in the field. Its core principle is to quantify the semantic consistency between the joint query vector and each knowledge node; the smaller the semantic distance, the higher the consistency. This will not be elaborated upon in this application either.

[0081] For example, in the home appliance industry scenario, the joint query vector corresponding to the refrigerator not cooling consultation is set as the starting point for path calculation. An initial cumulative distance record table is initialized to record the initial cumulative distance between all knowledge nodes and the starting point. Starting from the starting point, unvisited nodes directly connected to it, such as refrigerator cooling fault and refrigerator maintenance, are visited sequentially. When visiting the refrigerator cooling fault node, the semantic distance between the joint query vector and that node is calculated. This semantic distance is added to the direct distance from the starting point to that node to obtain the cumulative path distance, which is then written to the record table. When visiting the refrigerator maintenance node, the cumulative path distance is calculated in the same way.

[0082] If multiple paths lead to the same node, such as both the refrigerator refrigeration failure node and the refrigerator compressor repair node, the cumulative path distance of each path is calculated. The cumulative distance of the refrigerator refrigerant replenishment node in the record table is updated to the smaller value, and the corresponding predecessor node is recorded to update the path weight. This process of accessing, calculating, comparing, and updating is repeated until the optimal cumulative path distance of all nodes no longer changes, ultimately yielding the optimal path from the starting point to each node. The above example is merely one illustration of this application; in practical applications, the number of iterations can be adjusted according to the number of nodes, and this application does not impose any limitations on this.

[0083] Then, through S1043, the node with the smallest semantic distance to the joint query vector among all knowledge nodes is selected as the target knowledge node. The optimal path from the starting point to the target knowledge node is found by backtracking from the target knowledge node. All nodes on the path are arranged in the order of access and recorded as the query path.

[0084] For example, in the home appliance industry, after semantic distance calculation, the node representing a refrigerator refrigeration fault has the smallest semantic distance to the joint query vector, and is therefore selected as the target knowledge node. Tracing back from this node, the optimal path from the starting point of the joint query vector to this node is found. The node sequence on this path is the joint query vector—refrigerator refrigeration fault—insufficient refrigerant, and this node sequence is recorded as the query path. The above example is merely one illustration of this application. In practical applications, the selection of the target knowledge node can be adjusted based on the semantic distance calculation results, and this application does not impose any limitations on this.

[0085] Finally, the connection relationships between adjacent nodes on the query path are extracted through S1044, and these connection relationships are written into the high-speed cache as association rules to realize the temporary storage and fast reuse of association rules.

[0086] For example, in the home appliance industry scenario, the connection relationships between adjacent nodes on the joint query vector of the query path—refrigerator cooling failure—refrigerator refrigerant shortage are extracted. That is, the association relationship between the joint query vector and refrigerator cooling failure, and the association relationship between refrigerator cooling failure and refrigerator refrigerant shortage. These two connection relationships are used as association rules and written into the high-speed cache for quick access in subsequent similar queries. The above example is only one example of this application. In practical applications, the method of extracting association rules can be adjusted according to the number of nodes in the query path. This application does not limit this.

[0087] In another specific implementation, the initialization of the dynamic knowledge routing table can be combined with the characteristics of industry knowledge to set initial weights, thereby improving the efficiency of subsequent path calculation; the shortest path optimization algorithm can be an improved algorithm that is more adapted to the knowledge topology graph, thereby improving the speed of finding the optimal path; the high-speed cache can be configured with cache eviction rules to avoid cache redundancy and ensure the efficiency of association rule invocation.

[0088] This application addresses the issues of low knowledge matching accuracy, chaotic query paths, and poor reusability of association rules by constructing a dynamic knowledge routing table, optimizing path calculation and node matching, and storing association rules. It achieves accurate matching of query semantics and knowledge nodes, and efficient reuse of association rules. This improves the accuracy of knowledge matching and query efficiency. The dynamically stored association rules reduce redundant calculations and effectively avoid query inefficiencies caused by matching deviations and path redundancy.

