Search engine relation chain recommendation method based on knowledge graph

Abstract

The invention discloses a knowledge graph-based search engine relation chain recommendation method, which comprises the following steps of: receiving a relation chain recommendation request submitted by a user through a search engine to obtain initial query information; performing semantic analysis and intention recognition on the initial query information, and determining a corresponding entity type and a relationship type; calling a dynamic knowledge selection and evaluation mechanism according to the entity type and the relationship type, and dynamically screening the original knowledge graph to obtain effective knowledge sub-graphs; performing diffusion type reasoning based on a knowledge graph diffusion model to generate a candidate relation chain path set; evaluating the quality score of each candidate path in real time; and performing dynamic screening according to the path quality score, and outputting a relation chain path recommendation result. According to the method, accurate mining and dynamic optimization of the knowledge graph relation chain path are realized, and the accuracy of a recommendation result and the flexibility of user interaction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent recommendation, and in particular to a search engine relationship chain recommendation method based on a knowledge graph. BACKGROUND

[0002] With the rapid growth of Internet data, search engines have become an important way for users to efficiently obtain information. In order to improve information retrieval accuracy and user experience, the industry has gradually adopted knowledge graph technology to optimize search engine retrieval results. Knowledge graphs clearly present the semantic associations and deep relationships between retrieval content through structured and semantic entity and relationship connections, and have been widely used in information recommendation, semantic search, and knowledge mining fields. The existing knowledge graph recommendation technology of search engines generally implements search result optimization through entity recognition, semantic matching, or relationship chain reasoning methods.

[0003] The current common search engine relationship chain recommendation methods mainly include two types of recommendation based on rule reasoning and knowledge graph embedding. The method based on rule reasoning relies on pre-defined relationship rules to gradually reason about search entities, and the recommendation effect is limited by the rule coverage. It is difficult to effectively mine multi-level complex relationships, and the rule maintenance cost is high and the flexibility is poor. The method based on knowledge graph embedding realizes relationship prediction and link reasoning through vector embedding of entities and relationships, but the embedding model usually cannot fully capture the semantic granularity differences, resulting in insufficient accuracy of entity and relationship selection. In addition, the existing technology lacks a dynamic adjustment strategy for user query intent and knowledge granularity, making it difficult to realize real-time, efficient, and accurate filtering and evaluation of candidate relationship chain paths, resulting in low matching degree between relationship chain path recommendation results and actual user needs and more redundant information, thereby reducing the search engine recommendation efficiency and user satisfaction.

[0004] Therefore, how to provide a search engine relationship chain recommendation method based on a knowledge graph is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a search engine relationship chain recommendation method based on a knowledge graph. In order to solve the problem that the existing search engine cannot accurately mine high-quality entities and relationships required by users in real time when recommending relationship chain paths, a technical solution is proposed that dynamically selects and evaluates knowledge to filter knowledge graph subgraphs in real time, uses a diffusion model to perform path reasoning, and evaluates and dynamically optimizes the quality of candidate paths in real time. The present application has the effect of improving the accuracy of relationship chain path recommendation and improving user interaction experience.

[0006] According to the search engine relationship chain recommendation method based on the knowledge graph according to an embodiment of the present application, the method comprises: Receiving a relationship chain recommendation request submitted by a user through a search engine, obtaining initial query information; Performing semantic analysis and intent recognition on the initial query information, determining corresponding entity types and relationship types; According to the entity types and relationship types, calling a dynamic knowledge selection and evaluation mechanism to dynamically filter entities and relationships in the original knowledge graph, forming a knowledge graph subgraph associated with the initial query information; Calculating and filtering the effectiveness score and relevance score of each entity and each relationship in the knowledge graph subgraph, obtaining an effective knowledge graph subgraph; Based on the effective knowledge graph subgraph, performing diffusion reasoning through a knowledge graph diffusion model to generate a candidate relationship chain path set; For each candidate relationship chain path in the candidate relationship chain path set, calculating a path quality evaluation score of each candidate relationship chain path; According to the path quality evaluation score, dynamically filtering the candidate relationship chain path set again to form a relationship chain path recommendation result; Outputting the relationship chain path recommendation result to the user, and providing the user with an interactive interface for interactive exploration through a visual dynamic relationship chain.

[0007] Optionally, the receiving of the relationship chain recommendation request submitted by the user through the search engine and the obtaining of the initial query information are specifically: Receiving a relationship chain recommendation request submitted by a user through a search engine, the relationship chain recommendation request containing at least one initial query keyword input by the user; Fully matching the initial query keyword with entities and relationships in a pre-established entity index library and a pre-established relationship index library respectively, and determining target entities and target relationships that match successfully; For initial query keywords that do not achieve complete matching, calculating the semantic similarity score between the initial query keywords and candidate entities and candidate relationships in the pre-established entity index library and the pre-established relationship index library respectively using a semantic similarity calculation method; According to the semantic similarity score and a dynamically set semantic similarity threshold, filtering out candidate entities and candidate relationships; According to the target entities, the target relationships, the candidate entities and the candidate relationships, forming initial query information containing complete matching results and fuzzy matching results.

[0008] Optionally, the semantic analysis and intent recognition of the initial query information, and the determination of the corresponding entity types and relationship types are specifically: Based on the initial query information, performing semantic analysis on the target entities and the target relationships in the complete matching results, and determining the semantic types corresponding to the target entities and the target relationships respectively; The semantic analysis is performed on the candidate entities and the candidate relationships in the semantic fuzzy matching result, to determine candidate semantic types corresponding to the candidate entities and the candidate relationships respectively; According to the semantic types of the target entity and the target relationship and the candidate semantic types of the candidate entity and the candidate relationship, the semantic correlation degrees between the target entity and the target relationship and the candidate entity and the candidate relationship are calculated one by one through a semantic correlation degree calculation method; Based on the semantic correlation degree scores and a preset semantic correlation degree threshold, the candidate entities and the candidate relationships are screened to determine the candidate entities and the candidate relationships whose semantic correlation degree scores meet the threshold; Based on the determined target entity, the target relationship, and the candidate entities and the candidate relationships whose semantic correlation degree scores meet the threshold, corresponding entity types and relationship types are determined.

