Large-scale model-driven smart library management system

Abstract

This invention belongs to the technical field of library management systems, specifically relating to a large-scale model-driven intelligent library management system. It includes: a semantic analysis module for inputting book resource data into a large language model for analysis and extraction of key semantic knowledge; a clustering module for clustering semantic vectors transformed from the large language model using a quadtree segmentation algorithm; an RDF data generation module for generating triple RDF data and then creating URI links between books using a similarity matching method; and an RDF editor module for providing a user interface. The descriptive labels of the quadtree-segmented vector clusters are mapped to the attributes of the triple RDF data; URI links are generated based on the similarity matching of vector clusters of different books using the BIBFRAME system; this improves the knowledge relevance of the entire library collection.

Description

Technical Field

[0001] This invention belongs to the technical field of library management systems, specifically relating to a smart library management system driven by a large model. Background Technology

[0002] With the development of the digital age, libraries face the dual challenges of managing massive information resources and providing efficient knowledge services. Traditional library management systems have significant limitations in handling complex semantic relationships and knowledge organization, making it difficult to meet the needs of modern users for accurate knowledge discovery and related retrieval. In recent years, large language models, represented by Deepseek and other domestically developed open-source models, have made breakthrough progress in the field of natural language processing, providing technological possibilities for the innovation of library management systems. Furthermore, the BIBFRAME system, as a new generation of bibliographic framework application, is designed to replace the MARC format and is more suitable for the semantic web environment.

[0003] Currently, most library management systems use metadata standards such as MARC21 and Dublin Core to describe library collections. These descriptions are usually isolated and make it difficult to discover knowledge connections between book resources.

[0004] On the other hand, traditional classification methods often fail to accurately categorize similar semantic knowledge, neglect the multidimensional attributes of knowledge, and overlook a large number of potential knowledge connections, which severely limits users' ability to browse and retrieve related information. Summary of the Invention

[0005] To address the technical challenges of traditional library management systems in discovering knowledge connections between book resources and accurately classifying similar knowledge semantics, this invention provides a large-model-driven intelligent library management system.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] The big data model-driven smart library management system is characterized by comprising a data analysis layer, a BIBFRAME system management layer, and a smart service platform layer; the data analysis layer includes a data acquisition module, a semantic analysis module, and a clustering module.

[0008] The acquisition module is used to collect various types of book resource data and perform data format conversion and cleaning;

[0009] The semantic analysis module is used to input book resource data into a large language model for analysis and extraction of key semantic knowledge, and to convert the extracted key semantic knowledge into semantic vectors.

[0010] The clustering module is used to cluster and divide semantic vectors transformed from large language models using a quadtree partitioning algorithm;

[0011] The intelligent service platform layer sits above the BIBFRAME system management layer and is used to provide comprehensive reading services to end readers. These comprehensive reading services include book retrieval, personalized recommendations, borrowing management, and multi-terminal synchronization.

[0012] The intelligent service platform layer also includes a reader interaction module, an intelligent recommendation module, and a cross-system interconnection module:

[0013] Reader interaction module: Provides a unified UI across web, mobile, and self-service terminals;

[0014] Intelligent recommendation module: calls the vector clusters and URI links generated by this invention to output recommendation results in real time, including similar topics, related references and popular lists;

[0015] Cross-system interconnection module: Adhering to ISOILL and NCIP protocols, it enables interlibrary loan and single sign-on with other libraries and electronic resource platforms;

[0016] The BIBFRAME system management layer includes an RDF data generation module and an RDF editor module;

[0017] The RDF data generation module is used to generate triple RDF data, and then create URI association links between books through similarity matching methods;

[0018] The RDF editor module provides a user interface that enables book catalogers to create, edit, verify, and correct triple RDF data, and to create and modify URI association links between books.

[0019] Preferably, the key semantic knowledge includes book titles, content summaries, and application examples.

[0020] Preferably, the large language model is the Deepseek model.

