A knowledge slicing optimization method based on semantic analysis and semantic relation network

Abstract

The application relates to the technical field of knowledge management, in particular to a knowledge slicing optimization method based on semantic analysis and a semantic relation network. In one aspect, through multidimensional evaluation and standardized processing mechanism of heterogeneous data sources, semantic knowledge atoms of time sensitivity and domain labels are generated, the integration accuracy of multi-source heterogeneous data and the consistency of knowledge representation are significantly improved, a reliable data basis is provided for dynamic evolution of the semantic relation network, and the problems of data format conflict, weak semantic association and update lag in traditional knowledge management are solved. In another aspect, through an improved TransE graph embedding algorithm and a multi-layer constraint driven knowledge slicing division mechanism, adaptive structure optimization and multi-granularity knowledge slicing dynamic reorganization of the semantic relation network are realized, the problem of low user retrieval efficiency is effectively avoided, and the semantic retrieval efficiency and scene adaptability of a large-scale knowledge base are enhanced.

Description

Technical Field

[0001] This invention relates to the field of knowledge management technology, specifically to a knowledge slicing optimization method based on semantic analysis and semantic relation networks. Background Technology

[0002] With the accelerated development of the knowledge economy era, knowledge management technology has become a core driving force for enterprise digital transformation, and is widely used in fields such as intelligent decision-making, cross-domain collaboration, and information services. Currently, the scale and heterogeneity of knowledge resources are becoming increasingly prominent, especially in high-knowledge-density scenarios such as healthcare, finance, and law. However, existing technologies still have many problems:

[0003] In terms of integrating multi-source heterogeneous data, most solutions rely on local feature extraction of a single modality and lack the ability to dynamically prioritize and standardize the integration of cross-document libraries, databases and streaming data sources. Furthermore, the knowledge optimization mechanism lacks flexibility, relying on manually set partitioning rules and static weight allocation, and cannot dynamically adjust based on user behavior feedback, knowledge popularity and correlation strength, resulting in redundant and mixed knowledge slices and low retrieval efficiency.

[0004] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge slicing optimization method based on semantic analysis and semantic relation networks.

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

[0007] A knowledge slicing optimization method based on semantic analysis and semantic relation networks includes:

[0008] S1. Knowledge Data Acquisition and Slice Generation: Through data acquisition unit and heterogeneous data source, data communication connection is established to collect the original knowledge unit set, and each knowledge unit is structured to generate standardized knowledge slices.

[0009] S2. Semantic Feature Analysis: The semantic analysis unit performs in-depth semantic analysis on each knowledge slice and generates semantic feature vectors corresponding to each initial slice.

[0010] S3. Semantic Relationship Network Construction: By calculating the multidimensional semantic correlation between each knowledge slice, a dynamically evolving semantic relationship network is constructed.

[0011] S4. Knowledge Slicing and Dynamic Optimization: Based on semantic relationship networks, knowledge slices are divided and optimized, and dynamically adjusted to adapt to data updates and user interactions.

[0012] As a further improvement to the present invention, the specific collection process of the original knowledge unit set is as follows:

[0013] Heterogeneous data acquisition: Data is collected from heterogeneous data sources through the data acquisition unit to obtain a set of raw knowledge units, including academic paper paragraphs, technical document chapters, industry standard clauses, expert opinion records, and case analysis reports; heterogeneous data sources include structured databases, unstructured document libraries, and streaming data sources;

[0014] The timeliness of data is assessed based on a pre-built multidimensional dynamic evaluation model. Content value Data collection costs and compliance risks Perform real-time analysis and calculate the data source priority index, then normalize the data before inputting it into the formula. The data source priority index is obtained, where content value is obtained through a comprehensive evaluation of data relevance, data quality, scarcity, and timeliness. These are the weighting factors for the impact of timeliness, content value, collection costs, and compliance risks, respectively.

[0015] Heterogeneous data sources are sorted based on a data priority index, and data from heterogeneous data sources with higher priority indices is collected first.