[0009] Optionally, the dynamic knowledge selection and evaluation mechanism is called according to the entity types and the relationship types to dynamically screen entities and relationships in the original knowledge graph, to form a knowledge graph subgraph associated with the initial query information, specifically as follows: Based on the determined entity types and the relationship types, all entities and all relationships of the corresponding types are retrieved from the original knowledge graph to form an initial entity set and an initial relationship set; Based on the dynamic knowledge selection and evaluation mechanism, entity correlation scores between each entity in the initial entity set and the initial query information are calculated one by one; Based on the dynamic knowledge selection and evaluation mechanism, relationship correlation scores between each relationship in the initial relationship set and the initial query information are calculated one by one; According to the entity correlation scores and the relationship correlation scores, the initial entity set and the initial relationship set are screened by using a dynamically determined correlation score threshold to determine entities and relationships that meet the correlation score threshold; The entities that meet the correlation score threshold are taken as nodes, and the relationships that meet the correlation score threshold are taken as edges between the nodes, to construct a knowledge graph subgraph associated with the initial query information.

[0010] Optionally, the dynamic knowledge selection and evaluation mechanism specifically includes the following steps: Based on the initial query information, historical query appearance frequencies of each entity and each relationship in the original knowledge graph within a preset time window are counted in real time, and historical hotness weights of each entity and each relationship are calculated; Each entity and each relationship in the original knowledge graph and the initial query information are respectively mapped into corresponding semantic vectors to determine semantic relevance scores of each entity and each relationship and the initial query information; According to the historical hotness weights and the semantic relevance scores, comprehensive effectiveness scores of each entity and each relationship are calculated in real time through a dynamically adjusted weighting fusion ratio; Based on the comprehensive validity score obtained through real-time calculation and the dynamically determined validity score threshold, entities and relationships that meet the validity score threshold are selected in real time. During the reasoning process of the knowledge graph diffusion model, the comprehensive validity score of the entities and relationships involved in each candidate relation chain path generated by diffusion is obtained in real time, and the diffusion direction and diffusion results of the diffusion model are optimized. The candidate relationship chain paths, after being dynamically adjusted and optimized by diffusion weights, will be output as the final relationship chain path recommendation results.

[0011] Optionally, the calculation and filtering of the validity score and relevance score for each entity and each relation within the knowledge graph subgraph to obtain an effective knowledge graph subgraph specifically involves: Based on a dynamic knowledge selection and evaluation mechanism, the comprehensive effectiveness score of each entity and each relation in the knowledge graph subgraph is obtained under the current user query request. Based on a dynamic knowledge selection and evaluation mechanism, the semantic relevance score between each entity and each relation in the knowledge graph subgraph and the initial query information is obtained in real time. Based on the comprehensive effectiveness score and semantic relevance score, a dynamically adjusted weighted fusion ratio of effectiveness and relevance is used to calculate the comprehensive screening score of each entity and each relation in the knowledge graph subgraph in real time. The system analyzes the knowledge granularity requirements of the user's current query request in real time and dynamically determines the comprehensive screening score threshold based on the knowledge granularity requirements. Based on the comprehensive screening score and the dynamically determined comprehensive screening score threshold, entities and relationships with comprehensive screening scores below the threshold are filtered and removed in real time. By using the filtered entities as nodes and the filtered relationships as edges between nodes, an effective knowledge graph subgraph is generated.

[0012] Optionally, the step of generating a candidate relation chain path set by performing diffusion-based reasoning through a knowledge graph diffusion model based on an effective knowledge graph subgraph specifically involves: Using the target entity in the initial query information as the starting node for diffusion, the initial set of diffusion nodes is determined in real time based on the effective knowledge graph subgraph. Starting from each node in the initial set of diffusion nodes, determine the neighboring nodes corresponding to each node one by one, and obtain the relationship between each neighboring node and the current node in real time; Calculate the path diffusion score for the current node to spread to its neighboring nodes; Based on the path diffusion score obtained in real time and the path diffusion score threshold determined dynamically, nodes and corresponding relationships that meet the path diffusion score threshold are selected as candidate nodes and candidate relationships for the next round of diffusion. The number of diffusion iterations is determined in real time based on the user's query request, and the diffusion process is executed repeatedly until the number of diffusion iterations is reached; The diffusion path from the starting node to the current diffusion node is recorded and stored in real time, forming a set of candidate relationship chain paths.

[0013] Optionally, for each candidate relationship chain path in the candidate relationship chain path set, a path quality evaluation score is calculated for each candidate relationship chain path, specifically as follows: Based on a dynamic knowledge selection and evaluation mechanism, the comprehensive validity score of all nodes and relationships in the candidate relationship chain path is obtained in real time, and the association strength score is calculated in real time by weighting the comprehensive validity scores of each node and relationship in the path. Based on a dynamic knowledge selection and evaluation mechanism, the semantic vector cosine similarity between adjacent nodes and corresponding relationships in the candidate relationship chain path is calculated one by one, and the semantic consistency score of the candidate relationship chain path is determined in real time. The path length score is calculated in real time based on the difference between the actual path length of the candidate relationship chain path and the target path length explicitly stated in the user's current query request. The success frequency of each node and relationship in the candidate relationship chain path in the historical path query task is counted in real time, and the path confidence score of the candidate relationship chain path is calculated based on the success frequency. Based on the association strength score, semantic consistency score, path length score, and path confidence score, the weight ratio of the dynamically adjusted path quality assessment score is determined in real time. Based on the real-time dynamically adjusted weight ratios, the association strength score, semantic consistency score, path length score, and path confidence score are weighted and fused in real time to obtain the path quality evaluation score corresponding to each candidate relationship chain path.