[0021] Preferably, the specific implementation of process 3) includes:

[0022] 31) Convert the encoded text into the input format and feed it into the large language model for forward propagation computation. The input format includes lexical IDs, attention masks, and sentence type identifiers.

[0023] 32) The large language model outputs a tensor of shape [batch_size, sequence_length, hidden_size]. The output of the last four layers is extracted and a weighted average is calculated to generate a semantic vector.

[0024] 33) The PCA dimensionality reduction method is used to initially reduce the 768-dimensional semantic vector to 64 or 32 dimensions.

[0025] Preferably, the specific process of the clustering module includes:

[0026] a) Use the t-SNE dimensionality reduction method to reduce the dimensionality of all semantic vectors to a unified two-dimensional space;

[0027] b) Using the center point of the coordinates in the two-dimensional space as the root node of the quadtree, construct the quadtree, and then recursively divide the two-dimensional space into four quadrants, each of which can be further divided.

[0028] c) Set a threshold condition to trigger further splitting, and set a maximum depth to limit the splitting; the threshold condition includes the distance (such as covariance) between semantic vectors in the quadrant space exceeding a certain threshold or the number of vectors in the quadrant space exceeding a certain threshold.

[0029] d) After the quadtree is split, each quadrant space represents a leaf node, and each leaf node represents a vector cluster of a category;

[0030] e) The word embedding method automatically extracts the keywords that best represent the vector cluster as the descriptive label for subsequent mapping.

[0031] Preferably, the maximum depth is 6-8 layers.

[0032] Preferably, the specific process of the RDF data generation module includes:

[0033] Map the descriptive label of the vector cluster to which each book belongs to the corresponding triple RDF data attribute of that book in the BIBFRAME system;

[0034] Similarity matching is performed based on the RDF data attributes of the triplets mapped to each book;

[0035] Based on the matching results, create interconnected URI links between the books in the BIBFRAME system.

[0036] Preferably, the triple RDF data attributes include book title, content summary, and application examples; the triple RDF data attributes correspond to the key semantic knowledge extracted by the large language model;

[0037] Preferably, the specific process of the RDF editor module includes:

[0038] Load the book resource data to be edited from the RDF data generation module and provide a user interface;

[0039] Catalogers can select the corresponding book's triple RDF data attribute on the user interface for editing, including modifying descriptive tags or correcting URI-related link objects;

[0040] Perform SHACL or OWL rule validation on this edit to identify, merge, and split associated link objects of the same URI, and maintain the consistency of URI associated links.

[0041] Preferably, the intelligent recommendation module triggers a recommendation when the minimum similarity is greater than or equal to 0.75.

[0042] Preferably, the cross-system interconnection module is compatible with ISO10160 / 10161(ILL) and NCIP2.0 protocols.

[0043] Preferably, the reader interaction module supports unified identity authentication across the web, Android / iOS clients, and self-service borrowing and returning terminals.

[0044] The beneficial effects of this invention are:

[0045] 1. Utilizing the semantic understanding and analysis capabilities of large language models (such as Deepseek or other domestic open-source large models), it effectively captures fine-grained semantic relationships in book themes, abstracts, and application examples, achieving higher accuracy than traditional natural language processing (NLP) paradigms or TF IDF methods.

[0046] 2. By mapping high-dimensional semantic vectors to a fixed low-dimensional space and combining this with a quadtree segmentation algorithm, multi-level segmentation and clustering optimization of book resources is achieved. This clustering method breaks through the rigid limitations of traditional classification methods and can adaptively discover implicit relationships between books, forming an interconnected knowledge graph network.

[0047] 3. The descriptive labels of the vector clusters after quadtree segmentation are mapped to the RDF data attributes of triples; URI association links are generated based on the BIBFRAME system according to the similarity matching of vector clusters of different books; thus improving the knowledge relevance of the entire library collection.