[0089] S105. Based on the cumulative query path statistics, the transition probability between nodes is used to identify hot paths. Deep association rules are mined from the hot paths, and the nodes and connection relationships of the dynamic knowledge topology graph are adjusted in reverse according to the association rules to achieve adaptive construction and optimization of the industry knowledge base.

[0090] Among them, the cumulative query path refers to the set of query paths of all records in multiple knowledge queries; the transition probability refers to the probability of moving from one node to another between adjacent nodes in the query path, which is used to measure the frequency of association between nodes; the hot path refers to the path segment with a transition probability higher than a preset threshold and is frequently queried, reflecting the high-frequency query needs of users; the deep association rule is the inherent and regular connection logic between nodes mined from the hot path, which is different from the surface connection relationship.

[0091] The purpose of this step is to uncover user query patterns, optimize the knowledge topology, solve the problems of the knowledge base being unable to adapt to users' high-frequency needs and having a fixed and rigid structure, and improve the usability and query efficiency of the knowledge base.

[0092] Optionally, such as Figure 2 As shown, step S105 may specifically include the following steps: S1051. Collect accumulated historical query paths, calculate the transfer frequency between adjacent nodes in the query path, deduce the transfer probability based on the transfer frequency, and filter path segments with transfer probabilities higher than the threshold as hot paths.

[0093] Among them, historical query paths refer to every query path recorded in all past knowledge query processes, forming an accumulated set of query paths; transfer frequency refers to the number of times a transfer occurs between two adjacent nodes in all historical query paths, the more transfers, the higher the transfer frequency; transfer probability is a quantitative indicator derived from transfer frequency, reflecting the possibility of a transfer between adjacent nodes; threshold is a preset judgment standard used to filter out path segments with high transfer probabilities; hotspot paths are path segments with transfer probabilities higher than the threshold and high user query frequency, which are the core foundation for mining deep association rules.

[0094] S1052. Analyze the co-occurrence patterns and order relationships of nodes on hotspot paths, and extract deep association rules from the co-occurrence patterns and order relationships.

[0095] Specifically, step S1052 may include the following process: traversing hotspot paths, extracting multiple consecutively occurring nodes on the hotspot paths, and treating each extracted node sequence as a node combination to capture the co-occurrence pattern and order relationship of the nodes; counting the number of times each node combination appears in all historical query paths, and calculating the co-occurrence frequency of each node combination based on the number of occurrences; filtering all node combinations according to a preset frequency threshold, and retaining node combinations whose co-occurrence frequency exceeds the frequency threshold as valid node combinations; for each valid node combination, generating a deep association rule from the predecessor node to the successor node based on the order of the nodes in the sequence.

[0096] In the above steps, co-occurrence pattern refers to the combination pattern in which multiple nodes appear simultaneously and are jointly formed in the hot spot path, reflecting the closeness of the relationship between nodes; sequential relationship refers to the ordered relationship formed by the nodes in the hot spot path according to the order of query, reflecting the logical order of user query; deep association rule is the regular node connection logic mined from co-occurrence pattern and sequential relationship, which can reflect the inherent relationship between user query habits and knowledge, and is more practical than the surface connection relationship.

[0097] A node combination refers to an ordered set of multiple consecutively occurring nodes extracted from a hotspot path, used to capture the co-occurrence and sequence characteristics of nodes; co-occurrence frequency refers to the ratio of the number of times a node combination appears in all historical query paths to the total number of historical query paths, used to measure the frequency of node combinations; an effective node combination refers to a node combination whose co-occurrence frequency exceeds a preset frequency threshold, possessing high relevance and practicality; deep association rules are rules generated based on the sequential relationship of effective node combinations, clarifying the pointing relationship between predecessor and successor nodes, reflecting the deep logic between knowledge.

[0098] S1053. Apply deep association rules to add or delete nodes in the dynamic knowledge topology graph and update the connection relationships between nodes.

