Semantic adaptation and knowledge migration method, system and equipment based on large language model

Abstract

The embodiment of the invention provides a semantic adaptation and knowledge migration method, system and device based on a large language model, and the method comprises the steps: obtaining the feature data of a target task, carrying out the semantic annotation of the feature data, and obtaining the semantic annotation data; based on the semantic annotation data, establishing an interlayer semantic relation by adopting a graph construction technology, and generating a semantic network graph structure; based on the semantic network graph structure, screening key nodes and performing aggregation operation, and adjusting network hierarchy to form an adjusted semantic hierarchical structure; obtaining a description text of a target task, and extracting core semantic elements and task features as task semantic feature data; adjusting a model structure of the initial model based on the adjusted semantic hierarchical structure and the task semantic feature data to obtain a dynamic matching semantic model; and optimizing parameters of the dynamic matching semantic model to obtain a transferable model. By means of dynamic graph semantic representation and network level dynamic adjustment, accurate adaptation of the model and the target task semantic requirement is achieved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a semantic adaptation and knowledge transfer method, system and device based on a large language model. Background Technology

[0002] Existing deep learning and intelligent retrieval technologies fail to fully utilize the task-specific semantic structure and demand-optimized models when processing specific tasks. They lack specificity and flexibility, and rely heavily on pre-trained models and labeled data, making them difficult to play a role in data-scarce scenarios. At the same time, they lack effective dynamic adjustment mechanisms, resulting in weak model generalization ability and limited application scope, which affects the processing efficiency and effectiveness of complex tasks. Summary of the Invention

[0003] This application provides a semantic adaptation and knowledge transfer method, system, and device based on a large language model, which can solve the problems of insufficient model specificity, high dependence on labeled data, weak generalization ability, and lack of dynamic adjustment mechanism.

[0004] In a first aspect, embodiments of this application provide a semantic adaptation and knowledge transfer method based on a large language model. The method includes: acquiring feature data of a target task, semantically annotating the feature data, and acquiring semantically annotated data; based on the semantically annotated data, creating inter-layer semantic connections using graph construction technology to generate a semantic network graph structure; based on the semantic network graph structure, selecting key nodes and performing aggregation operations to adjust the network hierarchy and form an adjusted semantic hierarchy structure; acquiring the descriptive text of the target task, extracting core semantic elements and task features as task semantic feature data; adjusting the model structure of the initial model based on the adjusted semantic hierarchy structure and task semantic feature data to obtain a dynamically matching semantic model; and optimizing the parameters of the dynamically matching semantic model to obtain a transferable model.

[0005] In one technical solution of this application, the process of obtaining feature data of the target task and semantically annotating the feature data includes: identifying key semantic units based on the feature data of the target task, mapping the semantic categories of the key semantic units, and determining a semantic mapping table; analyzing the correlation strength between key semantic units in the semantic mapping table, adjusting the semantic classification, and obtaining an adjusted semantic mapping table; and annotating each semantic category in the adjusted semantic mapping table to obtain semantic annotation data.

[0006] In one technical solution of this application, based on semantically labeled data, a semantic network graph structure is generated by using graph construction technology to create inter-layer semantic connections. This includes: defining basic semantic nodes and semantic association edges based on semantically labeled data, and constructing a basic graph structure model, wherein the basic semantic nodes represent differential concepts and their attributes, and the semantic association edges reflect the semantic association between differential concepts and attributes; analyzing the semantic flow between basic semantic nodes in the basic graph structure model, identifying key nodes and key semantic association edges in the graph structure, adjusting the graph structure, and obtaining a graph structure transition model; and adjusting the semantic hierarchy of the graph structure transition model to obtain the semantic network graph structure.

[0007] In one technical solution of this application, the process of selecting key nodes and performing aggregation operations based on a semantic network graph structure to adjust the network hierarchy and form an adjusted semantic hierarchy structure includes: identifying core semantic nodes in the semantic network graph structure and calculating the comprehensive influence score of the core semantic nodes; clustering the core semantic nodes according to the comprehensive influence score, the semantic connection strength between core semantic nodes, and the mutual information threshold to obtain an aggregation model structure; evaluating the semantic representation efficiency of the aggregation model structure; and adjusting the connection relationship and aggregation degree between core semantic nodes based on the comprehensive influence score of the core semantic nodes to obtain the adjusted semantic hierarchy structure.

[0008] In one technical solution of this application, obtaining the description text of the target task and extracting core semantic elements and task features as task semantic feature data includes: parsing the description text of the target task, capturing core semantic elements and classifying and mapping them to predefined semantic categories, and obtaining a basic element table; evaluating the relevance between the core semantic elements in the basic element table and the requirements of the target task, adjusting the matching degree, and obtaining a semantic element matching table; and determining the connection point between the core semantic elements in the semantic element matching table and the model capabilities, as task semantic feature data.

[0009] In one technical solution of this application, adjusting the model structure of the initial model based on the adjusted semantic hierarchy and task semantic feature data to obtain a dynamic matching semantic model includes: mapping task semantic features to the adjusted semantic hierarchy, optimizing the data flow path, and obtaining a semantic matching model; adjusting the internal connections of the semantic matching model to obtain an intermediate transition model; and adjusting the semantic processing level of the intermediate transition model to strengthen the expression of core semantic elements and obtain a dynamic matching semantic model.

[0010] In one technical solution of this application, optimizing the parameters of a dynamic matching semantic model to obtain a transferable model includes: identifying key parameters of the dynamic matching semantic model, setting the parameter adjustment range, and constructing a parameter adjustment model; optimizing the key parameters of the parameter adjustment model to obtain a parameter optimization model; and adjusting the learning rate and regularization coefficient of the parameter optimization model to obtain a transferable model.

[0011] In one technical solution of this application, the semantic category mapping is achieved through an entropy threshold iterative optimization algorithm; the entropy threshold iterative optimization algorithm includes: calculating the interval entropy value to construct an entropy value array, initializing the threshold T, marking the interval exceeding the threshold, iteratively adjusting the threshold T until the preset condition is met, and obtaining the semantic mapping table.

[0012] In one technical solution of this application, the analysis of the correlation strength between key semantic units is achieved by the weighted entropy method; the weight coefficients of the weighted entropy method are calculated based on the co-occurrence frequency of key semantic units.

[0013] In one technical solution of this application, adjusting the semantic level includes: constructing an adjacency matrix and a degree matrix, constructing a semantic similarity matrix, balancing the adjacency matrix information and semantic similarity information through preset coefficients, and performing nonlinear transformation on the basic semantic node features.

[0014] In one technical solution of this application, the comprehensive influence score of the core semantic node comprehensively includes the degree, centrality and entropy value of the core semantic node; the mutual information threshold is determined through iterative optimization.

[0015] In one technical solution of this application, key parameters include the number of layers, the number of nodes per layer, the learning rate, and the regularization coefficient; optimization of key parameters is achieved through grid search or Bayesian optimization algorithms.

[0016] Secondly, embodiments of this application provide a semantic adaptation and knowledge transfer system based on a large language model, comprising: a data annotation module for acquiring feature data of a target task, semantically annotating the feature data, and acquiring semantically annotated data; a graph structure construction and optimization module for creating inter-layer semantic connections based on the semantically annotated data using graph construction technology, and generating a semantic network graph structure; a semantic hierarchy matching optimization module for filtering key nodes and performing aggregation operations based on the semantic network graph structure, adjusting the network hierarchy, and acquiring an adjusted semantic hierarchy structure; a task semantic feature extraction module for acquiring the descriptive text of the target task, extracting core semantic elements and task features as task semantic feature data; a dynamic matching model generation module for adjusting the model structure of the initial model based on the adjusted semantic hierarchy structure and task semantic feature data, and acquiring a dynamic matching semantic model; and a parameter optimization and transfer module for optimizing the parameters of the dynamic matching semantic model and acquiring a transferable model.

[0017] Thirdly, embodiments of this application provide an electronic device, the device including: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the semantic adaptation and knowledge transfer method based on a large language model as described in any of the above technical solutions.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the semantic adaptation and knowledge transfer method based on a large language model as described in any of the above technical solutions.

[0019] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the semantic adaptation and knowledge transfer method based on a large language model as described in any of the above technical solutions.