[0014] Optionally, based on the path quality assessment score, the candidate relationship chain path set is dynamically filtered again to form a relationship chain path recommendation result, specifically as follows: Based on a dynamic knowledge selection and evaluation mechanism, the path quality evaluation score of each candidate relationship chain path in the candidate relationship chain path set is obtained in real time, and the average path quality evaluation score of the candidate relationship chain path set is calculated in real time. Analyze the knowledge granularity requirements of the user's current query request in real time, and calculate the degree of difference between the knowledge granularity of the current query request and the knowledge granularity of historical user query requests; Based on the degree of difference, the adjustment ratio of the path quality assessment score threshold is dynamically determined; Based on the average path quality assessment score and the dynamically determined path quality assessment score threshold adjustment ratio, the path quality assessment score threshold corresponding to the current query request is calculated in real time. In real time, candidate relationship chain paths with path quality evaluation scores lower than the path quality evaluation score threshold are filtered and removed from the candidate relationship chain path set, while candidate relationship chain paths with path quality evaluation scores equal to or higher than the path quality evaluation score threshold are retained; The retained candidate relationship chain paths are output in real time to form relationship chain path recommendation results.

[0015] Optionally, the step of outputting the relationship chain path recommendation results to the user and providing an interactive, visual, dynamic relationship chain interface for user exploration specifically involves: The relationship chain path recommendation results are visualized in real time in the form of a graphical knowledge graph; The system calculates and displays the comprehensive effectiveness score of each node and relationship in the relationship chain path recommendation results in real time. Get and respond to user interactive operations on any node or relationship in the relationship chain path recommendation results; After users interact with the nodes or relationships in the relationship chain path recommendation results in real time, the system updates the comprehensive effectiveness score of the selected nodes or relationships in real time based on the dynamic knowledge selection and evaluation mechanism, and updates the graphical display of the relationship chain path recommendation results in real time. Real-time statistics and recording of user interaction history on nodes and relationships in the dynamic relationship chain interaction interface; and real-time dynamic optimization and updating of relationship chain path recommendation results based on user interaction history. It receives real-time interactive feedback from users on recommended paths based on a dynamic relationship chain interaction interface.

[0016] The beneficial effects of this invention are: (1) This invention uses a dynamic knowledge selection and evaluation mechanism to dynamically filter high-quality knowledge subgraphs in real time, and combines a knowledge graph diffusion model to deeply mine the relationship paths between entities, which effectively improves the accuracy and real-time performance of the search engine relationship chain recommendation results, and significantly improves the efficiency of users obtaining information and the accuracy of the recommendation path.

[0017] (2) This invention achieves intelligent and precise optimization of candidate relationship chain paths through real-time dynamic calculation of path quality assessment scores and multi-dimensional screening mechanism, which significantly improves the quality and reliability of path recommendation results and shows better adaptability and recommendation effect in the relationship mining and intelligent recommendation scenarios of search engines.

[0018] (3) In terms of dynamic and accurate mining of search engine relationship chain paths, this invention effectively solves the shortcomings of insufficient accuracy of path recommendation and excessive redundant paths in the prior art by dynamically adjusting the path diffusion score threshold and the path quality evaluation score in real time. It breaks through the limitation of low flexibility of the fixed threshold screening mechanism in the prior art, realizes significant progress in real-time dynamic optimization of relationship chain path mining, and effectively improves the level of intelligent application of knowledge graph in search engines. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0020] Fig. 1 This is an overall flowchart of a knowledge graph-based search engine relationship chain recommendation method proposed in this invention; Fig. 2 This is a flowchart illustrating the dynamic knowledge selection and evaluation mechanism of a knowledge graph-based search engine relationship chain recommendation method proposed in this invention. Fig. 3 This is a schematic diagram of the dynamic relationship chain interaction interface structure of a knowledge graph-based search engine relationship chain recommendation method proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figs. 1-3 A knowledge graph-based search engine relationship chain recommendation method includes: Receive relationship chain recommendation requests submitted by users through the search engine to obtain initial query information; Perform semantic analysis and intent recognition on the initial query information to determine the corresponding entity type and relationship type; Based on entity type and relation type, a dynamic knowledge selection and evaluation mechanism is invoked to dynamically filter entities and relations in the original knowledge graph and form a knowledge graph subgraph associated with the initial query information. Based on a dynamic knowledge selection and evaluation mechanism, the effectiveness score and relevance score of each entity and each relation in the knowledge graph subgraph are calculated and screened in real time. Entities and relations with low effectiveness scores or relevance scores are filtered and removed to obtain effective knowledge graph subgraphs. Based on effective knowledge graph subgraphs, diffusion reasoning is performed through a knowledge graph diffusion model, with the initial query information as the diffusion center, and a set of candidate relation chain paths is gradually mined and generated. For each candidate relation chain path in the candidate relation chain path set, a dynamic knowledge selection and evaluation mechanism is used to calculate the path quality evaluation score of each candidate relation chain path in real time. The path quality evaluation score includes path length score, semantic consistency score, association strength score and confidence score. Based on the path quality assessment score, the candidate relationship chain path set is dynamically screened again using a dynamic knowledge selection and evaluation mechanism. Paths with low path quality assessment scores are removed, while paths with high path quality assessment scores are retained, thus forming the relationship chain path recommendation results. The system outputs the recommended relationship chain paths to the user, allowing for interactive exploration through a visual and dynamic relationship chain interface.

[0023] In this embodiment, receiving the relationship chain recommendation request submitted by the user through the search engine and obtaining the initial query information specifically includes: Receive a relationship chain recommendation request submitted by a user through a search engine, wherein the relationship chain recommendation request contains at least one initial query keyword input by the user; The initial query keywords are matched one by one with the entities and relationships in the pre-established entity index and the pre-established relationship index to determine the target entities and relationships that are successfully matched. The complete matching is to compare the initial query keywords entered by the user with the objects to be matched in the pre-established entity index and the pre-established relationship index word by word. Only when the keywords and the objects to be matched are exactly the same in words are they judged as a successful match. If they are not exactly the same in words, they are judged as unsuccessful matches. For initial query keywords that do not achieve a complete match, the semantic similarity calculation method is used to calculate the semantic similarity score between the initial query keywords and the candidate entities and candidate relationships in the pre-built entity index and the pre-built relationship index. Candidate entities and relationships are selected based on semantic similarity scores and dynamically set semantic similarity thresholds; Based on the target entities and relationships determined by full matching, and the alternative entities and relationships obtained through semantic similarity filtering, initial query information containing both full matching results and fuzzy matching results is formed: The target entities and target relationships that are successfully matched are identified as the first component of the initial query information; The candidate entities and candidate relationships selected through semantic similarity scoring and based on dynamically set semantic similarity thresholds are determined as the second component of the initial query information; The first component and the second component are merged according to a preset data structure to form initial query information that simultaneously contains both exact match results and semantically fuzzy match results.