[0099] Updating the connection relationships between nodes refers to adding or deleting connection lines between nodes and adjusting edge weights according to deep association rules, so that the structure of the topology graph is more in line with user query habits.

[0100] In this embodiment, historical query paths are first collected, transition probabilities are calculated, and hot paths are filtered. Then, deep association rules are mined from the hot paths. Finally, the dynamic knowledge topology graph is adjusted based on the deep association rules to form a complete adaptive optimization process for the industry knowledge base, thereby achieving dynamic improvement of the knowledge base.

[0101] First, historical query paths accumulated during all past knowledge queries are collected through S1051. The number of transfers between adjacent nodes in each query path is counted, the transfer frequency is calculated, and the transfer probability between adjacent nodes is derived based on the transfer frequency. A preset threshold is set, and path segments with a transfer probability higher than the threshold are selected as hot paths.

[0102] The calculation of the transition probability adopts conventional calculation methods known in the art. The core is to obtain the transition probability by the ratio of the transition frequency to the total number of transitions. The calculation parameters can be adjusted according to the actual needs of the scenario, which will not be elaborated on in this application. The threshold setting can be determined in combination with the characteristics of industry queries and actual needs. There is no fixed standard, and this application does not limit it.

[0103] For example, in the home appliance industry, all historical query paths from past user inquiries are collected. These paths include various node sequences such as refrigerator cooling malfunction—insufficient refrigerant, washing machine spin-drying noise—washing machine bearing damage. The number of transitions between adjacent nodes is counted. For instance, the number of transitions from refrigerator cooling malfunction to insufficient refrigerant is 120, and the total number of transitions between all adjacent nodes is 1000. The transition frequency is calculated as 120 / 1000 = 0.12, and the transition probability is derived to be 0.12. A preset threshold of 0.08 is set. Since 0.12 is higher than the threshold, the path segment "refrigerator cooling malfunction—insufficient refrigerant" is identified as a hotspot path. The above example is merely one illustration of this application. In practical applications, the calculation method can be adjusted based on historical query data, and this application does not impose any limitations on this.

[0104] Secondly, by traversing all the hot paths selected by S1052, the co-occurrence patterns and order relationships of nodes on the hot paths are analyzed. Through a series of processes, deep association rules are mined and extracted so that the association rules can reflect the user query patterns and the inherent logic between knowledge.

[0105] The co-occurrence frequency is calculated using conventional methods known in the field. The core method is to obtain the co-occurrence frequency by the ratio of the number of times the node combination appears to the total number of historical query paths. The calculation accuracy can be adjusted according to actual needs, which will not be elaborated in detail in this application.

[0106] For example, in the home appliance industry scenario, all hotspot paths are traversed, including refrigerator cooling failure—insufficient refrigerant, washing machine spin-drying noise—washing machine bearing damage, etc. Multiple consecutive nodes appearing on the hotspot paths are extracted, with 2-3 consecutive nodes extracted as a node combination each time, capturing the co-occurrence pattern and order relationship of the nodes. The number of times each node combination appears in all historical query paths is counted; for example, refrigerator cooling failure—insufficient refrigerant appears 80 times, and washing machine spin-drying noise—washing machine bearing damage appears 75 times. Based on the frequency, the co-occurrence frequency of each node combination is calculated. Assuming a total of 500 historical query paths, the co-occurrence frequencies of the two combinations are 0.16 and 0.15, respectively. A preset frequency threshold of 0.1 is set, and two node combinations with co-occurrence frequencies exceeding the threshold are selected as valid node combinations. For each valid node combination, deep association rules are generated based on the node order, i.e., refrigerator cooling failure corresponds to insufficient refrigerant, and washing machine spin-drying noise corresponds to washing machine bearing damage. The above example is only one example of this application. In practical applications, the extraction rules can be adjusted according to the length of the node combination, and this application does not limit this.

[0107] Finally, the deep association rules mined are applied to the dynamic knowledge topology graph through S1053. Nodes are added or deleted according to the rules, and the connection relationships between nodes are updated to optimize the dynamic knowledge topology graph, thereby completing the adaptive construction and optimization of the industry knowledge base.

