An industry economic state analysis method based on a dynamic objective weighting coefficient mechanism and multi-modal semantic alignment

Abstract

The application discloses an industry economic state analysis method based on a dynamic objective weighting coefficient mechanism and multi-modal semantic alignment, and relates to the technical field of data processing, wherein the method comprises the following steps: obtaining structured economic index data and unstructured industry knowledge data of a target industry, wherein the unstructured industry knowledge data comprises policy texts and industry chain graph information; performing cross-modal alignment processing on the structured economic index data and the unstructured industry knowledge data by using a semantic graph neural network, and establishing a dynamic semantic correlation between policy entities and industry chain nodes; automatically adjusting the contribution weights of the structured data and the unstructured knowledge in the analysis process according to the industry scene characteristics by using a dynamic knowledge weight distribution mechanism (combining the difference analysis capability of the Theil index method and the information measurement characteristics of the entropy weight index method, a dynamic objective weighting optimization model is constructed); and generating an industry economic state analysis result containing structured indexes and unstructured knowledge reasoning based on the dynamic semantic correlation and the contribution weights.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment. Background Technology

[0002] Industry economic status analysis, as a crucial decision-making support tool in the digital economy era, is widely applied in macroeconomic regulation, corporate strategy formulation, and investment risk assessment. With the breakthrough development of artificial intelligence technology, existing technologies, through the collaborative operation of pre-trained large language models, knowledge graphs, and workflow engines, have constructed a multi-source economic data fusion and analysis system. Specifically, this technical system covers the entire process from data collection and feature engineering to predictive modeling, including key aspects such as structured economic indicator processing, unstructured text parsing, and industrial chain relationship modeling.

[0003] However, existing economic analysis methods directly adopt modular pipeline designs without establishing dynamic feedback mechanisms, which may lead to task execution paths that cannot adapt to real-time data changes. Specifically, existing technologies typically use static workflow scheduling, which cannot effectively reflect the impact of information shocks. When sudden policies cause economic or market fluctuations, model strategies need to be manually reconfigured, and this involves a large amount of subjective experience settings. At the same time, although traditional knowledge bases construct the relationship between structured economic indicators and industry chain maps, they do not solve the problem of deep semantic parsing of unstructured policy texts, resulting in insufficient interpretability for decision support. In addition, in the key knowledge fusion and weight allocation stages, existing methods mostly rely on subjective experience settings or simple percentage superposition. Such approaches fail to deeply quantify the sample differences and information contributions within different data sources, resulting in a significant drop in accuracy of general analysis frameworks in complex scenarios such as manufacturing. Because their feature engineering does not embed industry-specific indicators, it further highlights the industry pain point of weak domain specificity. Summary of the Invention

[0004] The main objective of this invention is to provide a method for analyzing the economic status of an industry based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment.

[0005] Another objective of this invention is to propose an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment.

[0006] The third objective of this invention is to provide a computer device.

[0007] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0008] To achieve the above objectives, a first aspect of the present invention proposes a method for analyzing industry economic status based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, comprising: S1, acquire structured economic indicator data and unstructured industry knowledge data of the target industry, wherein the unstructured industry knowledge data includes policy texts and industry chain map information; S2, using a semantic graph neural network to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data, and establishing dynamic semantic associations between policy entities and industrial chain nodes; S3, based on industry scenario characteristics, uses a dynamic knowledge weight allocation mechanism, combined with the difference analysis capability of the Theil index method and the information measurement characteristics of the entropy weight index method, to construct a dynamic objective weighting optimization model, which automatically calculates and adjusts the contribution weights of structured data and unstructured knowledge in the analysis process. S4. Based on the dynamic semantic relationship, and using an objective weighting mechanism that integrates the Theil index method and the entropy weight index method, the contribution weight is dynamically calculated to generate an industry economic status analysis result that includes structured indicators and unstructured knowledge inferences.

[0009] In one embodiment of the present invention, the step of acquiring structured economic indicator data and unstructured industry knowledge data of the target industry further includes: S11 uses entity boundary recognition theory to clean the original economic text and extract standardized representations of industry entities. S12 extracts multimodal features from unstructured industry knowledge data, where policy texts are represented by TF-IDF weighted word vectors and industry chain graph information is encoded using adjacency matrix.

[0010] In one embodiment of the present invention, the step of using a semantic graph neural network to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data to establish dynamic semantic associations between policy entities and industrial chain nodes further includes: S21, use a large language model to generate semantic embedding vectors for the policy text, and use a graph neural network to calculate the semantic similarity between policy entities and industry chain nodes, wherein the semantic similarity meets the usual industry standards. S22 optimizes cross-modal alignment results based on information bottleneck theory, filtering redundant semantic associations to improve analysis efficiency.

[0011] In one embodiment of the present invention, the step of constructing a dynamic objective weighting optimization model based on industry scenario characteristics, through a dynamic knowledge weight allocation mechanism, combining the difference analysis capability of the Theil index method with the information content measurement characteristics of the entropy weight index method, to automatically calculate and adjust the contribution weights of structured data and unstructured knowledge in the analysis process, further includes: S31, using formula The inter-group difference index is calculated using the entropy weight index method and the Theil index method, and the information content is measured by combining the entropy weight index method and the Theil index method. The historical accuracy weight of the model in a specific industry is calculated through a dynamic objective weighting model. S32, perform secondary calibration of the weight allocation results by combining industry characteristic parameters.

[0012] In one embodiment of the present invention, based on the dynamic semantic association relationship, and using an objective weighting mechanism that integrates the Theil index method and the entropy weight index method to dynamically calculate the contribution weight, an industry economic status analysis result containing structured indicators and unstructured knowledge inferences is generated, which further includes: S41, the confidence coefficient threshold for the prediction results is calculated using the entropy weight index method and the Theil index method, whereby... ; S42 uses the nearest neighbor principle of case analogy reasoning to fill in the prompt word template and generate strategy suggestions that include historical event anchors and risk constraint clauses.

[0013] In one embodiment of the present invention, it further includes: S5 constructs a utility function based on data freshness, resource utilization, and task priority, and dynamically adjusts the workflow execution path. The task urgency metric algorithm satisfies... Based on the utility function, the workflow branch prediction engine is triggered to select the optimal task execution path.