[0024] In this embodiment, the step of performing semantic analysis and intent recognition on the initial query information to determine the corresponding entity type and relationship type specifically includes: Based on the initial query information, semantic parsing is performed on the target entities and target relationships in the complete matching results to determine the semantic type of each target entity and target relationship. Semantic parsing is performed on the candidate entities and candidate relations in the semantic fuzzy matching results to determine the candidate semantic types corresponding to each candidate entity and candidate relation; Based on the semantic types of the target entity and target relation, and the candidate semantic types of the alternative entities and alternative relations, a semantic relevance score is calculated for each alternative entity and alternative relation to the target entity and target relation using a semantic relevance calculation method. The semantic relevance calculation method is as follows: The semantic types of the target entity, target relation, and alternative entities and alternative relations are mapped to corresponding semantic vectors through a pre-trained semantic representation model. The vector space similarity calculation method is used to calculate the cosine similarity between the semantic vector of the target entity and the semantic vector of the candidate entity, and between the semantic vector of the target relationship and the semantic vector of the candidate relationship. The semantic relevance score between candidate entities and candidate relations and target entities and target relations is determined based on the calculated cosine similarity. The semantic relevance score is used as a quantitative indicator to determine the degree of semantic relevance between candidate entities and candidate relationships and target entities and target relationships. Based on the semantic relevance score and the preset semantic relevance threshold, the candidate entities and candidate relationships are screened to determine the candidate entities and candidate relationships whose semantic relevance scores meet the threshold. Based on the identified target entities, target relationships, and candidate entities and relationships whose semantic relevance scores meet the threshold, the corresponding entity types and relationship types are determined.

[0025] In this embodiment, the step of invoking a dynamic knowledge selection and evaluation mechanism based on entity type and relation type to dynamically filter entities and relations in the original knowledge graph and form a knowledge graph subgraph associated with the initial query information specifically involves: Based on the determined entity type and relation type, all entities and relations of the corresponding type are retrieved from the original knowledge graph to form an initial entity set and an initial relation set; Based on a dynamic knowledge selection and evaluation mechanism, the entity association score between each entity in the initial entity set and the initial query information is calculated one by one. The calculation method for the entity association score is as follows: Each entity in the initial entity set and the entity in the initial query information are mapped to a corresponding semantic vector through a pre-trained semantic representation model; Calculate the cosine similarity between the semantic vector of each entity in the initial entity set and the semantic vector of the entities in the initial query information; The cosine similarity is used as the entity association score; Based on a dynamic knowledge selection and evaluation mechanism, the relationship association score between each relation in the initial relation set and the initial query information is calculated one by one. The calculation method for the relationship association score is as follows: Each relation in the initial relation set and the relation in the initial query information are mapped to a corresponding semantic vector through a pre-trained semantic representation model; Calculate the cosine similarity between the semantic vector of each relation in the initial relation set and the semantic vector of the relation in the initial query information; The cosine similarity is used as a relationship association score; Based on entity association scores and relationship association scores, dynamically determined association score thresholds are used to filter the initial entity set and the initial relationship set to determine the entities and relationships that meet the association score thresholds. By using entities that meet the association score threshold as nodes and relationships that meet the association score threshold as edges between nodes, a knowledge graph subgraph associated with the initial query information is constructed.

[0026] In this embodiment, the dynamic knowledge selection and evaluation mechanism is specifically as follows: Based on the initial query information, the frequency of historical queries for each entity and relation in the original knowledge graph within a preset time window is counted in real time. The historical popularity weight of each entity and relation is calculated based on the frequency of historical queries. The historical popularity weight is calculated as follows: the number of times each entity and relation in the original knowledge graph is called by user query requests and actually accessed within the preset time window is counted in real time. The actual access count obtained is divided by the total access count within the preset time window as a normalization process to obtain the historical popularity weight corresponding to each entity and relation. Each entity and relation in the original knowledge graph is mapped to a corresponding semantic vector from the initial query information. The cosine similarity between the semantic vector of each entity and relation and the semantic vector of the initial query information is calculated one by one to determine the semantic relevance score between each entity and relation and the initial query information. Based on historical popularity weights and semantic relevance scores, a comprehensive effectiveness score for each entity and each relationship is calculated in real time using a dynamically adjusted weighted fusion ratio, which changes in real time according to user requests. Based on the comprehensive validity score obtained through real-time calculation and the dynamically determined validity score threshold, entities and relationships that meet the validity score threshold are selected in real time. During the reasoning process of the knowledge graph diffusion model, the comprehensive validity score of the entities and relationships involved in each candidate relation chain path generated by diffusion is obtained in real time. Based on the comprehensive validity score of the entities and relationships in each candidate relation chain path, the diffusion weight of the path is dynamically adjusted in real time, and the diffusion direction and diffusion result of the diffusion model are optimized. The candidate relationship chain paths, after being dynamically adjusted and optimized by diffusion weights, will be output as the final relationship chain path recommendation results.