[0108] For example, in the home appliance industry, deep association rules, such as refrigerator cooling malfunctions corresponding to insufficient refrigerant and washing machine spin-drying noises corresponding to damaged washing machine bearings, are applied to a dynamic knowledge topology graph. If the topology graph lacks a node for refrigerant replenishment, but the deep association rules contain a rule for refrigerant replenishment corresponding to insufficient refrigerant, a new node for refrigerant replenishment is added. If the topology graph contains a node for repairing older refrigerator models that has not been queried for a long time, it is deleted. Simultaneously, based on the deep association rules, new connections are added between refrigerator cooling malfunctions and insufficient refrigerant, and between washing machine spin-drying noises and damaged washing machine bearings, adjusting the corresponding edge weights to make the connections between nodes more aligned with user query habits. The above example is merely one illustration of this application; in practical applications, node and connection relationships can be adjusted according to deep association rules, and this application does not limit this.

[0109] In another specific implementation, the calculation of transition probability can be optimized by combining the query popularity of nodes to improve the accuracy of hot path filtering; the mining of deep association rules can be assisted by natural language processing technology to discover more logical relationships; the adjustment of dynamic knowledge topology graph can be set with a delayed adjustment mechanism to avoid structural instability caused by frequent adjustments.

[0110] This application addresses the problems of fixed industry knowledge base structures, inability to adapt to high-frequency user queries, and weak knowledge associations by statistically analyzing transition probabilities, identifying hotspot paths, mining deep association rules, and optimizing the knowledge topology graph. It achieves adaptive construction and optimization of the knowledge base. It accurately captures user query patterns, dynamically improves the knowledge structure, enhances the tightness of knowledge associations and query efficiency, and effectively avoids poor query experience and inaccurate knowledge matching caused by rigid knowledge base structures.

[0111] Figure 3 This application provides a schematic diagram illustrating a specific implementation of an industry knowledge base dynamic construction system based on multi-source heterogeneous data. (Refer to...) Figure 3 The system may include: The acquisition module 31 is used to acquire multi-source heterogeneous data in the target industry scenario. The multi-source heterogeneous data includes user historical behavior data and natural language expression text proposed by the user to the intelligent agent. The natural language expression text is processed to generate a query data stream. The mapping module 32 is used to uniformly map user questions in the query data stream into query semantic vectors, and at the same time convert user interaction behaviors extracted from user historical behavior data into behavior feature vectors to generate joint query vectors that can represent user intent. Module 33 is used to automatically extract industry knowledge units from multi-source heterogeneous data and associated external knowledge sources to construct a dynamic knowledge topology graph. The matching module 34 is used to construct a dynamic knowledge routing table. It uses the joint query vector as the routing request, calculates and matches the target knowledge node with the highest semantic consistency with the current query in the dynamic knowledge topology graph, records the nodes passed through by the query as the query path, and dynamically stores the association rules on the query path into the cache. The identification module 35 is used to identify hot paths based on the cumulative query path statistics of the transition probability between nodes, mine deep association rules from the hot paths, and adjust the nodes and connection relationships of the dynamic knowledge topology graph in reverse according to the association rules, so as to realize the adaptive construction and optimization of the industry knowledge base.

[0112] The industry knowledge base dynamic construction system based on multi-source heterogeneous data in this application embodiment is used to implement the aforementioned industry knowledge base dynamic construction method based on multi-source heterogeneous data. Therefore, the specific implementation of the industry knowledge base dynamic construction system based on multi-source heterogeneous data can be found in the embodiment section of the industry knowledge base dynamic construction method based on multi-source heterogeneous data above. The specific implementation can be referred to the description of the corresponding embodiment, which will not be repeated here.

[0113] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the above-described method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data.

[0114] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for dynamically constructing an industry knowledge base based on multi-source heterogeneous data.

[0115] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0116] Embodiments of the present invention also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the embodiments of the above-described method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data.

