A search result optimization method and system based on knowledge graph

By identifying and classifying candidate entities with the same name in the knowledge graph and adjusting the feature weights of related paths, the problem of redundant search results caused by ambiguous referencing in user queries is solved, resulting in more accurate search result optimization and improved relevance of search results and user satisfaction.

CN123115235APending Publication Date: 2026-11-17SHENZHEN YISOU BAIYING INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611173652.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-04
Publication Date
2026-11-17

Smart Images

  • Figure CN123115235A_ABST
    Figure CN123115235A_ABST
Patent Text Reader

Abstract

This invention relates to the technical field of search result optimization, and provides a method and system for optimizing search results based on a knowledge graph. The method includes: identifying related paths from a knowledge graph for candidate entities with the same name under different entity classification results, and recording the path source identifier and path classification identifier of the related paths; judging the consistency between the related paths and related paths in the context constraint information based on the path source identifier and path classification identifier, obtaining a path consistency judgment result, and adjusting the path feature weights of the related paths according to the path consistency judgment result, obtaining adjusted path feature weights; generating structured path feature information based on the adjusted path feature weights; and ranking the search results based on the structured path feature information to achieve search result optimization. This invention has the effect of improving the relevance of search results and user satisfaction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the technical field of search result optimization, specifically to a search result optimization method and system based on knowledge graphs. Background Technology

[0002] In the field of multi-source retrieval of scientific and technological intelligence, in order to effectively integrate heterogeneous data such as papers, patents, and projects and optimize search results, existing search engines typically introduce knowledge graphs as a semantic enhancement hub. During normal operation, the system maps entities in the user's query to corresponding nodes in the knowledge graph, and uses these nodes as starting points to mine relevant association path features, which are then injected into the ranking module to improve the relevance of search results.

[0003] However, in actual cross-domain science and technology intelligence retrieval, users often query in a vague manner, such as only entering the name of a researcher and a broad research direction. In this case, the names referenced in the query often lack a unique correspondence in the knowledge graph, thus introducing entity ambiguity in the starting point determination stage of path mining. The problem begins to emerge when such vague name references simultaneously activate multiple candidate entities with the same name in multiple different domains of the knowledge graph, and these candidate entities have highly similar research direction keywords. The system will simultaneously use these sets of potential entities with unclear referencing as valid starting points, and conduct parallel path mining outwards along the entity relationship edges of their respective domains. For example, a path in materials science might link shape memory alloys, a path in computer science might be associated with computational simulations, and a path in biomedicine might point to drug delivery.

[0004] Because existing conventional methods often fail to effectively distinguish the semantic context differences arising from different starting points when aggregating path features, these messy cross-domain path features are indiscriminately mixed together. This approach not only causes the final generated composite path features to lose semantic focus, but also, when driving the ranking module, it promotes a large amount of redundant cross-domain content unrelated to the user's true intent to the top of the search results, causing the search optimization goal to fail. Therefore, in multi-source retrieval scenarios for scientific and technological intelligence, when the ambiguous reference of a person's name in a user's query simultaneously activates multiple domain candidate entities in the knowledge graph, the path mining proceeds in parallel from entities starting from different semantic contexts and is ultimately aggregated indiscriminately, causing the path features to lose semantic focus. Consequently, the search result ranking introduces redundant cross-domain content unrelated to the user's true intent.

[0005] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0006] This application discloses a search result optimization method and system based on knowledge graphs, aiming to solve the problem in multi-source retrieval of scientific and technological intelligence. When the ambiguous reference of a person's name in a user's query simultaneously activates multiple candidate entities in the knowledge graph, the path mining starts from the entity starting point in different semantic contexts and is carried out in parallel and is eventually aggregated without distinction. This causes the path features to lose semantic concentration, which in turn introduces cross-domain redundant content that is unrelated to the user's true intent into the search result ranking.

[0007] The technical solution of this application is as follows: Firstly, this application discloses a search result optimization method based on knowledge graphs, comprising the following steps: The system obtains entity referential information and contextual constraint information from user queries, and identifies multiple candidate entities with the same name and their corresponding local association information from a pre-defined knowledge graph based on the entity referential information. Entity referential information includes the name, abbreviation, alias, or contextual pronoun used in the user query to point to the entity to be retrieved. Contextual constraint information includes the constraint information used in the user query to limit the meaning of the entity, the search field, the time range, the research direction, or the application scenario. Based on contextual constraint information and local association information, the degree of matching between each candidate entity with the same name and the contextual constraint information is determined, and the candidate entities with the same name are classified according to the degree of matching to obtain the entity classification result; For candidate entities with the same name under different entity classification results, identify the associated paths from the knowledge graph and record the path source identifier and path classification identifier of the associated paths; Based on the path source identifier and path classification identifier, the consistency between the associated path and the related associated path in the context constraint information is determined to obtain the path consistency judgment result. The path feature weights of the associated path are then adjusted according to the path consistency judgment result to obtain the adjusted path feature weights. Based on the adjusted path feature weights, structured path feature information is generated; Search results are sorted based on structured path feature information to optimize search results.

[0008] This technical solution effectively addresses the problems of decreased semantic concentration and introduction of redundant information in search results caused by entity ambiguity and indiscriminate aggregation of cross-domain paths in existing technologies. By classifying candidate entities with the same name, identifying related paths, and adjusting the weights of path features, it achieves precise optimization of search results, significantly improving the relevance of search results and user satisfaction.

[0009] Secondly, this application also discloses a knowledge graph-based search result optimization system for performing knowledge graph-based search result optimization, including: The local association recognition module is used to obtain entity referential information and contextual constraint information in the user query, and based on the entity referential information, to identify multiple candidate entities with the same name and their corresponding local association information from the preset knowledge graph; The entity classification execution module is used to determine the degree of matching between each candidate entity with the same name and the context constraint information based on the context constraint information and the local association information, and to classify the candidate entities with the same name according to the degree of matching to obtain the entity classification result; The association path identification module is used to identify association paths from the knowledge graph for candidate entities with the same name under different entity classification results, and record the path source identifier and path classification identifier of the association path. The feature weight adjustment module is used to determine the consistency between the associated path and the related associated path in the context constraint information based on the path source identifier and the path classification identifier, obtain the path consistency judgment result, and adjust the path feature weight of the associated path according to the path consistency judgment result to obtain the adjusted path feature weight. The feature information generation module is used to generate structured path feature information based on the adjusted path feature weights. The search result optimization module is used to sort search results based on structured path feature information in order to optimize search results.

[0010] This application provides a system-level solution through a modular design, which divides the functions of each key step in the search result optimization method, making the entire system structure clear, easy to implement and maintain, and capable of efficiently executing knowledge graph-based search result optimization, providing users with more accurate and relevant search results.

[0011] Beneficial Effects: This application discloses a search result optimization method and system based on knowledge graphs. By acquiring entity referential information and contextual constraint information from user queries, and identifying candidate entities with the same name and local association information from the knowledge graph, it effectively solves the problem of entity ambiguity caused by vague referential information in multi-source retrieval of scientific and technological intelligence. This method further determines the degree of matching between each candidate entity with the contextual constraint information and classifies the candidate entities based on the contextual constraint information, thereby avoiding the drawback of parallel path mining of entities from different semantic contexts in existing technologies, which are ultimately aggregated indiscriminately. For candidate entities with the same name under different entity classification results, this application identifies related paths from the knowledge graph and records the path source identifier and path classification identifier, providing a refined basis for subsequent adjustment of path feature weights. More importantly, this application judges the consistency between the related path and the related related paths in the contextual constraint information based on the path source identifier and path classification identifier, and adjusts the path feature weights of the related paths according to the judgment results, thereby effectively solving the problem of loss of semantic concentration of path features and avoiding the promotion of a large amount of cross-domain redundant content unrelated to the user's true intent to the front end of search results. Finally, by generating structured path feature information based on the adjusted path feature weights and sorting the search results according to this information, this application achieves accurate optimization of search results, significantly improves the relevance of search results and user satisfaction, and overcomes the shortcomings of existing technologies in failing to achieve search optimization goals. Attached Figure Description

