A knowledge graph-based book resource intelligent retrieval method and system

Abstract

The present application relates to the field of information retrieval, knowledge graph retrieval and search engine resource discovery, and particularly relates to a book resource intelligent retrieval method and system based on a knowledge graph. The method obtains multi-source book data and user behavior data, obtains target multi-source book data through standardization, cleaning and entity relationship structuring processing, constructs a book resource knowledge graph and generates relationship labels and relationship weights; when processing a search request, entity information and facet information are extracted and semantics are parsed to form a structured search request, target nodes and facet constraints are obtained by matching graph nodes and relationship labels, candidate book node sets are obtained based on relationship weights through correlation reasoning, and the final candidate book node set is obtained by calling source data for standardized preprocessing and semantic matching, and a matching score is generated to output a book resource list and a theoretical basis in descending order. The present application can improve the clarity of candidate book node sources, the stability of retrieval matching and the result interpretability.

Description

Technical Field

[0001] This invention relates to the fields of information retrieval, knowledge graph retrieval, and search engine-based resource discovery, and particularly to an intelligent book resource retrieval method and system based on knowledge graphs. Background Technology

[0002] In the fields of information retrieval, knowledge graph retrieval, and search engine-style resource discovery, existing solutions typically first collect multi-source book data, then establish search entry points based on book titles, authors, themes, keywords, classifications, abstract attributes, publication dates, library holdings, and digital resource content, and return a list of book resources based on the search request. Some solutions also incorporate user behavior data, entity extraction, relation extraction, attribute association processing, and book resource knowledge graphs to expand search requests or organize the relationships between book resources.

[0003] Existing methods largely rely on keywords, entity information, or faceted information in the search request for matching. A typical approach involves first extracting query terms from the search request, then matching book titles, authors, themes, keywords, or categories, and finally outputting search results according to search function options and resource retrieval scope. This type of method can complete the retrieval when the search request matches the source data fields, but when there is an indirect relationship between the search request and the book resource, it easily limits the retrieval process to direct field matching. For hypernym, hyponym, synonym, and related term relationship tags between themes, existing solutions typically treat them as simple extended conditions, failing to form a unified candidate book node generation process with target nodes, faceted constraints, and relationship weights.

[0004] In applications of knowledge graphs for book resources, existing technologies typically focus on the construction of the knowledge graph, node matching, entity information extraction, and relation tag organization. Due to the lack of coordinated processing between target nodes, faceted constraints, and relation weights, associative reasoning easily degenerates into ordinary path traversal. During path expansion, library status, publication time, resource type attributes, and search function options are often processed in the backend screening stage, failing to participate in boundary constraints during path expansion. This results in the inclusion of weakly related book nodes in the candidate book node set and makes it difficult for the nodes traversed by the path, relation tags, and relation weights to jointly support subsequent processing.

[0005] When search requests involve natural language expressions, topic associations, or limitations on the scope of resource retrieval, existing solutions are prone to the problem of candidate book node generalization. Some book nodes may enter the candidate scope through behavioral nodes, click results, or borrowing records, but the book titles, topics, keywords, categories, abstract attributes, and digital resource document content in their source data do not semantically match the structured search request. If such candidate book nodes are directly used to generate a list of book resources, the search process will be affected by the path expansion range and user behavior data, making it difficult to ensure a clear correspondence between candidate book nodes and the search request.

[0006] For the joint processing of target nodes, facet constraints, relation weights, candidate book node sets, and source data, existing technologies generally lack a continuous process from structured retrieval requests to relational reasoning and then to semantic matching of source data. Existing methods struggle to establish a consistent link in book resource retrieval scenarios: "structured retrieval request—target node—facet constraints—relation weights—candidate book node set—source data—final candidate book node set," resulting in unclear path origins, source data basis, and matching processes for candidate book nodes. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a knowledge graph-based intelligent book resource retrieval method, comprising: S100. Acquire multi-source book data and user behavior data, and obtain target multi-source book data through data standardization, data cleaning and entity relationship structuring. S200. Based on the target multi-source book data, a book resource knowledge graph is constructed through entity extraction, relation extraction, and attribute association processing, and relation labels and relation weights are generated. S300. Based on the book resource knowledge graph, process the search request, extract entity information and faceted information, parse semantics, and obtain a structured search request; S400. Based on the structured retrieval request, match the nodes and relationship tags in the book resource knowledge graph to obtain the target node and facet constraints; S500: Based on the target node, the facet constraint and the relation weight, perform association reasoning to obtain the set of nodes traversed by the path and the set of candidate book nodes; S600. Based on the candidate book node set, the source data is called, and after normalization preprocessing and semantic matching, the final candidate book node set is obtained; S700. Based on the final candidate book node set, the nodes traversed by the path, the relationship weights, and the source data, generate a matching score, and output the book resource list and theoretical basis in descending order of the matching score.

[0008] Furthermore, the multi-source book data includes book title, author, subject, keywords, classification, abstract attribute, publication time, collection status, and digital resource document content; The user behavior data includes historical search records, search session streams, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time.

[0009] Further, S100 includes: performing field standardization processing on the multi-source book data using a unified metadata standard to obtain standardized multi-source book data; The standardized multi-source book data is subjected to data mapping processing to obtain mapped multi-source book data; The mapped multi-source book data is subjected to data deduplication, missing value filling, formatting, and outlier handling to obtain cleaned multi-source book data. The cleaned multi-source book data is processed by identifier allocation, fuzzy matching, and entity relation structuring to obtain the target multi-source book data.

[0010] Furthermore, the book resource knowledge graph includes book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes; The relationship tags include tags for relationships between books and authors, between books and topics, between books and categories, between books and collection status, between superordinate terms between topics, between subordinate terms between topics, between synonyms between topics, between related terms between topics, between books and borrowing records, between books and click results, between search terms and click results, between click results and subsequent operations, and between user dwell time and books.

[0011] Further, S200 includes: extracting retrieval session stream, query terms, click results, subsequent operations, and borrowing records from the user behavior data; Generate sequence co-occurrence information based on the query term and the click results; Generate behavioral feedback information based on the click result, subsequent operations, borrowing records, and user dwell time; The sequence co-occurrence information and the behavioral feedback information are written into the book resource knowledge graph, and the relation weights are weighted and updated in real time.

[0012] Furthermore, the entity information includes the book title, author, subject, keywords, category, and resource type in the search request; The facet information includes publication date, collection status, search function options, and resource search scope; The structured search request includes the entity information, the faceted information, the semantics, and the search intent.

[0013] Further, S400 includes: matching the entity information with book nodes, author nodes, topic nodes, category nodes, and keyword nodes in the book resource knowledge graph; The faceted information is matched with the library collection nodes, publication time attributes, resource type attributes, and search function options in the book resource knowledge graph; The nodes that are successfully matched form the target node, and the attributes that are successfully matched form the faceting constraint. Unmatched entity information is assisted by hypernyms, hyponyms, synonyms, and related words. Nodes that are successfully matched are incorporated into the target node.

[0014] Further, S500 includes: Starting from the target node, filter the relational labels that participate in the association reasoning according to the facet constraints; The path is expanded along the filtered relation labels to obtain the nodes traversed by the path; The path expansion results are filtered based on the node types of the nodes traversed by the path, the relation weights of the relation labels, and the facet constraints. The book nodes pointed to by the filtered paths form the candidate book node set.

[0015] Further, S600 includes: Extract book titles, authors, themes, keywords, categories, abstract attributes, publication dates, library holdings status, and digital resource document content from the source data; The book title, subject, keywords, classification, and abstract attributes are preprocessed to obtain title terms, subject terms, keywords, and abstract terms. Construct a resource entity feature vector based on the title words, the subject words, the keywords, and the abstract words; Construct a knowledge retrieval feature vector based on the entity information, faceted information, and semantics in the structured retrieval request; The resource entity feature vector is semantically matched with the knowledge retrieval feature vector, and combined with the relationship tags in the corresponding paths of the candidate book nodes to obtain the final set of candidate book nodes.

[0016] Furthermore, a knowledge graph-based intelligent book resource retrieval system includes: a data preprocessing module, a knowledge graph construction module, a search request parsing module, a node matching module, an association reasoning module, a source data matching module, and a sorting output module; the system is used to implement the method described in any of the above embodiments.

[0017] The key innovations of this invention include: (1) First, the structured search request is matched with the nodes and relation tags in the book resource knowledge graph to obtain the target node and facet constraints. Then, based on the target node, the facet constraints and the relation weights, association reasoning is performed to obtain the nodes traversed by the path and the set of candidate book nodes.

[0018] (2) Introduce the relation weight into the association reasoning process, so that the relation weight, the target node, and the facet constraint jointly participate in the generation of the candidate book node set, and make the nodes traversed by the path retain the relation label source.

[0019] (3) Based on the candidate book node set, the source data is called, and after normalization preprocessing and semantic matching, the final candidate book node set is obtained, so that the candidate book node set is processed by the source data matching before entering the matching score generation.

[0020] The following are its main beneficial effects: (1) In view of the problem that existing methods rely on direct matching of keywords, entity information or faceted information, this invention forms a candidate book node set by structured search request, target node, faceted constraint and association reasoning, so that the search request and the nodes and relationship tags in the book resource knowledge graph form a continuous processing link, reducing the situation where the search process only stays within the scope of direct field matching.

[0021] (2) In view of the problem that the association reasoning in the existing book resource knowledge graph application is easy to turn into ordinary path traversal, the present invention writes the relation weight and facet constraint into the association reasoning process, so that the candidate book node set retains the nodes passed by the path, relation labels and relation weights, which makes it easier to generate the final candidate book node set and matching score based on the same data basis.