[0027] In this embodiment, the calculation and filtering of the validity score and relevance score for each entity and each relation within the knowledge graph subgraph to obtain an effective knowledge graph subgraph specifically involves: Based on a dynamic knowledge selection and evaluation mechanism, the comprehensive effectiveness score of each entity and each relation in the knowledge graph subgraph is obtained under the current user query request. Based on a dynamic knowledge selection and evaluation mechanism, the semantic relevance score between each entity and each relation in the knowledge graph subgraph and the initial query information is obtained in real time. Based on the comprehensive effectiveness score and semantic relevance score, a dynamically adjusted weighted fusion ratio of effectiveness and relevance is used to calculate the comprehensive screening score of each entity and each relation in the knowledge graph subgraph in real time. The method for adjusting the weighted fusion ratio of validity and relevance in the dynamic adjustment is as follows: perform real-time semantic parsing based on the knowledge granularity requirements specified in the user's current query request, determine the degree of difference between the knowledge granularity requirements and the knowledge granularity of historical queries, and dynamically adjust the fusion weight ratio of validity score and relevance score according to the degree of difference. Specifically, when the degree of difference is small, the weight ratio of validity score is increased, and when the degree of difference is large, the weight ratio of relevance score is increased. The comprehensive screening score is calculated as follows: based on the comprehensive effectiveness score and semantic relevance score of each entity and each relation calculated in real time, a dynamically adjusted weighted fusion ratio of effectiveness and relevance is used for real-time fusion. The comprehensive effectiveness score and semantic relevance score are multiplied by the corresponding real-time adjusted weight ratio and then added together to obtain the comprehensive screening score corresponding to each entity and each relation. The system analyzes the knowledge granularity requirements of the user's current query request in real time and dynamically determines the comprehensive screening score threshold based on the knowledge granularity requirements. Based on the comprehensive screening score and the dynamically determined comprehensive screening score threshold, entities and relationships with comprehensive screening scores below the threshold are filtered and removed in real time. By using filtered entities as nodes and filtered relationships as edges between nodes, an effective knowledge graph subgraph that satisfies user query requests is generated in real time.

[0028] In this embodiment, the step of generating a candidate relation chain path set by performing diffusion-based reasoning through a knowledge graph diffusion model based on an effective knowledge graph subgraph specifically involves: Using the target entity in the initial query information as the starting node for diffusion, the initial set of diffusion nodes is determined in real time based on the effective knowledge graph subgraph. Starting from each node in the initial set of diffusion nodes, determine the neighboring nodes corresponding to each node one by one, and obtain the relationship between each neighboring node and the current node in real time; Calculate the path diffusion score for the current node to spread to neighboring nodes. The path diffusion score is obtained by real-time weighted calculation based on the comprehensive effectiveness score of the current node, the comprehensive effectiveness score of neighboring nodes, and the comprehensive effectiveness score of the relationship between nodes. Based on the path diffusion score obtained in real time and the dynamically determined path diffusion score threshold, nodes and corresponding relationships that meet the path diffusion score threshold are selected in real time as candidate nodes and candidate relationships for the next round of diffusion. The dynamic determination method of the path diffusion score threshold is as follows: the average path length of the target diffusion path is calculated in real time based on the knowledge granularity requirements of the user's current query request, the degree of difference between the average path length and the average length of the historical diffusion path is analyzed, and the path diffusion score threshold is dynamically determined based on the degree of difference. When the degree of difference is small, the path diffusion score threshold is increased to ensure path quality, and when the degree of difference is large, the path diffusion score threshold is decreased to expand the diffusion range. The number of diffusion iterations is determined in real time based on the user's query request to meet the knowledge granularity requirements of the diffusion path, and the diffusion process is executed cyclically until the number of diffusion iterations is reached. The diffusion path from the starting node to the current diffusion node is recorded and stored in real time, forming a set of candidate relationship chain paths.

[0029] In this embodiment, calculating the path quality evaluation score for each candidate relationship chain path in the candidate relationship chain path set specifically involves: Based on a dynamic knowledge selection and evaluation mechanism, the comprehensive validity score of all nodes and relationships in the candidate relationship chain path is obtained in real time, and the association strength score is calculated in real time by weighting the comprehensive validity scores of each node and relationship in the path. The method for calculating the association strength score is as follows: the comprehensive validity score of each node and relationship in the candidate relationship chain path is obtained in real time, the comprehensive validity scores of each node and relationship are added together and then divided by the total number of nodes and relationships in the candidate relationship chain path to obtain the association strength score corresponding to the candidate relationship chain path. Based on a dynamic knowledge selection and evaluation mechanism, the semantic vector cosine similarity between adjacent nodes and corresponding relationships in the candidate relationship chain path is calculated one by one, and the semantic consistency score of the candidate relationship chain path is determined in real time. The semantic consistency score is calculated as follows: the cosine similarity between the semantic vectors of each pair of adjacent nodes and their corresponding relationships in the candidate relationship chain path is calculated one by one, and the cosine similarity between all adjacent nodes and relationships in the candidate relationship chain path is added together and divided by the total number of adjacent nodes and relationship pairs to obtain the semantic consistency score of the candidate relationship chain path. Based on the degree of difference between the actual path length of the candidate relationship chain path and the target path length specified in the user's current query request, a path length score is calculated in real time. The path length score is calculated as follows: the absolute difference between the actual path length of the candidate relationship chain path and the target path length specified in the user's current query request is calculated in real time; the degree of difference is determined based on the ratio of the absolute difference to the target path length; and the result after subtracting the degree of difference from 1 is used as the path length score. The system counts the successful access frequency of each node and relationship in the candidate relationship chain path in the historical path query task in real time, and calculates the path confidence score of the candidate relationship chain path based on the successful access frequency. The path confidence score is calculated as follows: the total historical successful access frequency of all nodes and relationships in the candidate relationship chain path is counted, and the total historical successful access frequency is divided by the total historical access frequency of all nodes and relationships in the candidate relationship chain path to obtain the path confidence score. Based on the association strength score, semantic consistency score, path length score, and path confidence score, the dynamic adjustment weight ratio of the path quality assessment score is determined in real time. The dynamic adjustment method of the weight ratio is as follows: analyze the knowledge granularity requirements of the current query request in real time, and dynamically adjust the weight of each score in the calculation of the path quality assessment score according to the degree of difference between the knowledge granularity requirements and the knowledge granularity requirements of historical queries. Based on the real-time dynamically adjusted weight ratios, the association strength score, semantic consistency score, path length score, and path confidence score are weighted and fused in real time to obtain the path quality evaluation score corresponding to each candidate relationship chain path.