[0012] Figure 1 This is a flowchart of a search result optimization method based on knowledge graphs in one embodiment of the present invention; Figure 2 This is a flowchart of a search result optimization method based on knowledge graphs according to another embodiment of the present invention; Figure 3 This is a system block diagram of a knowledge graph-based search result optimization system according to another embodiment of the present invention; Explanation of reference numerals in the attached figures: 1. Knowledge graph-based search result optimization system; 11. Local association recognition module; 12. Entity classification execution module; 13. Association path recognition module; 14. Feature weight adjustment module; 15. Feature information generation module; 16. Search result optimization module. Detailed Implementation

[0013] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0014] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0015] This application proposes a search result optimization method based on knowledge graphs, combining... Figure 1 As shown, it includes: S1. Obtain entity referential information and contextual constraint information from the user query, and based on the entity referential information, identify multiple candidate entities with the same name and their corresponding local association information from the preset knowledge graph; entity referential information includes the name, abbreviation, alias or contextual referent used in the user query to point to the entity to be retrieved; contextual constraint information includes the constraint information used in the user query to limit the meaning of the entity, the search field, the time range, the research direction or the application scenario; S2, based on contextual constraint information and local association information, determine the degree of matching between each candidate entity with the same name and the contextual constraint information, and classify the candidate entities with the same name according to the degree of matching to obtain the entity classification result; S3: For candidate entities with the same name under different entity classification results, identify the associated paths from the knowledge graph and record the path source identifier and path classification identifier of the associated paths; S4. Based on the path source identifier and path classification identifier, determine the consistency between the associated path and the related associated path in the context constraint information, obtain the path consistency judgment result, and adjust the path feature weight of the associated path according to the path consistency judgment result to obtain the adjusted path feature weight. S5. Generate structured path feature information based on the adjusted path feature weights; S6 sorts the search results based on structured path feature information to optimize the search results.

[0016] To better understand the method proposed in this application, it is necessary to explain some of the key terms involved.

[0017] "Entity referential information" refers to the various forms of expression used in user queries to point to the entity to be retrieved, such as names of people, places, organizations, etc., which can be full names, abbreviations, aliases, or even pronouns in the context.

[0018] "Contextual constraint information" refers to additional information used in user queries to limit the meaning of entities, the scope of retrieval, or the scenario, such as specific research fields, time ranges, and application scenarios.

[0019] A knowledge graph is a database that stores knowledge in a graph structure, where nodes represent entities and edges represent relationships between entities.

[0020] "Candidate entities with the same name" refers to multiple entities in a knowledge graph that have the same name but may represent different concepts or belong to different domains.

[0021] "Local association information" refers to the entity, attribute, and relationship information that are directly or indirectly related to a certain entity in the knowledge graph.

[0022] "Association path" refers to a series of relational chains that connect two or more entities in a knowledge graph, revealing the deep semantic connections between entities.

[0023] The “Path Source Identifier” is used to mark which candidate entity with the same name the associated path was mined from.

[0024] "Path classification identifier" is used to mark the entity classification result to which the associated path belongs.

[0025] "Path feature weight" is a numerical value that measures the importance or relevance of a related path.

[0026] "Structured path feature information" refers to a data format with a certain structure that, after processing and weight adjustment, can clearly express the semantic relationships between entities.

[0027] First, the system obtains entity referential information and contextual constraint information from the user query. Based on the entity referential information, it identifies multiple candidate entities with the same name and their corresponding local association information from a pre-defined knowledge graph. The user query can be a keyword, phrase, or a complete natural language statement. For example, when a user enters "Li Ming, machine learning," the system parses the query and identifies the entity referential information and contextual constraint information. The entity referential information can be "Li Ming," and the contextual constraint information can be "machine learning." Subsequently, based on the entity referential information, the system retrieves the entity corresponding to that entity referential information from the pre-defined knowledge graph. Since "Li Ming" may correspond to multiple entities with the same name, the knowledge graph may contain multiple entities with the same name but different fields, identities, or associated objects, such as a "Li Ming" in the field of computer science and a "Li Ming" in the field of biomedicine. These entities are the candidate entities with the same name. While identifying candidate entities with the same name, the system extracts local association information of each candidate entity in the knowledge graph. Local association information may include information such as papers, patents, projects, research directions, affiliated institutions, collaborators, technical topics or domain tags that are directly or closely related to the candidate entity with the same name.

[0028] Secondly, based on contextual constraints and local association information, the system determines the degree of matching between each candidate entity with the contextual constraints and the information. The candidate entities are then categorized according to this matching degree to obtain entity classification results. In practice, the system semantically matches the contextual constraints in the user query with the local association information of each candidate entity to determine whether the candidate entity matches the user's query intent. For example, for a computer science entity named "Li Ming," its local association information might include papers, projects, or research directions related to machine learning, deep learning, model training, and algorithm optimization; for a biomedical entity named "Li Ming," its local association information might involve gene editing, drug development, medical imaging, or clinical research. The system can calculate the degree of matching between contextual constraints and the local association information of each candidate entity using methods such as keyword matching, semantic similarity calculation, vector retrieval, or graph relationship matching. Based on the degree of matching, candidate entities can be classified into categories that highly match the contextual constraints, categories with strong associations with other domains, or uncertain categories with insufficient matching, thus obtaining entity classification results.

[0029] Secondly, for candidate entities with the same name under different entity classification results, the system identifies related paths from the knowledge graph and records the path source identifier and path classification identifier of the related paths. Specifically, after obtaining the entity classification results, the system starts with the candidate entity with the same name under the corresponding category and expands the path along the entity relationship edge in the knowledge graph to identify related paths related to contextual constraint information. For example, for the entity "Li Ming" classified in the field of machine learning, the system can start from this entity and identify related paths such as "Li Ming - published papers - paper topic - machine learning algorithm", "Li Ming - project participation - project name - machine learning application", and "Li Ming - co-author - research direction - deep learning". The related path can include the entity nodes on the path, the relationship type, the path length, and the path endpoint object. When identifying related paths, the system records the path source identifier and path classification identifier for each related path. The path source identifier is used to indicate which specific candidate entity with the same name the related path originates from; the path classification identifier is used to indicate the entity classification result to which the related path belongs. For example, if a certain association path originates from the entity "Li Ming" in the field of computer science, and this entity is classified in the field of machine learning, then the path source identifier of this association path is the corresponding entity "Li Ming" in the field of computer science, and the path classification identifier is the field of machine learning.

[0030] Next, based on the path source identifier and path category identifier, the consistency between the associated path and the relevant associated paths in the context constraint information is determined, resulting in a path consistency judgment result. The path feature weights of the associated paths are then adjusted based on this result, yielding the adjusted path feature weights. In specific implementation, the system compares the path category identifier, entities and relationships on the path, the path endpoint topic, and the context constraint information for each associated path to determine whether the associated path matches the user's query intent. Figure 1 For example, when the context constraint information is "machine learning," associated paths connecting to machine learning algorithms, machine learning projects, deep learning papers, or related technical topics can be judged as having high consistency with the context constraint information; while associated paths connecting to unrelated fields such as biomedicine and drug development can be judged as having low consistency or semantic conflicts. The system adjusts the path feature weights of associated paths based on the path consistency judgment results. For associated paths that are highly consistent with the context constraint information, their path feature weights are increased; for associated paths that are weakly related to the context constraint information, have mismatched path classification labels, or have domain conflicts, their path feature weights are decreased. Through this weight adjustment process, associated paths that are more relevant to the user's actual query intent can be highlighted, and the influence of unrelated paths caused by ambiguity due to entities with the same name can be suppressed.

[0031] Then, based on the adjusted path feature weights, structured path feature information is generated. Specifically, the system integrates each associated path and its adjusted path feature weights to form structured information usable by the search ranking module. Structured path feature information can be represented as a vector or as a structured list containing path source identifiers, path category identifiers, path type, path length, entity nodes on the path, relationship types, and corresponding path feature weights. As one implementation, the path feature weights of multiple associated paths can be weighted and summarized according to path type or entity category to form a path feature vector representing the user's query intent. Alternatively, the node sequence, relationship sequence, and adjusted path feature weights of the associated paths can be input into a graph neural network or other representation learning model to generate fixed-dimensional structured path feature information. Through this step, the system can uniformly encode entity disambiguation results, context matching results, and knowledge graph associated path information into feature representations that facilitate subsequent ranking processing.