[0048] 4. This invention transforms complex semantic relationships into an intuitive and personalized reading experience through a smart service platform layer, significantly improving readers' knowledge discovery efficiency and collection utilization. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a structural framework diagram of the intelligent library management system driven by the large model of the present invention;

[0051] Figure 2 This is a flowchart of the steps in the clustering module of the large-model-driven smart library management system of the present invention.

[0052] Figure 3 This is a flowchart of the steps in the RDF data generation module of the large-model driven smart library management system of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figures 1-3 As shown, the big model-driven smart library management system includes a data analysis layer, a BIBFRAME system management layer, and a smart library service platform; the data analysis layer includes a selection module, a semantic analysis module, and a clustering module.

[0055] (1) Overall process of the smart service platform:

[0056] a) The system receives user search requests or reading records;

[0057] b) Use the RDF triples and vector database to perform SPARQL+ANN similarity retrieval;

[0058] c) Return personalized recommendation results and synchronize them to the user's multiple devices.

[0059] (2) Intelligent Recommendation Sub-process:

[0060] 1) Read user profiles (interest topics, borrowing history, reading time);

[0061] 2) Rank the collections by cosine similarity of vector clusters and filter the borrowability status;

[0062] 3) Combine the three candidate sets of "theme similarity", "citation association", and "new book release", and output the Top N after Softmax normalization;

[0063] 4) The results are encapsulated in XML or JSON-LD format and pushed to the front end via REST API.

[0064] The acquisition module is used to collect various types of book resource data and perform data format conversion and cleaning;

[0065] In the specific implementation process, book resource data of different formats or sources, such as print books and e-books, are collected. Various format converters are developed to convert data from different sources (such as MARC21, UNIMARC, etc.) into XML or JSON format as an internal standard, including fields such as book title, content summary, keywords, application examples, and author. Simultaneously, a data cleaning process based on XML or JSON format is implemented, including regular expression or rule-based deduplication (ISBN deduplication, title similarity deduplication), filling in missing fields (calling external book metadata APIs such as Open Library), and standardizing date / language formats.

[0066] The semantic analysis module is used to input book resource data into a large language model for analysis and extraction of key semantic knowledge, and to convert the extracted key semantic knowledge into semantic vectors; the key semantic knowledge includes book titles, content summaries and application examples.

[0067] The clustering module is used to cluster the semantic vectors transformed from the large language model using a quadtree partitioning algorithm.

[0068] The BIBFRAME system management layer includes an RDF data generation module and an RDF editor module;

[0069] The RDF data generation module is used to generate triple RDF data, and then create URI association links between books through similarity matching methods;

[0070] The RDF editor module provides a user interface that enables book catalogers to create, edit, verify, and correct triple RDF data, and to create and modify URI association links between books.

[0071] Specifically, the intelligent library management system of this invention adopts a large model-driven design concept, divided into two core layers: a data analysis layer and a BIBFRAME system management layer. The data analysis layer borrows the idea of ​​using large language models for text semantic vectorization (already implemented in large language models).

[0072] The BIBFRAME (Bibliographic Framework) system originated from the Library of Congress's development and release in 2012. Based on linked data technology, it emphasizes establishing links between data within the semantic web context, using resource description models such as RDF format to provide information, and linking book resources via URIs. Its aim is to develop bibliographic description standards into a linked data model.

[0073] This invention designs a solution that combines a large language model with the BIBFRAME system, providing a modular process of "data acquisition - semantic analysis - clustering and classification - RDF generation - human-computer interaction" to achieve intelligent management of book resources.