[0022] (3) In view of the problem that candidate book nodes are easily affected by the path expansion range, the present invention calls source data based on the candidate book node set, and performs normalized preprocessing and semantic matching on the source data, so that the final candidate book node set has both path source and source data basis.

[0023] (4) For book nodes that enter the candidate range through behavior nodes, click results or borrowing records, the present invention combines book title, theme, keywords, classification, abstract attributes and digital resource document content for semantic matching during the source data matching process, so that user behavior data still maintains a corresponding relationship with the source data when participating in the retrieval link.

[0024] (5) In view of the problem that the theoretical basis in the output of the book resource list is not clear enough, the present invention generates a matching score based on the final candidate book node set, the nodes passed through the path, the relation weight and the source data, and outputs the book resource list and theoretical basis in descending order of the matching score, so that the book resources in the book resource list and the structured search request retain a traceable processing record. Attached Figure Description

[0025] Figure 1 A flowchart illustrating a knowledge graph-based intelligent book resource retrieval method provided in this application embodiment; Figure 2 This is a structural block diagram of a knowledge graph-based intelligent book resource retrieval system provided in an embodiment of this application. Detailed Implementation

[0026] Example 1: Refer to Figure 1 This is a flowchart illustrating a knowledge graph-based intelligent book resource retrieval method provided in an embodiment of the present invention. The process may include at least steps S100-S700: S100. Acquire multi-source book data and user behavior data, and obtain target multi-source book data through data standardization, data cleaning and entity relationship structuring. S200. Based on the target multi-source book data, after entity extraction, relation extraction and attribute association processing, a book resource knowledge graph is constructed, and relation labels and relation weights are generated; S300. Based on the book resource knowledge graph, process the search request, extract entity information and faceted information, parse semantics, and obtain a structured search request; S400. Based on the structured retrieval request, match the nodes and relationship tags in the book resource knowledge graph to obtain the target node and facet constraints; S500: Based on the target node, the facet constraint and the relation weight, perform association reasoning to obtain the set of nodes traversed by the path and the set of candidate book nodes; S600. Based on the candidate book node set, the source data is called, and after normalization preprocessing and semantic matching, the final candidate book node set is obtained; S700. Based on the final candidate book node set, the nodes traversed by the path, the relationship weights, and the source data, generate a matching score, and output the book resource list and theoretical basis in descending order of the matching score.

[0027] S100. Acquire multi-source book data and user behavior data, and obtain target multi-source book data through data standardization, data cleaning and entity relationship structuring. Specifically, the data preprocessing module receives data from heterogeneous data sources when the intelligent book resource retrieval method is launched. These heterogeneous data sources include library business databases, e-book platforms, digitized scanned text libraries, user-generated content platforms, retrieval logs, and borrowing systems. The multi-source book data describes the book resources, including book titles, authors, themes, keywords, classifications, abstract attributes, publication dates, collection status, and digital resource document content. The user behavior data describes user behavior during the retrieval, browsing, and borrowing process, including historical search records, search session flow, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time. After receiving the multi-source book data and the user behavior data, the data preprocessing module records the data source, collection time, field names, data format, and data identifier, generates an original access record, and binds and stores the original access record with the multi-source book data and the user behavior data.

[0028] Specifically, the data standardization is performed by the data preprocessing module calling a unified metadata standard. This unified metadata standard configures corresponding field names, field types, and field value rules for book titles, authors, themes, keywords, classifications, abstract attributes, publication dates, collection status, and digital resource document content. The data preprocessing module first reads the field names from the multi-source book data, then groups fields representing the same content from different heterogeneous data sources into a unified field. For example, the book title field from the e-book platform, the title field from the digitized scanned text library, and the book title field from the library's business database are grouped into the book title field; the in-library status from the collection system, the borrowable status from the borrowing system, and the resource status from the e-book platform are grouped into the collection status field. For the user behavior data, the data preprocessing module groups search terms from the search logs into the query terms field, clicked objects from the browsing records into the click results field, borrowing and returning records from the borrowing system into the borrowing records field, and dwell time from page access records into the user dwell time field, completing the field standardization process to obtain standardized multi-source book data.

[0029] Specifically, the data preprocessing module performs data mapping processing on the standardized multi-source book data. This data mapping process involves grouping records of the same book resource from different heterogeneous data sources into the same book resource identifier. The data preprocessing module reads the book title, author, publication date, category, and collection status, and compares these fields with existing book resource identifiers. If the comparison is successful, the standardized multi-source book data is merged into the corresponding book resource identifier; if the comparison fails, the data preprocessing module generates a new book resource identifier and writes the book title, author, subject, keywords, category, abstract attribute, publication date, collection status, and digital resource document content into the record corresponding to that book resource identifier. For the user behavior data, the data preprocessing module maps query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time to corresponding book resource identifiers, obtaining mapped multi-source book data. Understandably, in the mapped multi-source book data, each book record retains its corresponding data source and collection time for subsequent anomaly tracking and version management.

[0030] Furthermore, the data preprocessing module performs data cleaning on the mapped multi-source book data. The data cleaning process includes data deduplication, missing value imputation, formatting, and outlier handling. The data deduplication process reads the book resource identifier, book title, author, and publication date. When multiple records correspond to the same book resource identifier, the record with higher field completeness is retained, and the remaining records' themes, keywords, categories, abstract attributes, collection status, and digital resource document content are merged into the retained record. The missing value imputation process, when there are missing values ​​in the book title, author, theme, keywords, category, abstract attributes, publication date, collection status, or digital resource document content, first retrieves values ​​from other heterogeneous data source records under the same book resource identifier; if no value is found, the missing field is written to the missing record table, and the book resource identifier is retained. The formatting process writes the publication date into a unified date format, the category into a unified category field, the rating into a unified numerical field, and the user dwell time into a unified time field. The outlier processing reads the collection status, ratings, reviews, and user dwell time. When a field value does not conform to the field value rules, the record is written to the outlier record table, and the outlier field is set to a pending verification status, thus obtaining the cleaned multi-source book data.

[0031] Furthermore, the data preprocessing module performs identifier allocation and fuzzy matching on the cleaned multi-source book data. The identifier allocation process configures identifiers for book titles, authors, themes, keywords, categories, collection status, and behavioral records. Book titles and authors correspond to book resource identifiers, themes and keywords to theme identifiers, categories to category identifiers, collection status to collection identifiers, and historical search records, search session flows, query terms, click results, subsequent operations, borrowing records, collection records, ratings, reviews, and user dwell time to behavioral identifiers. The fuzzy matching process reads the book titles, authors, themes, keywords, and categories, and matches fields with synonyms, similar expressions, or format differences. Fields that match successfully are assigned to the same identifier, while fields that fail to match retain their original values ​​and are written to the unannotated record. The unannotated record, after manual annotation or verification according to cleaning rules, is written back to the cleaned multi-source book data, forming a new version record.

[0032] Furthermore, the data preprocessing module performs entity relation structuring on the data that has undergone identifier allocation and fuzzy matching. This entity relation structuring involves converting book titles, authors, themes, keywords, categories, collection status, and behavior records into a data structure that allows S200 to perform entity extraction, relation extraction, and attribute association processing. Specifically, the data preprocessing module writes book titles, authors, themes, keywords, categories, collection status, and digital resource document content into book resource records, and writes historical search records, search session flows, query terms, click results, subsequent operations, borrowing records, collection records, ratings, reviews, and user dwell time into behavior records, establishing a mapping relationship between the book resource records and the behavior records. This mapping relationship records the click results corresponding to query terms, the subsequent operations corresponding to click results, the book resource identifiers corresponding to borrowing records, the book resource identifiers corresponding to collection records, and the book resource identifiers corresponding to ratings and reviews. The data preprocessing module encapsulates the book resource records, the behavior records, the mapping relationship, the original access records, the missing record table, the abnormal record table, and the version records into target multi-source book data.

[0033] In one implementation, when the university library runs the data preprocessing module, the source interface unit triggers data access when new book resources are added to the database, the collection status changes, new search logs are added, new borrowing records are added, or new ratings or comments are added. The source interface unit receives book titles, authors, categories, and collection status from the library's business database; abstract attributes and digital resource document content from the e-book platform; historical search records, search session streams, query terms, click results, and user dwell time from the search logs; borrowing records from the borrowing system; and collection records, ratings, and comments from the user-generated content platform. The data preprocessing module completes field standardization, data mapping, data cleaning, identifier allocation, fuzzy matching, and entity relationship structuring according to the same book resource identifier. When encountering missing data collection time, conflicting collection status values, or inconsistent author fields, the data preprocessing module writes to an exception record table and retains the original access record; after the cleaning rules or manual annotation are completed, the version record saves the field values ​​before and after the update. The book resource records, behavior records, and mapping relationships in the target multi-source book data serve as inputs for constructing a book resource knowledge graph in S200. Book titles, authors, themes, keywords, categories, collection status, and digital resource document content are used by S200 for entity extraction, relation extraction, and attribute association processing. Historical search records, search session flows, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time are used by S200 to generate relation labels and relation weights.

[0034] In summary, this step's technical effects are as follows: This step integrates book titles, authors, topics, keywords, categories, abstract attributes, publication dates, library status, digital resource document content, and user behavior data scattered across heterogeneous data sources into a single target multi-source book dataset. This target multi-source book dataset retains original access records, missing record tables, abnormal record tables, and version records, ensuring clear data sources and field correspondences for subsequent book resource knowledge graph construction. The book resource records, behavior records, and mapping relationships within this target multi-source book dataset provide input for S200 to generate relation labels and relation weights.