[0030] In this embodiment, the step of dynamically filtering the candidate relationship chain path set again based on the path quality assessment score to form a relationship chain path recommendation result is as follows: Based on a dynamic knowledge selection and evaluation mechanism, the path quality evaluation score of each candidate relationship chain path in the candidate relationship chain path set is obtained in real time, and the average path quality evaluation score of the candidate relationship chain path set is calculated in real time. Analyze the knowledge granularity requirements of the user's current query request in real time, and calculate the degree of difference between the knowledge granularity of the current query request and the knowledge granularity of historical user query requests; Based on the degree of difference, the adjustment ratio of the path quality evaluation score threshold is dynamically determined. When the degree of difference is large, the path quality evaluation score threshold is reduced to expand the range of candidate path selection. When the degree of difference is small, the path quality evaluation score threshold is increased to ensure the quality of candidate path selection. Based on the average path quality assessment score and the dynamically determined path quality assessment score threshold adjustment ratio, the path quality assessment score threshold corresponding to the current query request is calculated in real time. In real time, candidate relationship chain paths with path quality evaluation scores lower than the path quality evaluation score threshold are filtered and removed from the candidate relationship chain path set, while candidate relationship chain paths with path quality evaluation scores equal to or higher than the path quality evaluation score threshold are retained; The retained candidate relationship chain paths are output in real time to form relationship chain path recommendation results.

[0031] In this embodiment, the step of outputting the relationship chain path recommendation results to the user and providing an interactive, visual, dynamic relationship chain interface for user exploration specifically involves: The relationship chain path recommendation results are visualized in real time in the form of a graphical knowledge graph. Nodes in each relationship chain path are labeled with graphical elements that identify entity types, and the relationships between nodes are labeled with lines that identify relationship types. The system calculates and displays the comprehensive effectiveness score of each node and relationship in the relationship chain path recommendation results in real time, and adjusts the graphical display style of nodes and relationships in real time based on the comprehensive effectiveness score. The system can acquire and respond in real time to interactive operations performed by users on any node or relationship in the relationship chain path recommendation results. These interactive operations include expanding a node, collapsing a relationship, viewing detailed node information, and zooming in or out of the path. After users interact with the nodes or relationships in the relationship chain path recommendation results in real time, the system updates the comprehensive effectiveness score of the selected nodes or relationships in real time based on the dynamic knowledge selection and evaluation mechanism, and updates the graphical display of the relationship chain path recommendation results in real time. Real-time statistics and recording of user interaction history on nodes and relationships in the dynamic relationship chain interaction interface; and real-time dynamic optimization and updating of relationship chain path recommendation results based on user interaction history. The system receives interactive feedback from users on recommended paths in real time through a dynamic relationship chain interface. This feedback includes user confirmation or rejection of the recommended path selection. Based on the feedback, the system optimizes the weighting of path quality evaluation scores in real time to further improve the relationship chain path recommendation results.

[0032] Example 1: To verify the feasibility of this invention in practice, it was applied to a user relationship chain path intelligent recommendation scenario on a general-purpose intelligent search platform. Facing large-scale knowledge graph dynamic reasoning and recommendation tasks in a real-world environment, the accuracy, real-time performance, and user interaction experience improvement of the proposed solution in complex search engine query scenarios were evaluated. In this application scenario, traditional methods typically employ a single filtering approach based on static rule reasoning or knowledge graph embedding. This makes it difficult to dynamically adapt to users' personalized query intentions and knowledge granularity requirements, resulting in excessive redundancy and low relevance of generated candidate relationship chain paths. Furthermore, it lacks real-time response to user feedback, severely limiting the performance of search engines in complex semantic association retrieval.

[0033] During implementation, the system first receives relationship chain recommendation requests submitted by users through the search engine interface, clarifying the user's query keywords. Semantic analysis and intent recognition are performed on the initial query information to automatically determine the target entity and relationship type. Subsequently, based on a dynamic knowledge selection and evaluation mechanism, the system filters high-quality entities and relationships from the original knowledge graph in real time. Based on indicators such as historical query popularity and semantic relevance to the query intent, a comprehensive effectiveness score is calculated using a dynamic weighting method. The system dynamically sets effectiveness score thresholds, eliminating redundant or low-quality nodes and relationships in real time to obtain a knowledge graph subgraph that highly matches the user's intent. Based on a knowledge graph diffusion model, starting with the target entity, multiple rounds of diffusion-style reasoning are performed within the effective knowledge subgraph. The comprehensive effectiveness score of each node and relationship in each round of diffusion is dynamically obtained, and the diffusion weights are adjusted to achieve progressive mining of candidate relationship chain paths.

[0034] During the relationship chain path quality assessment phase, the system calculates path length score, semantic consistency score, association strength score, and path confidence score for each candidate path, dynamically adjusting the score weights to adapt to the knowledge granularity needs of different users. Finally, all candidate paths are weighted and fused for selection. By statistically analyzing the average quality assessment score of the candidate path set in real time and combining it with historical user knowledge granularity preferences, the system dynamically determines the path selection threshold, effectively eliminating low-quality or redundant paths and retaining only highly relevant and highly confident accurate recommendation paths. The final recommendation results are then output to the interactive interface in a knowledge graph visualization format.

[0035] During the user interaction phase, the system responds in real-time to user interactions with nodes and relationships in the knowledge graph, including expanding nodes, viewing detailed information, and zooming in and out of paths. After each interaction, the system automatically adjusts the overall effectiveness score and path evaluation weights based on user feedback, achieving adaptive and dynamic optimization of the recommendation results. All user interactions and feedback are recorded in real-time and participate in the continuous updates and iterations of the subsequent path recommendation model, further enhancing the model's adaptability and robustness to diverse user needs.

[0036] In this embodiment, a leading Chinese open knowledge graph dataset was used on the test platform, covering 100,000 entities, 300,000 relationships, and diverse complex semantic retrieval requests. The recommendation engine ran in high-performance parallel mode on the server side. The initial node of the diffusion model was set to a single target entity, with a maximum of 5 diffusion rounds, a target path length range of 3-6, and a dynamic threshold adjustment step size of 0.05. The knowledge graph node embedding dimension was 64, and improved cosine similarity was used for semantic relevance scoring. The initial weighted ratio of the comprehensive effectiveness score was set to historical popularity 0.4 and semantic relevance 0.6, and the interaction feedback optimization weight was adaptively adjusted to no more than 0.05 in each iteration.