[0117] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0118] The above provides a detailed description of the method and system for dynamically constructing an industry knowledge base based on multi-source heterogeneous data provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data, characterized in that, include: Collect multi-source heterogeneous data in the target industry scenario. The multi-source heterogeneous data includes user historical behavior data and natural language expression text submitted by users to the intelligent agent. Process the natural language expression text to generate a query data stream. The user questions in the query data stream are uniformly mapped into query semantic vectors, and the user interaction behaviors extracted from the user's historical behavior data are converted into behavioral feature vectors to generate a joint query vector that can represent the user's intent. Industry knowledge units are automatically extracted from the multi-source heterogeneous data and associated external knowledge sources to construct a dynamic knowledge topology graph. A dynamic knowledge routing table is constructed, and the joint query vector is used as the routing request. The target knowledge node with the highest semantic consistency with the current query is calculated and matched in the dynamic knowledge topology graph. The nodes traversed by the query are recorded as the query path, and the association rules on the query path are dynamically stored in the cache. Based on the cumulative query path statistics, the transition probability between nodes is used to identify hot paths. Deep association rules are mined from the hot paths, and the nodes and connections of the dynamic knowledge topology graph are adjusted in reverse according to the association rules to achieve adaptive construction and optimization of the industry knowledge base.

2. The method according to claim 1, characterized in that, Based on the cumulative query path statistics, the transition probability between nodes is statistically analyzed to identify hotspot paths. Deep association rules are then mined from these hotspot paths, and the nodes and connections of the dynamic knowledge topology are adjusted in reverse according to these association rules. This enables the adaptive construction and optimization of the industry knowledge base, including: Collect accumulated historical query paths, calculate the transition frequency between adjacent nodes in the query path, deduce the transition probability based on the transition frequency, and filter path segments with the transition probability higher than a threshold as hotspot paths; Analyze the co-occurrence patterns and order relationships of nodes on the hotspot path, and extract deep association rules from the co-occurrence patterns and order relationships; By applying the deep association rules, nodes can be added or deleted in the dynamic knowledge topology graph, and the connection relationships between nodes can be updated.

3. The method according to claim 2, characterized in that, Analyze the co-occurrence patterns and order relationships of nodes on the hotspot path, and extract deep association rules from the co-occurrence patterns and order relationships, including: Traverse the hotspot path, extract multiple consecutively occurring nodes on the hotspot path, and combine each extracted node sequence as a node combination to capture the co-occurrence pattern and order relationship of the nodes. Count the number of times each node combination appears in all historical query paths, and calculate the co-occurrence frequency of each node combination based on the number of occurrences; All node combinations are filtered according to a preset frequency threshold, and node combinations whose co-occurrence frequency exceeds the frequency threshold are retained as valid node combinations; For each valid combination of nodes, a deep association rule is generated from the predecessor node to the successor node based on the order of the nodes in the sequence.

4. The method according to claim 1, characterized in that, A dynamic knowledge routing table is constructed, using the joint query vector as the routing request. The target knowledge node with the highest semantic consistency with the current query is calculated and matched in the dynamic knowledge topology graph. The nodes traversed by the query are recorded as the query path, and the association rules on the query path are dynamically stored in the cache, including: Construct and initialize a dynamic knowledge routing table, and input the joint query vector into the dynamic knowledge routing table; In the dynamic knowledge topology graph, starting from the joint query vector, the shortest path optimization algorithm is applied to traverse the nodes and connections of the dynamic knowledge topology graph, and the semantic distance between the joint query vector and each node is calculated. The path weight is updated based on the semantic distance, and the optimal path to each node is iteratively found. The node with the smallest semantic distance is selected as the target knowledge node, and the optimal path from the starting point to the target knowledge node is backtracked. The node sequence on the optimal path is recorded as the query path. Extract the connection relationships between adjacent nodes on the query path and write the connection relationships as association rules into the cache.