[0016] As a further improvement of the present invention, the specific process of generating the knowledge slice is as follows:

[0017] A data fusion middleware is built to convert the original set of knowledge units into a standard format and standardize the encoding, and extract metadata including creation time, data source type, version identifier, author information, knowledge domain classification and credibility score;

[0018] The domain-adaptive word segmentation algorithm, TF-IDF weight calculation method, and named entity recognition technology are used to perform refined structural processing on knowledge units and generate semantic knowledge atoms with time sensitivity and domain labels.

[0019] Based on the knowledge representation specification, the structured knowledge atoms are encapsulated into knowledge slices, including core content blocks, context description domains, and quality assessment matrices.

[0020] As a further improvement of the present invention, the specific analysis process of S2 is as follows:

[0021] Multi-layered language analysis: Deep learning-based natural language processing tools perform lexical, syntactic, and semantic analysis on knowledge slices;

[0022] Semantic disambiguation and coreference resolution: Solving role overlap and cross-sentence reference problems through a context-sensitive multimodal disambiguation model;

[0023] Semantic feature vector generation: Based on an improved multi-head attention mechanism, the contribution of semantic roles is dynamically quantified. Differentiated weight distributions are generated by combining role type, contextual association strength and domain ontology constraints. The weighted semantic role vectors are heterogeneously fused through an adaptive gating mechanism to preserve the core semantic relationships between roles and generate semantic feature vectors with configurable dimensions. The core semantics include entities, relationships and context.

[0024] The dimension of the semantic feature vector can be adjusted based on the application scenario requirements to balance computational efficiency and semantic expressive power.

[0025] As a further improvement of the present invention, the specific process of S3 is as follows:

[0026] Based on semantic feature vectors, multi-dimensional semantic relevance is calculated for each knowledge slice to obtain the corresponding semantic relevance value; the calculation formula is as follows: in, These represent the core content vectors of knowledge slices i and j, respectively; Represents the cosine similarity of vectors; This represents the similarity of the k-th path connecting knowledge slices i and j; Let be the weight factor for the k-th path; Represents a domain relevance function; These are adjustable weighting coefficients, and ;

[0027] The node weights within the corresponding semantic relationship network are initialized based on the semantic relevance value and then marked as node weights. The formula for calculating node weights is as follows:

[0028] in: This represents the number of nodes that have a semantic relationship with node i. Indicates the degree of node j; Used to balance the impact of highly connected nodes;

[0029] An improved TransE graph embedding algorithm is used to optimize the structure of the semantic relation network, strengthening connections between nodes with clear semantic relationships and weakening connections between nodes with no or very weak relationships. The optimized semantic relation network is then constructed. in: A set of nodes; For a weighted edge set, This is the weight matrix.

[0030] As a further improvement of the present invention, the specific process of S4 is as follows:

[0031] Constrained by the consistency score of the quality assessment matrix, the grouping criteria are adjusted in real time based on the overall scale and correlation density of the knowledge network, and a hierarchical partitioning strategy from macro to micro is adopted to achieve knowledge slice partitioning at different granularities.

[0032] A minimum node count threshold is set to traverse the knowledge slice set. Slices with fewer than the minimum node count threshold are merged into the adjacent slice set with the highest merging priority index. The merging priority index is calculated based on the number of edges between slices, the number of nodes, and topic consistency. When the merging priority indices are the same, slices with more nodes or higher topic consistency are merged first. After merging, the original slice is updated and removed. If the slice size exceeds the preset threshold and the internal semantic consistency is lower than the preset threshold, the semantic feature vector distribution faults are identified through the semantic relationship network topology. The split boundary is located by combining the edge weight gradient change. The split point is selected by semantic association strength and domain ontology constraints, and the slice is divided into semantically cohesive sub-slices. The sub-slices inherit the original slice's context description domain and quality evaluation matrix, and the version identifier and association rules are updated synchronously.

[0033] As a further improvement of the present invention, S4 further includes:

[0034] The knowledge slice set is updated periodically by setting a period T. New knowledge units are acquired through data collection and new knowledge slices are generated. The semantic relevance between the new knowledge slices and existing slices is calculated.

[0035] Based on the semantic relevance threshold range, new knowledge slices are integrated into the associated set, stored independently, or assigned through weighted voting.