[0020] The semantic adaptation and knowledge transfer method, system, and device based on a large language model in this application can integrate the semantic requirements of the target task into the data layer based on semantically labeled data, fundamentally solving the problem of missing semantic information and providing a foundation for subsequent adaptation. Secondly, it constructs a semantic association carrier through a semantic network graph structure, enabling the task semantics to be recognized and utilized by the model, providing structural support for adaptation. Next, the adjusted semantic hierarchy optimizes the semantic framework, strengthening core semantics, eliminating redundancy, and improving adaptation efficiency. Subsequently, it clarifies the core requirements of the task using task semantic feature data, providing precise direction for model optimization and avoiding blind generalization adjustments. Then, by combining the initial model to generate a dynamically matching semantic model, it achieves deep adaptation between the model structure and task semantics, avoiding insufficient model specificity. Finally, it obtains a transferable model through parameter optimization, enabling cross-task reuse and avoiding weak generalization ability. By leveraging dynamic graph semantic representation and dynamic adjustment of network layers, the model achieves precise adaptation to the semantic requirements of the target task, significantly reducing the dependence on large-scale labeled data, while improving the model's generalization ability across tasks. This effectively addresses the core shortcomings of existing technologies and enhances the model's understanding and matching ability for specific tasks. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is one of the flowcharts illustrating a semantic adaptation and knowledge transfer method based on a large language model provided in this application embodiment; Figure 2 This is the second flowchart illustrating a semantic adaptation and knowledge transfer method based on a large language model provided in this application embodiment; Figure 3 This is a schematic diagram of the specific process of step 202 of a semantic adaptation and knowledge transfer method based on a large language model provided in an embodiment of this application; Figure 4 This is a schematic diagram of the specific process of step 204 of a semantic adaptation and knowledge transfer method based on a large language model provided in an embodiment of this application; Figure 5 This is a schematic diagram of the specific process of step 206 of a semantic adaptation and knowledge transfer method based on a large language model provided in an embodiment of this application; Figure 6 This is a schematic diagram of the specific process of step 208 of a semantic adaptation and knowledge transfer method based on a large language model provided in an embodiment of this application; Figure 7 This is a schematic diagram of the specific process of step 210 of a semantic adaptation and knowledge transfer method based on a large language model provided in an embodiment of this application; Figure 8 This is a schematic diagram of the specific process of step 212 of a semantic adaptation and knowledge transfer method based on a large language model provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of a semantic adaptation and knowledge transfer system based on a large language model provided in an embodiment of this application; Figure 10 This is a neural network model architecture diagram of a semantic adaptation and knowledge transfer system based on a large language model provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0023] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0024] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0025] In order to solve existing technical problems, such as Figures 1 to 11 As shown in the embodiments of this application, a method, system, and device for semantic adaptation and knowledge transfer based on a large language model are provided.

[0026] The semantic adaptation and knowledge transfer method based on a large language model provided in the embodiments of this application will be introduced first.

[0027] Figure 1 This illustration shows a flowchart of a semantic adaptation and knowledge transfer method based on a large language model provided in one embodiment of this application. Figure 1 As shown, the method may include the following steps: Step 102: Obtain feature data of the target task, perform semantic annotation on the feature data, and obtain semantic annotation data; Step 104: Based on the semantically labeled data, graph construction technology is used to create semantic connections between layers and generate a semantic network graph structure; Step 106: Based on the semantic network graph structure, key nodes are selected and aggregated to adjust the network hierarchy and form the adjusted semantic hierarchy structure. Step 108: Obtain the description text of the target task, extract the core semantic elements and task features, and use them as task semantic feature data. Step 110: Adjust the model structure of the initial model based on the adjusted semantic hierarchy and task semantic feature data to obtain a dynamic matching semantic model; Step 112: Optimize the parameters of the dynamic matching semantic model to obtain a transferable model.

[0028] The semantic adaptation and knowledge transfer method based on a large language model provided in this application specifically involves step 102, which is the basic data preparation stage. First, feature data directly related to the target task is collected through business scenario input or data acquisition methods.

[0029] For example, when the target task is "generating medical literature summaries," the feature data may include the original medical literature text, a terminology database, etc. Subsequently, the semantic layers are subdivided according to the semantic requirements of the target task, and the core information in the feature data is deeply annotated. This clarifies the conceptual abstraction and semantic relevance of the labeled data, ultimately forming structured semantically annotated data. This approach overcomes the deficiency of ignoring the task-specific semantic structure, integrating the semantic requirements of the target task into the basic data. This provides a dedicated semantic basis for subsequent graph structure construction and model optimization, avoiding model adaptation bias caused by the generalization of feature data. The entire technical process revolves around the target task from the outset, improving the specificity of the model's semantic recognition.

[0030] Secondly, step 104 of this application is the core solution for semantic information structuring. This step uses semantically labeled data as input, defines nodes representing different concepts and attributes, and edges reflecting the relationships between concepts and attributes through graph construction technology. A basic graph structure is first constructed, then the semantic flow patterns and intensity between nodes are analyzed in depth. Key nodes and edges are identified and rearranged, and the graph structure is adjusted to eliminate semantic levels that do not match the requirements of the target task. Finally, a semantic network graph structure that reveals the semantic relationships between network layers is generated, solving the problem of not establishing semantic connections between layers. The graph structure intuitively presents the semantic relationships of feature data, enabling the model to accurately capture the task-specific semantic structure. Furthermore, unlike methods that construct graph networks based on text span vectors, this step is driven by task-specific semantically labeled data, resulting in a graph structure that better fits the requirements of the target task, providing a precise carrier for subsequent adjustments to network layers.

[0031] Step 106 is a crucial step in semantic framework optimization. This step first identifies core semantic nodes from the semantic network graph structure, determining node importance by calculating their degree, centrality, and entropy. Then, based on the strength of semantic connections and mutual information thresholds between nodes, core nodes are clustered to form a preliminary aggregated model structure. Finally, the semantic representation efficiency is evaluated, and the connection relationships and aggregation degree between nodes are adjusted to optimize the network hierarchy. The final result is a semantic hierarchical structure highly matched to the requirements of the target task. This achieves dynamic adjustment of the network hierarchy, breaking the limitations of a fixed network architecture and enabling the semantic hierarchical structure to flexibly adapt to the dynamic needs of the target task. Simultaneously, the aggregation of key nodes eliminates redundant information, strengthens the expressive power of core semantics, and solves the problems of complex model structure and low semantic representation efficiency, reducing computational costs for subsequent model optimization.

[0032] Step 108 is the core step in structuring task requirements. It involves parsing the description text of the target task, evaluating the relevance of these semantic elements to the task requirements, adjusting the matching degree, and clarifying the connection points between semantic elements and model capabilities. This ultimately generates task semantic feature data containing core semantic elements, relevance scores, and model connection points. This accurately parses the core requirements of the target task, transforming the natural language description of the task into structured features that the model can recognize, thus solving the problem of the disconnect between task requirements and model capabilities. Simultaneously, it provides a clear direction for subsequent model structure optimization, enabling model adjustments to accurately respond to the core semantic requirements of the target task and avoiding performance losses caused by generalization optimization.

[0033] Step 110 uses a pre-trained large language model as the initial model, integrating the adjusted semantic hierarchy and task semantic feature data into it. Through fine-grained adjustments such as mapping task features to the semantic hierarchy, optimizing data flow paths, adjusting internal model connections, and strengthening the expression of key semantic elements, the structure of the initial model is optimized for adaptability, ultimately generating a semantic model dynamically matched to the target task. This semantic structure adaptation significantly reduces the dependence on large-scale labeled data, ensuring model performance even in data-scarce scenarios. Simultaneously, it achieves the effect of adapting the model to the task rather than the task to the model, addressing the shortcomings of insufficient model specificity and unsatisfactory performance on new tasks, and greatly improving the model's adaptability to the target task.

[0034] Step 112 first identifies key parameters affecting task performance in the dynamic matching semantic model, including the number of layers, the number of nodes per layer, the learning rate, and the regularization coefficient, and sets reasonable parameter adjustment ranges. Then, it optimizes the key parameters using grid search or Bayesian optimization algorithms, collecting model performance data under different parameter settings. Finally, it evaluates model performance based on a standardized test set, dynamically adjusting the learning rate and regularization coefficient to balance convergence speed and overfitting risk, ultimately achieving model performance optimization and dynamic transfer across tasks, resulting in a transferable model. This allows the model to be reused across domains and related tasks, significantly expanding the application scope of the technical solution. Simultaneously, it further improves the stability of model performance, ensuring efficient operation in the target task and related tasks, reducing the cost and complexity of model deployment.

[0035] In summary, this application constructs a progressive solution logic. First, based on semantically labeled data, the semantic requirements of the target task are integrated into the data layer, fundamentally addressing the problem of missing semantic information and providing a foundation for subsequent adaptation. Second, a semantic network graph structure is used to construct a semantic association carrier, enabling the task semantics to be recognized and utilized by the model, providing structural support for adaptation. Third, the adjusted semantic hierarchy optimizes the semantic framework, strengthening core semantics, eliminating redundancy, and improving adaptation efficiency. Fourth, task semantic feature data is used to clarify the core requirements of the task, providing precise direction for model optimization and avoiding blind adjustments to generalization. Fifth, by combining the initial model to generate a dynamically matching semantic model, a deep adaptation between the model structure and task semantics is achieved, directly addressing the core problem of insufficient targeting. Finally, parameter optimization yields a transferable model, enabling cross-task reuse and avoiding the problem of weak model generalization ability.