5. The method according to claim 4, characterized in that, In the dynamic knowledge topology graph, starting from the joint query vector, a shortest path optimization algorithm is applied to traverse the nodes and connections of the dynamic knowledge topology graph, and the semantic distance between the joint query vector and each node is calculated. The path weights are updated based on the semantic distance, and the optimal path to each node is iteratively found, including: Set the joint query vector as the starting point for path calculation, and initialize a record table to store the cumulative distance between each node and the starting point; Starting from the starting point, in the dynamic knowledge topology graph, unvisited nodes that are directly connected to the currently visited nodes are visited sequentially; When visiting each of the unvisited nodes, the semantic distance between the joint query vector and the node is calculated, and the semantic distance is added to the cumulative distance to the corresponding predecessor node to obtain the cumulative path distance to the corresponding node. Compare the cumulative path distances of different paths to the same node, update the value of the corresponding node in the record table to the minimum cumulative path distance, and record the corresponding predecessor node to update the path weight. Repeat the process of accessing, calculating, comparing, and updating until the optimal cumulative path distance of all nodes in the record table no longer changes, thus obtaining the optimal path from the starting point to each node.

6. The method according to claim 1, characterized in that, The user questions in the query data stream are uniformly mapped to query semantic vectors, and the user interaction behaviors extracted from the user's historical behavior data are converted into behavioral feature vectors to generate a joint query vector that can represent the user's intent, including: Based on the query data stream, extract the text string from the user question, and convert the text string into a first numerical array using a text vector tool, which serves as the query semantic vector. From the user's historical behavior data, the log records of user interaction behavior are parsed, the operation type and time series in the log records are extracted, a behavior pattern representation is constructed based on the operation type and time series, and the behavior pattern representation is converted into a second numerical array as a behavior feature vector using a vectorization tool; The query semantic vector and the behavioral feature vector are fused together, and a joint query vector that can represent the user's intent is generated by array concatenation and weight adjustment.

7. The method according to claim 1, characterized in that, Automated extraction of industry knowledge units from the multi-source heterogeneous data and associated external knowledge sources to construct a dynamic knowledge topology graph, including: By traversing the multi-source heterogeneous data, core terms and descriptive fields in the data content are detected, and the core terms and descriptive fields are packaged into industry knowledge units. Connect to associated external knowledge sources, extract relevant information blocks from the external knowledge sources, and decompose the relevant information blocks as supplementary knowledge units; Using the industry knowledge units and the supplementary knowledge units as graph nodes, connecting lines are added based on the contextual relationships between the industry knowledge units, and the semantic relevance between the graph nodes is used as the initial edge weights to construct a dynamic knowledge topology graph.

8. A dynamic construction system for an industry knowledge base based on multi-source heterogeneous data, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous data in the target industry scenario. The multi-source heterogeneous data includes user historical behavior data and natural language expression text submitted by the user to the intelligent agent. The natural language expression text is processed to generate a query data stream. The mapping module is used to uniformly map user questions in the query data stream into query semantic vectors, and at the same time convert user interaction behaviors extracted from the user's historical behavior data into behavioral feature vectors to generate a joint query vector that can represent the user's intent. A construction module is used to automatically extract industry knowledge units from the multi-source heterogeneous data and associated external knowledge sources to construct a dynamic knowledge topology graph. The matching module is used to construct a dynamic knowledge routing table, using the joint query vector as a routing request, to calculate and match the target knowledge node with the highest semantic consistency with the current query in the dynamic knowledge topology graph, record the nodes traversed by the query as the query path, and dynamically store the association rules on the query path into the cache. The identification module is used to identify hotspot paths based on the cumulative query path statistics of the transition probability between nodes, mine deep association rules from the hotspot paths, and adjust the nodes and connection relationships of the dynamic knowledge topology graph in reverse according to the association rules, so as to realize the adaptive construction and optimization of the industry knowledge base.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the method for dynamically constructing an industry knowledge base based on multi-source heterogeneous data as described in any one of claims 1 to 7.

Images