[0036] Low-quality knowledge slices are identified based on the quality assessment matrix and stored in the queue to be corrected. Automatic repair is performed by combining an active learning model with a domain ontology. If the repair fails, a human-machine collaboration process is triggered, where experts review the work and provide feedback through a multimodal interface to optimize the model. Knowledge slices that fail to meet the standards after multiple corrections or cannot be corrected are marked as dormant and removed from the active knowledge base.

[0037] Based on user interaction with search results, the overall relevance of knowledge slices is quantified;

[0038] Interactive behaviors include the number of clicks. Duration of stay ,score Sharing frequency and citation count After normalization, input the formula as follows: Obtain the overall correlation ,in, These are the weighting factors for the number of clicks, dwell time, rating, sharing frequency, and citation frequency, respectively.

[0039] Different adjustment rhythms are set based on the different types of knowledge slices.

[0040] The beneficial effects of this invention are:

[0041] This invention uses dynamic priority evaluation to filter high-value data from heterogeneous data sources, and combines data fusion middleware to achieve standardized transformation and fine-grained domain segmentation. It extracts semantic roles through multi-level semantic parsing and context disambiguation, generates semantic feature vectors, and constructs a semantic relationship network. Based on a hierarchical partitioning strategy and a semantic fragmentation-driven mechanism, it optimizes knowledge slicing, effectively avoiding the problem of low user retrieval efficiency and enhancing the semantic retrieval efficiency and scenario adaptability of large-scale knowledge bases. Attached Figure Description

[0042] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0043] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0044] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0045] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0046] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0047] like Figure 1 As shown, a knowledge slicing optimization method based on semantic analysis and semantic relation networks includes knowledge data acquisition and slice generation, semantic feature analysis, semantic relation network construction, and knowledge slice division and dynamic optimization.

[0048] S1. Knowledge Data Acquisition and Slice Generation: Through data acquisition units and heterogeneous data sources, a set of raw knowledge units is acquired, and each knowledge unit is structured to generate standardized knowledge slices. The specific implementation process includes:

[0049] 101. Heterogeneous Data Acquisition: Acquire raw knowledge units from heterogeneous data sources through distributed crawler interfaces, open API gateways, and database connectors; heterogeneous data sources include structured databases, unstructured document libraries, and streaming data sources; raw knowledge unit sets include academic paper paragraphs, technical document chapters, industry standard clauses, expert opinion records, and case analysis reports.

[0050] In the process of heterogeneous data acquisition, the timeliness of data is assessed based on a pre-built multidimensional dynamic evaluation model. Content value Data collection costs and compliance risks Perform real-time analysis and calculate the data source priority index, then normalize the data before inputting it into the formula. The data source priority index is calculated, where content value is obtained by comprehensively evaluating data relevance, data quality, scarcity, and timeliness. These are the weighting factors for the impact of timeliness, content value, collection costs, and compliance risks, respectively.

[0051] Heterogeneous data sources are sorted based on data priority index, and data from heterogeneous data sources with higher data source priority index is collected first.

[0052] 102. Data Standardization Processing: Construct a data fusion middleware to uniformly convert the original set of knowledge units into a standard format and standardize the encoding, while extracting metadata; metadata includes creation time, data source type, version identifier, author information, knowledge domain classification, and credibility score;

[0053] During the data standardization process, a dynamic hierarchical system is constructed by setting a preset sub-data capacity threshold. When the number of knowledge domain classification sub-category data in the metadata exceeds the preset sub-data capacity threshold, a domain ontology-driven subdivision process is triggered, as follows:

[0054] Based on the domain ontology tree structure and semantic association strength, the excess subcategories are decomposed into fine-grained levels, such as splitting medicine into internal medicine and surgery, and the classification identifiers and association rules in the metadata are updated simultaneously. During the subdivision process, metadata attributes such as creation time and data source type are inherited; new identifiers are generated to mark the iteration relationship.

[0055] 103. Knowledge Structuring: The domain-adaptive word segmentation algorithm, TF-IDF weight calculation method, and named entity recognition technology are used to perform refined structuring of knowledge units. During the structuring process, based on the language features and domain characteristics of the knowledge units, the word segmentation granularity and entity recognition threshold are adjusted to generate semantic knowledge atoms with time sensitivity and domain labels.