[0036] like Figure 2 As shown, further, the semantic adaptation and knowledge transfer method based on a large language model proposed in this application, based on the above steps, can also have the following main steps in the semantic adaptation and knowledge transfer process: Step 202: Collect feature data of the target task, subdivide semantic levels, perform deep annotation on the data, mark the degree of conceptual abstraction and semantic relevance, perform semantic classification on the annotated data, and obtain semantic annotation data; Step 204: Based on semantically labeled data, use graph construction technology to create connections between layers. By analyzing and reorganizing graph nodes, reveal the semantic connections between different network layers and generate a semantic network graph structure. Step 206: Through the semantic network graph structure, key nodes are selected and aggregated. The network hierarchy is dynamically adjusted according to the target task to form an adjusted semantic hierarchy structure. Step 208: Parse the task description text, extract core semantic elements and task features, link text features with task requirements, and generate task semantic feature data. Step 210: Based on the task semantic feature data and the adjusted semantic hierarchy, optimize the model structure and fine-grainedly adjust the model's adaptability to the target task to obtain a dynamically matched semantic model. Step 212: Optimize the parameters of the dynamic matching semantic model. Dynamically adjust the learning rate and regularization parameters according to task requirements and model feedback to optimize model performance and perform dynamic transfer.

[0037] Step 202 involves acquiring structured semantic annotation data containing markers for conceptual abstraction and semantic relevance, providing a task-specific semantic foundation for subsequent graph structure construction. Step 204 transforms semantic information into a graph structure, revealing semantic connections between network layers and providing a visual semantic association carrier for the model. Step 206 optimizes the semantic hierarchy structure, strengthens core semantic expression, and achieves dynamic adaptation between network layers and the target task. Step 208 transforms natural language task requirements into structured features, clarifying the connection point between the model and task requirements. Step 210 achieves accurate matching between the model and the target task through fine-grained structural optimization, reducing dependence on annotation data. Step 212 improves model performance stability, enables dynamic transfer across tasks, and broadens the model's application scope.

[0038] In one embodiment of this application, obtaining feature data of a target task and semantically annotating the feature data includes: identifying key semantic units based on the feature data of the target task, mapping the semantic categories of the key semantic units, and determining a semantic mapping table; analyzing the correlation strength between key semantic units in the semantic mapping table, adjusting the semantic classification, and obtaining an adjusted semantic mapping table; and annotating each semantic category in the adjusted semantic mapping table to obtain semantic annotation data.

[0039] In this embodiment, identifying key semantic units based on the feature data of the target task, mapping the semantic categories of these key semantic units, and determining a semantic mapping table are fundamental structural steps in semantic annotation. This step takes the feature data of the target task as input. First, it extracts semantically representative key semantic units (i.e., keywords and phrases) from the feature data using text parsing technology. Then, using an entropy threshold iterative optimization algorithm, it calculates the entropy value of the feature data interval and constructs an entropy array. The threshold is initialized with the average value of the array and iteratively adjusted until the optimal entropy threshold is determined. Each key semantic unit is then mapped to its corresponding semantic category that is strongly related to the target task, ultimately forming a semantic mapping table that records the correspondence between units and categories. Transforming scattered target task feature data into structured semantic relationships avoids semantic ambiguity caused by the generalization of the original data. Simultaneously, determining the optimal threshold through the algorithm ensures the accuracy of semantic category mapping, providing a standardized basis for subsequent semantic classification adjustments.

[0040] This study analyzes the correlation strength between key semantic units in the semantic mapping table, adjusts the semantic classification, and obtains the adjusted semantic mapping table. Based on the constructed semantic mapping table, the correlation strength between key semantic units is quantitatively analyzed using a weighted entropy method, where the weight coefficients are calculated from the co-occurrence frequency between units. Subsequently, by defining an entropy adjustment function, the entropy value of key semantic units within each semantic category is compared with the optimal entropy threshold. Based on the comparison results, similar semantic categories are merged, overly broad categories are split, and the semantic classification system is reconfigured, ultimately resulting in the adjusted semantic mapping table. This addresses potential issues such as category redundancy and loose correlation in the initial semantic classification. The quantitative analysis of the correlation strength between units achieves precise optimization of semantic classification, making it more aligned with the semantic logic of the target task and providing a more reasonable classification framework for subsequent deep annotation.

[0041] Each semantic category in the adjusted semantic mapping table is labeled to obtain the semantic annotation data. Based on the adjusted semantic mapping table, a deep annotation program is executed. Combining the specific requirements of the target task, each semantic category and the key semantic units within the category are meticulously annotated. Emphasis is placed on marking the abstractness of concepts (such as the hierarchical division of first-level and second-level concepts) and the semantic relevance scores between key semantic units. At the same time, the weights are adjusted according to the relevance between the category and the requirements of the target task to ensure the annotation accuracy of highly relevant categories. Finally, structured semantic annotation data is generated, transforming the optimized semantic classification into annotation data rich in task-specific semantic information. This breaks the limitation of not fully exploring the task-specific semantic structure and provides accurate and targeted semantic input for the subsequent graph structure construction module, ensuring the effective use of semantic information in subsequent processes.

[0042] like Figure 3 As shown, step 202 specifically includes: obtaining feature data for the target task, semantically labeling the feature data, and obtaining semantically labeled data. Step 302: Based on the feature data of the target task, perform text parsing, identify keywords and phrases, map semantic categories, and generate a semantic mapping table; Step 304: Based on the semantic mapping table, analyze the strength of the association between words, reconfigure the semantic classification, optimize the accuracy of the semantic categories, and generate the modified semantic mapping table; Step 306: Based on the modified semantic mapping table, perform deep annotation on each semantic category, link it to the target task requirements, and generate semantic annotation data.

[0043] Understandably, based on the feature data of the target task, an entropy threshold setting program is first established. The purpose is to construct an entropy value array M by calculating the entropy value of each interval, where n1 represents the total number of intervals. Next, the average value m1 of the entropy value array M is calculated, and the entropy threshold T is initialized to m1. According to the set entropy threshold T, the program iterates through the entropy value array M, using an indicator function x to calculate whether each interval exceeds the entropy threshold T. If it does, the interval is marked. The indicator function x is defined as: 1 when the interval entropy value is greater than T, and 0 otherwise. Next, the program checks the number of marked intervals. If the number of marked intervals is lower than a preset threshold α1 or has not changed compared to the previous iteration, the iteration stops, and T at this point is considered the optimal entropy threshold. Otherwise, the entropy threshold T is decremented by a step size δ, and the calculation is repeated. This process is repeated until the optimal entropy threshold T is found, generating a semantic mapping table.

[0044] The semantic mapping table is analyzed to determine the association strength between key semantic units, and the semantic classification is adjusted to obtain the adjusted semantic mapping table (step 304). Based on the semantic mapping table, the word association strength analysis program is then executed. First, an entropy adjustment function is defined. The semantic classification is adjusted by comparing the entropy value of words in each semantic category with the optimal entropy threshold T. For each semantic category, the program calculates the entropy value of each word within the category and compares it with T. Based on the comparison results, the semantic classification of the words is adjusted to optimize the accuracy of the semantic categories. During the adjustment process, the interaction between words is considered, and the weighted entropy method is used to quantify the association strength between words. The weight coefficients are calculated based on the co-occurrence frequency between words. Through an iterative optimization process, the adjusted semantic mapping table is generated.

[0045] In step 306, each semantic category in the adjusted semantic mapping table is labeled to obtain semantic annotation data. Based on the modified semantic mapping table, a deep annotation program is executed to perform in-depth analysis and annotation of the vocabulary within each semantic category, linking it to the target task requirements, considering the subdivision of semantic categories, and the specific attributes and contextual relationships of the vocabulary within each category. By setting a set of annotation rules and parameters, such as annotation depth and semantic precision parameters, the semantic categories can be automatically and meticulously divided and labeled. During the annotation process, each semantic category is weighted according to its relevance to the target task requirements, ensuring that highly relevant semantic categories receive more accurate annotations. Finally, semantic annotation data containing deep annotation information is generated.

[0046] In one embodiment of this application, the creation of inter-layer semantic connections and the generation of a semantic network graph structure based on semantically labeled data and graph construction technology includes: defining basic semantic nodes and semantic association edges based on semantically labeled data, and constructing a basic graph structure model, wherein the basic semantic nodes represent differential concepts and their attributes, and the semantic association edges reflect the semantic association between differential concepts and attributes; analyzing the semantic flow between basic semantic nodes in the basic graph structure model, identifying key nodes and key semantic association edges of the graph structure, and generating a graph structure transition model; adjusting the semantic hierarchy of the graph structure transition model to obtain the semantic network graph structure.

[0047] In this embodiment, the basic semantic node is a basic unit of graph structure defined based on semantic annotation data, representing the difference concepts and their attributes. It is the specific carrier of semantic information in the graph, and its source is directly related to the semantic annotation data of the target task.

[0048] Semantic association edges are graph structure elements used to connect basic semantic nodes, specifically reflecting the semantic relationships between different concepts and attributes. Their construction is based on semantic association information from semantic annotation data.

[0049] The basic graph structure model is an initial graph structure built by defining basic semantic nodes and semantically related edges. It can initially reflect the semantic relationships in the semantically labeled data and serves as the basis for subsequent graph structure optimization.

[0050] Semantic flow refers to the process by which semantic information is transmitted and interacted between basic semantic nodes in a basic graph structure model through semantic association edges. Its pattern and strength can reflect the tightness of semantic association.

[0051] Key nodes in a graph structure are those nodes with high entropy values ​​that play a core supporting role in semantic flow among the basic semantic nodes, and they are the concentrated embodiment of the core semantics of the target task.

[0052] Key semantic association edges are those edges that significantly affect the efficiency of semantic flow and have high entropy values; they are the key channels for the transmission of core semantics.

