Enterprise and policy information matching recommendation system and method based on deep learning

Abstract

The invention belongs to the technical field of enterprise services, and particularly relates to a deep learning-based enterprise and policy information matching recommendation system and method.Enterprise feature vectors are constructed and converted into semantic embedding, so that enterprise attributes form continuous expression in a deep space; semantic unit division and context-aware vector generation are carried out on the policy text, and the policy semantic center is adjusted by fusing the overall distribution trend of the enterprise, so that the reality adaptability of the policy text is improved; a multi-view representation mechanism is introduced, and a policy support direction is described from different dimensions; and finally, generating a comprehensive matching score in combination with the confidence coefficient of semantic distance conversion and the explicit condition satisfaction degree. According to the method, the limitation of keyword matching is broken through, the alignment of enterprises and policies on the semantic level is realized, the discovery capability and matching accuracy of potential applicable policies are remarkably improved, and more enterprises which conform to conditions but are not completely consistent in expression form can be effectively identified and corresponding policies can be recommended.

Description

Technical Field

[0001] This invention belongs to the field of enterprise service technology, specifically relating to a deep learning-based enterprise and policy information matching and recommendation system and method. Background Technology

[0002] In the current context of digital governance and intelligent enterprise services, governments are issuing numerous policies to support industry development, provide tax incentives, and promote technological innovation, aiming to guide and support enterprise growth. To access these resources, enterprises need to sift through a vast array of policies to identify those applicable to their specific circumstances. Traditional information matching methods often rely on manual reading, keyword searches, or simple rule-based filtering. This involves comparing the enterprise attributes (such as size, industry, and qualifications) explicitly listed in the policy with the information submitted by the enterprise to achieve a preliminary match. While these methods offer some interpretability, they struggle to capture the deeper semantic requirements of policy texts and fail to identify potential connections between similar but worded statements. For example, a policy might encourage "small and medium-sized enterprises with continuous R&D capabilities," while the enterprise's information only indicates that "R&D expenditure has accounted for more than 5% of revenue in the past three years." Due to a lack of semantic understanding, existing systems often fail to establish an effective connection, leading to missed matches.

[0003] The above approach has a prominent technical problem: relying solely on explicit condition matching cannot effectively identify the implicit semantic fit between enterprise characteristics and policy intentions, resulting in incomplete coverage and insufficient matching accuracy in the recommendation results. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based enterprise and policy information matching and recommendation system and method, which achieves semantic alignment between enterprises and policies, significantly improves the ability to discover potential applicable policies and the accuracy of matching, thereby solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based method for matching and recommending enterprise and policy information, comprising: Collect multi-source information about enterprises and perform structured processing to generate enterprise feature vectors. Input the enterprise feature vectors into a deep neural network to obtain enterprise semantic embedding vectors. The policy text is segmented, semantic units are extracted and text vectors are generated, and the importance weight of the unit is determined based on the number of keywords and conditions in the semantic unit. By combining the text vector with the unit importance weight, a policy original semantic representation is generated. The group center trend of enterprise semantic embedding is then integrated, and the policy original semantic representation is adjusted to obtain the policy semantic center. The policy semantic center is projected onto multiple independent subspaces to form a multi-perspective policy representation. The semantic distance between the enterprise semantic embedding vector and the policy representations of each perspective is calculated and converted into a matching confidence score. By combining the explicit condition satisfaction in the policy semantic unit, a comprehensive matching score between enterprises and policies is generated. Based on the matching score, the enterprises are sorted and threshold-filtered, and the recommendation results are output.

[0006] Preferably, the structuring process includes: Obtain the company's registration information, financial data, and documents related to its innovation activities; The registration information is mapped to fields and missing values ​​are filled in. Keyword extraction is performed on the innovation activity documents to generate a set of technical tags, and the financial data is divided into intervals and labeled with levels. The processed information is integrated into a numerical vector that includes industry, scale, innovation intensity, region, and main business characteristics.

[0007] Preferably, it also includes: introducing a contrastive learning objective when generating enterprise semantic embedding vectors; The comparative learning objective includes selecting companies with similar industries, scales, and innovation levels as positive sample pairs and companies with significant differences as negative sample pairs. During training, the semantic distance of positive sample pairs is minimized while the semantic distance of negative sample pairs is maximized to optimize the embedding distribution.

[0008] Preferably, it also includes: after generating the enterprise semantic embedding vector, introducing attention weighting based on the industry to which the enterprise belongs, including: Adjustable attention weights are assigned to different industries to enable sensitive industries to obtain higher expression intensity in semantic embedding, and the attention weights are multiplied element-wise with the enterprise semantic embedding vector.

[0009] Preferably, it also includes: identifying the enterprise attribute descriptions in the policy semantic units during extraction and converting them into a structured set of conditions containing attributes, thresholds, and comparison operations; The comparison operations include numerical comparison and category classification.

[0010] Preferably, it further includes: using a weighted average method when generating the original semantic representation of the policy, wherein the weights of the weighted average are derived from the unit importance weights of the semantic units; The importance weight of the unit is determined by the number of conditions and keyword density contained in the semantic unit, so that semantic units containing more core policy content occupy a higher proportion in the aggregation process.