[0014] To achieve the above objectives, a second aspect of the present invention proposes an industry economic status analysis device based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, comprising: The structured and unstructured data acquisition module is used to acquire structured economic indicator data and unstructured industry knowledge data of the target industry. The unstructured industry knowledge data includes policy texts and industry chain map information. The cross-modal semantic alignment processing module is used to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data using a semantic graph neural network, and to establish dynamic semantic associations between policy entities and industrial chain nodes. The dynamic knowledge weight allocation module is used to construct a dynamic objective weighting optimization model based on industry scenario characteristics, through a dynamic knowledge weight allocation mechanism, combining the difference analysis capability of the Theil index method and the information content measurement characteristics of the entropy weight index method, and automatically calculate and adjust the contribution weight of structured data and unstructured knowledge in the analysis process. The economic status analysis result generation module is used to dynamically calculate contribution weights based on the dynamic semantic relationship and using an objective weighting mechanism that integrates the Theil index method and the entropy weight index method, and generate industry economic status analysis results containing structured indicators and unstructured knowledge inferences.

[0015] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment as described in the first aspect embodiment.

[0016] To achieve the above objectives, the fourth aspect of this application proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment as described in the first aspect embodiment.

[0017] The embodiments of the present invention have the following beneficial effects: The methods, intelligent agents, devices, electronic devices, and computer-readable storage media of the present invention enable deep integration and dynamic correlation of structured economic data and unstructured industry knowledge, thereby improving the decision interpretability and response timeliness of industry economic status analysis. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, provided for embodiments of the present invention; Figure 2 A schematic diagram of the original text processing for an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment provided in an embodiment of the present invention; Figure 3 A schematic diagram of a decision-making diversion mechanism for an industry economic status analysis method based on dynamic objective weighting coefficients and multimodal semantic alignment, provided in an embodiment of the present invention; Figure 4 A workflow execution flowchart for an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, provided in this embodiment of the invention; Figure 5 This is a structural diagram of an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, provided in an embodiment of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] The following description, with reference to the accompanying drawings, describes an industry economic status analysis method, intelligent agent, and device based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, according to embodiments of the present invention.

[0022] Example 1 This embodiment provides a method for analyzing industry economic status based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment. For example... Figure 1 As shown, the method includes the following steps: S1. Obtain structured economic indicator data and unstructured industry knowledge data for the target industry. The unstructured industry knowledge data includes policy texts and industry chain map information.

[0023] Specifically, this step aims to acquire structured economic indicator data and unstructured industry knowledge data for the target industry. The unstructured data includes policy texts and industry chain mapping information. This step is the fundamental input for the entire industry's intelligent economic status analysis system. Its technical implementation involves the collection, cleaning, structuring, and semantic modeling of multi-source heterogeneous data, providing high-quality data support for subsequent large-scale language model rating, multi-model decision-making, and knowledge enhancement strategy generation.

[0024] In this embodiment of the invention, structured economic indicator data typically originates from authoritative sources such as the National Bureau of Statistics, industry associations, and listed company financial reports, covering key indicators such as GDP growth rate, capacity utilization rate, import and export value, and inventory turnover rate. This data must meet the ISO 8601 time format specification and undergo standardization processing through an ETL (Extract-Transform-Load) process to ensure field consistency and data integrity. Unstructured industry knowledge data is extracted from policy documents, research reports, and industry chain analysis documents using web crawlers, API interfaces, or OCR technology. Preliminary analysis is performed using feature engineering methods based on entity boundary recognition (such as the BiLSTM-CRF model) to identify key information such as policy themes, industry entities, and timestamps.

[0025] Furthermore, the semantic parsing of policy texts must meet (typically industry standards) the following requirements: TF-IDF keyword extraction accuracy ≥85% and entity recognition F1 score ≥90%. The extraction of industry chain graph information relies on graph neural networks (GNNs) for node relationship modeling, requiring a node embedding dimension of 256. Edge weights are calculated using cosine similarity, with a similarity threshold set at ≥0.82 to ensure semantic alignment accuracy between unstructured knowledge and structured indicators.

[0026] Specifically, this step is widely used in scenarios such as macroeconomic monitoring, industry risk early warning, and investment decision support. For example, in semiconductor industry analysis, the system can capture export control policy texts from various countries in real time and establish semantic relationships with nodes such as wafers and chips in the industry chain map, providing a semantic basis for subsequent policy impact assessments.

[0027] The technical effect of this step is that, through the collaborative collection and processing of structured and unstructured data, a basic dataset for multimodal knowledge fusion is constructed, which solves the problem of insufficient unstructured knowledge fusion in existing technologies and significantly improves the interpretability and decision support capabilities of industry economic status analysis.

[0028] Furthermore, S1 includes: S11 uses entity boundary recognition theory to clean the original economic text and extract standardized representations of industry entities.

[0029] Specifically, the step of "cleaning the original economic text and extracting standardized representations of industry entities by means of entity boundary recognition theory based on Lample et al., 2016" is the core link of the input preprocessing module in the entire intelligent analysis system of industry economic status. Its technical implementation is based on the sequence labeling model in deep learning, which aims to transform unstructured economic text into structured and semantically clear industry entity information, providing high-quality input for the subsequent rating module and multi-model decision module.

[0030] The technical principle of this step is based on the combination of a bidirectional LSTM-CRF model and a CRF (Conditional Random Field) decoding mechanism. Its core lies in the accurate identification of entity boundaries in economic texts. Specifically, the implementation is as follows: First, the original text is segmented and part-of-speech labeled, then input into an entity recognition model based on Lample et al., 2016. Second, the model captures contextual semantic information through a bidirectional LSTM, and combines this with a CRF layer to perform globally optimal decoding of entity labels, outputting labeled sequences such as "B-INDUSTRY" and "I-INDUSTRY," thereby identifying complete industry entity boundaries. In terms of parameter settings, the model uses a hidden layer dimension of 256, the CRF layer uses the Viterbi algorithm for decoding, and the entity labeling system follows the BIO (Begin, Inside, Outside) labeling standard, supporting the recognition of multiple types of entities, such as policy events, industry names, and economic indicators.

[0031] In this embodiment of the invention, this step can process policy-related texts such as "a certain country plans to impose additional tariffs on semiconductors," extract "semiconductor" as an industry entity through an entity recognition model, and standardize it into a unified industry classification code (such as "C39 Computer, Communication and Other Electronic Equipment Manufacturing" in the GB / T 4754-2017 National Economic Industry Classification Standard). This process serves as a key step in data cleaning and structuring within the system, ensuring that subsequent modules can perform semantic modeling and decision analysis based on accurate industry entities.

[0032] The technical advantage of this step lies in significantly improving the parsability and semantic consistency of economic texts. Experimental data shows that the entity recognition module achieves an F1 score exceeding 85% in industries such as semiconductors, new energy, and finance, and supports multilingual input (such as mixed Chinese and English text). Through standardized entity extraction, the system can effectively integrate unstructured policy texts with structured economic indicators, providing high-quality input for the multimodal knowledge fusion engine, thereby enhancing the comprehensiveness and interpretability of industry economic status analysis.