[0032] Finally, the search results are ranked based on structured path feature information to optimize the search results. In practice, the search ranking module can use structured path feature information as one of the ranking features, and combine it with other ranking factors such as document content relevance, user behavior data, page quality, timeliness, and historical click feedback to re-rank the initially obtained search results. When a search result contains key entities, relationships, or path topics with high weight in the structured path feature information, its ranking position can be improved; when a search result is mainly associated with entity paths with low weight or inconsistent with contextual constraints, its ranking position can be reduced. For example, when a user queries "Li Ming, machine learning," papers, project introductions, academic homepages, or research results pages related to "Li Ming" in the field of machine learning can obtain a higher ranking because they match high-weight associated paths; while results related to the same "Li Ming" in other fields but unrelated to machine learning will have their ranking positions reduced accordingly. Through the above processing, this application can optimize search ranking results by utilizing associated paths and path feature weights in the knowledge graph when user queries contain entity ambiguity, mixed entities with the same name, or cross-domain expressions, thereby improving the matching accuracy between search results and user query intent.

[0033] Optionally, the steps to adjust the path feature weights of associated paths based on the path consistency judgment results to obtain the adjusted path feature weights include: When the path consistency judgment result indicates that there is a potential cross-class conflict between the associated path and the related associated path in the context constraint information, the feature vector distance between the two paths is calculated. When the feature vector distance is greater than the preset conflict distance threshold, the weight reduction action of the related paths in the context constraint information is paused, and the endpoint nodes of the related paths in the context constraint information are extracted in the knowledge graph. For the endpoint nodes of the relevant associated paths in the associated path and context constraint information, check whether there is a bridging entity between the two endpoint nodes in the local relation subgraph, and check whether there is a co-occurrence relationship between the two endpoint nodes in the global knowledge graph, and obtain the local topology verification result and the global co-occurrence verification result. When the local topology verification result indicates that there is a bridging entity in the local relational subgraph that connects two endpoint nodes at the same time, or the global co-occurrence verification result indicates that the co-occurrence frequency of two endpoint nodes as cited tags in the same document entity in the global knowledge graph reaches the preset condition, it is determined that the association path and the related association path in the context constraint information belong to the interdisciplinary integration path. When it is determined that the related paths in the context constraint information belong to interdisciplinary integration paths, an interdisciplinary collaboration tag is added to the related paths in the context constraint information. Calculate the corresponding interdisciplinary enhancement coefficient based on the interdisciplinary collaborative markers; The respective interdisciplinary enhancement coefficients are multiplied by the original path weights of the related paths in the associated paths and context constraint information to obtain the adjusted path feature weights. When both the local topology verification result and the global co-occurrence verification result are invalid, the related paths in the context constraint information are confirmed as true cross-class conflict paths. Conflict attenuation weighting is then applied to the related paths in the context constraint information to obtain the adjusted path feature weights. When the path consistency judgment result indicates that the associated path is consistent with the related associated path in the context constraint information, the original path weight of the associated path is retained as the adjusted path feature weight.

[0034] Specifically, when the path consistency assessment indicates a potential cross-class conflict between the associated path and the related paths in the context constraint information, it is necessary to calculate the feature vector distance between these two paths. This feature vector distance can be understood as the degree of difference between the two paths in the semantic space, used to initially assess the severity of their conflict. The preset conflict distance threshold is a configurable parameter used to define what level of distance is considered a potential conflict requiring further review.

[0035] Furthermore, when the feature vector distance exceeds a preset conflict distance threshold, the system pauses the deweighting of these paths and proceeds with a deeper analysis. At this point, the endpoint nodes of these two paths in the knowledge graph are extracted. The endpoint node refers to the final entity or concept that the path points to.

[0036] For these endpoint nodes, two types of verification are performed: first, local topology verification, which checks whether there is a bridging entity in the local relational subgraph that can connect the two endpoint nodes simultaneously, indicating that they are related in a certain local context; second, global co-occurrence verification, which checks whether the two endpoint nodes are both cited tags in the same document entity in the global knowledge graph, and whether their co-occurrence frequency meets the preset conditions, indicating that they have common concerns or research backgrounds in a more macro context.

[0037] When any verification result is true, i.e., a bridging entity exists or the co-occurrence frequency condition is met, these paths are determined to belong to interdisciplinary fusion paths. This means that although they may differ in classification, they are interconnected in deep semantics or structure. At this point, interdisciplinary collaboration labels are attached to these paths, and corresponding interdisciplinary enhancement coefficients are calculated based on these labels. These enhancement coefficients aim to increase the weights of these fusion paths to reflect their potential collaborative value. Finally, these enhancement coefficients are multiplied into the original path weights to obtain the adjusted path feature weights.

[0038] Conversely, if both the local topology check and the global co-occurrence check results are invalid, these paths are confirmed as genuine cross-class conflict paths. In this case, the system will perform conflict attenuation and weight reduction operations on these paths to reduce their influence on search result optimization, thereby avoiding the introduction of irrelevant or misleading information.

[0039] Furthermore, when the path consistency judgment result indicates that the associated path is consistent with the related associated path in the context constraint information, the original path weight of the associated path remains unchanged and is used as the adjusted path feature weight, because no additional adjustment is required at this time.

[0040] Optionally, for candidate entities with the same name under different entity classification results, the steps of identifying association paths from the knowledge graph and recording the path source identifier and path classification identifier of the association paths include: For each candidate entity with the same name under each entity classification result, identify the association path related to the context constraint information from the knowledge graph; Record the candidate entity identifier with the same name that is identified in the associated path as the initial path source identifier; Record the entity classification result of the associated path as the initial path classification identifier; Determine whether there is a potential cross-category semantic fusion relationship between the associated paths obtained from different candidate entities with the same name; When there is no potential cross-category semantic fusion relationship, the initial path source identifier is used as the path source identifier, and the initial path classification identifier is used as the path classification identifier. When there is a potential cross-category semantic fusion relationship, the initial path classification identifier is temporarily removed, the associated path is forcibly isolated, and the endpoint node of the associated path is extracted. Determine whether the endpoint nodes share common semantic connections in the knowledge graph; When there are no common semantic connection points, the initial path classification identifier is restored, and the initial path source identifier is used as the path source identifier. The restored initial path classification identifier is used as the path classification identifier. When common semantic connection points exist, the initial path classification identifier is updated to a cross-category fusion identifier; The initial path source identifier is used as the path source identifier, and the cross-category fusion identifier is used as the path classification identifier.

[0041] Specifically, for each candidate entity with the same name under each entity classification result, the system retrieves and identifies association paths from the knowledge graph that are highly relevant to the contextual constraints in the user's query. These association paths can be direct relationships between entities or indirect relationships connected through multiple intermediate nodes. The identifier of the candidate entity with the same name from which the association path was identified is recorded as the initial path source identifier, the purpose of which candidate entity with the same name initially led to the path. Simultaneously, the entity classification result to which the association path belongs is recorded as the initial path classification identifier, used to initially define the semantic scope of the path.

[0042] Furthermore, the system determines whether there is a potential cross-category semantic fusion relationship between the associated paths identified from different candidate entities with the same name. This fusion relationship means that although these paths may initially belong to different entity categories, they may overlap or merge at a deeper semantic level. For example, an entity may have different meanings in different domains, but its associated paths may point to the same concept in certain specific contexts. When there is no potential cross-category semantic fusion relationship, it indicates that the semantic boundaries of these paths are clear. In this case, the initial path source identifier and the initial path classification identifier can be directly used as the final path source identifier and path classification identifier.

[0043] However, when potential cross-category semantic fusion relationships exist, to avoid misjudgment or simplify processing, the system temporarily removes the initial path classification identifier and forcibly isolates the associated paths for more detailed analysis. Based on this, the system extracts the endpoint nodes of the associated paths, i.e., the final entity or concept that the path points to. Subsequently, the system determines whether the endpoint nodes share common semantic connections in the knowledge graph. Common semantic connections refer to shared nodes in the knowledge graph that can connect or explain multiple entities or concepts of different categories, such as an interdisciplinary concept or a multi-domain expert. When no common semantic connections exist, it indicates that although potential cross-category semantic fusion relationships exist, there is no explicit shared semantic basis between the endpoint nodes of these paths. Therefore, the system restores the initial path classification identifier and uses it as the final path classification identifier, while also using the initial path source identifier as the final path source identifier. Conversely, when common semantic connections exist, this strongly indicates that these associated paths do indeed exhibit cross-category semantic fusion. In this case, the initial path classification identifier is updated to a cross-category fusion identifier to explicitly mark that the path has cross-category fusion characteristics. Meanwhile, the initial path source identifier remains the path source identifier, while the updated cross-category fusion identifier becomes the final path classification identifier.