[0053] The graph structure transition model is an intermediate graph structure that has undergone key node and edge filtering and layout adjustment. Compared with the basic graph structure model, it has reduced redundancy and more prominent core semantic connections.

[0054] Semantic network graph structure is the final graph structure obtained by making targeted adjustments to the semantic hierarchy of the graph structure transition model. It can accurately reveal the semantic relationships between network layers and is highly adapted to the requirements of the target task.

[0055] Transforming task-specific semantic annotation data into a semantic network graph structure that accurately reveals inter-layer semantic relationships and is highly adapted to the target task solves the defects of not establishing inter-layer semantic connections and the disconnect between graph structure and task requirements. This provides a high-quality semantic carrier for subsequent key node aggregation, dynamic adjustment of network hierarchy, and model structure optimization, ensuring the effective transmission and utilization of semantic information throughout the entire technical solution. A three-layer progressive logic of "basic graph construction, key element selection, and semantic hierarchy adjustment" is constructed: First, by defining basic semantic nodes and semantic relationship edges, the structured semantic annotation data is transformed into a graph structure, solving the problem that the semantic information form is not conducive to subsequent analysis and providing a foundation for semantic relationship mining.

[0056] Secondly, by analyzing semantic flow and filtering key elements, redundant information is eliminated, core semantic connections are strengthened, the problem of low efficiency in the initial graph structure is solved, and the relevance of the graph structure is improved.

[0057] Finally, by adjusting the semantic level and optimizing graph convolution, the graph structure is made to fit the requirements of the target task, thus solving the problem of the graph structure being out of touch with the task.

[0058] Among them, such as Figure 4 As shown, the steps described above—creating inter-layer semantic connections based on semantically labeled data, adjusting the graph structure, and obtaining the semantic network graph structure—that is, step 204, include: Step 402: Based on semantically labeled data, create nodes and edges. Nodes represent differentiated concepts and attributes, and edges represent semantic connections between concepts and attributes. Construct a graph structure that represents basic semantic connections between layers and generate a basic graph structure model. Step 404: Based on the basic graph structure model, analyze the semantic flow between nodes, identify key nodes and edges, reconfigure nodes and edges according to the importance of semantic flow, adjust the graph structure, and generate a graph structure transition model. Step 406: Based on the semantic network graph structure, identify the semantic levels that do not match the target task requirements, reconstruct the semantic levels in the graph according to the task requirements, optimize the semantic levels of the model, match the current target task requirements, and generate the semantic network graph structure. In step 402 of this application, graph structure construction is initiated based on semantically labeled data. The goal is to transform the semantically labeled data into a graph structure to better represent the semantic relationships between concepts and attributes. First, nodes and edges are defined. Nodes represent differentiated concepts and attributes in the labeled data, and edges represent semantic connections between concepts and attributes. The importance of each node is determined by calculating its entropy value and comparing these values ​​with the previously obtained optimal entropy threshold T. Based on this step, a graph structure representing basic semantic connections between layers is constructed. The interlexical association strength calculated in previous steps is used as the basis for edge construction to ensure that the graph structure accurately reflects the semantic connections between concepts and attributes. Finally, a basic graph structure model is generated. In step 404, semantic flow analysis is performed based on the basic graph structure model. The pattern and intensity of semantic flow are analyzed by calculating the path length and number of paths between nodes in the graph. An entropy-based method is used to identify key nodes and edges, i.e., by calculating the entropy values ​​of nodes and edges to measure their importance in semantic flow. Based on the importance of semantic flow, the layout of nodes and edges is reconfigured, and the graph structure is adjusted to improve the efficiency and accuracy of information flow. Through iterative optimization, a graph structure transition model is generated. Step 406 involves semantic hierarchy optimization of the model, according to formula (1): (1); Based on Format 1-1, compute the new representation of the nodes and generate a semantic network graph structure that matches the requirements of the current target task. For the first Layer node feature representation, This represents the node features after a single layer of graph convolution. , is the adjacency matrix of the graph. Add the identity matrix , for The degree matrix, For the first The weight matrix of the layer, It is a non-linear activation function. The adjacency matrix and the self-connection are the assemblage coefficients. The weighted coefficient for semantic similarity. This is a semantic similarity matrix. This is the weight matrix for semantic similarity.

[0059] The specific execution process of the above steps is as follows: First, calculate the adjacency matrix. Sum-degree matrix To construct the basic structure of the graph, and then utilize Node features are updated to capture direct interactions between nodes in the constructed graph; a semantic similarity matrix is ​​introduced. and the corresponding weight matrix ,pass Strengthen content-based semantic connections between nodes, here and These are coefficients used to balance the importance of adjacency matrix information and semantic similarity information. The optimal values ​​of these coefficients are determined through cross-validation or automatic adjustment methods based on task performance. The results from both parts are then processed using the coefficients... and Weighted merging, and through activation function A nonlinear transformation is performed to obtain the final new representation of the nodes. The process not only considers the structural information of the structure graph, but also integrates the semantic relationships between nodes, thereby more accurately adjusting and optimizing the semantic level of the model to match the needs of the current target task.

[0060] The semantic level optimization of the model is calculated according to formula (2): (2) Where V represents the length of the dataset dictionary, i represents the i-th item to be removed from vector z, n(t) represents the i-th item to be removed from z (i.e., the i-th word wi=t), the number of times topic k is assigned to word t, n(k) represents the i-th item to be removed from vector z, the number of times topic k is assigned to words in document m, βt represents the Dirichlet prior of word t, and αk represents the Dirichlet prior of topic k.

[0061] After determining the implicit topic of each word, the parameters θm and φk of the LDA topic model are calculated using the formula shown in (3): (3) Where θm,k represents the probability that document m is assigned to topic k, φk,i represents the probability that topic k is assigned to word wi=t, n(k) represents the number of times document m is assigned to topic k, n(t) represents the number of times topic k is assigned to word t, k=1,2,…,K, i=1,2,…,V,m=1,2,…,M.

[0062] A new representation of the computing nodes is generated to create a semantic network graph structure that matches the requirements of the current target task. For the first Layer node feature representation, This represents the node features after a single layer of graph convolution. , is the adjacency matrix of the graph. Add the identity matrix , for The degree matrix, For the first The weight matrix of the layer, It is a non-linear activation function. The adjacency matrix and the self-connection are the assemblage coefficients. The weighted coefficient for semantic similarity. This is a semantic similarity matrix. This is the weight matrix for semantic similarity.

[0063] In one embodiment of this application, the process of filtering key nodes and performing aggregation operations based on a semantic network graph structure to adjust the network hierarchy and form an adjusted semantic hierarchy structure includes: identifying core semantic nodes in the semantic network graph structure and calculating the comprehensive influence score of the core semantic nodes; clustering the core semantic nodes according to the comprehensive influence score, the semantic connection strength between core semantic nodes, and the mutual information threshold to obtain an aggregation model structure; evaluating the semantic representation efficiency of the aggregation model structure; and adjusting the connection relationship and aggregation degree between core semantic nodes based on the comprehensive influence score of the core semantic nodes to obtain the adjusted semantic hierarchy structure.

[0064] In this embodiment, the core semantic node is a node selected from the semantic network graph structure that corresponds to the key concepts and attributes of the target task. It is the core carrier of semantic information, and its selection is based on its supporting role in the semantic expression of the target task.

[0065] The overall influence score is a quantitative indicator that measures the importance of core semantic nodes. It is calculated by combining the node's degree, centrality, and entropy value, and directly reflects the weight of the node's role in semantic flow.

[0066] Semantic connection strength refers to the degree of semantic association between core semantic nodes, and is one of the key criteria for judging whether a node is suitable for clustering.

[0067] The mutual information threshold is a critical condition for clustering core semantic nodes. It refers to the minimum standard of shared information between two nodes. It is determined through iterative optimization and is used to filter out loosely semantically related node combinations.

[0068] The aggregation model structure is an intermediate model structure formed by integrating core semantic nodes that meet the clustering conditions, reflecting the association and clustering relationship of the core semantic nodes.

[0069] Semantic representation efficiency is an indicator that measures the ability of an aggregation model structure to convey and express semantic information. It is evaluated by combining path length and clustering coefficient.

[0070] The adjusted semantic hierarchy is the final semantic framework formed after semantic representation efficiency evaluation and optimization of node connection relationships and aggregation degree. It includes the priority of key concept nodes and the configuration of inter-level connections, and is highly adapted to the requirements of the target task.

[0071] By identifying the core semantic nodes in the semantic network graph structure, the comprehensive influence score of the core semantic nodes is calculated. Using the semantic network graph structure as input, the core semantic nodes that play a core supporting role in the semantic expression of the target task are first screened out from the graph. Then, the comprehensive influence score of each core semantic node is determined through quantitative calculation. The calculation dimensions clearly include the degree, centrality, and entropy of the node. The three together constitute the core indicators for measuring the importance of the node in semantic flow.