[0033] S12 extracts multimodal features from unstructured industry knowledge data, where policy texts are represented by TF-IDF weighted word vectors and industry chain graph information is encoded using adjacency matrix.

[0034] Specifically, in the technical solution of this invention, multimodal feature extraction of unstructured industry knowledge data is one of the key steps in realizing intelligent analysis of industry economic status. This step transforms data of different modalities (such as policy texts and industry chain maps) into a unified mathematical representation, providing structured input for subsequent model decision-making and knowledge reasoning.

[0035] In this embodiment of the invention, the feature extraction of policy texts employs the TF-IDF weighted word vector representation method. This method quantifies the importance of words by calculating the product of word frequency and inverse document frequency, thereby constructing a vector representation that reflects the semantic features of the policy. Specifically, for a given set of policy texts... Each word In the document The TF-IDF weights in the equation can be expressed as:

[0036] in, Indicator In the document Frequency of occurrence in Total number of documents For containing words The number of documents. Using this method, keywords in policy texts (such as "tariff" and "review cycle") are given higher weights, thus highlighting their impact on the industry's economic state in semantic modeling.

[0037] For industry chain graph information, this invention employs an adjacency matrix encoding method to transform the node and edge relationships in the graph into matrix form, facilitating subsequent processing by a graph neural network (GNN). Let the industry chain graph contain... If there are n nodes, then its adjacency matrix is... elements in Represents a node With nodes The connection strength between nodes is typically represented by binary or real values. This encoding method follows the storage specifications of graph structures, ensuring the integrity and computability of graph information during the model input stage.

[0038] Specifically, this step primarily processes unstructured information such as policy changes and industry events, integrating it with structured economic indicators (such as capacity utilization and inventory levels) to enhance the interpretability and decision support capabilities of the analysis system. For example, in semiconductor industry analysis, the policy text "A certain country plans to impose additional tariffs on semiconductors" will be transformed into a TF-IDF vector, while nodes such as "wafers" and "chips" in the industry chain map will represent their upstream and downstream relationships through an adjacency matrix.

[0039] The technical advantage of this step lies in its ability to simultaneously capture the semantic information of policy texts and the topological features of the industrial chain structure through multimodal feature extraction, providing high-quality input for subsequent multi-model decision-making and knowledge enhancement. In practical tests, this method has improved the model's response speed and accuracy to unstructured information in policy-sensitive industries (such as semiconductors and new energy), laying a solid foundation for dynamic workflow scheduling and domain-adaptive analysis.

[0040] S2, using a semantic graph neural network to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data, and establish dynamic semantic associations between policy entities and industrial chain nodes.

[0041] Specifically, in some implementations, semantic graph neural networks are used to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data. The establishment of dynamic semantic relationships between policy entities and industry chain nodes is achieved based on multimodal information fusion and deep semantic modeling techniques using graph neural networks (GNNs). This step aims to address the problems of insufficient fusion of structured and unstructured data, lagging knowledge graph updates, and shallow semantic relationships between policies and industry nodes in existing technologies, thereby improving the interpretability and real-time response capabilities of industry economic status analysis.

[0042] In this embodiment of the invention, this step employs a Semantic Graph Neural Network (SGNN) as the core model to perform cross-modal alignment between structured economic indicators (such as GDP growth rate, capacity utilization rate, and inventory turnover rate) and unstructured industry knowledge (such as policy texts, research reports, and news announcements). Specifically, structured data is first mapped to node attributes in a graph structure, while unstructured text is extracted using pre-trained language models such as Qwen3 to extract semantic embedding vectors. Subsequently, the SGNN performs semantic alignment between the text embeddings and graph node features through graph convolution operations, calculates the semantic similarity between nodes, and establishes dynamic edge weights. The edge weights can be calculated based on cosine similarity, with a threshold set. To filter out significant semantic associations.

[0043] Furthermore, the key parameters involved in this step include: semantic embedding dimension (e.g., 768 dimensions for Qwen3 output), graph node feature dimension (e.g., 128 dimensions for normalized economic indicators), edge weight update frequency (e.g., updated every 24 hours), and semantic similarity threshold (e.g., cosine similarity). Furthermore, the model weight allocation mechanism employs a dynamic weight formula:

[0044] in Representation Model In the industry The weights in The model's historical accuracy in this industry, This represents the total number of models. This formula ensures that the contribution of different models to different industries can be dynamically adjusted during cross-modal alignment, improving the adaptability and accuracy of the overall analysis.

[0045] Specifically, this step is widely applicable to scenarios such as financial investment consulting, supply chain risk warning, and policy impact assessment. For example, when semiconductor industry policies change, the system can automatically identify policy entities such as "export licenses" and "review cycles," and establish semantic associations between them and supply chain nodes (such as wafers, chips, packaging and testing) through SGNN, thereby generating analysis results with causal explanations, such as "export license review cycle extended → wafer supply delay → chip price fluctuations."

[0046] This step achieves deep integration of structured and unstructured data through a semantic graph neural network, significantly improving the accuracy of semantic association and dynamic response capability between policy entities and industry nodes. Experiments show that in semiconductor industry testing, this step improves semantic alignment accuracy by 22% compared to traditional methods, and automatically updates the knowledge graph within 24 hours of policy release, meeting real-time analysis needs.

[0047] Furthermore, S2 includes: S21, use a large language model to generate semantic embedding vectors for the policy text, and calculate the semantic similarity between policy entities and industry chain nodes using a graph neural network, wherein the semantic similarity satisfies .

[0048] Specifically, in some implementations, the steps of this invention generate semantic embedding vectors for policy texts using a large language model and combine this with a graph neural network (GNN) to calculate the semantic similarity between policy entities and industry chain nodes, thereby achieving structured identification and quantitative assessment of policy impact. This step, based on a fusion mechanism of natural language processing (NLP) and graph structure modeling, is a key link in achieving multimodal knowledge fusion throughout the system.

[0049] It should be noted that the large language model should select a currently mainstream pre-trained language model and apply the Transformer architecture within the large language model to perform context-aware semantic encoding on the input text. Specifically, the policy text first undergoes word segmentation and standardization before being input into the large language model, outputting a 768-dimensional (or less) semantic embedding vector for each word or entity. Further, a module for entity recognition extracts key entities from the policy text (such as "tariff," "wafer," and "chip"), and their corresponding embedding vectors are used as input node features for the graph neural network.

[0050] In this embodiment of the invention, a graph neural network (GNN) is used to model the semantic relationships between industry chain nodes and policy entities. Each industry chain node (e.g., "wafer manufacturing," "chip packaging") is also represented as a 768-dimensional (or less) vector, which can be derived from pre-trained industry knowledge graph embeddings or extracted from a structured database through entity linking technology. The GNN calculates the semantic similarity between policy entities and industry chain nodes by aggregating the semantic information of neighboring nodes, and uses a cosine similarity formula for measurement, with a threshold set at [value missing]. (Generally selected according to industry standards or actual business needs) to ensure that only highly semantically matching node relationships are retained.