[0056] 104. Knowledge Slice Encapsulation: Based on the knowledge representation specification, structured knowledge atoms are encapsulated into knowledge slices. A knowledge slice includes: a core content block storing vectorized semantic topics and key concepts; a context description domain recording original contextual information and related entities to ensure knowledge traceability; and a quality assessment matrix containing scores for completeness, consistency, and accuracy. The knowledge representation specification refers to the ontology construction standard jointly published by the International Organization for Standardization (ISO) and the International Electrotechnical Commission (IEC), officially named Top-Level Ontology.

[0057] S2. Semantic Feature Analysis: A semantic analysis unit performs in-depth semantic analysis on each knowledge slice and generates semantic feature vectors corresponding to each initial slice. The specific implementation process is as follows:

[0058] 201. Multi-layered Language Analysis: Based on deep learning-based natural language processing tools, knowledge slices are analyzed at three levels: lexical, syntactic, and semantic. Lexical analysis involves accurate word segmentation and part-of-speech tagging, identifying entity boundaries and grammatical attributes in the text. Syntactic analysis constructs a grammatical relation tree through dependency parsing, locating core predicate nodes and their governing argument structures. Semantic analysis maps grammatical roles to a semantic role system based on a pre-trained predicate and argument framework model, forming a set of predicate, role, and entity triples.

[0059] 202. Semantic Disambiguation and Core Reference Resolution: This model addresses role overlap and cross-sentence reference issues through a context-sensitive multimodal disambiguation model. It integrates linguistic knowledge, domain ontology, and external knowledge bases, and calculates the core reference probability between entities through a bidirectional attention mechanism.

[0060] 203. Semantic Feature Vector Generation: Based on an improved multi-head attention mechanism, the contribution weight of semantic roles is dynamically quantified. The multi-head attention mechanism combines role type, contextual association strength, and domain ontology constraints to generate a differentiated weight distribution. The weighted semantic role vectors are heterogeneously fused through an adaptive gating mechanism to preserve the core semantic relationships between roles and generate semantic feature vectors with configurable dimensions. The core semantics include entities, relationships, and context.

[0061] The semantic feature vector dimension can be adjusted based on application scenario requirements to balance computational efficiency and semantic expressive power. Based on the knowledge domain classification and data source type in metadata, the baseline semantic feature vector dimension for different scenarios is preset: for domains with high semantic complexity, high-dimensional vectors are used to enhance fine-grained semantic representation, and more implicit contextual relationships and domain-specific features are captured by expanding the dimension; for streaming data processing scenarios with high real-time requirements, the semantic feature vector dimension is reduced to reduce computational complexity, shorten response latency, prioritize the rapid extraction of core semantic relationships, and discard some fine-grained features to adapt to hardware resource limitations.

[0062] S3. Semantic Relationship Network Construction: By calculating the multidimensional semantic correlation between each knowledge slice, a dynamically evolving semantic relationship network is constructed; the specific implementation process is as follows:

[0063] 301. Semantic Relevance Calculation: Based on the semantic feature vector, multi-dimensional semantic relevance is calculated for each knowledge slice to obtain the corresponding semantic relevance value; the calculation formula is as follows:

[0064] ,in, These represent the core content vectors of knowledge slices i and j, respectively; Represents the cosine similarity of vectors; This represents the similarity of the k-th path connecting knowledge slices i and j; Let be the weight factor for the k-th path; Represents a domain relevance function; These are adjustable weighting coefficients, and ;

[0065] 302. Node Weight Initialization: Initialize the node weights within the corresponding semantic relationship network based on the semantic relevance value, and mark them as node weights. The formula for calculating node weights is as follows:

[0066] ,in: This represents the number of nodes that have a semantic relationship with node i. Indicates the degree of node j; Used to balance the impact of highly connected nodes;

[0067] 303. Semantic Relationship Network Optimization: An improved TransE graph embedding algorithm is used to optimize the structure of the semantic relationship network, strengthening connections between nodes with clear semantic associations and weakening connections between nodes with no or weak associations. The optimized semantic relationship network G is represented as follows: in: A set of nodes, representing a slice of knowledge; This is a set of weighted edges, representing the semantic relationships between knowledge slices; The weight matrix contains node weights and edge weights; the semantic relation network G supports incremental updates and dynamic evolution, adaptively adjusting its structure and weights based on changes in knowledge data;