[0044] Optional, combined Figure 2 As shown, the steps for determining the matching degree between each candidate entity with the same name and the context constraint information, based on context constraint information and local association information, include: A1, parse the time range words and research direction words in the context constraint information to identify the specific time period and research direction pointed to by the user query; A2. When the start or end time of a specific time period is later than the most recent update time of the knowledge graph, the real-time information supplementation process is initiated. A3 calls an external real-time database interface through a real-time information supplementation process and uses entity reference information and research direction words as search conditions to obtain newly published relevant text content within a specific time period. A4 performs entity recognition and keyword extraction on relevant text content to generate real-time supplementary related information; A5 integrates the real-time supplemented association information with the local association information of each candidate entity with the same name in the knowledge graph to obtain the integrated association information; A6. Adjust the weight of real-time supplementary related information in the fused related information based on the interval between the release time of the real-time supplementary related information and the current query time; A7. Based on the weight of the adjusted real-time supplementary association information in the fused association information, the matching degree is calculated according to the fused association information and the context constraint information to obtain the matching degree.

[0045] Specifically, when determining the degree of matching between each candidate entity with the same name and the contextual constraint information, the contextual constraint information is first parsed to identify potential time-range terms and research direction terms. These terms are used to clarify the specific time period and research direction that the user is interested in. For example, time-range terms could be "latest in 2023" or "development in the last five years," while research direction terms could be "frontiers of artificial intelligence" or "advances in quantum computing."

[0046] Furthermore, when the start or end time of a specific identified time period is later than the most recent update time of the preset knowledge graph, it indicates that the knowledge graph may lack the latest information for that time period. In this case, the system will initiate a real-time information supplementation process. This process aims to compensate for the timeliness deficiencies of the knowledge graph.

[0047] As a preferred implementation, the real-time information replenishment process is achieved by calling an external real-time database interface. This interface can connect to real-time updated data sources such as news databases, academic paper repositories, and social media data streams. During the call, the entity references in the user's query and identified research direction terms are used as search criteria to retrieve newly published relevant text content from these external data sources within a specific time period.

[0048] After acquiring the relevant text content, entity recognition and keyword extraction are performed to generate real-time supplementary related information. Entity recognition aims to identify other entities related to the user's query entity from the text, while keyword extraction is used to capture the core themes and key concepts of the text. This real-time supplementary related information reflects the latest developments related to the query entity and research direction within a specific time period.

[0049] Subsequently, this real-time supplementary association information is merged with the existing local association information of each candidate entity with the same name in the knowledge graph to obtain merged association information. The merged association information combines the structured and historical information of the knowledge graph with the timeliness information of external real-time data, providing a more comprehensive and timely view of entity associations.

[0050] To make more efficient use of real-time supplementary related information, its weight in the fused related information is adjusted based on the interval between its publication time and the current query time. Generally, the closer the publication time of real-time supplementary related information is to the current query time, the stronger its timeliness, and therefore it will be given a higher weight; conversely, the weight may be appropriately reduced.

[0051] Finally, based on the adjusted weights of the real-time supplementary association information in the fused association information, the matching degree is calculated according to the fused association information and the contextual constraint information, thus obtaining the final matching degree. The matching degree calculation can employ various semantic similarity algorithms, such as methods based on vector space models, deep learning models, or graph neural network models, to quantify the relevance between the fused association information and the user's contextual constraints.

[0052] Optionally, the step of determining the degree of matching between each candidate entity with the same name and the context constraint information based on context constraint information and local association information specifically includes: Semantically expand the constraint words in the context constraint information to generate an expanded word set; For each candidate entity with the same name, extract the core keywords and related entities from the corresponding local association information; Calculate the semantic similarity between the expanded word set and the core topic words and related entities; By combining the domain affiliation information of candidate entities with the same name in the knowledge graph, the matching relationship between the domain to which the candidate entity with the same name belongs and the domain pointed to by the context constraint information is determined, and the semantic similarity is corrected according to the matching relationship to obtain the domain matching corrected similarity. Analyze the co-occurrence patterns of each constraint word in the contextual constraint information and identify co-occurring constraint word pairs; When the local association information of candidate entities with the same name simultaneously covers the co-occurrence constraint word pair, the domain matching similarity is enhanced based on the co-occurrence constraint word pair to obtain the collaborative enhancement score; When the local association information of candidate entities with the same name does not simultaneously cover the co-occurrence constraint word pairs, the domain matching correction similarity is used as the basic matching score; The matching degree is generated based on the collaborative enhancement score or the basic matching score.

[0053] Specifically, semantic expansion of constraint words in contextual constraint information refers to using tools such as word vector models, thesaurus, or domain ontology to identify words semantically related to the original constraint words, thereby generating an expanded word set containing broader semantic information. Its purpose is to compensate for potential lexical sparsity in user queries and capture deeper semantic intent.

[0054] For each candidate entity with the same name, core keywords and related entities are extracted from the corresponding local association information. This can be understood as identifying the key terms and related entities that best represent the core concepts of the entity and its knowledge graph context from the attributes, relationships, and descriptive texts that are directly or indirectly related to the entity in the knowledge graph. The purpose is to provide rich and representative features for subsequent semantic similarity calculations.

[0055] In practical applications, the semantic similarity between the expanded word set and the core topic words and related entities is calculated. For example, cosine similarity, Jaccard similarity, or deep learning-based semantic matching models can be used to quantify the semantic association strength between the expanded word set and entity features. The purpose is to preliminarily assess the relevance between the user's query intent and candidate entities with the same name.

[0056] Furthermore, by combining the domain affiliation information of candidate entities with the same name in the knowledge graph, the matching relationship between the domain to which the candidate entity belongs and the domain pointed to by the contextual constraints is determined. The semantic similarity is then corrected based on the matching relationship to obtain a domain-corrected similarity. This means that if a candidate entity with the same name is identified as belonging to a specific domain (e.g., "computer science"), and the contextual constraints of the user's query also explicitly point to that domain, the semantic similarity between the entity and the query will be enhanced; conversely, if the domains do not match, the similarity may be weakened. The aim is to introduce domain knowledge to improve the domain specificity and accuracy of the matching.

[0057] Furthermore, analyzing the co-occurrence patterns of constraint words in contextual constraint information and identifying co-occurring constraint word pairs refers to discovering word combinations that frequently appear together in user queries and have specific semantic relationships through statistical analysis or pre-set rules. For example, in a query for "Apple Inc.", "apple" and "industry" constitute a co-occurrence constraint word pair. The purpose is to capture the more complex semantic structure and intent in user queries.

[0058] When the local association information of candidate entities with the same name simultaneously covers co-occurrence constraint word pairs, the domain matching similarity is enhanced based on the co-occurrence constraint word pairs, resulting in a collaborative enhancement score. This means that if a candidate entity with the same name not only matches the query semantically and in terms of domain, but its local association information also reflects the collaborative relationship of key constraint words in the query, then its matching degree will be significantly improved. The aim is to reward entities that can comprehensively and structurally respond to the user's complex query intent.

[0059] When the local association information of candidate entities with the same name does not simultaneously cover the co-occurrence constraint word pairs, the domain-corrected similarity is used as the base matching score. This means that if an entity fails to reflect a specific co-occurrence pattern in the query, its matching degree will be based on the domain-corrected semantic similarity.

[0060] Finally, a matching degree is generated based on either the collaborative enhancement score or the basic matching score. This matching degree is a comprehensive quantitative indicator that reflects the final degree of matching between the candidate entities with the same name and the contextual constraints of the user query, and is used for subsequent entity classification and search result optimization.