[0037] Table 1 shows the performance comparison data of the method of the present invention and mainstream traditional methods (static rule reasoning method and knowledge graph embedding method) on typical test samples in terms of path recommendation accuracy, average response latency, path redundancy rate, user interaction adaptability and average user satisfaction.

[0038] Table 1 Comparison of Key Performance Indicators of Search Engine Relationship Chain Path Recommendation

[0039] As shown in Table 1 above, the method of this invention significantly outperforms existing technologies in path recommendation accuracy, with multiple samples achieving a recommendation accuracy exceeding 95.5%. In comparison, the highest accuracies of the traditional rule-based reasoning method and the embedding method in the same test group were 89.2% and 92.4%, respectively. Regarding average response latency, the system, through efficient dynamic diffusion reasoning and parallel evaluation mechanisms, stably controls the recommendation latency to within 350ms, significantly better than the traditional method's average of approximately 470ms. In terms of path redundancy, the method of this invention, through multiple rounds of dynamic screening and weight adaptation, maintains a redundancy rate below 13%, far lower than the embedding method's average redundancy of 22%. The user interaction adaptability score (out of 5) and average user satisfaction both reach above 4.7, reflecting the high adaptability and intelligent optimization capability of the interface and model feedback mechanism to complex and changing user needs.

[0040] Based on the above embodiments and experimental data, this invention not only achieves efficient and accurate response to users' complex relationship chain query intent, but also breaks through the bottlenecks of existing technologies in personalized recommendation, real-time adaptation and high-quality filtering. It has significant technical advantages such as rigorous structure, efficient operation, accurate recommendation results and excellent interactive experience, and can be widely applied to practical application scenarios such as large-scale search engines, intelligent question answering and complex semantic knowledge recommendation.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based search engine relationship chain recommendation method, characterized in that, include: Receive relationship chain recommendation requests submitted by users through the search engine to obtain initial query information; Perform semantic analysis and intent recognition on the initial query information to determine the corresponding entity type and relationship type; Based on entity type and relation type, a dynamic knowledge selection and evaluation mechanism is invoked to dynamically filter entities and relations in the original knowledge graph and form a knowledge graph subgraph associated with the initial query information. Calculate and filter the validity score and relevance score of each entity and each relation within the knowledge graph subgraph to obtain valid knowledge graph subgraphs; Based on effective knowledge graph subgraphs, diffusion reasoning is performed through a knowledge graph diffusion model to generate a set of candidate relation chain paths. For each candidate relationship chain path in the candidate relationship chain path set, calculate the path quality evaluation score for each candidate relationship chain path; Based on the path quality assessment score, the candidate relationship chain path set is dynamically filtered again to form the relationship chain path recommendation result; The system outputs the recommended relationship chain paths to the user, allowing for interactive exploration through a visual and dynamic relationship chain interface.

2. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, The process of receiving a relationship chain recommendation request submitted by a user through a search engine to obtain initial query information specifically includes: Receive a relationship chain recommendation request submitted by a user through a search engine, wherein the relationship chain recommendation request contains at least one initial query keyword input by the user; The initial query keywords are matched one by one with the entities and relationships in the pre-built entity index and the pre-built relationship index to determine the target entities and target relationships that are successfully matched. For initial query keywords that do not achieve a complete match, the semantic similarity calculation method is used to calculate the semantic similarity score between the initial query keywords and the candidate entities and candidate relationships in the pre-built entity index and the pre-built relationship index. Candidate entities and relationships are selected based on semantic similarity scores and dynamically set semantic similarity thresholds; Based on the target entity, target relationship, alternative entities, and alternative relationships, initial query information containing both exact match results and fuzzy match results is formed.

3. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, The step of performing semantic analysis and intent recognition on the initial query information to determine the corresponding entity type and relationship type specifically involves: Based on the initial query information, semantic parsing is performed on the target entities and target relationships in the complete matching results to determine the semantic type of each target entity and target relationship. Semantic parsing is performed on the candidate entities and candidate relations in the semantic fuzzy matching results to determine the candidate semantic types corresponding to each candidate entity and candidate relation; Based on the semantic types of the target entity and target relation and the candidate semantic types of the alternative entities and alternative relations, the semantic relevance score between the alternative entities and alternative relations and the target entity and target relation is calculated one by one using the semantic relevance calculation method. Based on the semantic relevance score and the preset semantic relevance threshold, the candidate entities and candidate relationships are screened to determine the candidate entities and candidate relationships whose semantic relevance scores meet the threshold. Based on the identified target entities, target relationships, and candidate entities and relationships whose semantic relevance scores meet the threshold, the corresponding entity types and relationship types are determined.

4. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, The process of invoking a dynamic knowledge selection and evaluation mechanism based on entity type and relation type to dynamically filter entities and relations in the original knowledge graph and form a knowledge graph subgraph associated with the initial query information is as follows: Based on the determined entity type and relation type, all entities and relations of the corresponding type are retrieved from the original knowledge graph to form an initial entity set and an initial relation set; Based on a dynamic knowledge selection and evaluation mechanism, the entity association score between each entity in the initial entity set and the initial query information is calculated one by one. Based on a dynamic knowledge selection and evaluation mechanism, the relationship score between each relation in the initial relation set and the initial query information is calculated one by one; Based on entity association scores and relationship association scores, dynamically determined association score thresholds are used to filter the initial entity set and the initial relationship set to determine the entities and relationships that meet the association score thresholds. By using entities that meet the association score threshold as nodes and relationships that meet the association score threshold as edges between nodes, a knowledge graph subgraph associated with the initial query information is constructed.

5. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, The dynamic knowledge selection and evaluation mechanism is as follows: Based on the initial query information, the frequency of historical queries of each entity and relation in the original knowledge graph within a preset time window is counted in real time, and the historical popularity weight of each entity and relation is calculated. Each entity and relation in the original knowledge graph is mapped to a corresponding semantic vector from the initial query information, and the semantic relevance score between each entity and relation and the initial query information is determined. Based on historical popularity weights and semantic relevance scores, the comprehensive effectiveness score of each entity and each relationship is calculated in real time through dynamically adjusted weighted fusion ratios. Based on the comprehensive validity score obtained through real-time calculation and the dynamically determined validity score threshold, entities and relationships that meet the validity score threshold are selected in real time. During the reasoning process of the knowledge graph diffusion model, the comprehensive validity score of the entities and relationships involved in each candidate relation chain path generated by diffusion is obtained in real time, and the diffusion direction and diffusion results of the diffusion model are optimized. The candidate relationship chain paths, after being dynamically adjusted and optimized by diffusion weights, will be output as the final relationship chain path recommendation results.

6. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, The calculation and filtering of the validity score and relevance score for each entity and each relation within the knowledge graph subgraph to obtain an effective knowledge graph subgraph specifically involves: Based on a dynamic knowledge selection and evaluation mechanism, the comprehensive effectiveness score of each entity and each relation in the knowledge graph subgraph is obtained under the current user query request. Based on a dynamic knowledge selection and evaluation mechanism, the semantic relevance score between each entity and each relation in the knowledge graph subgraph and the initial query information is obtained in real time. Based on the comprehensive effectiveness score and semantic relevance score, a dynamically adjusted weighted fusion ratio of effectiveness and relevance is used to calculate the comprehensive screening score of each entity and each relation in the knowledge graph subgraph in real time. The system analyzes the knowledge granularity requirements of the user's current query request in real time and dynamically determines the comprehensive screening score threshold based on the knowledge granularity requirements. Based on the comprehensive screening score and the dynamically determined comprehensive screening score threshold, entities and relationships with comprehensive screening scores below the threshold are filtered and removed in real time. By using the filtered entities as nodes and the filtered relationships as edges between nodes, an effective knowledge graph subgraph is generated.

7. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, The process of generating a candidate relation chain path set based on an effective knowledge graph subgraph and through diffusion reasoning using a knowledge graph diffusion model is as follows: Using the target entity in the initial query information as the starting node for diffusion, the initial set of diffusion nodes is determined in real time based on the effective knowledge graph subgraph. Starting from each node in the initial set of diffusion nodes, determine the neighboring nodes corresponding to each node one by one, and obtain the relationship between each neighboring node and the current node in real time; Calculate the path diffusion score for the current node to spread to its neighboring nodes; Based on the path diffusion score obtained in real time and the path diffusion score threshold determined dynamically, nodes and corresponding relationships that meet the path diffusion score threshold are selected as candidate nodes and candidate relationships for the next round of diffusion. The number of diffusion iterations is determined in real time based on the user's query request, and the diffusion process is executed repeatedly until the number of diffusion iterations is reached; The diffusion path from the starting node to the current diffusion node is recorded and stored in real time, forming a set of candidate relationship chain paths.

8. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, For each candidate relationship chain path in the candidate relationship chain path set, a path quality evaluation score is calculated, specifically as follows: Based on a dynamic knowledge selection and evaluation mechanism, the comprehensive validity score of all nodes and relationships in the candidate relationship chain path is obtained in real time, and the association strength score is calculated in real time by weighting the comprehensive validity scores of each node and relationship in the path. Based on a dynamic knowledge selection and evaluation mechanism, the semantic vector cosine similarity between adjacent nodes and corresponding relationships in the candidate relationship chain path is calculated one by one, and the semantic consistency score of the candidate relationship chain path is determined in real time. The path length score is calculated in real time based on the difference between the actual path length of the candidate relationship chain path and the target path length explicitly stated in the user's current query request. The success frequency of each node and relationship in the candidate relationship chain path in the historical path query task is counted in real time, and the path confidence score of the candidate relationship chain path is calculated based on the success frequency. Based on the association strength score, semantic consistency score, path length score, and path confidence score, the weight ratio of the dynamically adjusted path quality assessment score is determined in real time. Based on the real-time dynamically adjusted weight ratios, the association strength score, semantic consistency score, path length score, and path confidence score are weighted and fused in real time to obtain the path quality evaluation score corresponding to each candidate relationship chain path.

9. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, Based on the path quality assessment score, the candidate relationship chain path set is dynamically filtered again to form the relationship chain path recommendation result, specifically as follows: Based on a dynamic knowledge selection and evaluation mechanism, the path quality evaluation score of each candidate relationship chain path in the candidate relationship chain path set is obtained in real time, and the average path quality evaluation score of the candidate relationship chain path set is calculated in real time. Analyze the knowledge granularity requirements of the user's current query request in real time, and calculate the degree of difference between the knowledge granularity of the current query request and the knowledge granularity of historical user query requests; Based on the degree of difference, the adjustment ratio of the path quality assessment score threshold is dynamically determined; Based on the average path quality assessment score and the dynamically determined path quality assessment score threshold adjustment ratio, the path quality assessment score threshold corresponding to the current query request is calculated in real time. In real time, candidate relationship chain paths with path quality evaluation scores lower than the path quality evaluation score threshold are filtered and removed from the candidate relationship chain path set, while candidate relationship chain paths with path quality evaluation scores equal to or higher than the path quality evaluation score threshold are retained; The retained candidate relationship chain paths are output in real time to form relationship chain path recommendation results.

10. The knowledge graph-based search engine relationship chain recommendation method according to claim 1, characterized in that, The step of outputting the relationship chain path recommendation results to the user and providing an interactive, visual, dynamic relationship chain interface for user exploration is as follows: The relationship chain path recommendation results are visualized in real time in the form of a graphical knowledge graph; The system calculates and displays the comprehensive effectiveness score of each node and relationship in the relationship chain path recommendation results in real time. Get and respond to user interactive operations on any node or relationship in the relationship chain path recommendation results; After users interact with the nodes or relationships in the relationship chain path recommendation results in real time, the system updates the comprehensive effectiveness score of the selected nodes or relationships in real time based on the dynamic knowledge selection and evaluation mechanism, and updates the graphical display of the relationship chain path recommendation results in real time. Real-time statistics and recording of user interaction history on nodes and relationships in the dynamic relationship chain interaction interface; and real-time dynamic optimization and updating of relationship chain path recommendation results based on user interaction history. It receives real-time interactive feedback from users on recommended paths based on a dynamic relationship chain interaction interface.

Images