[0051] Furthermore, the embedding dimension of the large language model is 768 (no less), consistent with the node feature dimension of the graph neural network, to ensure the feasibility of cross-modal alignment. Cosine similarity threshold. It is based on the matching experiment settings of historical policy events and industrial chain responses (generally selected according to industry standards or actual business needs) to ensure that the similarity calculation results have sufficient semantic relevance and avoid noise interference.

[0052] Specifically, this step is widely used in scenarios such as policy impact analysis, supply chain risk warning, and industry trend forecasting. For example, when semiconductor industry policies change, the system can automatically identify policy entities such as "export licenses" and "review cycles," and use GNN to calculate their semantic similarity with supply chain nodes (such as "wafers" and "chips"), thereby determining the potential impact of the policies on each link of the supply chain.

[0053] The technical effect of this step lies in achieving a deep integration of unstructured policy texts and structured industry chain knowledge through the joint application of a large language model and a Generative Neural Network (GNN), providing semantic-level decision-making basis for subsequent dynamic workflow scheduling and strategy generation. Simultaneously, by setting... The strict threshold ensures the usability of semantic matching performance in specific application industries (such as the economic field), and improves the system's analytical depth and interpretability in complex economic scenarios.

[0054] S22 optimizes cross-modal alignment results based on information bottleneck theory, filtering redundant semantic associations to improve analysis efficiency.

[0055] Specifically, this step optimizes the cross-modal alignment results based on the information bottleneck theory, aiming to filter redundant semantic associations and improve the efficiency and accuracy of industry economic status analysis. The information bottleneck theory is an optimization method for feature compression and information preservation within the framework of information theory. Its core idea is to minimize the redundancy of input information while maximizing its relevance to the target variable. In this invention, this theory is applied to the semantic alignment process between multimodal data (such as structured economic indicators and unstructured policy texts) to remove irrelevant or inefficient information associations, thereby improving the model's inference efficiency and decision-making quality.

[0056] In this embodiment of the invention, the system first uses a semantic graph neural network to perform cross-modal alignment between key entities in the policy text (such as "tariff increase" and "export license") and nodes in the industry chain graph (such as "wafer" and "chip"). During the alignment process, the system calculates the semantic similarity between entities, typically using cosine similarity as the metric, and sets a similarity threshold. The system filters out highly relevant associations (generally selected according to industry standards or business requirements). For semantic edges below a certain threshold, the system treats them as redundant information and filters them, thereby reducing unnecessary computational overhead and semantic noise.

[0057] Furthermore, the system introduces an information bottleneck objective function, and the optimization objective of the alignment result can be expressed as:

[0058] in, This represents the input multimodal data (such as policy texts and economic indicators). For intermediate semantic representation, The target variable is (e.g., industry risk level). Indicates mutual information, This is a hyperparameter for controlling the balance between information compression and retention. Through this objective function, the system retains key semantic information while compressing redundant features unrelated to the objective, thereby improving the response speed and decision accuracy of subsequent analysis modules.

[0059] This step has significant value in practical applications, particularly in scenarios such as financial investment consulting and supply chain risk warning. For example, in analyzing policy changes in the semiconductor industry, the system can quickly identify the strong correlation between "extended export license review period" and "wafer supply chain," while ignoring weakly correlated information such as "chip packaging technology," which is irrelevant to the current task. By reducing redundant semantic paths, the system improves reasoning efficiency by approximately 30% when handling complex economic events, while enhancing the interpretability of the decision chain, thus meeting financial audit and compliance requirements.

[0060] S3, based on industry scenario characteristics, constructs a dynamic objective weighting optimization model by combining the difference analysis capability of the Theil index method with the information content measurement characteristics of the entropy weight index method through a dynamic knowledge weight allocation mechanism. This model automatically calculates and adjusts the contribution weights of structured data and unstructured knowledge in the analysis process.

[0061] Specifically, the steps of this invention are the core links in realizing knowledge fusion and decision optimization in an intelligent analysis system for industry economic status. Based on multimodal information fusion theory and dynamic decision optimization theory, this step achieves dynamic weighted fusion of structured economic indicators and unstructured industry knowledge (such as policy texts and industry chain maps) by introducing an industry-model fit evaluation system and a confidence-weighted heterogeneous model output aggregation method.

[0062] In this embodiment of the invention, this step first relies on the model's historical accuracy index in a specific industry i. Where m represents the model index (e.g., QWEN, DeepSeek, Llama, etc.). The dynamic weight formula is:

[0063] When calculating the contribution weights of each model in different industries, this embodiment of the invention uses the entropy weight index method and the Theil index method to measure the inter-group difference index respectively. To combine the complementary advantages of these two methods in measuring information content and difference distribution, the normalized entropy weight index and the Theil index are generally averaged to construct a fused difference index. The entropy weight index objectively reflects the degree of indicator variation based on information entropy, while the Theil index focuses on capturing inter-group imbalance. Both are calculated based on the characteristics of the data itself, ensuring the objectivity of weight quantification. During indicator normalization, this embodiment of the invention further calibrates the weights based on the actual contributions of each model in different industries to ensure that the weight allocation truly reflects its discriminative power and explanatory power. This fusion method enables the model weights to dynamically match industry characteristics, thereby improving the accuracy of prediction and analysis in multi-model collaborative decision-making. To adapt to the dynamic characteristics of model performance changing over time, the system subsequently introduces a weight update mechanism:

[0064] in This is the historical weight decay factor (generally selected according to industry standards or business characteristics). This indicates the model's performance on the most recent task. This mechanism ensures that the model weights can be dynamically adjusted with the input of new data, avoiding decision bias caused by model performance drift.

[0065] Specifically, this step can be applied to specific economic sectors such as financial investment consulting, supply chain risk warning, and policy impact assessment. For example, in semiconductor industry analysis, the system dynamically allocates the weights of policy sensitivity models (such as QWEN) and supply chain impact models (such as DeepSeek) in the current task based on their historical performance. This includes considering the semantic similarity between the current policy text and the supply chain map. In this case, the weight of unstructured knowledge will be significantly increased to enhance its explanatory power regarding policy impact.

[0066] The technical value of this step lies in its ability to more accurately capture industry characteristics and risk transmission paths by scientifically and dynamically adjusting the contribution ratio of structured and unstructured knowledge, thereby improving the interpretability and practicality of the analysis results. In actual testing, this mechanism achieved an F1 score of 88% for manufacturing risk prediction, significantly outperforming traditional static fusion methods.