[0074] Furthermore, the specific process of the semantic analysis module includes:

[0075] 1) Input the book resource data into a large language model for analysis and extract key semantic knowledge;

[0076] 2) Use the Tokenizer class tool that comes with the large language model to encode the extracted key semantic knowledge;

[0077] 3) The encoded text is processed by the forward computation of the Deepseek model to obtain a semantic vector of fixed dimensions (such as 64 or 32 dimensions);

[0078] Furthermore, the large language model is the Deepseek model or other domestically developed open-source large-scale models; preferably, other domestically developed open-source large-scale models. Specifically, unlike the traditional Deepseek model, which can only model context from left to right or right to left, other domestically developed open-source large-scale models "see" the words to the left and right of a word in the sequence simultaneously in all layers, thus understanding the semantics of a word or sentence more accurately; this is more conducive to summarizing and extracting the keywords and thematic essence of a book.

[0079] In the specific implementation process, key semantic knowledge about book resources is analyzed and extracted using a large language model (Deepseek or other domestically developed open-source large models). This key semantic knowledge is structured or semi-structured, for example:

[0080] Book title: including title, subtitle, etc.

[0081] Content summary: Includes chapter information, keywords, and topic concepts.

[0082] Application examples: including application areas, industries, or instances mentioned in the text.

[0083] Then, a set of tokenizer tools (such as BertTokenizerFast or BertTokenizer) specifically responsible for converting arbitrary text into a discrete TokenID sequence acceptable to the model are used in the large language model; it is called to complete a series of processing steps from word segmentation → mapping → packaging; among them, word segmentation (tokenization): for initial splitting, processing of spaces, punctuation, case normalization, etc., the original text is cut into "pre-tokens".

[0084] Mapping (convert_tokens_to_ids): Assign a unique integer ID to each token; including special markers such as [CLS], [SEP], [MASK], etc.

[0085] Packaging (build_inputs_with_special_tokens): Automatically adds [CLS] and [SEP] to both ends of the sequence; handles padding ([PAD]), truncation, and attention mask.

[0086] After the above word segmentation, mapping, and packaging, a sequence of integer token IDs is output for use in the forward computation of the Deepseek model. The forward computation can generate semantic vectors using a weighted average of the outputs of the last four layers of Deepseek, improving the robustness of semantic representation. Vector compression techniques are used to reduce the high-dimensional original vectors (768-dimensional) of the large language model to 64 or 32-dimensional vectors via PCA for efficient clustering. A vector database (such as FAISS or Milvus) is also established to support efficient similarity matching calculations.

[0087] Furthermore, the specific implementation of process 3) includes:

[0088] 31) Convert the encoded text into the input format and feed it into the large language model for forward propagation computation. The input format includes lexical IDs, attention masks, and sentence type identifiers.

[0089] 32) The large language model outputs a tensor of shape [batch_size, sequence_length, hidden_size]. The output of the last four layers is extracted and a weighted average is calculated to generate a semantic vector.

[0090] 33) The PCA dimensionality reduction method is used to initially reduce the 768-dimensional semantic vector to 64 or 32 dimensions.

[0091] In practice, the encoded text is converted into an input format acceptable to the Deepseek model, including input ID, attention mask, and token type ID. These encoded tensors are then fed into a pre-trained Deepseek model for forward propagation computation. The model processes the input information layer by layer, capturing deep semantic features of the text through self-attention mechanisms and feedforward neural networks. For long texts (such as content summaries), the system employs a segmentation strategy, with each segment limited to the maximum Deepseek sequence length (typically 128K tokens). Each segment of text is processed independently by the Deepseek model, and then the system integrates the representations of these segments to ensure that no important semantic information is lost.

[0092] The output of the Deepseek model is typically a high-dimensional vector, requiring feature extraction and dimensionality reduction to obtain a specified low-dimensional semantic vector. For single-labeled text (such as [CLS] tags), the system directly extracts the [CLS] tag vector from the last layer of Deepseek as the representation of the entire sentence. For multi-labeled text, the system extracts the output of the last four layers of Deepseek and calculates a weighted average to enhance the richness of the semantic representation. The weight of each tag can be determined by the attention score during the pre-training stage or by using uniform weights. To improve the quality of semantic representation, the system implements a pooling operation selection mechanism, dynamically selecting average pooling or max pooling strategies based on different types of text content.