[0035] S200. Based on the target multi-source book data, after entity extraction, relation extraction and attribute association processing, a book resource knowledge graph is constructed, and relation labels and relation weights are generated; Specifically, the graph construction module receives the target multi-source book data output by S100 and reads the book resource records, behavior records, mapping relationships, original access records, missing record table, abnormal record table, and version records from the target multi-source book data. The book resource records include book title, author, subject, keywords, classification, abstract attribute, publication time, collection status, and digital resource document content. The behavior records include historical search records, search session flow, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time. The mapping relationship records the click results corresponding to the query terms, the subsequent operations corresponding to the click results, the book resource identifiers corresponding to the borrowing records, the book resource identifiers corresponding to the collection records, and the book resource identifiers corresponding to the ratings and comments. The graph construction module starts processing when it receives new target multi-source book data, version record changes, new search session flow entries, or new borrowing records, and writes the processing batch, processing time, and data source into the graph construction record.

[0036] Specifically, the graph construction module calls a preset ontology model and a preset relation extraction model to process the target multi-source book data. The preset ontology model is a configuration set of entity types, attribute fields, and relation tags. The entity types include book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes. The attribute fields include book title, abstract attribute, publication time, collection status, digital resource document content, rating, reviews, and user dwell time. The preset relation extraction model is a processing model that extracts relationships between entities from the book resource records and the behavior records. Its inputs are book title, author, topic, keywords, category, abstract attribute, collection status, query terms, click results, subsequent operations, and borrowing records. Its outputs are entity pairs, relation tags, and relation sources. The relation sources are the source records of corresponding fields in the original access records, mapping relationships, or behavior records.

[0037] Specifically, entity extraction involves extracting node data that can be written into a graph database from the target multi-source book data. The graph construction module extracts book nodes from book titles, author nodes from authors, theme nodes from topics, category nodes from classifications, collection nodes from collection status, keyword nodes from keywords, and behavior nodes from historical search records, search session flows, query terms, click results, subsequent operations, borrowing records, collection records, ratings, reviews, and user dwell time. The graph construction module reads the book resource identifier, theme identifier, category identifier, collection identifier, and behavior identifier generated in S100 and writes fields corresponding to the same identifier into the same node. When a node field is missing, the graph construction module reads the missing record table and writes the node status to be supplemented; when a node field has abnormal records, the graph construction module reads the abnormal record table and writes the abnormal field into the verification field in the node attributes.

[0038] Specifically, the relationship extraction involves extracting relationship labels between nodes from the target multi-source book data. The graph construction module generates relationship labels between books and authors based on book titles and authors in the book resource records; between books and topics based on book titles and themes; between books and categories based on book titles and classifications; and between books and collection status based on book titles and collection status. The graph construction module also generates hypernym relationship labels, hyponym relationship labels, synonym relationship labels, and related word relationship labels between topics based on the mapping relationship between topics, keywords, and categories. Furthermore, the graph construction module generates relationship labels between books and borrowing records based on book resource identifiers and borrowing records; between books and click results based on book resource identifiers and click results; between query terms and click results; between click results and subsequent operations based on click results; and between user dwell time and book resource identifiers.

[0039] Specifically, the attribute association processing involves binding the attribute fields of book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes to their corresponding nodes and relationship tags. The graph construction module writes abstract attributes, publication time, digital resource document content, and collection status into the attribute fields of book nodes; authors into the attribute fields of author nodes; topics and keywords into the attribute fields of topic and keyword nodes; categories into the attribute fields of category nodes; and historical search records, search session flow, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time into the attribute fields of behavior nodes. Understandably, the graph construction module writes the relationship source, generation time, book resource identifier, and version record into each relationship tag, ensuring that the relationship tag corresponds to the original access record and version record in S100.

[0040] Furthermore, the graph construction module writes the extracted nodes, attribute fields, and relationship tags into the graph database to form the book resource knowledge graph. The book resource knowledge graph consists of book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, behavior nodes, attribute fields, and relationship tags. The graph database is a data storage structure for storing nodes, attribute fields, and relationship tags. Each book node in the book resource knowledge graph is bound to a corresponding book resource identifier. Before writing a node, the graph construction module reads existing nodes. If an existing node and the new node identifier are the same, attribute fields are merged; if they are different, a new node is created. Before writing a relationship tag, the graph construction module reads existing relationship tags. If the start node, end node, and relationship type of an existing relationship tag are the same as those of a new relationship tag, the relationship source is merged; otherwise, a new relationship tag is created and an initial relationship weight value is written.

[0041] Further, the relation weight is the weight field of the relation label in the book resource knowledge graph. The relation weight records the association strength between book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes. The graph construction module extracts retrieval session streams, query terms, click results, subsequent operations, and borrowing records from the user behavior data. The retrieval session stream is a behavioral time series formed by query terms, click results, user dwell time, and subsequent operations in the same retrieval process. The graph construction module generates sequence co-occurrence information based on the query terms and click results. The sequence co-occurrence information records the order of occurrence, frequency of occurrence, and corresponding book resource identifiers of query terms and click results in the same retrieval session stream. The graph construction module generates behavioral feedback information based on the click results, subsequent operations, borrowing records, and user dwell time. The behavioral feedback information records the book resource identifiers corresponding to the clicked collection records, ratings, comments, borrowing records, and user dwell time.

[0042] Furthermore, the graph construction module writes the sequence co-occurrence information and the behavioral feedback information into the book resource knowledge graph, and performs real-time weighting and updating operations on the relationship weights. Specifically, when the same query term points to the same click result in multiple search session streams, the graph construction module increases the relationship weight between the query term and the click result; when the click result corresponds to subsequent operations, collection records, ratings, comments, or borrowing records, the graph construction module increases the relationship weight between the book and the click result and the relationship weight between the book and the borrowing record; when the user's dwell time is lower than a preset time threshold and no subsequent operations are generated, the graph construction module decreases the relationship weight between the corresponding book and the click result. The preset time threshold can be set to 30 seconds, or it can be generated by the average user dwell time in the search session stream. The relationship weights are written into the relationship tags using normalized values ​​from 0 to 1, and the relationship weights before and after the update are written into the version record.

[0043] In one implementation, the university library uses the graph construction module to process the target multi-source book data output by S100. The graph construction module first extracts book nodes, author nodes, category nodes, and collection nodes from the library's business database (book titles, authors, categories, and collection status). Then, it extracts topic nodes and keyword nodes from the abstract attributes of the e-book platform and the content of digital resource documents. Finally, it extracts behavior nodes from the search log and the borrowing system. After a user enters a query term in the search interface, the search log generates a search session flow. After a user clicks on a book resource, the search log generates the click result and the user's dwell time. After a user favorites a book resource or creates a borrowing record, the borrowing system and the user-generated content platform generate subsequent operations, favorite records, ratings, and comments. The graph construction module writes the above behavior records into the book resource knowledge graph and generates relationship labels between query terms and click results, relationship labels between click results and subsequent operations, and relationship labels between books and borrowing records. The graph construction module updates the relationship weights based on sequence co-occurrence information and behavior feedback information, and uses the book resource knowledge graph, relationship labels, and relationship weights as input for S300 to process the search request. S300 invokes the book resource knowledge graph when parsing the search request; S400 invokes the relationship label when matching nodes and relationship labels; and S500 invokes the relationship weight during association reasoning.

[0044] In summary, this step transforms the target multi-source book data output by S100 into a book resource knowledge graph containing nodes, attribute fields, and relationship labels. The relationship weights are written into the retrieval session flow, query terms, click results, subsequent operations, and borrowing records, and are updated with each version record. The book resource knowledge graph, the relationship labels, and the relationship weights form the continuous inputs to S300, S400, and S500, providing a common data foundation for search request processing, node matching, and association reasoning.

[0045] S300. Based on the book resource knowledge graph, process the search request, extract entity information and faceted information, parse semantics, and obtain a structured search request; Specifically, the search request parsing module receives the book resource knowledge graph, the relationship tags, and the relationship weights output by S200, and initiates processing when the user inputs a search request on the search interface. The search request can be a natural language query, keyword query, or resource retrieval condition input by the user. The search request parsing module reads the book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes in the book resource knowledge graph, and reads the relationship tags in the following order: relationship tags between books and authors, relationship tags between books and topics, relationship tags between books and categories, relationship tags between books and collection status, relationship tags between topics (hypernyms, hyponyms, related words, borrowing records, click results, query terms, click results and subsequent operations, and relationship tags between user dwell time and books). The search request parsing module uses the book resource knowledge graph, the relationship tags, and the relationship weights as the processing basis to generate a search request processing record, which includes the input time, query terms, search function options, and processing batch.

[0046] Specifically, the search request parsing module performs normalization preprocessing on the search request. This normalization preprocessing includes removing blank characters without searchable meaning, standardizing the field formats for book title, author, subject, keywords, category, publication time, and collection status, and segmenting continuous input content into matchable query terms. For search requests containing multiple query terms, the search request parsing module performs initial classification according to the field order of book title, author, subject, keywords, category, and resource type. For search requests containing publication time, collection status, search function options, and resource search scope, the search request parsing module categorizes the corresponding content into faceted information. For query terms that cannot be classified, the search request parsing module writes them into a record to be matched, and uses hypernym relationship tags, hyponym relationship tags, synonym relationship tags, and related word relationship tags in the book resource knowledge graph for auxiliary parsing during subsequent processing.