[0067] To further explain, S3 includes: S31, using formula The inter-group difference index is calculated using the entropy weight index method and the Theil index method, and the information content is measured by combining the entropy weight index method and the Theil index method. The historical accuracy weight of the model in a specific industry is calculated through a dynamic objective weighting model.

[0068] Specifically, in the multi-model decision module of this invention, the step "using formula" The calculation of historical accuracy weights for specific industries is a core step in achieving collaborative decision-making using heterogeneous large language models. This step involves calculating inter-group difference indices using both the entropy weight index and the Theil index. To integrate the complementary advantages of these two methods in measuring information content and difference distribution, the normalized entropy weight index and the Theil index are typically averaged to construct a fused difference index. The entropy weight index objectively reflects the degree of indicator variation based on information entropy, while the Theil index focuses on capturing inter-group imbalance. Both are calculated based on the inherent characteristics of the data, ensuring the objectivity of weight quantification. During indicator normalization, this embodiment further calibrates the weights based on the actual contributions of each model in different industries, ensuring that the weight allocation truly reflects its discriminative power and explanatory power. Furthermore, by dynamically adjusting the weights of each model in different industries, a more accurate analysis of industry economic conditions is achieved.

[0069] In this embodiment of the invention, the first step is to extract each model from historical training data. In specific industries Accuracy index This accuracy is typically based on evaluation metrics (such as F1 score, RMSE, etc.) for the model's classification or regression tasks within that industry. Subsequently, a fusion indicator (entropy weight indicator and Theil indicator) is constructed for the accuracy of all models within that industry, and its representation in the model set is calculated based on the fusion indicator. The relative weights in the formula. In the formula, the denominator... This represents the sum of the accuracies of all models for the current task, used for standardization to ensure the weight values ​​are accurate. Falling Within the interval, and the sum of all model weights is 1.

[0070] Furthermore, model index Optional models include mainstream large language models such as QWEN, DeepSeek, and Llama, and industry categories. This is further categorized based on actual application scenarios, such as finance, semiconductors, and manufacturing. Historical accuracy index. The update mechanism follows the formula ,in This is a historical weight decay factor used to balance historical performance with recent performance, ensuring the timeliness and adaptability of model weight allocation.

[0071] In practical applications, particularly in intelligent analysis systems for industry economic conditions, this step is used to dynamically select the optimal model combination to address the differences in semantic understanding, numerical reasoning, and policy sensitivity across various industries. For example, in semiconductor industry analysis, if the QWEN model has a historical accuracy index of 0.95 in the policy sensitivity task and DeepSeek has 0.93 in the supply chain impact analysis, then their comprehensive weights in this industry would be 0.4 and 0.4 respectively, resulting in a final comprehensive score of [missing value]. If the score is lower than a preset threshold (such as 0.80, which is the industry standard or business requirement), a manual review mechanism will be triggered to ensure the reliability of the analysis results.

[0072] S32, perform secondary calibration of the weight allocation results by combining industry characteristic parameters.

[0073] Specifically, in some implementations, a secondary calibration of the weight allocation results using industry-specific parameters (such as policy sensitivity coefficients) is a key optimization step in the multi-model decision-making module of this invention. This aims to improve the model's prediction accuracy and decision interpretability across different industry scenarios. This step is based on an extension of the dynamic weighted voting model, introducing industry-specific parameters to correct the initial model weights, thereby achieving more refined model selection and result fusion.

[0074] Specifically, initial model weights The calculation formula is:

[0075] in, Indicates the model index (such as QWEN, DeepSeek, Llama). Indicates industry category, Representation Model In the industry The historical accuracy index is used. This formula, through normalization, ensures that the sum of the weights of each model in a specific industry is 1, thereby achieving preliminary model selection.

[0076] Furthermore, to enhance the model's adaptability to industry characteristics, this invention introduces a policy sensitivity coefficient. This coefficient is calculated by examining the correlation strength between historical policy events and changes in industry economic indicators, typically quantified using the Pearson correlation coefficient or mutual information. During the secondary calibration phase, the model weights will be adjusted as follows:

[0077] in, To calibrate the gain factor, it is usually set to This is used to control the sensitivity of industry characteristics to weight adjustments. The value range is $[0,1]$, which is especially relevant when an industry is highly sensitive to policy changes (such as the semiconductor industry). When the value approaches 1, the weighting of policy-sensitive models (such as QWEN) will increase significantly.

[0078] In practical applications, such as when dealing with policy events like "a country extending the review period for semiconductor export licenses to 120 days," the system will adjust the process based on the relevant industry... Value (e.g.) The model weights are dynamically adjusted to prioritize models that are good at handling policy-related inputs, thereby improving the accuracy and interpretability of the analysis results.

[0079] Furthermore, this secondary calibration mechanism effectively addresses the problem of insufficient model generalization ability in existing technologies, especially in policy-driven industries (such as finance, energy, and semiconductors), significantly improving the system's responsiveness to unstructured information and decision-making quality.

[0080] S4. Based on the dynamic semantic relationship, and using an objective weighting mechanism that integrates the Theil index method and the entropy weight index method, the contribution weight is dynamically calculated to generate an industry economic status analysis result that includes structured indicators and unstructured knowledge inferences.

[0081] Specifically, in some implementations, generating industry economic status analysis results containing structured indicators and unstructured knowledge inferences based on the aforementioned dynamic semantic associations and contribution weights is the core step in the knowledge enhancement and decision-making fusion of the technical solution of this invention. This step dynamically associates structured economic data (such as GDP growth rate, capacity utilization rate, inventory turnover rate, etc.) with unstructured knowledge (such as policy texts, industry research reports, expert comments, etc.) at the semantic level, and combines this with the historical accuracy index of each analytical model in a specific industry. Weighted fusion is performed to generate industry economic status analysis results with high interpretability and accuracy.

[0082] In this embodiment of the invention, the specific implementation includes: First, using a semantic graph neural network (GNN) to represent key entities in unstructured knowledge (such as "export license," "review cycle," and "wafer") as vectors, and then performing cross-modal alignment with nodes in a structured knowledge graph (such as industry chain nodes and economic indicator nodes). During the alignment process, the edge weights are determined by semantic similarity (such as cosine similarity). Dynamic calculations ensure that the economic impact implied in policy texts can be quantified and mapped onto a structured indicator system. Secondly, based on the specific industry... Historical accuracy index Through the normalization formula Assign model contribution weights to achieve weighted aggregation of outputs from multiple models.