[0093] After obtaining the semantic vector representation of the Deepseek model, the system needs to convert it into a semantic vector of a specified dimension in advance: the PCA dimensionality reduction method can be used to initially reduce the 768-dimensional semantic vector to 64 or 32 dimensions, retaining about 95% of the variance information.

[0094] In a more preferred embodiment, after dimensionality reduction, the system performs L2 normalization on the semantic vectors to ensure that the Euclidean length of all vectors is 1, which facilitates subsequent similarity calculation.

[0095] Furthermore, the specific process of the clustering module includes:

[0096] a) Use the t-SNE dimensionality reduction method to reduce the dimensionality of all semantic vectors to a unified two-dimensional space;

[0097] b) Using the center point of the coordinates in the two-dimensional space as the root node of the quadtree, construct the quadtree, and then recursively divide the two-dimensional space into four quadrants, each of which can be further divided.

[0098] c) Set a threshold condition to trigger further splitting, and set a maximum depth or minimum space size to limit splitting; the threshold condition includes the distance (such as covariance) between semantic vectors in the quadrant space exceeding a certain threshold or the number of vectors in the quadrant space exceeding a certain threshold.

[0099] d) After the quadtree is split, each quadrant space represents a leaf node, and each leaf node represents a vector cluster of a category;

[0100] e) Word embedding-based methods (such as ZhKeyBERT) automatically extract the keywords that best represent the vector cluster as the descriptive label for subsequent mapping;

[0101] In the specific implementation process, the high-dimensional semantic vectors are first mapped to a two-dimensional space using t-SNE or UMAP nonlinear dimensionality reduction algorithms. The t-SNE algorithm is preferred because it supports efficient mapping of new data without recalculating all the data.

[0102] Secondly, a vectorized two-dimensional coordinate system is established to ensure the unique positioning of semantic vectors in two-dimensional space. Simultaneously, based on the distribution characteristics of two-dimensional vectors, the optimal position of the initial root node is dynamically calculated (which may not be a strict coordinate center). Threshold conditions for leaf node splitting are set: for example, splitting is triggered when the number of vectors within a node exceeds 50-200 or the maximum distance between vectors within a node exceeds a preset threshold; this prevents efficiency issues caused by severe imbalance in the tree structure.

[0103] In a specific example, the maximum depth can be set to 6-8 layers or the minimum quadrant space size can be no less than 1 / 512 of the total space.

[0104] After the quadtree splits, ZhKeyBERT and TF-IDF technologies are combined to extract 6-8 of the most representative keyword groups from the book text in each leaf node as descriptive tags; the tags of the main leaf nodes represent a broader topic, and the child leaf nodes represent more specialized sub-topics, thus achieving hierarchical descriptive tags.

[0105] Furthermore, the specific process of the RDF data generation module includes:

[0106] Map the descriptive label of the vector cluster to which each book belongs to the corresponding triple RDF data attribute of that book in the BIBFRAME system;

[0107] Similarity matching is performed based on the RDF data attributes of the triplets mapped to each book;

[0108] Based on the matching results, create interconnected URI links between the books in the BIBFRAME system;

[0109] Furthermore, the triple RDF data attributes include book title, content summary, and application examples; the triple RDF data attributes correspond to the key semantic knowledge extracted by the large language model.

[0110] In practical implementation, the RDF data generation module establishes a mapping rule base for the BIBFRAME system, for example:

[0111] Map the bf:Topic property to its corresponding book title, including the book title, heading, subtitle, etc.

[0112] Map the bf:work attribute to its corresponding content summary, including chapter information, keywords, and topic concepts.

[0113] Map the bf:instance property to its corresponding application instance: including the application area, industry, or instance in the text.