[0072] Based on the comprehensive influence score, the semantic connection strength between the core semantic nodes, and the mutual information threshold, the core semantic nodes are clustered to obtain an aggregated model structure. The comprehensive influence score of the core semantic nodes is used as the priority criterion, combined with the semantic connection strength between nodes (i.e., the degree of association between the concepts corresponding to the core semantic nodes), and a mutual information threshold (the critical value of shared information between two nodes, determined through iterative optimization) is introduced. Clustering and integration are only performed when the mutual information between core semantic nodes exceeds this threshold, grouping nodes with close semantic associations and similar influence levels into one category, ultimately forming an aggregated model structure. This ensures that the clustering results focus on the core semantics and conform to the semantic logic of the target task, avoiding semantic bias caused by the mis-aggregation of irrelevant nodes, and making the aggregated model structure more aligned with the core requirements of the task.

[0073] The semantic representation efficiency of the aggregated model structure is evaluated. Based on the comprehensive influence score of the core semantic nodes, the connection relationships and aggregation degree between the core semantic nodes are adjusted to obtain the adjusted semantic hierarchy. This step first uses an improved path length and clustering coefficient evaluation method to quantitatively evaluate the semantic representation efficiency of the aggregated model structure. Then, based on the evaluation results and combined with the comprehensive influence score of the core semantic nodes, the connection relationships and aggregation degree between nodes are finely adjusted. The connection weights of high-scoring nodes are enhanced to strengthen the core semantic expression, the aggregation degree or position of low-scoring nodes is adjusted to reduce redundancy interference, and the hierarchical relationship is reconfigured to place the core semantic nodes in more critical hierarchical positions, ultimately forming the adjusted semantic hierarchy. The technical effect of this step is that it achieves dynamic optimization of the network hierarchy.

[0074] Among them, such as Figure 5 As shown, the above-mentioned process, based on the semantic network graph structure, involves filtering key nodes and performing aggregation operations to adjust the network hierarchy and form an adjusted semantic hierarchical structure. Step 206 includes: Step 502: Based on the semantic network graph structure, identify key nodes in the graph, aggregate nodes according to the strength and importance of semantic connections between nodes, and generate an aggregation model structure. Step 504: Based on the aggregated model structure, evaluate the efficiency of the structural representation model semantics, enhance the representation ability of key concepts in the model and adjust the position of secondary concepts to generate an enhanced model structure; Step 506: Based on the enhanced model structure, perform detailed analysis, identify the locations in the structure that differ from the requirements of the target task, reconfigure the connection relationships and aggregation degree between nodes, strengthen the semantic representation of key nodes, and generate an adjusted semantic hierarchy structure.

[0075] In step 502, based on the semantic network graph structure, node aggregation is initiated. First, key nodes in the graph are identified by calculating the comprehensive influence score of each node. This score considers the node's degree (number of connections), centrality (important position in the graph), and entropy (a measure of information content). Next, based on the strength of semantic connections between nodes and their importance, an entropy-based aggregation strategy is used to aggregate the nodes. This strategy considers the mutual information between nodes—the amount of information shared between two nodes—to determine the aggregation priority. By setting a threshold, nodes are aggregated only when the mutual information between them exceeds this threshold, thus generating the aggregation model structure.

[0076] Step 504: Based on the aggregated model structure, perform a model representation efficiency evaluation. First, based on the current state of the aggregated model structure, use an improved path length and clustering coefficient evaluation method to assess the semantic representation efficiency of the model. Path length evaluation focuses on the speed and efficiency of information transmission in the model, while clustering coefficient evaluation focuses on the density of nodes in the model, serving as an indicator of the model's semantic representation capability. Based on the evaluation results, structural adjustments are made, including enhancing the representation capability of key concepts—achieved by increasing the weight of edges associated with key nodes—and adjusting the position of secondary concepts. This is achieved by reducing the weight of their connections to central nodes or moving them to lower levels, resulting in an enhanced model structure.

[0077] Step 506 involves several optimization measures to enhance the accuracy and efficiency of the model in processing complex semantic information. Specifically, this includes adjusting the semantic processing hierarchy to ensure the model can more accurately and effectively parse complex semantic information.

[0078] In one embodiment of this application, obtaining the description text of the target task and extracting core semantic elements and task features as task semantic feature data includes: parsing the description text of the target task, capturing core semantic elements and classifying and mapping them to predefined semantic categories, and obtaining a basic element table; evaluating the relevance between the core semantic elements in the basic element table and the requirements of the target task, adjusting the matching degree, and obtaining a semantic element matching table; and determining the connection points between the core semantic elements in the semantic element matching table and the model capabilities as task semantic feature data.

[0079] In this embodiment, the description text of the target task refers to the user's description of the requirements for the task to be completed, such as task instructions or business requirement descriptions, which is the direct basis for extracting core semantic elements.

[0080] Core semantic elements are key information units that are selected from the description text of the target task and are directly related to the core requirements of the task. They include the task theme, core functions, constraints, etc., and are the core carriers of task semantics.

[0081] Predefined semantic categories are semantic classification frameworks set in advance based on the domain and functional type of the task. They are used to classify and organize core semantic elements to ensure the systematicness and standardization of semantic elements.

[0082] The basic element table is a preliminary structured table that records the correspondence between core semantic elements and predefined semantic categories. It is an intermediate product in the transformation of task requirements from natural language to structured data.

[0083] Relevance refers to the degree of connection between core semantic elements and the requirements of the target task, which is quantitatively evaluated through weighted similarity calculation.

[0084] The semantic element matching table is a structured table formed after relevance evaluation and matching degree adjustment. It retains only the core semantic elements that are highly relevant to the target task requirements and clarifies the matching priority of each element.

[0085] The docking point is the correspondence between the core semantic element and the model capability, that is, the functional module or processing logic in the model that can process the semantic element, which is the key to connecting task requirements with model capabilities.

[0086] Task semantic feature data is the final structured data formed by integrating core semantic elements, relevance priorities, and model docking points. It is the direct input for subsequent model structure optimization and can be recognized and utilized by the model to adapt to the target task.

[0087] Based on this, firstly, by parsing the descriptive text and mapping it to predefined semantic categories, unstructured natural language is transformed into a preliminary structured basic element table, which solves the problem of difficulty in quantifying task requirements and provides a foundation for subsequent processing.

[0088] Secondly, by adjusting the matching degree through weighted similarity evaluation, highly relevant core semantic elements are selected, which solves the problem of redundancy and mixing in the initial element table and improves the accuracy of semantic features.

[0089] Finally, by identifying the interface points with the model's capabilities, the core semantic elements were transformed into task semantic feature data that the model could recognize. This solved the problem of the disconnect between task requirements and model capabilities, building a bridge between them. The entire process, from requirement analysis to feature optimization and model integration, ensured the integrity of the core task requirements while achieving seamless integration with model capabilities. The resulting task semantic feature data effectively improved the matching degree between tasks and models through semantic analysis.

[0090] like Figure 6 As shown, obtaining the description text of the target task, extracting core semantic elements and task features, and obtaining task semantic feature data can be understood as parsing the task description text, extracting core semantic elements and task features, linking text features with task requirements, and generating task semantic feature data, i.e., step 208. This step specifically includes: Step 602: Parse the task description text, capture the core words and phrases in the text through semantic recognition technology, classify the core elements and map them to predefined semantic categories, identify key semantic elements directly related to the task description, and generate a basic element table. Step 604: Based on the basic element table, analyze each semantic element in the basic element table, evaluate its relevance to the task requirements, adjust the matching degree between the semantic elements and the task requirements, and form a semantic element matching table. Step 606: Based on the semantic element matching table, determine the connection points between the semantic elements in the semantic element matching table and the model capabilities, identify the semantic elements that the model can directly respond to and the elements that need to be transformed or mapped, and generate task semantic feature data.

[0091] Step 602 transforms the task description in natural language into a preliminary structured basic element table, identifying key semantic elements directly related to the task and providing standardized materials for subsequent relevance assessment. Step 604 filters highly relevant semantic elements, removes redundant information, optimizes the matching degree between semantic elements and task requirements, and forms a semantic element matching table focusing on core requirements. Step 606 builds a bridge between task requirements and model capabilities, generating structured task semantic feature data, providing a clear basis for subsequent model adaptation.

[0092] Specifically, the assessment in step 604 is related to the task requirements, according to formula (4): (4) The above formula calculates the weighted similarity between semantic elements and task requirements, generating an optimized semantic element matching table, where, This is a vector representation of the initial semantic elements. This is a vector representation of the task requirements. Represents the dot product of two vectors. and They represent and The Euclidean norm, For context-dependent parameter vectors, This is the angular difference vector between the context and the semantic element. For the number of context parameters, and These are the weighting coefficients.

[0093] In one embodiment of this application, adjusting the model structure of the initial model based on the adjusted semantic hierarchy and task semantic feature data to obtain a dynamic matching semantic model includes: mapping task semantic features to the adjusted semantic hierarchy, optimizing the data flow path, and obtaining a semantic matching model; adjusting the internal connections of the semantic matching model to obtain an intermediate transition model; and adjusting the semantic processing level of the intermediate transition model to strengthen the expression of core semantic elements and obtain a dynamic matching semantic model.