[0083] Furthermore, the model weight update employs an exponential smoothing mechanism, with the update formula being: ,in A historical weight decay factor is used to ensure that model weights can be dynamically adjusted based on the latest task performance. Furthermore, the system supports injecting domain feature slots (such as "supply chain characteristics" and "policy sensitivity") into different industry scenarios, and combines prompt engineering to guide the large language model to generate analytical inferences that conform to the industry context.

[0084] In practical applications, such as semiconductor industry analysis, when the input is "a certain country has extended the review period for semiconductor export licenses to 120 days," the system can automatically identify key entities and generate a hybrid analysis result combining structured indicators (e.g., $50 billion in direct exports in this sector will be blocked) and unstructured knowledge inferences (e.g., "given the 2021 export control event") with historical event anchors (e.g., "the country is vigorously strengthening the layout of domestic alternative supply chains"). This result not only includes quantifiable economic indicator changes but also provides strategic suggestions based on knowledge graphs and analogical reasoning, meeting the interpretability and decision support requirements of scenarios such as financial auditing and policy making, and allowing for inferences about deeper, more likely, indirect impacts.

[0085] Specifically, this step effectively solves the problems of shallow knowledge integration and black box decision-making in existing technologies. Through dynamic semantic association and model weight allocation mechanism, it significantly improves the comprehensiveness, timeliness and domain depth adaptability of industry economic status analysis.

[0086] The industry economic status analysis method based on multimodal semantic alignment in this invention achieves deep integration of structured economic data and unstructured industry knowledge, improving the interpretability and accuracy of the analysis results; it enhances the system's responsiveness to real-time data and task priorities through a dynamic workflow scheduling mechanism, improving analysis efficiency and adaptability; and it constructs a domain-adaptive analysis module to support the rapid generation of industry-specific analysis logic, enhancing the model's practicality and generalization ability in specific scenarios.

[0087] Furthermore, S4 includes: S41, calculate the confidence level of the prediction result using the entropy method, where .

[0088] Specifically, in the multi-model decision module of this invention, calculating the confidence level of the prediction result using the entropy method is a key technical step, and its core formula is:

[0089] The technical implementation principle of this step is based on the entropy theory in information theory, used to quantify the uncertainty and illusion problems of prediction distributions. In the multi-model collaborative decision-making mechanism, each model (such as QWEN, DeepSeek, and Llama) predicts the economic state of the same industry and outputs a probability distribution. Entropy represents the confidence level of the model in various prediction outcomes. The smaller the value, the more concentrated the prediction results and the lower the uncertainty, therefore the higher the confidence level. The higher.

[0090] First, the predictions output by each model are normalized into a probability distribution, typically using the softmax function to ensure that the sum of the probabilities of all predicted categories is 1. Then, the Shannon entropy of this distribution is calculated:

[0091] in Indicates the first The probability of each predicted category, The predicted total number of categories. In this invention, the predicted categories typically include economic status labels such as "high risk," "medium risk," and "low risk," and their quantity... Generally, the value should not exceed 5 to ensure computational efficiency and interpretability.

[0092] Furthermore, confidence level The value range of is [0, 1]. When the prediction result is completely certain, it means that the prediction result is completely certain. ;when When the time is specified, it indicates that the prediction result is completely uncertain. This indicator is used to determine whether a manual review mechanism needs to be initiated, for example, in the semiconductor industry. If this happens, the workflow engine will skip the automatic decision-making module and proceed directly to the manual intervention process.

[0093] This step plays a crucial role in the entire system, particularly in dynamic workflow scheduling and risk control scenarios. By quantifying the uncertainty of the model output, the system can dynamically assess the trust in the prediction results, thus providing a scientific basis for model service degradation or knowledge base retrieval expansion. Furthermore, this method, combined with a multimodal knowledge fusion engine, ensures that the system maintains robustness and interpretability even when unstructured knowledge is involved in decision-making, providing precise decision support for specific scenarios such as financial investment consulting and industry cycle analysis.

[0094] S42 uses the nearest neighbor principle of case analogy reasoning to fill in the prompt word template and generate strategy suggestions that include historical event anchors and risk constraint clauses.

[0095] Specifically, in the strategy generation module, using the nearest neighbor principle of case analogy reasoning to fill in the prompt word template is one of the key steps in realizing intelligent analysis of the industry's economic status. This step is based on the case analogy reasoning theory in cognitive science. By retrieving the semantic similarity between historical event anchors and the current input event, it extracts a strategic framework with reference value and combines it with risk constraint clauses to generate industry strategy recommendations that are interpretable and compliance-oriented.

[0096] In this embodiment of the invention, the system first retrieves data from the knowledge base that has a semantic similarity higher than a threshold (e.g., ...). Historical event anchors are stored in a graph structure, encompassing multi-dimensional information such as policy texts, market reactions, and industry impacts. A semantic graph neural network is used to perform cross-modal alignment between policy entities and industry chain nodes, calculating edge weights to reflect the strength of associations between events. Subsequently, the system fills the "historical event anchor" slot in the prompt word template with the historical event ID with the highest matching degree, for example: "Refer to the strategy in {event ID}".

[0097] Meanwhile, based on the loss aversion principle in prospect theory, the system injects risk constraint clauses into the prompts. These clauses are driven by key risk indicators (such as inventory backlog rate and policy sensitivity) output by the risk assessment module and are embedded in the form of "Risk: Be wary of {indicator} fluctuations," ensuring that strategy recommendations are generated within a controllable risk range. The selection of risk indicators is based on the uncertainty of the prediction distribution calculated using the entropy method, i.e. When the confidence level is lower than the set threshold, the system will automatically increase the weight of the risk constraint.

[0098] Specifically, this step is widely used in scenarios such as financial investment advisory and supply chain risk warning. For example, in semiconductor industry analysis, when the system receives the input "export license review period extended," it will retrieve "2021 export control events" as an anchor point and, combined with indicators such as current wafer inventory and policy sensitivity, generate a strategy suggestion such as "Q2 wafer inventory increased by 30% → triggering supply chain risk warning → recommend switching to automotive chip production." In this strategy, 40% are structured indicators and 60% are unstructured knowledge inferences, meeting the compliance and traceability requirements of financial audits.

[0099] The technical value of this step lies in its ability to semantically align historical experience with the current context through a dynamic filling mechanism, significantly improving the interpretability and applicability of strategy recommendations. Simultaneously, the injection of risk constraint clauses enables the system to possess loss aversion capabilities when generating strategies, enhancing its practicality and robustness in high-risk economic decision-making scenarios.

[0100] S5 constructs a utility function based on data freshness, resource utilization, and task priority, and dynamically adjusts the workflow execution path. The task urgency metric algorithm satisfies... Based on the utility function, the workflow branch prediction engine is triggered to select the optimal task execution path.