[0047] Further, the entity information refers to the content in the search request that corresponds to a node in the book resource knowledge graph. The entity information includes book title, author, theme, keywords, category, and resource type. The search request parsing module first compares the query terms in the search request with the book node, author node, theme node, category node, and keyword node. If the comparison is successful, the search request parsing module writes the corresponding query term into the entity information field and records the matched node type and node identifier. If the comparison fails, the search request parsing module reads the hypernym relationship label, hyponym relationship label, synonym relationship label, and related word relationship label from the relationship labels to perform auxiliary matching on the query term. If the auxiliary matching is successful, the search request parsing module writes the node obtained from the auxiliary matching into the entity information field and writes the auxiliary matching source into the search request processing record. If the auxiliary matching fails, the search request parsing module retains the query term in the pending matching record and marks it as a mismatched query term.

[0048] Furthermore, the faceted information is data that limits the search scope in the search request. The faceted information includes publication time, collection status, search function options, and resource search scope. The search request parsing module reads the time field, collection field, resource field, and function field from the search request and writes them into the publication time, collection status, search function options, and resource search scope, respectively. The search function options include comprehensive search, content search, rich abstract search, and resource search. The resource search scope includes book resources, digital resource document content, and collection resources. The search request parsing module pre-matches the faceted information with the collection node, publication time attribute, resource type attribute, and search function options. When the pre-match is successful, the faceted information is written into a structured field. When the pre-match is unsuccessful, the search request parsing module writes the corresponding field into an error record and retains the field content from the original search request for use by S400 during faceted constraint matching.

[0049] Specifically, the semantics refers to the textual meaning in the search request that reflects the user's search intent. The search request parsing module reads entity information, faceted information, and query term combinations from the search request through semantic parsing to generate search intent. The search intent includes comprehensive search, content search, rich abstract search, and resource search. When processing comprehensive search, the search request parsing module writes book titles, authors, topics, keywords, and categories into semantic fields; when processing content search, it uses abstract attributes, keywords, topics, and digital resource document content as semantic parsing objects; when processing rich abstract search, it uses abstract attributes and keywords as semantic parsing objects; and when processing resource search, it uses collection status, resource type, and publication time as semantic parsing objects. Understandably, the semantic parsing process does not directly generate a list of book resources, but rather generates structured fields for S400 matching nodes and relationship labels.

[0050] Furthermore, the search request parsing module encapsulates the entity information, faceted information, semantics, and search intent into a structured search request. The structured search request includes entity information fields, faceted information fields, semantic fields, search intent fields, hit node fields, auxiliary matching source fields, records to be matched, and search request processing records. The entity information fields record the book title, author, subject, keywords, category, and resource type. The faceted information fields record the publication time, library status, search function options, and resource search scope. The semantic fields record the semantic content generated by the search request parsing module. The search intent fields record comprehensive search, content search, rich summary search, or resource search. The hit node fields record the node identifiers in the book resource knowledge graph that match the search request. The auxiliary matching source fields record hypernym relationship tags, hyponym relationship tags, synonym relationship tags, or related word relationship tags. The records to be matched include fields for missing query terms and abnormal faceted fields.

[0051] In one implementation, a user of a university library enters "Knowledge Graph Introduction Available Books" as a search request on the search interface. Upon receiving this search request, the search request parsing module first compares "Knowledge Graph" with the topic nodes and keyword nodes, and writes the matching results into the entity information field; then, it writes "Introduction" into the semantic field and records the comprehensive search in the search intent field; subsequently, it pre-matches "Available" with the collection nodes and collection status, and writes it into the faceted information field; finally, it writes "Books" into the resource type field. If the book resource knowledge graph contains synonym tags or related term tags for "Knowledge Graph," the search request parsing module writes the corresponding auxiliary matching source into the structured search request. After the structured search request is generated, it serves as the input for S400 to match nodes and relationship tags in the book resource knowledge graph. The entity information field is used by S400 to match book nodes, author nodes, topic nodes, category nodes, and keyword nodes. The faceted information field is used by S400 to match library collection nodes, publication time attributes, resource type attributes, and search function options. The semantic field and search intent field are used by S400 to generate target nodes and faceted constraints.

[0052] In summary, this step transforms the user's search request into a structured retrieval request containing entity information, faceted information, semantics, and retrieval intent. The structured retrieval request retains the fields of the hit nodes, auxiliary matching source fields, and records to be matched, providing clear input for subsequent node matching and faceted constraint generation. This structured retrieval request inherits the book resource knowledge graph, relation tags, and relation weights generated in S200 and serves as input for S400 to obtain the target nodes and faceted constraints.

[0053] S400. Based on the structured retrieval request, match the nodes and relationship tags in the book resource knowledge graph to obtain the target node and facet constraints; Specifically, the node matching module receives the structured retrieval request output by S300 and reads the book resource knowledge graph, the relation tags, and the relation weights output by S200. The structured retrieval request includes entity information fields, faceted information fields, semantic fields, retrieval intent fields, hit node fields, auxiliary matching source fields, records to be matched, and search request processing records. Upon receiving a new structured retrieval request, the node matching module initiates matching processing and writes the search request processing record, processing batch, matching time, and the version of the book resource knowledge graph invoked into the node matching record. The node matching record is used to record the matching source, matching status, and node identifier for each field in this step and is transmitted to S500 along with the output of this step.

[0054] Specifically, the target node is a book node, author node, subject node, category node, or keyword node that matches the entity information field. The faceted constraints are the collection node, publication time attribute, resource type attribute, and search function options that match the faceted information field. The node matching module reads the book title, author, subject, keyword, category, and resource type from the entity information field and performs field comparisons on the book nodes, author nodes, subject nodes, category nodes, and keyword nodes in the book resource knowledge graph, respectively. The field comparison uses book resource identifiers, subject identifiers, category identifiers, collection identifiers, and behavior identifiers. When a field comparison is successful, the node matching module writes the successfully matched node into the target node field and records the node type, node identifier, field source, and relationship tag source in the hit node field.

[0055] Furthermore, when the query term in the entity information field does not directly match a node, the node matching module reads the superordinate relation tags, hyponym relation tags, synonym relation tags, and related word relation tags between topics from the relation tags, and performs auxiliary matching on the unmatched query term. During auxiliary matching, the node matching module first searches for nodes in the keyword nodes that are identical to the query term field, and then reads adjacent topic nodes along the synonym relation tags and related word relation tags; if adjacent topic nodes have superordinate relation tags or hyponym relation tags, the node matching module continues to read the corresponding superordinate or hyponym nodes. When the auxiliary matching is successful, the successfully matched node is merged into the target node, and the corresponding synonym relation tags, related word relation tags, superordinate relation tags, or hyponym relation tags are written into the auxiliary matching source field. When the auxiliary matching is unsuccessful, the node matching module writes the query term into the record to be matched and retains the original field content of the query term in the structured retrieval request.

[0056] Specifically, the node matching module reads the publication time, collection status, search function options, and resource search range from the faceted information field, and matches them with the collection nodes, publication time attributes, resource type attributes, and search function options in the book resource knowledge graph. When the collection status matches the collection node successfully, the node matching module writes the collection node into the faceted constraint field. When the publication time matches the publication time attribute successfully, the node matching module writes the corresponding time range into the faceted constraint field. When the resource search range matches the resource type attribute successfully, the node matching module writes the corresponding resource type into the faceted constraint field. When the search function option matches successfully, the node matching module writes comprehensive search, content search, rich summary search, or resource search into the faceted constraint field. When a faceted information field fails to match, the node matching module writes the field into an exception record and retains the original field content for use when filtering S500 path expansion results.

[0057] Furthermore, when the same entity information field matches multiple nodes, the node matching module reads the relation tags and relation weights connected to each matching node, and performs node filtering in conjunction with the semantic field and the search intent field. When the semantic field matches the attribute field of a topic node, keyword node, or book node, the corresponding node is written into the target node field. When the search intent field is a comprehensive search, the node matching module retains the matching results of book nodes, author nodes, topic nodes, category nodes, and keyword nodes; when the search intent field is a content search, the node matching module prioritizes retaining book nodes, topic nodes, and keyword nodes related to abstract attributes, keywords, and digital resource document content; when the search intent field is a resource search, the node matching module prioritizes retaining nodes related to collection nodes, resource type attributes, and publication time attributes. Matching nodes with relation weights lower than a preset threshold are written into the target node candidate record, but not directly into the target node field. The preset threshold uses a normalized value from 0 to 1, generated from the historical distribution of the relation weights, and is written into the version record.

[0058] Furthermore, the node matching module handles conflicting fields and duplicate nodes during the matching process. For cases where the same book title corresponds to multiple author nodes, the node matching module reads the publication time, category, and abstract attributes, and verifies the author nodes based on field consistency. For cases where the same topic corresponds to multiple category nodes, the node matching module reads the hypernym relationship tags, hyponym relationship tags, synonym relationship tags, and related word relationship tags between topics, and merges them according to the search intent field in the structured search request. For cases of conflicting collection status, the node matching module reads the version records and exception records in S100, writes the collection node corresponding to the latest version record into the faceted constraint field, and writes the conflicting fields into the node matching record. Understandably, the node matching module does not change the nodes and relationship tags in the book resource knowledge graph; it only records the matching results in the target node field, faceted constraint field, and node matching record generated in this step.

[0059] In one implementation, the structured search request input by a university library user includes entity information fields such as "knowledge graph" and "introduction," faceted information fields such as "available for borrowing" and "books," and a search intent field of "comprehensive search." The node matching module first compares the "knowledge graph" with topic nodes and keyword nodes. If a field match is successful, the corresponding topic node and keyword node are written into the target node field. The node matching module then reads "introduction" as a semantic field and searches for adjacent topic nodes along synonym and association tags. If an auxiliary match is successful, the adjacent topic nodes are merged into the target node field. Subsequently, the node matching module matches "available for borrowing" with collection nodes and "books" with resource type attributes, writing the successfully matched collection nodes and resource type attributes into the faceted constraint field. If "introduction" does not form a node in the book resource knowledge graph, the node matching module writes it into the record to be matched, while retaining the matched topic nodes, keyword nodes, collection nodes, and resource type attributes.