[0094] In this embodiment, task semantic features are mapped to an adjusted semantic hierarchy, data flow paths are optimized, and a semantic matching model is obtained, achieving a preliminary and accurate connection between the core requirements of the task and the dedicated semantic framework. This step uses the adjusted semantic hierarchy (a semantic framework tailored to the target task) as a carrier to accurately map structured task semantic feature data to the corresponding levels of the model. At the same time, redundant levels are merged, and unnecessary data flow branches are simplified, allowing the model structure to revolve around the core semantics of the task. This eliminates redundant structures unrelated to the task and ensures more efficient data flow, providing a concise and targeted basic framework for subsequent model optimization.

[0095] Adjusting the internal connections of the semantic matching model and obtaining an intermediate transition model effectively addresses the model's semantic response weaknesses, improving its ability to parse and respond to task semantics. By analyzing the semantic matching model's response efficiency to task semantics, we can accurately pinpoint weak points in semantic parsing and unreasonable internal connections. By strengthening the connection strength of key semantic paths and adjusting the configuration of inefficient connections, we can improve the model's speed and accuracy in capturing and parsing the core semantics of the task, enabling the model to transition from initial adaptation to deep adaptation, significantly improving adaptability and response efficiency.

[0096] Adjusting the semantic processing hierarchy of the intermediate transition model and strengthening the expression of core semantic elements to obtain a dynamic matching semantic model achieves a precise fit between the model and the target task, endowing the model with dedicated semantic processing capabilities. This step, based on the overall model performance evaluation results, optimizes the priority of semantic processing hierarchy (placing the hierarchy corresponding to core semantics in a more critical position), while strengthening the expression of core semantic elements through weight adjustments and other methods. This allows the model to focus on the core requirements of the task, avoiding a fixed model structure that is out of touch with task needs, ultimately forming a dynamic matching semantic model highly adapted to the target task.

[0097] like Figure 7 As shown, based on the adjusted semantic hierarchy and task semantic feature data, as well as the initial model, a dynamic matching semantic model is obtained. This means that based on the task semantic feature data and the adjusted semantic hierarchy, the model structure is optimized. By fine-grained adjustment of the model's adaptability to the target task, a dynamic matching semantic model is obtained. Specifically, step 210 includes: Step 702: Based on the task semantic feature data and the adjusted semantic hierarchy, map the task features to the corresponding levels of the model and adjust the levels, merge levels with similar functions and optimize the data flow path to generate a semantic matching model. Step 704: Based on the semantic matching model, analyze the response efficiency of the semantic matching model to the semantics of the task, adjust the parts with insufficient response, strengthen the semantic parsing ability and adjust the internal connections to generate an intermediate transition model. Step 706: Based on the intermediate matching model, conduct an overall performance evaluation of the model, adjust the semantic processing level and strengthen the expression of key semantic elements according to the evaluation results, and generate a dynamic matching semantic model.

[0098] Step 702 achieves initial adaptation between task features and model layers, merging redundant functional layers, optimizing data flow paths, and generating a simplified semantic matching model that fits the task's semantics. Step 704 specifically addresses the model's semantic response shortcomings, strengthening semantic parsing capabilities, optimizing internal connections, improving the model's response efficiency to task semantics, and generating a more adaptable intermediate transition model. Step 706 optimizes the semantic processing layers through overall performance evaluation, strengthens the expression of key semantic elements, and finally generates a dynamic matching semantic model that accurately matches the target task. The intermediate matching model is the intermediate transition model.

[0099] In one embodiment of this application, optimizing the parameters of a dynamic matching semantic model to obtain a transferable model includes: identifying key parameters of the dynamic matching semantic model, setting a parameter adjustment range, and constructing a parameter adjustment model; optimizing the key parameters of the parameter adjustment model to obtain a parameter optimization model; and adjusting the learning rate and regularization coefficient of the parameter optimization model to obtain a transferable model.

[0100] In this embodiment, key parameters of the dynamic matching semantic model are identified, parameter adjustment ranges are set, and a parameter adjustment model is constructed. The core objects of parameter optimization are clarified (such as the number of layers, the number of nodes per layer, the learning rate, and the regularization coefficient). By reasonably defining the adjustment boundaries, the blindness of parameter optimization is avoided, providing a standardized framework for subsequent accurate optimization. At the same time, the range is set by combining the model's adaptability and performance to ensure that parameter adjustments do not deviate from the target task requirements, laying a stable foundation for optimization work.

[0101] Optimizing parameters involves adjusting key parameters of the model and obtaining an optimized model. This is achieved by using algorithms such as grid search or Bayesian optimization to select the optimal parameter combination that best suits the target task. This improves the model's core performance on the target task (such as semantic parsing accuracy and response efficiency) while preserving the model's ability to adapt to the task's semantics, thus solving the performance bottleneck problem caused by coarse parameter adjustment.

[0102] By adjusting the learning rate and regularization coefficient of the model and obtaining a transferable model, the convergence speed and overfitting risk of the model are precisely balanced. By dynamically adjusting the key regularization parameters, the model's adaptability to new domains and tasks is enhanced, and the model's cross-task dynamic transfer is realized, breaking the limitation of the model adapting to a single task.

[0103] Specifically, for this embodiment, such as Figure 8 As shown, the steps of adjusting the parameters of the dynamic matching semantic model to obtain a transferable model can be understood as optimizing the parameters of the dynamic matching semantic model, dynamically adjusting the learning rate and regularization parameters according to task requirements and model feedback, optimizing model performance, and performing dynamic transfer. Specifically, step 212 includes: Step 802: Perform preliminary parameter analysis on the semantic model of dynamic matching, identify key parameters that affect the performance of the model, determine the parameter adjustment range, and set the parameter adjustment range in combination with the model's adaptability and efficiency balance to obtain the parameter adjustment model. Step 804: Based on the parameter tuning model, conduct experiments on the model, collect performance data, and adjust the learning rate, regularization coefficient, and layer configuration according to the data feedback to generate a parameter-optimized model. Step 806: Based on the parameter optimization model, evaluate the model performance, reconfigure the learning rate and regularization coefficient, optimize the model performance, and perform dynamic transfer.

[0104] Step 802 involves performing preliminary parameter analysis on the dynamic matching semantic model to identify key parameters affecting model performance. Through detailed calculations and simulations, key parameters can be determined, including the number of layers, the number of nodes per layer, the learning rate, and the regularization coefficient.

[0105] Subsequently, based on a comprehensive consideration of the model's adaptability and performance balance, a reasonable range for parameter adjustment was set, and a parameter adjustment model was constructed, laying the foundation for subsequent optimization work.

[0106] Step 804 involves parameter tuning of the model. Based on experimental data, performance data of the model under different parameter settings is collected. Specifically, the impact of key parameters such as learning rate, regularization coefficient, and layer configuration on model performance is analyzed. Algorithms such as grid search or Bayesian optimization are used to meticulously optimize the key parameters, aiming to find the optimal parameter combination to improve model performance. This step generates a parameter-optimized model that, while maintaining its original adaptability, achieves a significant performance improvement through precise adjustment of key parameters.

[0107] Step 806: Based on parameter optimization, the learning rate and regularization coefficient were adjusted using a series of standardized test sets and evaluation results to find the optimal balance point, ensuring that the model converges quickly while avoiding overfitting. Ultimately, model performance was optimized while retaining its ability to dynamically adapt to different task requirements. Simultaneously, dynamic transfer of model performance was achieved, ensuring the model's stability and efficiency in changing environments.

[0108] In one embodiment of this application, the semantic category mapping is implemented through an entropy threshold iterative optimization algorithm; the entropy threshold iterative optimization algorithm includes: calculating interval entropy values ​​to construct an entropy value array, initializing a threshold T, marking intervals exceeding the threshold, iteratively adjusting the threshold T until a preset condition is met, and obtaining a semantic mapping table.

[0109] In this embodiment, semantic category mapping is achieved through an entropy threshold iterative optimization algorithm. The core process of this algorithm involves calculating the entropy value of each interval to construct an entropy array, initializing the threshold T, marking intervals exceeding the threshold, iteratively adjusting the threshold T until a preset condition is met, and finally obtaining a semantic mapping table to accurately lock the boundaries of semantic categories strongly related to the target task. The algorithm converts the degree of semantic confusion into a quantifiable indicator by quantifying the entropy value of each interval. Then, using the average value of the entropy array as the initial threshold, iterative adjustments are made to ensure that the threshold always conforms to the semantic distribution pattern of the feature data. Iteration stops when the number of marked intervals is lower than the preset threshold or no longer changes. The optimal threshold at this point can accurately distinguish different semantic categories. This process not only avoids the problem of semantic category division being too coarse or too fine, but also provides objective data support for the mapping relationship between key semantic units and semantic categories. The generated semantic mapping table has a high degree of structure and strong semantic correlation, providing a standardized and accurate basis for subsequent analysis of the correlation strength between key semantic units and adjustment of semantic classification, ensuring that the entire semantic annotation process revolves around the semantic needs of the target task from the very beginning.

[0110] In one embodiment of this application, the analysis of the correlation strength between key semantic units is achieved by the weighted entropy method; the weight coefficients of the weighted entropy method are calculated based on the co-occurrence frequency of key semantic units.