[0101] Specifically, the core of this step lies in constructing a utility function based on data freshness, resource utilization, and task priority, and dynamically adjusting the workflow execution path accordingly to achieve intelligent and real-time task scheduling. This method plays a crucial role in multi-model collaborative decision-making and dynamic workflow scheduling systems, solving problems such as slow response, low resource utilization, and inflexible task priority handling in existing static scheduling mechanisms.

[0102] In this embodiment of the invention, the quantification of task urgency is modeled using a linear weighted method, and the formula is as follows:

[0103] in, To indicate the freshness of data, it is usually normalized to the difference between the timestamp and the current time. The time interval reflects the impact of data timeliness on the value of the task; This indicates the utilization rate of system resources, such as CPU, memory, and GPU utilization, and its value range is... This is used to measure the load status of the current execution environment; The priority of a task is set by the user or the system based on the urgency and scope of its impact. It is usually expressed in discrete levels (e.g., 1-5) or continuous values ​​(e.g., 0.1-1.0).

[0104] Furthermore, weighting coefficients , , The value needs to be dynamically adjusted based on industry characteristics. For example, in a financial investment advisory scenario, It can be set to 0.4. It is 0.3. The value is set at 0.3 to emphasize the balance between data timeliness and task priority; however, in manufacturing supply chain analysis, It can be improved to 0.6 to prioritize responses to real-time data changes. Weight adjustments can be optimized using an online learning mechanism based on statistical feedback of historical task completion efficiency and resource consumption.

[0105] This utility function serves as input to the workflow branch prediction engine, evaluating the overall value of different execution paths. During task scheduling, the system compares the current utility value with a preset threshold (such as...). The system determines whether workflow refactoring is necessary and selects the optimal path to execute tasks. For example, when policy changes lead to a decrease in data timeliness, the system can skip redundant feature engineering modules and directly enter the strategy generation stage, significantly improving response speed.

[0106] The technical value of this step lies in: by quantifying task urgency, it enables dynamic optimization of workflow execution paths, thereby improving the system's real-time decision-making capabilities and resource utilization in complex economic scenarios. Its innovation lies in unifying multi-dimensional scheduling factors into a utility function and combining it with reinforcement learning mechanisms for path prediction, providing an interpretable and scalable scheduling framework for industry economic status analysis.

[0107] Example 2 This embodiment provides a system for analyzing industry economic status based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment. The system specifically includes: The input preprocessing module is used to perform deep structuring and noise removal on the original economic text.

[0108] Specifically, the embodiments of this invention follow the entity boundary recognition theory in feature engineering (Lample et al., 2016). The original economic text contains a large amount of noise and non-standard expressions (such as "tight wafer capacity"), such as... Figure 2 As shown, the original text needs to be processed to form a standardized data vector. For example, "A certain country plans to impose additional tariffs on semiconductors" can be transformed into: {"domain": "semiconductors", "action": "tariff increase", "entity": ["wafers", "chips"]}.

[0109] The large language model rating module is used for the quantitative representation of industry trends.

[0110] Specifically, this module is based on attention-based economic semantic modeling (Devlin et al., 2018). (1) The model design principle is expressed as:

[0111] in, For industry classification weight indicator matrix, The input vector.

[0112] (2) The confidence generation mechanism (entropy method to measure prediction certainty) can be expressed as:

[0113] A multi-model decision module is used for dynamic optimization through ensemble learning.

[0114] Specifically, this module extends the integrated learning weighted voting model (Dietterich, 2000) to multiple domain scenarios, with the dynamic weight formula:

[0115] in, For model indexing (QWEN / DeepSeek / Llama). By industry category For the model in the industry The historical accuracy index (the mean of the entropy weight index and the Theil index, i.e., the fusion difference index).

[0116] like Figure 3As shown, the decision-making mechanism of this module divides the input into automatic analysis paths and manual intervention paths.

[0117] Strategy generation module, knowledge-enhanced reasoning engine Specifically, this module integrates case-based analogy reasoning (Kolodner, 1992) with cue learning (Brown et al., 2020). The scientific design of the cue word templates is shown in Table 1.

[0118] Table 1

[0119] like Figure 4 The diagram shows the theoretical mapping of the workflow execution process. The raw data is processed through feature engineering, decision optimization, and knowledge enhancement phases to form the final report.

[0120] The feature engineering stage addresses the problem of integrating unstructured knowledge. The theoretical basis for this stage is the information bottleneck theory (Tishby et al., 1999).

[0121] The decision optimization stage is used to overcome the limitations of static scheduling. The theoretical support for this stage is the dynamic threshold setting of the multi-armed gambling machine model (Auer et al., 2002).

[0122] The knowledge enhancement stage is used to strengthen domain specificity. The theoretical support for this stage is the template construction of frame semantics theory (Fillmore, 1982).

[0123] In one embodiment of this invention, theoretical verification specifically includes: Regarding the information report that "a certain country has extended the review period for semiconductor export licenses to 120 days," the following theoretical application analysis is conducted: First, feature engineering was performed to identify three entities: ["export license", "review cycle", and "semiconductor"], all of which conform to the named entity boundary criteria. Second, model weights were assigned, such as the weights for the policy sensitivity model. =0.4, Weights of the supply chain impact model =0.4, and the calculated comprehensive score is 0.4×0.95+0.4×0.93=0.752. This score is lower than the preset threshold of 0.8 in the semiconductor industry, triggering the manual review mechanism. Finally, knowledge enhancement is implemented, and the first prompt is dynamically filled with "As a semiconductor industry analyst, in response to the [export license extension] event, referring to the historical case [2021 export control event] (match degree 82%), a risk control-oriented strategy is generated". This process conforms to the nearest neighbor principle of case analogy reasoning.

[0124] Furthermore, the present invention also includes the following embodiments: (1) The model weight update mechanism is expressed as:

[0125] Where α = 0.7 is the historical weight decay factor.

[0126] (2) Theoretical basis of exception handling: Model service degradation: A Byzantine Generals Problem solution that follows fault-tolerant computation.

[0127] Knowledge base retrieval extension: semantic similarity metric learning (cosine ≥ 0.75).

[0128] (3) As shown in Table 2, the knowledge base construction principles use different storage structures and cognitive theoretical bases for two different types of historical events and policies and regulations.

[0129] Table 2

[0130] The embodiments of this invention, through the combination of the aforementioned theoretical framework and engineering implementation, provide rigorous academic support for system design while maintaining technical feasibility. Those skilled in the art can reproduce the system based on the description, and theoretical guidance avoids empirical deviations during the implementation process.

[0131] In summary, this invention implements a multimodal economic knowledge fusion engine. By constructing a fusion mechanism that supports deep association between structured data (economic indicators) and unstructured knowledge (policy texts / industrial chain maps), it overcomes the limitation of existing knowledge bases that solely rely on structured data. Specifically, it includes: (1) A cross-modal alignment method for policy text and industry chain map based on semantic graph neural network.