[0060] Furthermore, the target node fields output in this step include node type, node identifier, matching field, auxiliary matching source field, and relation tag source field. The faceted constraint fields include library collection nodes, publication time attribute, resource type attribute, search function options, and abnormal records. The node matching records include search request processing records, book resource knowledge graph version, relation weight version, target node candidate records, records to be matched, and conflict fields. The target node and the faceted constraints serve as input to S500. Starting from the target node, S500 filters relation tags participating in association reasoning according to the faceted constraints, and calls the relation weights to perform path expansion and path expansion result filtering.

[0061] In summary, this step maps the structured search request generated by S300 to target nodes and faceted constraints in a book resource knowledge graph. The target nodes retain entity information fields, auxiliary matching source fields, and relation tag source fields. The faceted constraints retain library collection nodes, publication time attributes, resource type attributes, and search function options. The target nodes and faceted constraints inherit the book resource knowledge graph, relation tags, and relation weights from S200, and provide starting nodes and filtering conditions for S500 associative reasoning.

[0062] S500: Based on the target node, the facet constraint and the relation weight, perform association reasoning to obtain the set of nodes traversed by the path and the set of candidate book nodes; Specifically, the association reasoning module receives the target node, facet constraints, and node matching records output by S400, and reads the book resource knowledge graph, relation tags, and relation weights generated by S200. The target node is a book node, author node, topic node, category node, or keyword node successfully matched by S400. The facet constraints are the library collection nodes, publication time attributes, resource type attributes, and search function options successfully matched by S400. The relation weights are the normalized values ​​written into the relation tags by S200. The association reasoning module initiates processing when a node identifier exists in the target node field, and writes the structured search request, node matching records, book resource knowledge graph version, relation weight version, and processing time into the association reasoning record.

[0063] Specifically, the association reasoning involves extending a path from the target node to adjacent nodes in the book resource knowledge graph according to the relationship tags. The path extension consists of a starting node, relationship tags, nodes traversed by the path, and an ending node. The starting node is the target node. The nodes traversed by the path are author nodes, topic nodes, category nodes, collection nodes, keyword nodes, behavior nodes, or book nodes encountered during the path extension process. The ending node is the book node pointed to by the filtered path. The candidate book node set is a dataset formed by multiple book nodes pointed to by the filtered paths. The association reasoning module first reads the node type of the target node, then filters the relationship tags participating in the association reasoning according to the faceted constraints, and writes the filtered relationship tags into the path configuration record.

[0064] Furthermore, when the target node is a topic node or keyword node, the association reasoning module prioritizes reading the hypernym relationship tags between topics, the hyponym relationship tags between topics, the synonym relationship tags between topics, the related word relationship tags between topics, the relationship tags between books and topics, and the relationship tags between books and click results, and expands the path along these relationship tags. When the target node is an author node, the association reasoning module reads the relationship tags between books and authors, the relationship tags between books and topics, and the relationship tags between books and categories, and connects author nodes, book nodes, topic nodes, and category nodes along these relationship tags. When the target node is a book node, the association reasoning module reads the relationship tags between books and topics, the relationship tags between books and categories, the relationship tags between books and library status, the relationship tags between books and borrowing records, the relationship tags between books and click results, and the relationship tags between user dwell time and books, and connects adjacent nodes along these relationship tags. When the target node is a category node, the association reasoning module reads the relationship tags between books and categories and the relationship tags between books and topics, and connects category nodes, book nodes, and topic nodes along these relationship tags.

[0065] Specifically, the association reasoning module performs boundary control on path expansion according to the facet constraints. When a collection node is written into the facet constraints, the association reasoning module retains the path that passes through the corresponding collection node or the corresponding collection status attribute. When the publication time attribute is written into the facet constraints, the association reasoning module retains the path where the publication time of the book node is within the corresponding time range. When the resource type attribute is written into the facet constraints, the association reasoning module retains the path where the resource type corresponding to the book node is consistent. When the search function option is content search, the association reasoning module retains the path that passes through the topic node, keyword node, and book node. When the search function option is resource search, the association reasoning module retains the path that passes through the collection node, resource type attribute, and book node. When the search function option is rich summary search, the association reasoning module retains the path that passes through the keyword node, topic node, and book node, and writes a summary attribute call flag in the association reasoning record for S600 to read when calling source data.

[0066] Further, the association reasoning module filters the path expansion results based on the relationship weights. The relationship weights are read from the relationship tags and written into the path weight record along with the nodes traversed by the path. When a path contains multiple relationship tags, the association reasoning module reads the relationship weights in the order of the relationship tags and generates path relationship weights. The path relationship weight is formed by the average of the relationship weights contained in the path. When the path relationship weight is lower than a preset weight threshold, the path is written into the low-weight path record and not into the candidate path record. The preset weight threshold uses a normalized value from 0 to 1 and is written into the version record along with the relationship weight version. In one embodiment, the preset weight threshold is 0.3. For multiple candidate paths passing through the same book node, the association reasoning module retains the path with the higher path relationship weight and writes the remaining paths into the candidate path record.

[0067] Specifically, the association reasoning module filters the nodes traversed by the path by node type. If the path traversed only contains behavior nodes and does not connect to book nodes, the path is written as a record of no book nodes. If the path traversed passes through a library node but the library node is inconsistent with the facet constraint, the path is written as a facet inconsistency record. If the path traversed passes through a topic node but the corresponding topic node is in a state of pending supplementation or pending verification, the path is written as a path to be verified record. If there are abnormal records among the book nodes traversed by the path, the association reasoning module reads the abnormal record table in S100 and writes the book node into the abnormal candidate record. Understandably, the low-weight path records, records of no book nodes, facet inconsistency records, path records to be verified, and abnormal candidate records are all written into the association reasoning record and do not enter the candidate book node set.

[0068] Furthermore, the association reasoning module handles duplicate book nodes during path expansion. When different target nodes expand to the same book node, the association reasoning module retains the book node once and merges multiple target nodes, multiple relation labels, and multiple nodes traversed by the path into the same candidate book node record. When the same target node expands to the same book node via hypernym relation labels, synonym relation labels, and related word relation labels, the association reasoning module writes the corresponding path source into the auxiliary path source field. When a candidate book node corresponds to multiple collection nodes, the association reasoning module retains matching collection nodes according to the faceted constraints and writes unmatched collection nodes into the candidate collection record. The candidate book node record includes a book node identifier, a target node identifier, nodes traversed by the path, relation labels, relation weights, path relation weights, faceted constraints, and a book resource knowledge graph version.

[0069] In one implementation, a university library user's structured search request is processed by S400 to obtain a topic node "knowledge graph," a keyword node "knowledge graph," a collection node "available for borrowing," and a resource type attribute "book." The association reasoning module starts from the topic node "knowledge graph," reads adjacent topic nodes along synonym relationship tags, reads related topic nodes along association word relationship tags, and then reads book nodes along the relationship tags between books and topics. The association reasoning module also reads the collection node "available for borrowing" and retains the paths passing through the corresponding collection nodes. If a book node is connected to the target node simultaneously through topic nodes, keyword nodes, and click results, the association reasoning module merges the nodes traversed by these paths and writes them into the candidate book node record. If another book node is connected to the target node only through a low-relationship-weight behavior node, and the path relationship weight is less than 0.3, then this path is written into the low-weight path record and does not enter the candidate book node set.

[0070] Furthermore, the nodes traversed by the path output in this step include author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes traversed between the target node and the candidate book node. The candidate book node set includes candidate book node records, candidate path records, path relationship weights, facet constraints, auxiliary path source fields, and association reasoning records. The candidate book node set is not directly output as a book resource list, but rather as input for S600 to call source data. S600 calls the source data corresponding to each candidate book node based on the candidate book node set, and performs normalization preprocessing and semantic matching on the source data.

[0071] In summary, this step transforms the target node and faceted constraints obtained in S400 into a set of nodes traversed by the path and a set of candidate book nodes in the book resource knowledge graph. The candidate book node set retains relation labels, relation weights, path relation weights, and faceted constraints, ensuring that the candidate book nodes have a traceable path origin. This candidate book node set inherits from the book resource knowledge graph of S200 and the target node of S400, and serves as input for source data retrieval and semantic matching in S600.

[0072] S600. Based on the candidate book node set, the source data is called, and after normalization preprocessing and semantic matching, the final candidate book node set is obtained; Specifically, the source data matching module receives the candidate book node set output by S500 and reads the candidate book node records, candidate path records, path relationship weights, facet constraints, auxiliary path source fields, and association reasoning records from the candidate book node set. The candidate book node records include book node identifiers, target node identifiers, nodes traversed by the path, relationship tags, relationship weights, path relationship weights, facet constraints, and the version of the book resource knowledge graph. When a book node identifier exists in the candidate book node set, the source data matching module initiates a source data call and writes the call time, the version of the book resource knowledge graph, the relationship weight version, and the candidate book node record into the source data matching record. The source data consists of the book resource records and original access records corresponding to the candidate book nodes in S100, including book title, author, subject, keywords, classification, abstract attributes, publication time, library status, and digital resource document content.