[0061] Optionally, the steps of combining the domain affiliation information of candidate entities with the same name in the knowledge graph to determine the matching relationship between the domain to which the candidate entity with the same name belongs and the domain pointed to by the contextual constraint information, and correcting the semantic similarity according to the matching relationship to obtain the domain matching corrected similarity include: Obtain multiple domain identifiers from the domain affiliation information of candidate entities with the same name in the knowledge graph; For each domain identifier, extract the semantic association strength between the candidate entities with the same name in the corresponding domain and the context constraint information; Analyze the semantic overlap between multiple domain identifiers to identify cross-domain pairs or domain pairs with ambiguous boundaries; Calculate the co-occurrence frequency and semantic distance of cross-domain pairs or fuzzy boundary domain pairs in the knowledge graph; Domain fusion weights are determined based on co-occurrence frequency and semantic distance; The semantic association strength is adjusted based on the domain fusion weight to obtain the fused and adjusted semantic association strength. Based on the semantic association strength after fusion and correction, the semantic similarity is corrected by domain matching to obtain the domain matching corrected similarity.

[0062] In this context, obtaining multiple domain identifiers refers to identifying and extracting all relevant domains that an entity may be involved in from the domain affiliation information associated with candidate entities with the same name in the knowledge graph. For example, an entity may belong to both the domains of "computer science" and "artificial intelligence," in which case both domains will be obtained as domain identifiers.

[0063] Furthermore, for each domain identifier, the semantic association strength between the corresponding candidate entities with the same name and the contextual constraint information within that domain is extracted. The purpose is to quantify the semantic relevance between the candidate entities with the same name and the contextual constraint information in the user query within a specific domain context. This can be achieved by calculating the semantic similarity between entity descriptions, attributes, relationships, and contextual constraint information within that domain, for example, using word vector models or graph embedding techniques.

[0064] Building upon this foundation, the semantic overlap between multiple domain identifiers is analyzed to identify cross-domain pairs or fuzzy-boundary domain pairs. Semantic overlap refers to the degree of sharing in concepts, terminology, or research scope between different domains. Cross-domain pairs refer to two or more domains that share significant and explicit common research themes or application scenarios; for example, "bioinformatics" is an intersection of "biology" and "computer science." Fuzzy-boundary domain pairs refer to domains where the boundaries are unclear, with a large number of shared concepts but no clear interdisciplinary framework yet formed. Identifying these domain pairs helps to understand the multi-domain attributes of entities more precisely.

[0065] Subsequently, the co-occurrence frequency and semantic distance of cross-domain pairs or fuzzy boundary domain pairs in the knowledge graph are calculated. Co-occurrence frequency refers to the number of times these domain pairs appear together in the knowledge graph, such as the frequency of co-occurrence in the same document, the same project, or the same entity description. Semantic distance measures the proximity of these domains in the semantic space; the smaller the distance, the closer the semantics.

[0066] Domain fusion weights are determined based on co-occurrence frequency and semantic distance. Domain fusion weight is a quantitative indicator used to represent the degree of fusion or collaboration between different domains. Domain pairs with higher co-occurrence frequency and closer semantic distance generally have larger domain fusion weights, indicating a more semantically cohesive relationship between these domains.

[0067] Therefore, the semantic association strength is corrected based on the domain fusion weights to obtain the fused and corrected semantic association strength. When candidate entities with the same name involve multiple domains, and there are overlapping or fuzzy boundaries between these domains, the previously extracted semantic association strength can be weighted or adjusted by the domain fusion weights to more accurately reflect the true degree of association between the entity and the contextual constraint information in a multi-domain context.

[0068] Finally, based on the semantic association strength after fusion and correction, the semantic similarity is corrected by domain matching to obtain the domain matching corrected similarity. This means that when calculating the overall matching degree between candidate entities with the same name and contextual constraint information, the semantic association characteristics of entities in multi-domain contexts will be fully considered, thus obtaining a more comprehensive and accurate semantic similarity value.

[0069] Optionally, the steps of analyzing the co-occurrence patterns of each constraint word in the contextual constraint information and identifying co-occurring constraint word pairs include: Identify domain-specific abbreviations or polysemous words in contextual constraint information; Extracting semantic interpretations of domain-specific abbreviations or polysemous words in different domains; By combining the domain affiliation information of candidate entities with the same name in the knowledge graph, the semantic interpretation is domain-limited to obtain the domain-limited semantic interpretation. Based on the semantic interpretation after domain limitation, semantic disambiguation and expansion are performed on the constraint words in the context constraint information to generate a domain semantic representation; Based on the domain semantic representation, calculate the semantic association strength of each constraint word pair in the specific domain; Based on the semantic association strength, identify co-occurrence constraint word pairs.

[0070] Specifically, when identifying domain-specific abbreviations or polysemous words in contextual constraints, pre-built domain dictionaries, abbreviation dictionaries, or polysemous word dictionaries can be used for matching. For example, for a user query "AI applications in the medical field," the system will identify that "AI" is likely a domain-specific abbreviation, while "applications" is likely a polysemous word. Extracting the semantic interpretation of domain-specific abbreviations or polysemous words in different domains refers to obtaining the specific meanings of these words in different contexts or domains from knowledge graphs or external semantic resources. For example, "AI" may refer to "artificial intelligence" in computer science, but may have other meanings in other fields; "applications" may refer to "application applications" in software engineering, but may refer to "applied research" in biology.

[0071] In practical applications, the semantic interpretation is domain-limited by combining the domain affiliation information of candidate entities with the same name in the knowledge graph, resulting in a domain-limited semantic interpretation. This means that if a candidate entity with the same name is assigned to the "Computer Science" domain, then the semantic interpretation of "AI" will be limited to "Artificial Intelligence," thus excluding other irrelevant interpretations. The purpose is to ensure that subsequent semantic processing is performed within the correct domain context. Furthermore, based on the domain-limited semantic interpretation, semantic disambiguation and expansion are performed on the constraint words in the contextual constraint information to generate a domain semantic representation. Semantic disambiguation aims to eliminate the ambiguity of polysemous words, giving them a unique and definite meaning within a specific domain; semantic expansion aims to capture broader semantic information related to the constraint words, such as synonyms, hypernyms, or related concepts, to enrich their semantic representation. This results in a more accurate and comprehensive domain semantic representation. Based on this, the semantic association strength of each constraint word pair within the specific domain is calculated according to the domain semantic representation. This can be achieved through various methods, such as calculating the cosine similarity of word vectors, similarity based on knowledge graph paths, or co-occurrence frequency. A higher semantic association strength indicates a greater likelihood that the two constraint words will co-occur within a specific domain, or a closer semantic relationship between them. Ultimately, co-occurring constraint word pairs are identified based on the semantic association strength. For example, a threshold can be set; when the semantic association strength between two constraint words exceeds this threshold, they are identified as co-occurring constraint word pairs.

[0072] Optionally, the steps for calculating the matching degree based on the weight of the adjusted real-time supplementary association information in the fused association information, and the matching degree between the fused association information and the context constraint information, include: Read the real-time supplementary association information from the fused association information, and obtain the local association information of the knowledge graph corresponding to the fused association information; Obtain the data source identifier, publication time, and current query time for real-time supplementary related information, and obtain the data source identifier and graph update time for local related information in the knowledge graph; Obtain the reliability level corresponding to the data source identifier; Calculate the real-time publishing interval between the publishing time and the current query time, and calculate the map update interval between the map update time and the current query time; Calculate the real-time information quality score based on the reliability level and the real-time release interval; Calculate the graph information quality score of local association information in the knowledge graph based on the reliability level and graph update interval; When the real-time information quality score is lower than the preset quality threshold, the real-time supplementary related information is filtered or downweighted to obtain the real-time supplementary related information after quality processing. When the quality score of the knowledge graph information is lower than the preset quality threshold, the local association information of the knowledge graph is filtered or downweighted to obtain the local association information of the knowledge graph after quality processing. The real-time supplement weight is obtained based on the real-time release interval, the map update interval, and the weight of the adjusted real-time supplementary related information in the fused related information. Based on the real-time supplementation weights, feature extraction and weighted fusion are performed on the real-time supplementation association information after quality processing and the local association information of the knowledge graph after quality processing to generate fused association feature information. Feature extraction is performed on context constraint information to generate context constraint feature information; The matching degree is calculated based on the fusion of related feature information and contextual constraint feature information.