[0114] This process automatically generates RDF data attributes based on triples and uses the Turtle (TTL) format as the standard for RDF storage and exchange. Based on the vector cluster similarity association mechanism of different books, a minimum similarity threshold of 0.75-0.9 is set. A unique URI identifier is generated for each book's associated link, following a unified naming convention. Multi-level URI association link types are established, including "theme similarity," "keyword similarity," and "reference citation" link types.

[0115] Furthermore, the specific process of the RDF editor module includes:

[0116] Load the book resource data to be edited from the RDF data generation module and provide a user interface;

[0117] Catalogers can select the corresponding book's triple RDF data attribute on the user interface for editing, including modifying descriptive tags or correcting URI-related link objects;

[0118] Perform SHACL or OWL rule validation on this edit to identify, merge, and split associated link objects of the same URI, and maintain the consistency of URI associated links.

[0119] In practice, the RDF editor module is an associated data (RDF) editor for cataloging, supporting collaborative cataloging across multiple institutions. Users can use the RDF editor within their resource templates, editing books as the object of editing. Descriptive tags are used to reveal and optimize the book's triple RDF data attributes and their relationships. The created RDF data can be converted to MARC or XML format. It can also be integrated into third-party systems and loaded into ILS and LSP.

[0120] In summary, this invention, based on a large model-driven intelligent library management system, leverages the semantic understanding and analysis capabilities of large language models (such as Deepseek or other domestically developed open-source large models) to effectively capture fine-grained semantic relationships in book topics, abstracts, and application examples, achieving higher accuracy than traditional Natural Language Processing (NLP) paradigms or TF IDF methods. Furthermore, by mapping high-dimensional semantic vectors to a fixed low-dimensional space and combining this with a quadtree segmentation algorithm, multi-level segmentation and clustering optimization of book resources is achieved. This clustering method breaks through the rigid limitations of traditional classification methods, adaptively discovering implicit relationships between books and forming an interconnected knowledge graph network.

[0121] Furthermore, it incorporates the triple RDF data model of the BIBFRAME system framework, mapping the descriptive labels of the vector clusters after quadtree segmentation to triple RDF data attributes; and generates URI association links based on the similarity matching of vector clusters of different books using the BIBFRAME system. This improves the knowledge relevance of the entire library collection and constructs a high-quality book resource association data network.

[0122] In the description of this specification, the terms "further," "in practice," "specific example," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. The above description is merely an example and illustration of the structure of the present invention. Various modifications or additions to the described specific embodiments, or substitutions made by those skilled in the art, as long as they do not depart from the structure of the invention or exceed the scope defined by the claims, should fall within the protection scope of the present invention.

Claims

1. A smart library management system driven by a large model, characterized by: It includes a data analysis layer, a BIBFRAME system management layer, and a smart service platform layer; the data analysis layer includes a data collection module, a semantic analysis module, and a clustering module. The acquisition module is used to collect various types of book resource data and perform data format conversion and cleaning; The semantic analysis module is used to input book resource data into a large language model for analysis and extraction of key semantic knowledge, and to convert the extracted key semantic knowledge into semantic vectors. The clustering module is used to cluster and divide semantic vectors transformed from large language models using a quadtree partitioning algorithm; The intelligent service platform layer sits above the BIBFRAME system management layer and is used to provide comprehensive reading services to end readers. These comprehensive reading services include book retrieval, personalized recommendations, borrowing management, and multi-terminal synchronization. The BIBFRAME system management layer includes an RDF data generation module and an RDF editor module; The RDF data generation module is used to generate triple RDF data, and then create URI association links between books through similarity matching methods; The RDF editor module provides a user interface that enables book catalogers to create, edit, verify, and correct triple RDF data, and then create and modify URI association links between books. The specific process of the clustering module includes: a) Use the t-SNE dimensionality reduction method to reduce the dimensionality of all semantic vectors to a unified two-dimensional space; b) Using the center point of the coordinates in the two-dimensional space as the root node of the quadtree, construct the quadtree, and then recursively divide the two-dimensional space into four quadrants, each of which can be further divided. c) Set a threshold condition to trigger further splitting, and set a maximum depth to limit the splitting; the threshold condition includes the distance between semantic vectors in the quadrant space exceeding a certain threshold or the number of vectors in the quadrant space exceeding a certain threshold. d) After the quadtree is split, each quadrant space represents a leaf node, and each leaf node represents a vector cluster of a category; e) The word embedding method automatically extracts the keywords that best represent the vector cluster as descriptive labels for subsequent mapping; The specific process of the RDF data generation module includes: Map the descriptive label of the vector cluster to which each book belongs to the corresponding triple RDF data attribute of that book in the BIBFRAME system; Similarity matching is performed based on the RDF data attributes of the triplets mapped to each book; Based on the matching results, create interconnected URI links between the books in the BIBFRAME system; The intelligent service platform layer also includes a reader interaction module, an intelligent recommendation module, and a cross-system interconnection module: Reader interaction module: Provides a unified UI across web, mobile, and self-service terminals; The intelligent recommendation module calls the generated vector clusters and URI links to output recommendation results in real time. The recommendation results include similar topics, related references, and popular lists. Cross-system interconnection module: Following the ISOILL and NCIP protocols, it enables interlibrary loan and single sign-on with other libraries and electronic resource platforms.