[0073] Specifically, the source data matching module calls the source data according to the book node identifier. During the calling process, the book node identifier is first read from the book resource knowledge graph, and then the book resource record, original access record, missing record table, abnormal record table, and version record in S100 are retrieved according to the book node identifier. If the book resource record contains complete book title, author, subject, keywords, category, abstract attribute, publication time, collection status, and digital resource document content, the source data matching module writes the above fields into the source data field. If the abstract attribute or digital resource document content is missing, the missing record table is read, and the missing fields are written into the source data matching record. If there are abnormal fields in the collection status, author, or category, the abnormal record table and version record are read, and the fields in the latest version record are written into the source data field. The source data field is bound to the candidate book node record to form the source data to be preprocessed.

[0074] Furthermore, the standardized preprocessing involves uniformly processing the text fields and attribute fields in the source data to be preprocessed. The source data matching module extracts book titles, themes, keywords, categories, and abstract attributes from the source data to be preprocessed, and performs field format standardization, punctuation cleanup, redundant word string cleanup, and word segmentation on these fields. The book titles are segmented to form title terms. Themes are standardized to form theme terms. Keywords are retained as keywords after redundant word string cleanup. Categories are standardized to form category fields. Abstract attributes are cleaned up with punctuation and segmented to form abstract terms. Publication time and collection status are not processed by word segmentation; their field values ​​are directly written into the source data attribute fields. The digital resource document content is read from text paragraphs and written into the source data content field for semantic matching.

[0075] Specifically, the resource entity feature vector is a representation of the resource entity formed based on title terms, subject terms, keywords, and abstract terms. The source data matching module reads the title terms, subject terms, keywords, and abstract terms, and records the field source, term position, and the corresponding book node identifier, respectively. The title terms correspond to the book title field, the subject terms correspond to the subject field, the keywords correspond to the keyword field, and the abstract terms correspond to the abstract attribute field. The source data matching module merges the title terms, subject terms, keywords, and abstract terms according to the book node identifier to generate the resource entity feature vector, and writes the publication time, collection status, classification, and digital resource document content into the attribute fields of the resource entity feature vector. Understandably, the field sources in the resource entity feature vector correspond to the original access records in S100 and are bound to the candidate path records in S500.

[0076] Further, the knowledge retrieval feature vector is a retrieval request representation formed based on the structured retrieval request output by S300. The source data matching module reads entity information, faceted information, and semantics from the structured retrieval request. The entity information includes book title, author, subject, keywords, category, and resource type. The faceted information includes publication time, collection status, retrieval function options, and resource retrieval scope. The semantics is the textual meaning in the search request that reflects the retrieval intent. The source data matching module writes entity information into the entity field, faceted information into the faceted field, semantics into the semantic field, and retrieval intent into the retrieval intent field, generating a knowledge retrieval feature vector. The knowledge retrieval feature vector maintains a correspondence with the target node and faceted constraints output by S400.

[0077] Specifically, the semantic matching involves performing field matching and semantic field comparison between the resource entity feature vector and the knowledge retrieval feature vector. The source data matching module first compares the entity fields with title words, subject words, keywords, and abstract words. When an entity field matches a title word, subject word, or keyword, a direct matching record is generated. When an entity field does not match any of these fields, the source data matching module reads the relationship tags in the corresponding path of the candidate book node and performs auxiliary matching by combining hypernym relationship tags, hyponym relationship tags, synonym relationship tags, and related word relationship tags. When the auxiliary matching is successful, an auxiliary matching record is generated. The source data matching module then compares the faceted fields with publication time, collection status, resource type attributes, and search function options. When the faceted fields match the source data attribute fields, a faceted matching record is generated. When the faceted fields do not match the source data attribute fields, the candidate book node is written into a faceted inconsistency record.

[0078] Furthermore, the source data matching module performs semantic matching verification by combining the relationship tags in the corresponding paths of the candidate book nodes. If a candidate book node is connected to the target node through a topic node, keyword node, book node, or author node, and the resource entity feature vector and the knowledge retrieval feature vector have a direct matching record or an auxiliary matching record, then the candidate book node is written into the final candidate book node set. If a candidate book node is only connected to the target node through a behavior node, the source data matching module reads the relationship tags between the query terms and the click results, the relationship tags between the click results and subsequent operations, the relationship tags between books and borrowing records, and the relationship tags between user dwell time and books in the path, and checks whether there are title words, topic words, keywords, or abstract words in the source data fields that correspond to the semantic fields. When the corresponding fields exist, the candidate book node is written into the final candidate book node set; when the corresponding fields do not exist, the candidate book node is written into a semantic mismatch record.

[0079] Specifically, the source data matching module merges duplicate candidate book nodes. If multiple candidate paths point to the same book node and the source data fields are identical, one final candidate book node record is retained, and the nodes traversed by the multiple paths, multiple relationship labels, multiple path relationship weights, and multiple matching records are merged and written into this final candidate book node record. If the same book node corresponds to multiple source data versions, the source data matching module reads the version records in S100, selects the latest version record, and writes the corresponding book title, author, subject, keywords, classification, abstract attributes, publication time, collection status, and digital resource document content into the candidate source data record. If the source data corresponding to a book node is missing, the source data matching module writes the book node into the missing source data record and does not write it into the final candidate book node set.

[0080] In one implementation, after the university library's S500 outputs a set of candidate book nodes, the source data matching module retrieves book resource records based on the book node identifier. For candidate book node A, the source data matching module extracts the book title, author, subject, keywords, category, abstract attribute, publication time, collection status, and digital resource document content from the book resource records, forming title terms, subject terms, keywords, and abstract terms. The source data matching module constructs the title terms, subject terms, keywords, and abstract terms into a resource entity feature vector, and constructs the entity information, faceted information, and semantics from S300 into a knowledge retrieval feature vector. If the subject in the entity information matches the subject term, the collection status in the faceted information matches the source data attribute field, and there is a relationship label between the book and the subject in the candidate path, then candidate book node A is written into the final candidate book node set. For candidate book node B, if it is only connected to the target node through the result of a click, and the title terms, subject terms, keywords, and abstract terms do not match the semantic fields, then candidate book node B is written into a semantic mismatch record.

[0081] Further, the final candidate book node set output in this step includes final candidate book node records, resource entity feature vectors, knowledge retrieval feature vectors, direct matching records, auxiliary matching records, faceted matching records, semantic mismatch records, missing source data records, alternative source data records, and source data matching records. The final candidate book node record includes a book node identifier, a target node identifier, nodes traversed by the path, a relation label, a relation weight, a path relation weight, source data fields, and a matching record. The final candidate book node set serves as input for S700 to generate a matching score. S700 reads the final candidate book node set, the nodes traversed by the path, the relation weights, and the source data, generates a matching score, and outputs a list of book resources and theoretical basis in descending order of the matching score.

[0082] In summary, the technical effects of this step are as follows: This step maps the candidate book node set obtained in S500 to the source data and forms resource entity feature vectors through normalization preprocessing. The resource entity feature vectors are then semantically matched with the knowledge retrieval feature vectors to form direct matching records, auxiliary matching records, and faceted matching records. The final candidate book node set retains candidate paths, source data fields, and matching records, providing input for S700 to generate matching scores and theoretical basis.

[0083] S700: Generate a matching score based on the final candidate book node set, the nodes traversed by the path, the relationship weight, and the source data; and output the book resource list and theoretical basis in descending order of the matching score. Specifically, the sorting output module receives the final candidate book node set output by S600 and reads the final candidate book node records, resource entity feature vectors, knowledge retrieval feature vectors, direct matching records, auxiliary matching records, faceted matching records, semantic mismatch records, missing source data records, alternative source data records, and source data matching records from the final candidate book node set. The sorting output module also reads the nodes traversed by the path, candidate path records, path relationship weights, and association reasoning records output by S500, and reads the relationship weights written in the relationship tags by S200. The source data consists of the book titles, authors, topics, keywords, categories, abstract attributes, publication dates, library status, and digital resource document content that S600 has called and completed normalized preprocessing of. The sorting output module initiates processing when a final candidate book node record exists in the final candidate book node set, and writes the book resource knowledge graph version, relationship weight version, source data matching record, and processing time into the sorting output record.

[0084] Specifically, the matching score is the ranking field of the final candidate book node under the current search request. The ranking output module generates a semantic matching score, a path relationship score, and a source data score for each final candidate book node, and combines the semantic matching score, the path relationship score, and the source data score into the matching score. The semantic matching score is generated from direct matching records, auxiliary matching records, and faceted matching records. Direct matching records are derived from field matches between entity information and title words, subject words, keywords, or abstract words. Auxiliary matching records are derived from auxiliary matches corresponding to hypernym relationship tags, hyponym relationship tags, synonym relationship tags, or related word relationship tags. Faceted matching records are derived from matches between publication time, collection status, search function options, and resource search scope and source data attribute fields.

[0085] Further, the path relationship score is generated by the nodes traversed by the path and the relationship weights. The sorting output module reads the target node, the nodes traversed by the path, the relationship label, and the path relationship weights corresponding to each final candidate book node, and reads the relationship labels between the book and the author, the relationship labels between the book and the topic, the relationship labels between the book and the category, the relationship labels between the book and the collection status, the hypernym relationship labels between topics, the hyponym relationship labels between topics, the synonym relationship labels between topics, the related word relationship labels between topics, the relationship labels between the book and the borrowing record, the relationship labels between the book and the click result, the relationship labels between the query term and the click result, the relationship labels between the click result and subsequent operations, and the relationship labels between the user's dwell time and the book into the path relationship score generation process. When the same final candidate book node corresponds to multiple paths, the sorting output module retains the path with the higher path relationship weight as the primary path and writes the remaining paths into the candidate path field. When there are nodes to be verified or abnormal candidate records in the path, the sorting output module reads the verification status in the source data matching record and writes the path into the candidate path field.