[0073] Specifically, when reading real-time supplementary related information from the fused related information, the text content, entity recognition results, and keyword extraction results obtained from the real-time information supplementation process can be parsed from the fused related information. Simultaneously, local related information extracted from the knowledge graph corresponding to this real-time supplementary related information is obtained. This information typically includes the attributes, relationships, and domains of candidate entities with the same name. The data source identifier can be understood as the original source of the real-time supplementary related information or the local related information in the knowledge graph, such as a news website, academic journal database, social media platform, or specific knowledge base. The reliability level refers to the assessment of the trust level of different data sources. It can be pre-set to multiple levels such as high, medium, and low, or represented by a continuous numerical value between 0 and 1, aiming to quantify the authority and accuracy of the data source. The publication time refers to the timestamp when the real-time supplementary related information was published, while the graph update time refers to the time when the corresponding local related information in the knowledge graph was last updated. The current query time refers to the moment when the user initiated the search request. In practical applications, the real-time publication interval refers to the time difference between the publication time and the current query time, reflecting the timeliness of the real-time information. The graph update interval refers to the time difference between the graph update time and the current query time, reflecting the timeliness of knowledge graph information. These time intervals can be calculated in seconds, minutes, hours, or days.

[0074] Furthermore, the calculation of real-time information quality scores and knowledge graph information quality scores aims to comprehensively consider the reliability and timeliness of the information. For example, a weighted average model can be used, with the reliability level and the reciprocal of the time interval as weighting factors. When the quality score of any information source is lower than a preset quality threshold, the information can be filtered out, i.e., completely rejected, or its weight can be reduced, i.e., its influence in subsequent matching degree calculations can be decreased. The purpose is to eliminate the interference of low-quality or outdated information on the matching results. Therefore, after obtaining the quality-processed real-time supplementary association information and knowledge graph local association information, it is necessary to obtain the real-time supplementary weight based on the real-time release interval, the knowledge graph update interval, and the adjusted weight of the real-time supplementary association information in the fused association information. This real-time supplementary weight can be a dynamically adjusted coefficient used to balance the contributions of real-time information and knowledge graph information in the final feature fusion. For example, when the real-time information is very new and of high quality, its weight can be appropriately increased. Finally, feature extraction and weighted fusion are performed on the quality-processed real-time supplementary association information and the quality-processed knowledge graph local association information to generate fused association feature information. These features can include word vectors, topic model features, entity relationship features, etc. Simultaneously, context constraint information is extracted to generate context constraint feature information. Finally, based on these fused related feature information and context constraint feature information, the degree of matching between candidate entities with the same name and the context constraint information is calculated using cosine similarity, neural network models, or other machine learning algorithms.

[0075] Optionally, the steps for calculating the corresponding matching degree based on the fused association feature information and contextual constraint feature information include: Identify whether there are sentiment words or subjective bias words in the contextual constraint information; When sentiment words or subjective bias words are present, extract text fragments related to research attitudes, viewpoints, or evaluations from the integrated and associated information; Sentiment analysis is performed on text fragments to obtain sentiment scores for candidate entities with the same name on specific issues; Based on the sentiment tendency score, the fused correlation feature information is corrected to obtain sentiment-corrected fused correlation feature information; The matching degree is calculated based on the sentiment correction fusion of related feature information and context constraint feature information; When no sentiment words or subjective bias words are available, the corresponding matching degree is calculated directly based on the fusion of related feature information and contextual constraint feature information.

[0076] Specifically, before calculating the matching degree, the contextual constraints are first analyzed to identify whether they contain sentiment words or subjective bias words. Sentiment words can refer to words that express emotions such as joy, anger, sorrow, and happiness, such as "like," "dislike," "excellent," and "terrible." Subjective bias words can refer to words that express personal opinions, evaluations, or attitudes, such as "think," "feel," "suggest," and "criticize." This identification process can be achieved through a pre-set sentiment dictionary, a subjective bias dictionary, or a machine learning-based sentiment classification model.

[0077] When the system detects sentiment words or subjective bias words in the contextual constraints, it extracts text fragments related to these sentiments or biases from the fused association information. The fused association information is a comprehensive set of information combining real-time supplementary association information and local association information from the knowledge graph, which may contain various descriptions, comments, evaluations, and other textual content about entities or events. The extracted text fragments should reflect the research attitudes, viewpoints, or evaluations of the candidate entity with the same name on a specific issue. For example, if the query involves "positive evaluations of a certain technology," then positive evaluation text about that technology will be extracted from the fused association information.

[0078] Subsequently, sentiment analysis is performed on the extracted text fragments. Sentiment analysis is a natural language processing technique that aims to identify whether the sentiment expressed in the text is positive, negative, or neutral, and to quantify its intensity, thereby obtaining a sentiment score for candidate entities with the same name on a specific issue. This analysis can employ rule-based methods, dictionary-based methods, or deep learning model-based methods. For example, the ratio of positive to negative words in the text can be calculated, or a pre-trained sentiment classifier can be used to score the text.

[0079] Based on the obtained sentiment score, the fused correlation feature information is revised. This revision aims to incorporate the dimension of sentiment or subjective tendency into the original fused correlation feature, making it more comprehensively reflect the user's query intent. For example, if the sentiment score shows a strong positive tendency, the weight of the dimension related to positive evaluation in the fused correlation feature can be increased, or a new sentiment feature dimension can be introduced. The revised feature information is called sentiment-revised fused correlation feature information.

[0080] Finally, based on sentiment correction and the fusion of related feature information and contextual constraint feature information, the corresponding matching degree is calculated. This calculation process can use various methods such as cosine similarity, dot product, and neural network models, but its input features already include considerations of sentiment or subjective bias.

[0081] On the other hand, when the contextual constraints do not contain sentiment words or subjective bias words, it indicates that the user's query may focus more on objective facts or neutral information. In this case, the system will directly calculate the corresponding matching degree based on the original fused association feature information and contextual constraint feature information, without the need for sentiment correction.

[0082] This application also discloses a knowledge graph-based search result optimization system for performing knowledge graph-based search result optimization. Combined with... Figure 3 As shown, the knowledge graph-based search result optimization system 1 includes: The local association recognition module 11 is used to obtain entity referential information and context constraint information in the user query, and based on the entity referential information, to identify multiple candidate entities with the same name and their corresponding local association information from the preset knowledge graph. The entity classification execution module 12 is used to determine the degree of matching between each candidate entity with the same name and the context constraint information based on the context constraint information and the local association information, and to classify the candidate entities with the same name according to the degree of matching to obtain the entity classification result. The association path identification module 13 is used to identify association paths from the knowledge graph for candidate entities with the same name under different entity classification results, and record the path source identifier and path classification identifier of the association path; The feature weight adjustment module 14 is used to determine the consistency between the associated path and the related associated path in the context constraint information based on the path source identifier and the path classification identifier, obtain the path consistency judgment result, and adjust the path feature weight of the associated path according to the path consistency judgment result to obtain the adjusted path feature weight. The feature information generation module 15 is used to generate structured path feature information based on the adjusted path feature weights. The search result optimization module 16 is used to sort the search results based on structured path feature information in order to optimize the search results.

[0083] To better understand the system proposed in this application, it is necessary to explain some of the key modules involved.

[0084] The local association identification module is designed to acquire entity referential information and contextual constraint information from user queries, and based on the entity referential information, identify multiple candidate entities with the same name and their corresponding local association information from a pre-defined knowledge graph. Specifically, this module can receive user-input queries, such as "Li Ming, machine learning". The module first parses the user query, identifying the entity referential information (e.g., "Li Ming") and contextual constraint information (e.g., "machine learning"). Then, based on the identified entity referential information, the module searches the pre-defined knowledge graph to identify all candidate entities with the same name "Li Ming," which may belong to different fields. Simultaneously, the module extracts the local association information of each candidate entity in the knowledge graph, such as papers, patents, projects, research directions, or collaborators directly associated with that entity.