[0132] (2) Dynamic knowledge weight allocation mechanism, which is used to automatically adjust the contribution weight of structured and unstructured knowledge according to industry scenarios.

[0133] This invention implements an objectively weighted dynamic workflow scheduling system. By proposing a three-dimensional decision-making model of real-time feedback, resource awareness, and task priority, it achieves dynamic optimization of workflow paths, addressing the problem of failing to deeply quantify the sample differences and information contribution within different data sources. Specifically, it includes: (1) Industry-model fit evaluation system, which is used to construct an objective scoring matrix by combining semantic understanding differences and model reasoning differences.

[0134] (2) Multi-expert result fusion algorithm for the aggregation method of heterogeneous model output based on confidence weighting.

[0135] Example 3 This invention also provides an industry economic status analysis device based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, such as... Figure 5 As shown, the device 10 includes: The structured and unstructured data acquisition module 100 is used to acquire structured economic indicator data and unstructured industry knowledge data of the target industry. The unstructured industry knowledge data includes policy texts and industry chain map information. The cross-modal semantic alignment processing module 200 is used to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data using a semantic graph neural network, and to establish dynamic semantic associations between policy entities and industrial chain nodes. The dynamic knowledge weight allocation module 300 is used to automatically adjust the contribution weights of structured data and unstructured knowledge in the analysis process based on industry scenario characteristics through a dynamic knowledge weight allocation mechanism. The economic status analysis result generation module 400 is used to generate industry economic status analysis results containing structured indicators and unstructured knowledge inferences based on the dynamic semantic association and contribution weight.

[0136] Example 4 To implement the methods of the above embodiments, the present invention also provides a computer device, which includes a memory and a processor; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, so as to implement the various steps of the methods described above.

[0137] Example 5 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0139] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0140] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. An industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, the specific steps of which are as follows: S1, acquire structured economic indicator data and unstructured industry knowledge data of the target industry, wherein the unstructured industry knowledge data includes policy texts and industry chain map information; S2, using a semantic graph neural network to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data, and establishing dynamic semantic associations between policy entities and industrial chain nodes; S3, based on industry scenario characteristics, uses a dynamic knowledge weight allocation mechanism, combined with the difference analysis capability of the Theil index method and the information measurement characteristics of the entropy weight index method, to construct a dynamic objective weighting optimization model, which automatically calculates and adjusts the contribution weights of structured data and unstructured knowledge in the analysis process. S4. Based on the dynamic semantic relationship, and using an objective weighting mechanism that integrates the Theil index method and the entropy weight index method, the contribution weight is dynamically calculated to generate an industry economic status analysis result that includes structured indicators and unstructured knowledge inferences.

2. The method according to claim 1, characterized in that, The acquisition of structured economic indicator data and unstructured industry knowledge data of the target industry also includes: S11 uses entity boundary recognition theory to clean the original economic text and extract standardized representations of industry entities. S12 extracts multimodal features from unstructured industry knowledge data, where policy texts are represented by TF-IDF weighted word vectors and industry chain graph information is encoded using adjacency matrix.

3. The method according to claim 1, characterized in that, The method of using a semantic graph neural network to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data to establish dynamic semantic relationships between policy entities and industrial chain nodes also includes: S21, use a large language model to generate semantic embedding vectors for the policy text, and use a graph neural network to calculate the semantic similarity between policy entities and industry chain nodes, wherein the semantic similarity meets the usual industry standards. S22 optimizes cross-modal alignment results based on information bottleneck theory, filtering redundant semantic associations to improve analysis efficiency.

4. The method according to claim 1, characterized in that, Based on industry scenario characteristics, a dynamic knowledge weight allocation mechanism is used, combining the difference analysis capabilities of the Theil index method with the information content measurement characteristics of the entropy weight index method, to construct a dynamic objective weighting optimization model. This model automatically calculates and adjusts the contribution weights of structured data and unstructured knowledge in the analysis process. Specifically: S31, using formula The inter-group difference index is calculated using the entropy weight index method and the Theil index method, and the information content is measured by combining the entropy weight index method and the Theil index method. The historical accuracy weight of the model in a specific industry is calculated through a dynamic objective weighting model. S32, perform secondary calibration of the weight allocation results by combining industry characteristic parameters.

5. The method according to claim 1, characterized in that, Based on the aforementioned dynamic semantic relationships, and using an objective weighting mechanism that integrates the Theil index method and the entropy weight index method to dynamically calculate contribution weights, an industry economic status analysis result containing structured indicators and unstructured knowledge inferences is generated, specifically: S41, calculate the confidence coefficient threshold of the prediction result using the entropy method, where... ; S42 uses the nearest neighbor principle of case analogy reasoning to fill in the prompt word template and generate strategy suggestions that include historical event anchors and risk constraint clauses.

6. The method according to claim 1, characterized in that, Also includes: S5 constructs a utility function based on data freshness, resource utilization, and task priority, and dynamically adjusts the workflow execution path. The task urgency metric algorithm satisfies... Based on the utility function, the workflow branch prediction engine is triggered to select the optimal task execution path.

7. An industry economic status analysis device based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment, characterized in that, include: The structured and unstructured data acquisition module is used to acquire structured economic indicator data and unstructured industry knowledge data of the target industry. The unstructured industry knowledge data includes policy texts and industry chain map information. The cross-modal semantic alignment processing module is used to perform cross-modal alignment processing on the structured economic indicator data and unstructured industry knowledge data using a semantic graph neural network, and to establish dynamic semantic associations between policy entities and industrial chain nodes. The dynamic knowledge weight allocation module is used to construct a dynamic objective weighting optimization model based on industry scenario characteristics, through a dynamic knowledge weight allocation mechanism, combining the difference analysis capability of the Theil index method and the information content measurement characteristics of the entropy weight index method, and automatically calculate and adjust the contribution weight of structured data and unstructured knowledge in the analysis process. The economic status analysis result generation module is used to dynamically calculate contribution weights based on the dynamic semantic relationship and using an objective weighting mechanism that integrates the Theil index method and the entropy weight index method, and generate industry economic status analysis results containing structured indicators and unstructured knowledge inferences.

8. A computer device, characterized in that, Including processor and memory; The processor reads the executable program code stored in the memory to run the program corresponding to the executable program code, so as to implement the industry economic status analysis method based on dynamic objective weighting coefficient mechanism and multimodal semantic alignment as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements an industry economic status analysis method based on a dynamic objective weighting coefficient mechanism and multimodal semantic alignment as described in any one of claims 1-6.

Images