[0086] Specifically, the source data score is generated from the field completeness and field matching status of the source data. The sorting output module reads the book title, author, subject, keywords, category, abstract attributes, publication time, collection status, and digital resource document content, and checks their field sources in the source data matching records. When the book title, author, subject, keywords, and category correspond to the entity information in the structured search request, the sorting output module writes an entity field hit record to the source data score. When the publication time and collection status correspond to faceted information, the sorting output module writes a faceted field hit record to the source data score. When the abstract attributes and digital resource document content correspond to semantic fields, the sorting output module writes a content field hit record to the source data score. If the final candidate book node corresponds to a missing source data record, the sorting output module writes the corresponding missing fields into the source data score record and retains the candidate status of the final candidate book node.

[0087] Furthermore, the sorting output module reads historical search records, search session streams, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time, and writes these fields as behavioral fields into the matching score generation process. When the query terms and click results form a sequence co-occurrence information, the sorting output module reads the corresponding relationship weight. When the click results have subsequent operations, borrowing records, collection records, ratings, or comments, the sorting output module reads the corresponding behavioral feedback information. When the user dwell time has a relationship tag with the book node, the sorting output module reads the relationship weight corresponding to the user dwell time. Understandably, the behavioral fields do not generate a separate book resource list, but are written into the matching score record, and together with the semantic matching score, path relationship score, and source data score, form the final candidate book node matching score.

[0088] Specifically, the sorting output module processes the semantic matching score, the path relationship score, the source data score, and the behavior field through a semantic sorting model. The semantic sorting model is a sorting processing model that receives multiple sorting fields and outputs a matching score. Its inputs include the final candidate book node record, resource entity feature vector, knowledge retrieval feature vector, direct matching record, auxiliary matching record, faceted matching record, nodes traversed by the path, relationship weights, and the source data field. Its output is the matching score. The matching score is written to the final candidate book node record using a normalized value from 0 to 1. The sorting output module sorts the matching scores of multiple final candidate book nodes in descending order. When matching scores are the same, the sorting output module sequentially reads the number of direct matching records, path relationship weights, faceted matching records, and the completeness of the source data field, and generates a tie-breaking sorting record.

[0089] Furthermore, the book resource list is output data of book resources sorted in descending order according to the matching score. The sorting output module reads the book node identifier, book title, author, subject, keywords, category, abstract attribute, publication time, collection status, and digital resource document content from the final candidate book node record, and writes the above fields into the book resource list. Each book resource in the book resource list is bound to a corresponding matching score, target node, nodes traversed by the path, relation label, relation weight, source data field, and source data matching record. If the final candidate book node corresponds to a semantically mismatched record or a missing source data record, the sorting output module does not write the final candidate book node into the book resource list. If the final candidate book node corresponds to a candidate source data record, the sorting output module reads the corresponding fields of the latest version record and writes them into the book resource list, and writes the candidate source data record into the sorting output record.

[0090] Specifically, the theoretical basis is the relevance data between each book resource in the book resource list and the search request. The sorting output module reads entity information, faceted information, semantics, and search intent from the structured search request; reads the target node and faceted constraints generated by S400; reads the nodes traversed by the path generated by S500 and relational tags; and reads the source data fields, direct matching records, auxiliary matching records, and faceted matching records generated by S600. The sorting output module merges the above content into the theoretical basis. The theoretical basis includes entity information and faceted information in the search request, the hit target node, the nodes traversed by the path from the target node to the book node, relational tags in the path, book titles, themes, keywords, abstract attributes, or digital resource document content matching content in the source data, and relational weights and behavioral fields involved in generating the matching score. The theoretical basis is bound to each book resource in the book resource list and written into the sorting output record.

[0091] In one implementation, the structured search request of a university library corresponds to the topic node "knowledge graph," the collection node "available for borrowing," and the resource type attribute "book." The final candidate book node set output by S600 includes book node A, book node B, and book node C. The sorting output module reads the direct matching record corresponding to book node A, finds that its topic words match entity information, its collection status matches faceted information, and there are relationship tags between books and topics and synonym relationship tags between topics in the path. The sorting output module reads the borrowing records, click results, and user dwell time corresponding to book node A, and writes the corresponding relationship weights into the matching score record. For book node B, if it enters the final candidate book node set through a behavior node, but the title words, topic words, keywords, and abstract words in the source data only form auxiliary matching records, the sorting output module generates a lower matching score based on the auxiliary matching records and path relationship weights. For book node C, if it has missing source data records, the sorting output module does not write it into the book resource list. Subsequently, the sorting output module arranges book node A and book node B in descending order of matching score, and writes the target node, the nodes traversed by the path, the relation label, the source data field, and the theoretical basis for each item.

[0092] Furthermore, the book resource list output in this step includes book node identifiers, book titles, authors, themes, keywords, categories, abstract attributes, publication dates, library status, digital resource document content, and matching scores. The theoretical basis includes entity information, faceted information, target nodes, nodes traversed by the path, relation labels, relation weights, source data fields, direct matching records, auxiliary matching records, and faceted matching records. The sorted output records include the book resource knowledge graph version, relation weight version, source data matching records, tie-breaking records, alternative path fields, and alternative source data records. The book resource list and the theoretical basis serve as the output of this method and are stored together with the version records, node matching records, association reasoning records, and source data matching records formed in steps S100 to S600.

[0093] In summary, this step transforms the final candidate book node set output by S600 into a book resource list with matching scores. The matching scores are generated from semantic matching scores, path relationship scores, source data scores, and behavioral fields, and correspond to the relationship weights and source data fields. The theoretical basis for this step is the recording of target nodes, nodes traversed by the path, relationship labels, and source data fields, ensuring that each book resource in the list has a corresponding retrieval process record.

[0094] Example 2: Figure 2 This diagram illustrates a structural block diagram of a knowledge graph-based intelligent book resource retrieval system according to an embodiment of the present invention. Figure 2 As shown, the structure may include: The data preprocessing module 01 is used to acquire multi-source book data and user behavior data. After data standardization, data cleaning, and entity relationship structuring, it obtains target multi-source book data and sends this data to the graph construction module. Specifically, the data preprocessing module receives multi-source book data and user behavior data from the library's business database, e-book platform, digitized scanned text library, user-generated content platform, search logs, and borrowing system. The multi-source book data includes book titles, authors, themes, keywords, classifications, abstract attributes, publication dates, collection status, and digital resource document content. The user behavior data includes historical search records, search session flows, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time. Specifically, the data preprocessing module performs field standardization according to a unified metadata standard, classifying book title fields and subject titles from different data sources into book titles, collection status and borrowable status into collection status, search terms into query terms, and clicked objects into click results. Subsequently, the data preprocessing module performs data cleaning, including deduplication, missing value imputation, formatting, and outlier handling, writing outlier fields into anomaly records. After cleaning, the preprocessing module performs identifier allocation, fuzzy matching, and entity relationship structuring, encapsulating book resource records, behavior records, and mapping relationships into target multi-source book data. This target multi-source book data includes book resource identifiers, book resource records, behavior records, mapping relationships, and version records, and is then sent to the graph construction module.

[0095] The graph construction module 02 receives the target multi-source book data, performs entity extraction, relation extraction, and attribute association processing to construct a book resource knowledge graph, and generates relation tags and relation weights. The book resource knowledge graph, relation tags, and relation weights are then sent to the search request parsing module. Specifically, the graph construction module receives the target multi-source book data from the data preprocessing module and reads the book resource records, behavior records, and mapping relationships within it. Through entity extraction processing, the graph construction module extracts book nodes from book titles, author nodes from authors, topic nodes from themes, category nodes from classifications, collection nodes from collection status, keyword nodes from keywords, and behavior nodes from historical search records, search session streams, query terms, clicked results, subsequent operations, borrowing records, collection records, ratings, reviews, and user dwell time. Specifically, the graph construction module generates relationship tags between books and authors, books and topics, books and categories, books and library status, hypernym relationships between topics, hyponym relationships between topics, synonym relationships between topics, related word relationships between topics, relationship tags between books and borrowing records, relationship tags between books and click results, relationship tags between query terms and click results, relationship tags between click results and subsequent operations, and relationship tags between user dwell time and books. Subsequently, the graph construction module binds abstract attributes, publication time, library status, and digital resource document content to corresponding nodes, completing attribute association processing. The graph construction module generates sequence co-occurrence information and behavioral feedback information based on the search session flow, query terms, click results, subsequent operations, and borrowing records, and updates the relationship weights accordingly. The book resource knowledge graph, the relationship tags, and the relationship weights are sent to the search request parsing module.

[0096] The search request parsing module 03 receives the book resource knowledge graph, processes the search request, extracts entity information and faceted information, parses semantics, obtains a structured search request, and sends the structured search request to the node matching module. Specifically, the search request parsing module receives the book resource knowledge graph, relation tags, and relation weights from the graph construction module and initiates processing when the user inputs a search request. The search request parsing module performs standardized preprocessing on the search request, unifying the field formats of book title, author, theme, keywords, category, publication time, and collection status, and segments the input content into query terms. Specifically, the search request parsing module compares the query terms with book nodes, author nodes, theme nodes, category nodes, and keyword nodes, extracting book title, author, theme, keywords, category, and resource type as entity information. The search request parsing module reads the publication time, collection status, search function options, and resource search scope from the search request and writes faceted information. For query terms not directly matched, the search request parsing module reads hypernym relationship tags, hyponym relationship tags, synonym relationship tags, and related word relationship tags for auxiliary parsing. Subsequently, the search request parsing module encapsulates entity information, faceted information, semantics, and search intent into a structured search request. The structured search request includes entity information fields, faceted information fields, semantic fields, search intent fields, matched node fields, auxiliary matching source fields, and records to be matched, and is sent to the node matching module.