[0068] S4. Knowledge Slicing and Dynamic Optimization: Based on semantic relationship networks, knowledge slices are segmented and optimized, and dynamically adjusted to adapt to data updates and user interactions; the specific implementation process is as follows:

[0069] 401. Knowledge Slice Partitioning: Knowledge slice sets are partitioned using multiple constraints, dividing the semantic relation network into highly cohesive knowledge slice sets; specific steps include:

[0070] Constraints are applied based on the consistency score of the quality assessment matrix to ensure that the semantic similarity within each subset of knowledge slices is higher than that outside. The grouping criteria are adjusted in real time based on the overall size and association density of the knowledge network. When the network size increases, the partitioning sensitivity is automatically reduced to avoid over-segmentation. When the association density increases, the partitioning accuracy is improved to prevent hybrid redundancy. A hierarchical partitioning strategy from macro to micro is adopted, first forming a large range of basic knowledge groups, and then refining and splitting them layer by layer based on actual needs to achieve knowledge slice partitioning of different granularities.

[0071] 402. Knowledge Slice Merging and Splitting: Setting a Minimum Node Count Threshold Iterate through the knowledge slice set; if the number of nodes in a knowledge slice is less than... If the knowledge slice is merged, it will be merged into a subset of the adjacent knowledge slice set with the highest merging priority index; the merging priority index is calculated as follows:

[0072] ,in, Indicates connecting slices and The number of edges; These represent knowledge slices. and The number of nodes; Represents a topic consistency function; The weighting coefficient for thematic consistency;

[0073] When multiple knowledge slices have the same merging priority index, the knowledge slice with more nodes or higher topic consistency is merged first; after merging, the knowledge slice set is updated and the original knowledge slice is removed.

[0074] If the size of a knowledge slice exceeds a preset threshold and its internal semantic consistency is lower than a preset threshold, a slice splitting operation is triggered. The distribution discontinuity of semantic feature vectors within the knowledge slice is identified through the topological structure of the semantic relation network. Potential splitting boundaries are located based on the gradient changes of the edge weights of the semantic relation network. Through the dual constraints of semantic association strength and domain ontology, regions with significant differences in semantic role contribution weights are selected as splitting points and divided into multiple semantically cohesive sub-slices. The split sub-slices re-establish cross-domain connections through the semantic relation network, inherit the context description domain and quality evaluation matrix of the original slice, and update the version identifier and association rules simultaneously.

[0075] 403. Knowledge Slice Update and Maintenance: Set a time period T to periodically update the knowledge slice set; within the update period:

[0076] New original knowledge units are acquired through the data acquisition unit. The processing flow from S1 to S3 is repeated to generate new knowledge slices. The semantic correlation between the new knowledge slice and the existing knowledge slice set is calculated. If the semantic correlation between the new knowledge slice and the existing knowledge slice exceeds a preset threshold, the new knowledge slice is integrated into the knowledge slice set with the highest correlation.

[0077] If the semantic relevance of a new knowledge slice to all existing knowledge slices is below a preset threshold, it is added to the knowledge slice set as an independent new knowledge slice.

[0078] If the semantic relevance between a new knowledge slice and multiple existing knowledge slices is within a preset threshold range, the ownership will be determined by a weighted voting mechanism.

[0079] Meanwhile, the quality of the knowledge slice set is evaluated based on the quality assessment matrix. Based on the scores of completeness, consistency and accuracy, low-quality knowledge slices are identified and processed. If the score is lower than the preset threshold, the knowledge slice is stored in the queue to be corrected.

[0080] For slices to be corrected, the active learning model, based on an uncertainty sampling strategy and domain ontology constraints, prioritizes slices with low prediction confidence or ambiguous semantic relationships. It then uses a pre-trained semantic model to automatically repair their semantic roles and structures. After repair, the quality is reassessed. If the knowledge slice meets the standards, it is updated; otherwise, a human-machine collaborative process is triggered. During human-machine collaborative correction, the system generates task priorities based on the domain importance of the knowledge slice, user feedback frequency, and the number of automatic correction failures. A multimodal interactive interface displays the original content of the knowledge slice, automatic correction suggestions, and a score comparison, allowing domain experts to review, modify, and supplement information. Expert operation data is fed back to the active learning model in real time to optimize the sampling strategy and correction algorithm. Knowledge slices that fail to meet the standards after multiple corrections or cannot be corrected are marked as dormant and removed from the active knowledge base.