[0085] The entity classification execution module determines the degree of matching between each candidate entity with the same name and the contextual constraints based on contextual constraint information and local association information. It then classifies the candidate entities based on this matching degree to obtain the entity classification result. After the local association identification module obtains multiple candidate entities with the same name and their local association information, the entity classification execution module evaluates the degree of matching between each candidate entity and the contextual constraints (e.g., "machine learning") in the user query. For example, for a computer scientist named "Li Ming," the local association information might include a large number of papers, projects, and research directions related to "machine learning"; while for a biomedical expert named "Li Ming," the local association information might be more related to "gene editing" or "drug development." This module can determine the degree of matching by calculating the semantic similarity between the contextual constraint information and the local association information of each candidate entity with the same name. For example, a word vector model can be used to calculate the similarity between "machine learning" and the related words of the computer scientist "Li Ming," as well as the similarity between "machine learning" and the related words of the biomedical expert "Li Ming." Based on these matching degrees, the module can classify the candidate entities with the same name. For example, "Li Ming" with a high degree of matching with "machine learning" can be classified into one category, "Li Ming" with a high degree of matching with "biomedicine" can be classified into another category, or entities with a low degree of matching can be classified into the "uncertain" category.

[0086] The association path identification module identifies association paths from the knowledge graph for candidate entities with the same name under different entity classification results, and records the path source identifier and path classification identifier of the association paths. After the entity classification execution module classifies the candidate entities with the same name, the association path identification module identifies association paths related to contextual constraints from the knowledge graph for each category of candidate entities with the same name. For example, for the entity "Li Ming" classified as "machine learning field", this module starts from this entity and follows the relation edges in the knowledge graph to mine paths connecting this entity with entities such as concepts, technologies, and applications related to "machine learning". These paths may include "Li Ming - published papers - paper topics - machine learning algorithms", "Li Ming - project participation - project name - machine learning applications", etc. While identifying these association paths, the module records the path source identifier of each path, i.e., which specific candidate entity with the same name the path started mining from, and the path classification identifier, i.e., the entity classification result to which the path belongs. For example, if the path is mined starting from "Computer Scientist Li Ming" and the entity belongs to the "Machine Learning" category, then the path source is identified as "Computer Scientist Li Ming" and the path category is identified as "Machine Learning".

[0087] The feature weight adjustment module determines the consistency between associated paths and related paths in the context constraint information based on path source identifiers and path classification identifiers. It then adjusts the path feature weights of the associated paths based on this consistency judgment, resulting in adjusted path feature weights. After the associated path identification module identifies the associated paths and records their source and classification identifiers, the feature weight adjustment module needs to determine the consistency between these associated paths and the related paths implied in the context constraint information of the user query. For example, if the context constraint information of the user query is "machine learning," then paths highly related to "machine learning" are considered "consistent." If the path classification identifier of a path highly matches the domain indicated by the context constraint information, it is considered to have high consistency. Conversely, if the path classification identifier conflicts with the context constraint information, such as a user querying "Li Ming, machine learning," but a path's classification identifier is "biomedical field," then inconsistency exists. Based on this consistency judgment result, the module adjusts the path feature weights of the associated paths. For example, paths that are highly consistent with contextual constraints may have their weight increased, while paths that are inconsistent or conflicting may have their weight decreased. This adjustment helps to highlight paths that are more relevant to the user's true intent and suppress irrelevant paths.

[0088] The feature information generation module generates structured path feature information based on the adjusted path feature weights. After the feature weight adjustment module adjusts the feature weights of all identified associated paths, the feature information generation module integrates this path information with the adjusted weights to generate structured path feature information. Structured path feature information can be a vector representation or a list containing path type, path length, entities and relationships on the path, and corresponding weights. For example, all adjusted path feature weights can be weighted and summed, or the path information can be encoded into a fixed-dimensional vector using a deep learning model. This structured information can be more effectively utilized by the subsequent search ranking module because it already contains path features that are highly relevant to the user's query intent after semantic analysis and weight adjustment.

[0089] The search result optimization module sorts search results based on structured path feature information to optimize the search results. After the feature information generation module generates structured path feature information, the search result optimization module inputs it as one of the important ranking features into the search ranking module. The search ranking module comprehensively considers various factors such as document content relevance, user behavior, and page quality, and combines them with the structured path feature information generated in this application to re-rank the initially retrieved search results. For example, search results that highly match the structured path feature information (such as documents containing key entities and relationships in the path) will be ranked higher. In this way, search results can more accurately reflect the user's query intent, especially when the user's query contains entity ambiguity or cross-domain needs. The system of this application can effectively improve the relevance and accuracy of search results, thereby optimizing the search results.