[0097] Node matching module 04 is used to receive the structured search request, match nodes and relationship tags in the book resource knowledge graph to obtain target nodes and facet constraints, and send the target nodes and facet constraints to the association reasoning module. Specifically, the node matching module receives the structured search request from the search request parsing module and reads book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes in the book resource knowledge graph. The node matching module matches the book title, author, topic, keyword, category, and resource type in the entity information field with the corresponding nodes. Successfully matched book nodes, author nodes, topic nodes, category nodes, and keyword nodes are written to the target node field. Specifically, the node matching module matches the publication time, collection status, search function options, and resource search scope in the facet information field with collection nodes, publication time attributes, resource type attributes, and search function options. Successfully matched collection nodes, publication time attributes, resource type attributes, and search function options are written to the facet constraint field. For entity information that is not matched, the node matching module uses hypernym relation tags, hyponym relation tags, synonym relation tags, and related word relation tags for auxiliary matching, and merges the nodes that are successfully matched into the target node field. The target node and the faceted constraint, along with the node matching record, are sent to the association reasoning module.

[0098] The association reasoning module 05 receives the target node, the facet constraints, and the relationship weights, performs association reasoning to obtain the nodes traversed by the path and a set of candidate book nodes, and sends the set of candidate book nodes to the source data matching module. Specifically, the association reasoning module receives the target node and facet constraints from the node matching module, and reads the relationship labels and relationship weights from the graph construction module. Starting from the target node, the association reasoning module selects relationship labels for association reasoning according to the node type of the target node. When the target node is a topic node or keyword node, the association reasoning module extends the path along the hypernym relationship labels, hyponym relationship labels, synonym relationship labels, related word relationship labels, and the relationship label between the book and the topic. When the target node is an author node, the association reasoning module extends the path along the relationship labels between the book and the author, the relationship label between the book and the topic, and the relationship label between the book and the category. Specifically, the association reasoning module filters the path extension results according to the facet constraints. Collection nodes, publication time attributes, resource type attributes, and search function options all participate in path boundary constraints. The association reasoning module reads the relation weights in the path, generates path relation weights, and writes low-weight paths, inconsistent facet paths, and paths without book nodes into the association reasoning record. The book nodes pointed to by the filtered paths form a candidate book node set. This candidate book node set includes book node identifiers, target node identifiers, nodes traversed by the path, relation labels, relation weights, path relation weights, and facet constraints, and is sent to the source data matching module.

[0099] The source data matching module 06 receives the candidate book node set and calls the source data. After normalization preprocessing and semantic matching, it obtains the final candidate book node set and sends the final candidate book node set to the sorting output module. Specifically, the source data matching module receives the candidate book node set from the association reasoning module and calls the source data according to the book node identifiers in the candidate book node set. The source data includes book title, author, subject, keywords, category, abstract attribute, publication time, collection status, and digital resource document content. Specifically, the source data matching module performs normalization preprocessing on the source data, processing the book title into title words, the subject into subject words, retaining keywords as keywords, processing the abstract attribute into abstract words, and writing the publication time, collection status, category, and digital resource document content into the source data fields. The source data matching module constructs a resource entity feature vector based on the title words, subject words, keywords, and abstract words, and constructs a knowledge retrieval feature vector based on the entity information, faceted information, and semantics in the structured retrieval request. Subsequently, the source data matching module performs semantic matching between the resource entity feature vector and the knowledge retrieval feature vector, and generates direct matching records, auxiliary matching records, and faceted matching records by combining the relation tags in the corresponding paths of the candidate book nodes. Candidate book nodes that satisfy both semantic matching and path consistency are written into the final candidate book node set. The final candidate book node set is then sent to the sorting output module.

[0100] The sorting output module 07 is used to receive the final candidate book node set, the nodes traversed by the path, the relation weights, and the source data, generate a matching score, and output the book resource list and theoretical basis in descending order of the matching score. Specifically, the sorting output module receives the final candidate book node set from the source data matching module and reads the final candidate book node records, resource entity feature vectors, knowledge retrieval feature vectors, direct matching records, auxiliary matching records, faceted matching records, and source data fields. The sorting output module also reads the nodes traversed by the path, relation labels, relation weights, and path relation weights generated by the association reasoning module. Specifically, the sorting output module generates a semantic matching score based on the direct matching records, auxiliary matching records, and faceted matching records; generates a path relation score based on the nodes traversed by the path and relation weights; and generates a source data score based on the book title, author, theme, keywords, category, abstract attribute, publication time, collection status, and digital resource document content. The sorting output module reads historical search records, search session flow, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time, and writes the corresponding behavior fields into the matching score generation process. After the matching score is generated, the sorting output module arranges the final candidate book nodes in descending order of the matching score and outputs a list of book resources. The sorting output module reads the target node, the nodes traversed by the path, the relationship label, the relationship weight, the source data field, the direct matching record, the auxiliary matching record, and the faceted matching record to generate theoretical basis, and binds the theoretical basis to each book resource in the book resource list.

Claims

1. A knowledge graph-based intelligent retrieval method for book resources, characterized in that, include: S100. Acquire multi-source book data and user behavior data, and obtain target multi-source book data through data standardization, data cleaning and entity relationship structuring. S200. Based on the target multi-source book data, a book resource knowledge graph is constructed through entity extraction, relation extraction, and attribute association processing, and relation labels and relation weights are generated. S300. Based on the book resource knowledge graph, process the search request, extract entity information and faceted information, parse semantics, and obtain a structured search request; S400. Based on the structured retrieval request, match the nodes and relationship tags in the book resource knowledge graph to obtain the target node and facet constraints; S500: Based on the target node, the facet constraint and the relation weight, perform association reasoning to obtain the set of nodes traversed by the path and the set of candidate book nodes; S600. Based on the candidate book node set, the source data is called, and after normalization preprocessing and semantic matching, the final candidate book node set is obtained; S700. Based on the final candidate book node set, the nodes traversed by the path, the relationship weights, and the source data, generate a matching score, and output the book resource list and theoretical basis in descending order of the matching score.

2. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, The multi-source book data includes book title, author, subject, keywords, classification, abstract attributes, publication time, collection status, and digital resource document content; The user behavior data includes historical search records, search session streams, query terms, click results, subsequent operations, borrowing records, collection records, ratings, comments, and user dwell time.

3. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, S100 includes: performing field standardization processing on the multi-source book data using a unified metadata standard to obtain standardized multi-source book data; The standardized multi-source book data is subjected to data mapping processing to obtain mapped multi-source book data; The mapped multi-source book data is subjected to data deduplication, missing value filling, formatting, and outlier handling to obtain cleaned multi-source book data. The cleaned multi-source book data is processed by identifier allocation, fuzzy matching, and entity relation structuring to obtain the target multi-source book data.

4. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, The book resource knowledge graph includes book nodes, author nodes, topic nodes, category nodes, collection nodes, keyword nodes, and behavior nodes; The relationship tags include tags for relationships between books and authors, between books and topics, between books and categories, between books and collection status, between superordinate terms between topics, between subordinate terms between topics, between synonyms between topics, between related terms between topics, between books and borrowing records, between books and click results, between search terms and click results, between click results and subsequent operations, and between user dwell time and books.

5. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, S200 includes: extracting retrieval session stream, query terms, click results, subsequent operations, and borrowing records from the user behavior data; Generate sequence co-occurrence information based on the query term and the click results; Generate behavioral feedback information based on the click result, subsequent operations, borrowing records, and user dwell time; The sequence co-occurrence information and the behavioral feedback information are written into the book resource knowledge graph, and the relation weights are weighted and updated in real time.

6. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, The entity information includes the book title, author, theme, keywords, category, and resource type in the search request; The facet information includes publication date, collection status, search function options, and resource search scope; The structured search request includes the entity information, the faceted information, the semantics, and the search intent.

7. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, S400 includes: matching the entity information with book nodes, author nodes, topic nodes, category nodes, and keyword nodes in the book resource knowledge graph; The faceted information is matched with the library collection nodes, publication time attributes, resource type attributes, and search function options in the book resource knowledge graph; The nodes that are successfully matched form the target node, and the attributes that are successfully matched form the faceting constraint. Unmatched entity information is assisted by hypernyms, hyponyms, synonyms, and related words. Nodes that are successfully matched are incorporated into the target node.

8. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, The S500 includes: Starting from the target node, filter the relational labels that participate in the association reasoning according to the facet constraints; The path is expanded along the filtered relation labels to obtain the nodes traversed by the path; The path expansion results are filtered based on the node types of the nodes traversed by the path, the relation weights of the relation labels, and the facet constraints. The book nodes pointed to by the filtered paths form the candidate book node set.

9. The intelligent book resource retrieval method based on knowledge graphs according to claim 1, characterized in that, The S600 includes: Extract book titles, authors, themes, keywords, categories, abstract attributes, publication dates, library holdings status, and digital resource document content from the source data; The book title, subject, keywords, classification, and abstract attributes are preprocessed to obtain title terms, subject terms, keywords, and abstract terms. Construct a resource entity feature vector based on the title words, the subject words, the keywords, and the abstract words; Construct a knowledge retrieval feature vector based on the entity information, faceted information, and semantics in the structured retrieval request; The resource entity feature vector is semantically matched with the knowledge retrieval feature vector, and combined with the relationship tags in the corresponding paths of the candidate book nodes to obtain the final set of candidate book nodes.

10. A knowledge graph-based intelligent book resource retrieval system, characterized in that, include: The module includes: data preprocessing module, graph construction module, search request parsing module, node matching module, association reasoning module, source data matching module, and sorting output module. The system is used to implement the method as described in any one of claims 1-9.

Images