[0081] 404. Dynamic Adjustment of Knowledge Slice Weights: By integrating user behavior feedback and timeliness factors, the weights of knowledge slices and semantic relationship network edges are dynamically adjusted to achieve personalized and timeliness optimization; the specific process is as follows:

[0082] Z1. User Behavior Feedback Analysis: Based on user interaction with search results, quantify the comprehensive relevance of knowledge slices; interaction behavior includes click count. Duration of stay ,score Sharing frequency and citation count After normalization, input the formula as follows: Calculate the overall correlation degree ,in, These are the weighting factors for the number of clicks, dwell time, rating, sharing frequency, and citation frequency, respectively.

[0083] Z2. Based on the different types of knowledge slices, different adjustment rhythms are set. For example, the importance of knowledge slices with high timeliness will be reduced quickly, while the importance of knowledge slices with stable timeliness will be adjusted slowly. The importance of frequently interacting knowledge slices will be increased based on the results of user behavior feedback analysis, while the presence of knowledge slices that have not been paid attention to for a long time will be reduced. The system automatically monitors the importance distribution of all knowledge slices to prevent knowledge slices from being too singular or overly concentrated. When an imbalance trend is detected, it will be corrected through built-in rules.

[0084] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A knowledge slicing optimization method based on semantic analysis and semantic relation networks, characterized in that, include: S1. Knowledge Data Acquisition and Slice Generation: Through data acquisition unit and heterogeneous data source, data communication connection is established to collect the original knowledge unit set, and each knowledge unit is structured to generate standardized knowledge slices. S2. Semantic Feature Analysis: The semantic analysis unit performs in-depth semantic analysis on each knowledge slice and generates semantic feature vectors corresponding to each initial slice. S3. Semantic Relationship Network Construction: By calculating the multidimensional semantic correlation between each knowledge slice, a dynamically evolving semantic relationship network is constructed. S4. Knowledge Slicing and Dynamic Optimization: Based on semantic relationship networks, knowledge slices are divided and optimized, and dynamically adjusted to adapt to data updates and user interactions. The specific process of S4 is as follows: Constrained by the consistency score of the quality assessment matrix, the grouping criteria are adjusted in real time based on the overall scale and correlation density of the knowledge network, and a hierarchical partitioning strategy from macro to micro is adopted to achieve knowledge slice partitioning at different granularities. A minimum node count threshold is set to traverse the knowledge slice set. Slices with fewer than the minimum node count threshold are merged into the adjacent slice set with the highest merging priority index. The merging priority index is calculated based on the number of edges between slices, the number of nodes, and topic consistency. When the merging priority indices are the same, slices with more nodes or higher topic consistency are merged first. After merging, the original slice is updated and removed. If the slice size exceeds the preset threshold and the internal semantic consistency is lower than the preset threshold, the semantic feature vector distribution faults are identified through the semantic relationship network topology. The split boundary is located by combining the edge weight gradient change. The split point is selected by semantic association strength and domain ontology constraints, and the slice is divided into semantically cohesive sub-slices. The sub-slices inherit the original slice's context description domain and quality evaluation matrix, and the version identifier and association rules are updated synchronously.

2. The knowledge slicing optimization method based on semantic analysis and semantic relation networks according to claim 1, characterized in that, The specific collection process for the original knowledge unit set is as follows: Heterogeneous data acquisition: Data is collected from heterogeneous data sources through the data acquisition unit to obtain a set of raw knowledge units, including academic paper paragraphs, technical document chapters, industry standard clauses, expert opinion records, and case analysis reports; Heterogeneous data sources include structured databases, unstructured document libraries, and streaming data sources; The timeliness of data is assessed based on a pre-built multidimensional dynamic evaluation model. Content value Data collection costs and compliance risks Perform real-time analysis and calculate the data source priority index, then normalize the data before inputting it into the formula. The data source priority index is obtained, where content value is obtained through a comprehensive evaluation of data relevance, data quality, scarcity, and timeliness. These are the weighting factors for the impact of timeliness, content value, collection costs, and compliance risks, respectively. Heterogeneous data sources are sorted based on a data priority index, and data from heterogeneous data sources with higher priority indices is collected first.