[0111] In this embodiment, the analysis of the correlation strength between key semantic units is achieved through a weighted entropy method. The weight coefficients of this method are calculated based on the co-occurrence frequency of key semantic units, enabling precise quantitative analysis of correlation strength and effectively addressing potential issues such as category redundancy and loose correlations in initial semantic classification. The weighted entropy method determines the weight coefficients based on the co-occurrence frequency of key semantic units; a higher co-occurrence frequency indicates a stronger semantic correlation between units, resulting in a larger weight proportion. This design makes the correlation strength analysis more aligned with the semantic logic of the target task. By defining an entropy adjustment function, the entropy value of key semantic units within each semantic category is compared with the optimal entropy threshold. Combined with the quantified correlation strength data, this accurately identifies combinations of semantically closely related units and redundant, loosely defined categories. This allows for the merging of similar categories, the splitting of broad categories, and the reconfiguration of the semantic classification system. The resulting adjusted semantic mapping table exhibits more logical and targeted semantic classification, avoiding interference from irrelevant nodes in subsequent deep annotation and providing a reasonable classification framework for generating annotation data rich in task-specific semantic information.

[0112] In one embodiment of this application, adjusting the semantic hierarchy includes: constructing an adjacency matrix and a degree matrix, constructing a semantic similarity matrix, balancing the adjacency matrix information and semantic similarity information through preset coefficients, and performing nonlinear transformation on the basic semantic node features.

[0113] In this embodiment, adjusting the semantic level includes constructing an adjacency matrix and a degree matrix, constructing a semantic similarity matrix, balancing the adjacency matrix information and semantic similarity information through preset coefficients, and performing nonlinear transformation on the basic semantic node features. This achieves precise adaptation of the semantic level, allowing the graph structure to retain complete semantic associations while also meeting the core requirements of the target task.

[0114] The adjacency matrix and degree matrix can capture the direct connections between basic semantic nodes, reflecting the topological features of the graph structure. The semantic similarity matrix focuses on the semantic associations at the content level of nodes, making up for the semantic fit problem that may be ignored when relying solely on structural information.

[0115] By balancing two types of information with preset coefficients α and β, and then transforming node features through a nonlinear activation function σ, the stability of structural connections between nodes is ensured, while the relevance of core semantics is strengthened. This effectively solves the defect of graph structures that only focus on connection relationships and ignore semantic essence. It can accurately identify semantic levels in graph structure transition models that do not match the requirements of the target task. Through reconstruction and optimization, the semantic levels are made more in line with the semantic logic of the task. The generated semantic network graph structure can clearly reveal the semantic connections between different network layers, providing an efficient and targeted semantic carrier for subsequent key node screening and aggregation operations.

[0116] In one embodiment of this application, the comprehensive influence score of the core semantic node comprehensively includes the degree, centrality, and entropy value of the core semantic node; the mutual information threshold is determined through iterative optimization.

[0117] In this embodiment, the comprehensive influence score of the core semantic node comprehensively includes the degree, centrality and entropy of the core semantic node, and the mutual information threshold is determined through iterative optimization to ensure the accuracy of key node screening and clustering operations, so that the aggregation model structure focuses on the core semantics of the target task.

[0118] Degree reflects the number of connections a node has, centrality reflects the key position of a node in the graph structure, and entropy measures the information content carried by a node. The comprehensive influence score calculated by combining these three factors quantifies the importance of a node in semantic flow from multiple dimensions, avoiding the problem of missing key nodes or misjudging secondary nodes due to screening with a single indicator.

[0119] The mutual information threshold is determined through iterative optimization, dynamically adapting to the distribution patterns of core semantic nodes. Clustering is only performed when the shared information between two nodes exceeds this threshold, effectively filtering out loosely semantically related node combinations. This solves the problems of lack of specificity and susceptibility to semantic bias. The generated aggregation model structure ensures the integrity of the core semantics while eliminating redundant interference, laying a solid foundation for subsequent evaluation of semantic representation efficiency and adjustment of the semantic hierarchy, ensuring that semantic hierarchy optimization always revolves around the core requirements of the target task.

[0120] In one embodiment of this application, key parameters include the number of layers, the number of nodes per layer, the learning rate, and the regularization coefficient; optimization of key parameters is achieved through grid search or Bayesian optimization algorithms.

[0121] In this embodiment, the key parameters explicitly include the number of layers, the number of nodes per layer, the learning rate, and the regularization coefficient. The optimization of the key parameters is achieved through grid search or Bayesian optimization algorithms, which realizes the precise optimization of the dynamic matching semantic model parameters and balances the model's performance on the target task with its cross-task transferability.

[0122] Since the number of layers and the number of nodes in each layer determine the structural capacity of the model, the learning rate affects the convergence speed of the model, and the regularization coefficient is related to the risk of overfitting, clarifying these key parameters and setting reasonable adjustment ranges avoids the blind optimization of parameters.

[0123] The grid search algorithm iterates through all parameter combinations, while the Bayesian optimization algorithm iterates through historical performance data to find the optimal solution. Both algorithms can efficiently select parameter combinations suitable for the target task. This ensures that the model achieves optimal core performance such as semantic parsing accuracy and response efficiency on the target task, and also enables the model to have good generalization ability by optimizing the learning rate and regularization coefficient.

[0124] This process not only solves the problem of lack of basis for adjusting the parameters of dynamic matching semantic models, but also enables the model to play a stable role in different domains and related tasks, realizes cross-task dynamic migration, broadens the scope of model application, and reduces the cost of model deployment and reuse.

[0125] Figure 9 This is a schematic diagram of the structure of a system provided in an embodiment of this application. For example... Figure 9 As shown, the semantic adaptation and knowledge transfer system 900 based on a large language model may include a data annotation module 902, a graph structure construction and optimization module 904, a semantic hierarchy matching optimization module 906, a task semantic feature extraction module 908, a dynamic matching model generation module 910, and a parameter optimization and transfer module 912. The data annotation module 902 is used to obtain feature data of the target task, perform semantic annotation on the feature data, and obtain semantic annotation data. The graph structure construction and optimization module 904 is used to create inter-layer semantic connections based on the semantic annotation data using graph construction technology, generating a semantic network graph structure. The semantic hierarchy matching optimization module 906 is used to filter key nodes and perform aggregation operations based on the semantic network graph structure, adjust the network hierarchy, and obtain the adjusted semantic hierarchy structure. The task semantic feature extraction module 908 is used to obtain the descriptive text of the target task, extract core semantic elements and task features as task semantic feature data. The dynamic matching model generation module 910 is used to optimize the model structure based on the adjusted semantic hierarchy structure and task semantic feature data, obtaining a dynamic matching semantic model. The parameter optimization and transfer module 912 is used to optimize the parameters of the dynamic matching semantic model, obtaining a transferable model.

[0126] The semantic adaptation and knowledge transfer system 900 based on a large language model provided in this application includes a data annotation module 902 that performs text parsing based on the feature data of the target task, identifies keywords and phrases, maps semantic categories, analyzes the strength of associations between words, adjusts semantic classifications, performs deep annotation on each semantic category, and generates semantic annotation data. A graph structure construction and optimization module 904, based on the semantic annotation data, creates nodes and edges, constructs a graph structure representing basic semantic connections between layers, analyzes the semantic flow between nodes, identifies key nodes and edges, adjusts the graph structure, and generates a semantic network graph structure. A semantic hierarchy matching and optimization module 906, based on the semantic network graph structure, identifies semantic levels that do not match the requirements of the target task, optimizes the model's semantic hierarchy, matches the current target task requirements, aggregates nodes according to the strength and importance of semantic connections between nodes, and generates a preliminary aggregated model structure.

[0127] The semantic adaptation and knowledge transfer system 900 based on a large language model in this application also includes a model structure aggregation and adjustment module 914. Based on the preliminary aggregation of the model structure, it evaluates the efficiency of the structure representation of the model semantics, adjusts the hierarchical relationship between nodes, and performs detailed analysis to adjust the aggregation degree of nodes, reconfigure the connection relationship between nodes, strengthen the semantic representation of key nodes, and generate an adjusted semantic hierarchical structure.

[0128] The task semantic feature extraction module 908 parses the task description text, identifies key semantic elements directly related to the task description, analyzes each semantic element, evaluates its relevance to task requirements, adjusts the matching degree between semantic elements and task requirements, determines the connection point between semantic elements and model capabilities, and generates task semantic feature data. The dynamic matching model generation module 910, based on the task semantic feature data and the adjusted semantic hierarchy, merges the hierarchy of similar functions and optimizes the data flow path, strengthens semantic parsing capabilities and adjusts internal connections, adjusts the semantic processing hierarchy and strengthens the expression of key semantic elements, and generates a dynamic matching semantic model. The parameter optimization and transfer module 912 performs preliminary parameter analysis on the dynamic matching semantic model, sets the preliminary range of parameter adjustment, conducts experiments on the model, optimizes key parameters, evaluates and optimizes model performance, and performs dynamic transfer.