2. The large-model-driven smart library management system according to claim 1, characterized in that: The key semantic knowledge includes book titles, content summaries, and application examples.

3. The large-model-driven smart library management system according to claim 1, characterized in that: The specific process of the semantic analysis module includes: 1) Input the book resource data into a large language model for analysis and extract key semantic knowledge; 2) Use the Tokenizer class tool that comes with the large language model to encode the extracted key semantic knowledge; 3) The encoded text is processed by the forward computation of the Deepseek model to obtain a fixed-dimensional semantic vector.

4. The large-model-driven smart library management system according to claim 3, characterized in that: The specific implementation methods for obtaining fixed-dimensional semantic vectors by performing forward computation on the encoded text using the Deepseek model include: 31) Convert the encoded text into the input format and feed it into the large language model for forward propagation computation. The input format includes lexical IDs, attention masks, and sentence type identifiers. 32) The large language model outputs a tensor of shape [batch_size, sequence_length, hidden_size]. The output of the last four layers is extracted and a weighted average is calculated to generate a semantic vector. 33) The PCA dimensionality reduction method is used to initially reduce the 768-dimensional semantic vector to 64 or 32 dimensions.

5. The large-model-driven smart library management system according to claim 3 or 4, characterized in that: The large language model is the Deepseek model.

6. The large-model-driven smart library management system according to claim 1, characterized in that: The maximum depth is 6-8 layers.

7. The large-model-driven smart library management system according to claim 1, characterized in that: The triple RDF data attributes include book title, content summary, and application examples; the triple RDF data attributes correspond to the key semantic knowledge extracted by the large language model.

8. The large-model-driven smart library management system according to claim 1, characterized in that: The specific process of the RDF editor module includes: Load the book resource data to be edited from the RDF data generation module and provide a user interface; Catalogers can select the corresponding book's triple RDF data attribute on the user interface for editing, including modifying descriptive tags or correcting URI-related link objects; Perform SHACL or OWL rule validation on this edit to identify, merge, and split associated link objects of the same URI, and maintain the consistency of URI associated links.

9. The large-model-driven smart library management system according to claim 1, characterized in that: The intelligent recommendation module triggers recommendations when the minimum similarity is greater than or equal to 0.

75.

10. The large-model-driven smart library management system according to claim 9, characterized in that: The cross-system interconnection module is compatible with ISO10160 / 10161 and NCIP2.0 protocols.

11. The large-model-driven smart library management system according to claim 9, characterized in that: The reader interaction module supports unified identity authentication across the web, Android / iOS clients, and self-service borrowing and returning terminals.

Images