3. The knowledge slicing optimization method based on semantic analysis and semantic relation networks according to claim 1, characterized in that, The specific process for generating the knowledge slices is as follows: A data fusion middleware is built to convert the original set of knowledge units into a standard format and standardize the encoding, and extract metadata including creation time, data source type, version identifier, author information, knowledge domain classification and credibility score; The domain-adaptive word segmentation algorithm, TF-IDF weight calculation method, and named entity recognition technology are used to perform refined structural processing on knowledge units and generate semantic knowledge atoms with time sensitivity and domain labels. Based on the knowledge representation specification, the structured knowledge atoms are encapsulated into knowledge slices, including core content blocks, context description domains, and quality assessment matrices.

4. The knowledge slicing optimization method based on semantic analysis and semantic relation networks according to claim 1, characterized in that, The specific analysis process for S2 is as follows: Multi-layered language analysis: Deep learning-based natural language processing tools perform lexical, syntactic, and semantic analysis on knowledge slices; Semantic disambiguation and coreference resolution: Solving role overlap and cross-sentence reference problems through a context-sensitive multimodal disambiguation model; Semantic feature vector generation: Based on an improved multi-head attention mechanism, the contribution of semantic roles is dynamically quantified. Differentiated weight distributions are generated by combining role type, contextual association strength and domain ontology constraints. The weighted semantic role vectors are heterogeneously fused through an adaptive gating mechanism to preserve the core semantic relationships between roles and generate semantic feature vectors with configurable dimensions. Core semantics include entities, relationships, and context; The dimension of the semantic feature vector can be adjusted based on the application scenario requirements to balance computational efficiency and semantic expressive power.

5. The knowledge slicing optimization method based on semantic analysis and semantic relation networks according to claim 1, characterized in that, The specific process of S3 is as follows: Based on semantic feature vectors, multi-dimensional semantic relevance is calculated for each knowledge slice to obtain the corresponding semantic relevance value; the calculation formula is as follows: in, These represent the core content vectors of knowledge slices i and j, respectively; Represents the cosine similarity of vectors; This represents the similarity of the k-th path connecting knowledge slices i and j; Let be the weight factor for the k-th path; Represents a domain relevance function; These are adjustable weighting coefficients, and ; The node weights within the corresponding semantic relationship network are initialized based on the semantic relevance value and then marked as node weights. The formula for calculating node weights is as follows: in: This represents the number of nodes that have a semantic relationship with node i. Indicate the degree of node j; Used to balance the impact of highly connected nodes; An improved TransE graph embedding algorithm is used to optimize the structure of the semantic relation network, strengthening connections between nodes with clear semantic relationships and weakening connections between nodes with no or very weak relationships. The optimized semantic relation network is then constructed. in: A set of nodes; For a weighted edge set, This is the weight matrix.

6. The knowledge slicing optimization method based on semantic analysis and semantic relation networks according to claim 1, characterized in that, S4 further includes: The knowledge slice set is updated periodically by setting a period T. New knowledge units are acquired through data collection and new knowledge slices are generated. The semantic relevance between the new knowledge slices and existing slices is calculated. Based on the semantic relevance threshold range, new knowledge slices are integrated into the associated set, stored independently, or assigned through weighted voting. Low-quality knowledge slices are identified based on the quality assessment matrix and stored in the queue to be corrected. Automatic repair is performed by combining an active learning model with a domain ontology. If the repair fails, a human-machine collaboration process is triggered, where experts review the work and provide feedback through a multimodal interface to optimize the model. Knowledge slices that fail to meet the standards after multiple corrections or cannot be corrected are marked as dormant and removed from the active knowledge base. Based on user interaction with search results, the overall relevance of knowledge slices is quantified; Interactive behaviors include the number of clicks. Duration of stay ,score Sharing frequency and citation count After normalization, input the formula as follows: Obtain the overall correlation ,in, These are the weighting factors for the number of clicks, dwell time, rating, sharing frequency, and citation frequency, respectively. Different adjustment rhythms are set based on the different types of knowledge slices.

Images