[0129] Figure 10 This is a neural network model architecture diagram for a semantic adaptation and knowledge transfer system based on a large language model. The diagram illustrates the hierarchical structure and data flow logic of the system, consisting of a "topic embedding layer + neural network layer". The training document set D is the input source for the model's training data. The LDA topic model is the core model of the topic embedding layer, used to extract document topic information; the topic word distribution φ represents the topic and word probability distribution results output by the LDA topic model. Topic similarity γ... iThis calculates the similarity between the topic and word distribution and the target word. Target word w i This refers to the core word currently being processed. The context word set (w) i-c ~w i+c ): target word w i The model consists of: context neighbors; context embedding word vectors W: vector representations of context words after embedding; topic embedding word vector matrix Ψ: output matrix after fusing topic similarity and context embedding vectors; input layer: the input layer of the model, carrying the topic embedding word vector matrix; hidden layer: the intermediate computation layer of the model, responsible for feature transformation and transmission; output layer: the result output layer of the model, outputting the processed words and feature results; and the topic embedding layer is responsible for fusing topic information and word embedding information.

[0130] Figure 11 A schematic diagram of the hardware structure of the electronic device provided in an embodiment of this application is shown.

[0131] The electronic device may include a processor 1101 and a memory 1102 storing computer program instructions.

[0132] Specifically, the processor 1101 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0133] Memory 1102 may include mass storage for data or instructions. For example, and not limitingly, memory 1102 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 1102 may include removable or non-removable (or fixed) media, or memory 1102 may be non-volatile solid-state memory. Memory 1102 may be internal or external to the integrated gateway disaster recovery device.

[0134] In one instance, memory 1102 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0135] The processor 1101 reads and executes computer program instructions stored in memory 1102 to achieve... Figures 1 to 8 The semantic adaptation and knowledge transfer method based on a large language model is shown in the embodiment.

[0136] In one example, the electronic device may also include a communication interface 1103 and a bus 1104. For example, Figure 11 As shown, the processor 1101, memory 1102, and communication interface 1103 are connected through bus 1104 and complete communication with each other.

[0137] The communication interface 1103 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0138] Bus 1104 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not as a limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 1104 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0139] This electronic device can execute the semantic adaptation and knowledge transfer method based on a large language model as described in this embodiment of the invention, thereby achieving a combination of... Figures 1 to 8 The paper describes a semantic adaptation and knowledge transfer method based on a large language model.

[0140] Furthermore, in conjunction with the semantic adaptation and knowledge transfer methods based on large language models in the above embodiments, this invention can be implemented using a computer storage medium. This computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the semantic adaptation and knowledge transfer methods based on large language models in the above embodiments.

[0141] This application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the semantic adaptation and knowledge transfer methods based on a large language model described in the above embodiments.

[0142] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0143] The functional blocks shown in the above-described structural diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0144] It should also be noted that the exemplary embodiments mentioned in this invention describe methods or systems based on a series of steps or apparatus. However, this invention is not limited to the order of the steps described above; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0145] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0146] The above description is merely a specific embodiment of the present invention. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the protection scope of the present invention.

Claims

1. A semantic adaptation and knowledge transfer method based on a large language model, characterized in that, include: Obtain feature data of the target task, perform semantic annotation on the feature data, and obtain semantic annotation data; Based on the semantically labeled data, graph construction techniques are used to create semantic connections between layers and generate a semantic network graph structure. Based on the semantic network graph structure, key nodes are selected and aggregated to adjust the network hierarchy and form an adjusted semantic hierarchy structure. Obtain the description text of the target task, and extract the core semantic elements and task features as task semantic feature data; Based on the adjusted semantic hierarchy and the task semantic feature data, the model structure of the initial model is adjusted to obtain a dynamic matching semantic model; Optimize the parameters of the dynamic matching semantic model to obtain a transferable model.

2. The method according to claim 1, characterized in that, The step of obtaining feature data for the target task and semantically annotating the feature data includes: Based on the feature data of the target task, identify key semantic units, map the semantic categories of the key semantic units, and determine a semantic mapping table; Analyze the correlation strength between key semantic units in the semantic mapping table, adjust the semantic classification, and obtain the adjusted semantic mapping table; Each semantic category in the adjusted semantic mapping table is labeled to obtain the semantic labeling data.

3. The method according to claim 1, characterized in that, The step of creating inter-layer semantic connections and generating a semantic network graph structure based on the semantically labeled data using graph construction techniques includes: Based on the semantic annotation data, basic semantic nodes and semantic association edges are defined to construct a basic graph structure model, wherein the basic semantic nodes represent different concepts and their attributes, and the semantic association edges reflect the semantic association between the different concepts and the attributes; Analyze the semantic flow between basic semantic nodes in the basic graph structure model, identify key nodes and key semantic association edges of the graph structure, and generate a graph structure transition model; Adjust the semantic hierarchy of the graph structure transition model to obtain the semantic network graph structure.

4. The method according to claim 1, characterized in that, The step of filtering key nodes and performing aggregation operations based on the semantic network graph structure to adjust the network hierarchy and form an adjusted semantic hierarchy structure includes: Identify the core semantic nodes in the semantic network graph structure and calculate the comprehensive influence score of the core semantic nodes; Based on the comprehensive influence score, the semantic connection strength and mutual information threshold between the core semantic nodes, the core semantic nodes are clustered to obtain the aggregation model structure. The semantic representation efficiency of the aggregation model structure is evaluated, and the connection relationship and aggregation degree between the core semantic nodes are adjusted based on the comprehensive influence score of the core semantic nodes to obtain the adjusted semantic hierarchy structure.

5. The method according to claim 1, characterized in that, The process of obtaining the description text of the target task and extracting core semantic elements and task features as task semantic feature data includes: Parse the description text of the target task, capture the core semantic elements and classify and map them to predefined semantic categories to obtain the basic element table; Evaluate the relevance of the core semantic elements in the basic element table to the target task requirements, adjust the matching degree, and obtain a semantic element matching table; Determine the connection points between the core semantic elements in the semantic element matching table and the model capabilities, and obtain the task semantic feature data.

6. The method according to claim 1, characterized in that, The step of adjusting the model structure of the initial model based on the adjusted semantic hierarchy and the task semantic feature data to obtain a dynamic matching semantic model includes: Map the semantic features of the task to the adjusted semantic hierarchy, optimize the data flow path, and obtain a semantic matching model; Adjust the internal connections of the semantic matching model to obtain an intermediate transition model; Adjust the semantic processing hierarchy of the intermediate transition model, strengthen the expression of core semantic elements, and obtain the dynamic matching semantic model.

7. The method according to claim 1, characterized in that, Optimizing the parameters of the dynamic matching semantic model to obtain a transferable model includes: Identify the key parameters of the dynamic matching semantic model, set the parameter adjustment range, and construct a parameter adjustment model; Optimize the key parameters of the parameter adjustment model to obtain the parameter optimization model; Adjust the learning rate and regularization coefficient of the optimized model to obtain the transferable model.

8. The method according to claim 2, characterized in that, The semantic categories are mapped using an entropy threshold iterative optimization algorithm. The entropy threshold iterative optimization algorithm includes: calculating interval entropy values ​​to construct an entropy value array, initializing a threshold T, marking intervals exceeding the threshold, iteratively adjusting the threshold T until a preset condition is met, and obtaining a semantic mapping table.

9. The method according to claim 2, characterized in that, The strength of the association between the key semantic units is analyzed using the weighted entropy method. The weighting coefficients of the weighted entropy method are calculated based on the co-occurrence frequency of the key semantic units.

10. The method according to claim 3, characterized in that, Adjusting the semantic hierarchy includes: constructing an adjacency matrix and a degree matrix, constructing a semantic similarity matrix, balancing the adjacency matrix information and the semantic similarity information through preset coefficients, and performing a nonlinear transformation on the basic semantic node features.

11. The method according to claim 4, characterized in that, The comprehensive influence score of the core semantic node comprehensively includes the degree, centrality, and entropy value of the core semantic node; The mutual information threshold is determined through iterative optimization.

12. The method according to claim 7, characterized in that, The key parameters include the number of layers, the number of nodes per layer, the learning rate, and the regularization coefficient. The key parameters are optimized using grid search or Bayesian optimization algorithms.

13. A semantic adaptation and knowledge transfer system based on a large language model, characterized in that, include: The data annotation module is used to acquire feature data of the target task, perform semantic annotation on the feature data, and acquire semantic annotation data. The graph structure construction and optimization module is used to create inter-layer semantic connections based on the semantic annotation data and generate a semantic network graph structure using graph construction technology. The semantic hierarchy matching optimization module is used to filter key nodes and perform aggregation operations based on the semantic network graph structure, adjust the network hierarchy, and obtain the adjusted semantic hierarchy structure. The task semantic feature extraction module is used to obtain the description text of the target task, extract core semantic elements and task features, and use them as task semantic feature data. The dynamic matching model generation module is used to adjust the model structure of the initial model based on the adjusted semantic hierarchy and the task semantic feature data to obtain a dynamic matching semantic model. The parameter optimization and transfer module is used to optimize the parameters of the dynamic matching semantic model and obtain a transferable model.

14. An electronic device, characterized in that, The device includes: a processor and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the semantic adaptation and knowledge transfer method based on a large language model as described in any one of claims 1-12.

15. A computer-readable storage medium, characterized in that, The computer storage medium stores computer program instructions, which, when executed by a processor, implement the semantic adaptation and knowledge transfer method based on a large language model as described in any one of claims 1-12.

16. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the semantic adaptation and knowledge transfer method based on a large language model as described in any one of claims 1-12.

Images