[0090] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1.A knowledge graph based search result optimization method, characterized in that, include: The system obtains entity referential information and contextual constraint information from the user query, and identifies multiple candidate entities with the same name and their corresponding local association information from a preset knowledge graph based on the entity referential information. The entity referential information includes the name, abbreviation, alias, or contextual referent used in the user query to point to the entity to be retrieved. The contextual constraint information includes the constraint information used in the user query to limit the meaning of the entity, the search field, the time range, the research direction, or the application scenario. Based on the context constraint information and the local association information, the degree of matching between each candidate entity with the same name and the context constraint information is determined, and the candidate entities with the same name are classified according to the degree of matching to obtain the entity classification result; For candidate entities with the same name under different entity classification results, identify the associated paths from the knowledge graph and record the path source identifier and path classification identifier of the associated paths; Based on the path source identifier and the path classification identifier, the consistency between the associated path and the related associated path in the context constraint information is determined to obtain the path consistency judgment result. The path feature weight of the associated path is then adjusted according to the path consistency judgment result to obtain the adjusted path feature weight. Based on the adjusted path feature weights, structured path feature information is generated; Based on the structured path feature information, the search results are sorted to optimize the search results. 2.The knowledge graph-based search result optimization method of claim 1, wherein, The step of adjusting the path feature weights of the associated paths based on the path consistency judgment result to obtain the adjusted path feature weights includes: When the path consistency judgment result indicates that there is a potential cross-class conflict between the associated path and the related associated path in the context constraint information, the feature vector distance between the two paths is calculated. When the feature vector distance is greater than a preset conflict distance threshold, the weight reduction action of the associated path and the related associated path in the context constraint information is paused, and the endpoint node of the associated path and the related associated path in the context constraint information in the knowledge graph is extracted. For the endpoint nodes of the associated paths and the related associated paths in the context constraint information, check whether there is a bridging entity between the two endpoint nodes in the local relation subgraph, and check whether there is a co-occurrence relationship between the two endpoint nodes in the global knowledge graph, to obtain the local topology verification result and the global co-occurrence verification result; When the local topology verification result indicates that there is a bridging entity in the local relation subgraph that connects two endpoint nodes at the same time, or when the global co-occurrence verification result indicates that the co-occurrence frequency of the two endpoint nodes as citation tags in the same document entity in the global knowledge graph reaches a preset condition, it is determined that the association path and the related association path in the context constraint information belong to the interdisciplinary integration path. When it is determined that the associated path and the related associated path in the context constraint information belong to the interdisciplinary integration path, an interdisciplinary collaboration tag is added to the associated path and the related associated path in the context constraint information; Calculate the corresponding interdisciplinary enhancement coefficient based on the interdisciplinary collaborative markers; The respective interdisciplinary enhancement coefficients are multiplied by the original path weights of the associated paths and the related associated paths in the context constraint information to obtain the adjusted path feature weights. When both the local topology verification result and the global co-occurrence verification result are invalid, it is confirmed that the associated path and the related associated path in the context constraint information are real cross-class conflict paths, and conflict attenuation weighting is performed on the associated path and the related associated path in the context constraint information to obtain the adjusted path feature weights. When the path consistency judgment result indicates that the associated path is consistent with the related associated path in the context constraint information, the original path weight of the associated path is maintained as the adjusted path feature weight. 3.The knowledge graph-based search result optimization method of claim 1, wherein, The steps of identifying related paths from the knowledge graph for candidate entities with the same name under different entity classification results, and recording the path source identifier and path classification identifier of the related paths, include: For each candidate entity with the same name under each entity classification result, identify the association path related to the context constraint information from the knowledge graph; Record the candidate entity identifier with the same name that was identified in the associated path as the initial path source identifier; Record the entity classification result of the associated path as the initial path classification identifier; Determine whether there is a potential cross-category semantic fusion relationship between the associated paths obtained from different candidate entities with the same name; When no potential cross-category semantic fusion relationship exists, the initial path source identifier is used as the path source identifier, and the initial path classification identifier is used as the path classification identifier; When the potential cross-category semantic fusion relationship exists, the initial path classification identifier is temporarily eliminated, the associated path is forcibly isolated, and the endpoint node of the associated path is extracted; Determine whether the endpoint nodes have common semantic connections in the knowledge graph; When the common semantic connection point does not exist, the initial path classification identifier is restored, and the initial path source identifier is used as the path source identifier, and the restored initial path classification identifier is used as the path classification identifier; When the common semantic connection point exists, the initial path classification identifier is updated to a cross-category fusion identifier; The initial path source identifier is used as the path source identifier, and the cross-category fusion identifier is used as the path classification identifier. 4.The knowledge graph-based search result optimization method of claim 1, wherein, The step of determining the matching degree between each candidate entity with the same name and the context constraint information based on the context constraint information and the local association information includes: Parse the time range words and research direction words in the context constraint information to identify the specific time period and research direction pointed to by the user query; When the start or end time of the specific time period is later than the most recent update time of the knowledge graph, the real-time information supplementation process is initiated. The real-time information supplementation process calls an external real-time database interface and uses the entity reference information and the research direction keywords as query conditions to obtain newly published relevant text content within the specific time period. Entity recognition and keyword extraction are performed on the relevant text content to generate real-time supplementary related information; The real-time supplementary association information is fused with the local association information of each candidate entity with the same name in the knowledge graph to obtain fused association information; The weight of the real-time supplementary related information in the fused related information is adjusted based on the interval between the publication time of the real-time supplementary related information and the current query time. Based on the adjusted weight of the real-time supplementary association information in the fused association information, the matching degree is calculated according to the fused association information and the context constraint information to obtain the matching degree. 5.The knowledge graph based search result optimization method of claim 1, wherein, The step of determining the matching degree between each candidate entity with the same name and the context constraint information based on the context constraint information and the local association information includes: Semantically expand the constraint words in the context constraint information to generate an expanded word set; For each candidate entity with the same name, extract the core keywords and related entities from the corresponding local association information; Calculate the semantic similarity between the extended word set and the core topic words and the associated entities; By combining the domain affiliation information of the candidate entities with the same name in the knowledge graph, the matching relationship between the domain to which the candidate entities with the same name belong and the domain pointed to by the context constraint information is determined, and the semantic similarity is corrected according to the matching relationship to obtain the domain matching corrected similarity. Analyze the co-occurrence patterns of each constraint word in the context constraint information to identify co-occurring constraint word pairs; When the local association information of the candidate entities with the same name simultaneously covers the co-occurrence constraint word pair, the domain matching correction similarity is enhanced based on the co-occurrence constraint word pair to obtain a collaborative enhancement score. When the local association information of the candidate entities with the same name does not simultaneously cover the co-occurrence constraint word pair, the domain matching correction similarity is used as the basic matching score; A matching degree is generated based on the collaborative enhancement score or the basic matching score. 6.The knowledge graph-based search result optimization method of claim 5, wherein, The step of combining the domain affiliation information of the candidate entities with the same name in the knowledge graph, determining the matching relationship between the domain to which the candidate entities with the same name belong and the domain pointed to by the context constraint information, and correcting the semantic similarity according to the matching relationship to obtain the domain matching corrected similarity includes: Multiple domain identifiers are obtained from the domain affiliation information of the candidate entities with the same name in the knowledge graph; For each domain identifier, extract the semantic association strength between the candidate entity with the same name in the corresponding domain and the context constraint information; Analyze the semantic overlap between multiple domain identifiers to identify cross-domain pairs or domain pairs with ambiguous boundaries; Calculate the co-occurrence frequency and semantic distance of the cross-domain pair or the fuzzy boundary domain pair in the knowledge graph; The domain fusion weight is determined based on the co-occurrence frequency and the semantic distance; The semantic association strength is corrected according to the domain fusion weight to obtain the fused and corrected semantic association strength; Based on the semantic association strength after fusion correction, the semantic similarity is corrected by domain matching to obtain the domain matching corrected similarity. 7.The knowledge graph-based search result optimization method of claim 5, wherein, The step of analyzing the co-occurrence patterns of each constraint word in the context constraint information and identifying co-occurring constraint word pairs includes: Identify domain-specific abbreviations or polysemous words in the context constraint information; Extract the semantic interpretations of domain-specific abbreviations or polysemous words in different domains; By combining the domain affiliation information of the candidate entities with the same name in the knowledge graph, the semantic interpretation is domain-limited to obtain the domain-limited semantic interpretation; Based on the semantic interpretation after domain limitation, semantic disambiguation and expansion are performed on the constraint words in the context constraint information to generate a domain semantic representation; Based on the domain semantic representation, calculate the semantic association strength of each constraint word pair in the specific domain; Based on the semantic association strength, identify co-occurrence constraint word pairs. 8.The knowledge graph based search result optimization method of claim 4, wherein, The step of calculating the matching degree based on the weight of the adjusted real-time supplementary association information in the fused association information, and the matching degree between the fused association information and the context constraint information, includes: Read the real-time supplementary association information from the fused association information, and obtain the local association information of the knowledge graph corresponding to the fused association information; Obtain the data source identifier, publication time, and current query time of the real-time supplementary related information, and obtain the data source identifier and graph update time of the local related information of the knowledge graph; Obtain the reliability level corresponding to the data source identifier; Calculate the real-time publishing interval between the publication time and the current query time, and calculate the map update interval between the map update time and the current query time; Calculate the real-time information quality score of the real-time supplementary related information based on the reliability level and the real-time release interval; Calculate the graph information quality score of the local association information of the knowledge graph based on the reliability level and the graph update interval; When the real-time information quality score is lower than a preset quality threshold, the real-time supplementary related information is filtered or downweighted to obtain the real-time supplementary related information after quality processing. When the quality score of the knowledge graph information is lower than a preset quality threshold, the local association information of the knowledge graph is filtered or downweighted to obtain the local association information of the knowledge graph after quality processing. The real-time supplement weight is obtained based on the real-time release interval, the map update interval, and the adjusted weight of the real-time supplementary association information in the fused association information; Based on the real-time supplementation weights, feature extraction and weighted fusion are performed on the quality-processed real-time supplementation association information and the quality-processed local association information of the knowledge graph to generate fused association feature information. Feature extraction is performed on the context constraint information to generate context constraint feature information; The matching degree is calculated based on the fused association feature information and the context constraint feature information. 9.The knowledge graph-based search result optimization method of claim 8, wherein, The step of calculating the corresponding matching degree based on the fused association feature information and the context constraint feature information includes: Identify whether there are sentiment words or subjective bias words in the context constraint information; When the emotional words or subjective bias words are present, extract text fragments related to research attitudes, viewpoints, or evaluations from the fused association information; Sentiment analysis is performed on the text fragments to obtain the sentiment scores of candidate entities with the same name on specific issues; Based on the sentiment tendency score, the fused correlation feature information is corrected to obtain sentiment-corrected fused correlation feature information; Based on the sentiment correction fusion association feature information and the context constraint feature information, the corresponding matching degree is calculated; When the sentiment word or the subjective tendency word is not present, the corresponding matching degree is calculated directly based on the fused association feature information and the context constraint feature information. 10.A knowledge graph based search result optimization system for performing knowledge graph based search result optimization, characterized in that, include: The local association recognition module is used to obtain entity referential information and context constraint information in the user query, and based on the entity referential information, to identify multiple candidate entities with the same name and their corresponding local association information from a preset knowledge graph; The entity classification execution module is used to determine the degree of matching between each candidate entity with the same name and the context constraint information based on the context constraint information and the local association information, and to classify the candidate entities with the same name according to the degree of matching to obtain the entity classification result; The association path identification module is used to identify association paths from the knowledge graph for candidate entities with the same name under different entity classification results, and record the path source identifier and path classification identifier of the association path; The feature weight adjustment module is used to determine the consistency between the associated path and the related associated path in the context constraint information based on the path source identifier and the path classification identifier, obtain the path consistency judgment result, and adjust the path feature weight of the associated path according to the path consistency judgment result to obtain the adjusted path feature weight. The feature information generation module is used to generate structured path feature information based on the adjusted path feature weights. The search result optimization module is used to sort the search results based on the structured path feature information in order to optimize the search results.