[0011] Preferably, it further includes: introducing a group mean vector of enterprise semantic embeddings when adjusting the original semantic representation of the policy, wherein the group mean vector is obtained by the arithmetic mean of all enterprise semantic embedding vectors, and the adjustment process is achieved by linearly fusing the original semantic representation of the policy with the group mean vector, so that the policy semantic center can adapt to the current overall distribution characteristics of enterprises while maintaining the original theme.

[0012] Preferably, it also includes: using a negative exponential decay method when generating the matching confidence score, wherein the negative exponential decay takes semantic distance as input, the smaller the semantic distance, the higher the matching confidence score, and the larger the semantic distance, the closer the matching confidence score is to zero, ensuring that enterprises with highly semantic fit obtain a confidence output close to the maximum value.

[0013] Preferably, the explicit condition satisfaction is determined by judging whether the actual attributes of the enterprise meet the conditions listed in the policy, and then participating in the final matching score calculation after being weighted by unit weights, including: For each semantic unit, the proportion of conditions met by enterprises is statistically analyzed, and then a weighted sum is calculated based on the importance weight of the semantic units. The results are then combined with the matching confidence score according to a preset ratio to generate a comprehensive matching score.

[0014] On the other hand, this invention proposes a deep learning-based enterprise and policy information matching and recommendation system, comprising: The information acquisition unit is used to acquire and structure information from multiple sources within the enterprise. The semantic embedding unit is used to perform the mapping from enterprise feature vectors to semantic embedding vectors; The text parsing unit is used to perform segmentation of policy text and generation of semantic unit vectors; The weight calculation unit is used to determine the importance weights of semantic units; The policy representation unit is used to generate policy semantic centers and multi-perspective representations. The matching calculation unit is used to perform a comprehensive calculation of semantic distance, confidence, and condition satisfaction. The recommendation generation unit is used to perform matching score sorting and recommendation result output; The feedback update unit is used to collect user behavior feedback and adjust the parameters of the semantic embedding unit to optimize subsequent matching.

[0015] Technical effects and advantages of the present invention: The enterprise and policy information matching and recommendation system and method based on deep learning proposed in this invention have the following advantages compared with the prior art: This invention constructs enterprise feature vectors and transforms them into semantic embeddings, enabling continuous expression of enterprise attributes in deep space. It performs semantic unit segmentation and context-aware vector generation on policy texts, and adjusts the policy semantic center by integrating the overall distribution trend of enterprises, thus improving its real-world adaptability. A multi-perspective representation mechanism is introduced to characterize policy support directions from different dimensions. Finally, a comprehensive matching score is generated by combining the confidence level of semantic distance transformation and the explicit condition satisfaction level. This method overcomes the limitations of keyword matching, achieving semantic alignment between enterprises and policies, significantly improving the ability to discover potentially applicable policies and the accuracy of matching. It enables more enterprises that meet the conditions but whose expressions are not entirely consistent to be effectively identified and recommended with corresponding policies. Attached Figure Description

[0016] Figure 1 This is a flowchart of the deep learning-based enterprise and policy information matching and recommendation method of the present invention; Figure 2 This is a block diagram of the deep learning-based enterprise and policy information matching and recommendation system of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention provides, for example Figure 1 This paper presents a deep learning-based method for matching and recommending enterprise and policy information, enabling efficient, dynamic, and semantic-level matching and recommendation between enterprise information and government-issued policy documents. The method constructs multi-dimensional association paths between enterprise feature vectors and policy semantic representations, combined with a context-aware semantic alignment mechanism, to achieve accurate identification and ranking of policy applicability.

[0019] The deep learning-based enterprise and policy information matching and recommendation method in this embodiment includes the following steps: The system collects multi-source enterprise information and performs structured processing to generate enterprise feature vectors. Specific structured processing includes: acquiring enterprise registration information, financial data, and innovation activity documents; mapping fields and filling in missing values ​​in the registration information; extracting keywords from the innovation activity documents and generating a set of technical tags; dividing the financial data into intervals and assigning levels; and integrating the processed information into a numerical vector containing industry, scale, innovation intensity, region, and main business characteristics.

[0020] By collecting enterprise registration information, financial data, and innovation activity documents, and performing structured processing such as field mapping, missing value completion, keyword extraction, interval division, and level labeling, the originally scattered, heterogeneous, and non-standardized enterprise information is transformed into a unified format of numerical vector representation. This process achieves deep integration and format unification of multi-source enterprise data, eliminates information silos and semantic ambiguities, and ensures the comparability and operability of various attributes in subsequent calculations. The generated enterprise feature vectors comprehensively cover key dimensions such as industry affiliation, business scale, innovation investment intensity, geographical distribution, and main business direction, providing a complete and consistent data foundation for constructing high-fidelity enterprise semantic representations. This enhances the ability of subsequent semantic mapping to characterize the enterprise's development status and improves the comprehensiveness and reliability of matching policy requirements.

[0021] The enterprise feature vector is input into a deep neural network to obtain the enterprise semantic embedding vector; Specifically, a contrastive learning objective is introduced when generating enterprise semantic embedding vectors. The contrastive learning objective includes selecting enterprises with similar industries, scales, and innovation levels as positive sample pairs and selecting enterprises with significant differences as negative sample pairs. During training, the semantic distance of positive sample pairs is minimized while the semantic distance of negative sample pairs is maximized to optimize the embedding distribution.

[0022] Furthermore, after generating the enterprise semantic embedding vector, attention weighting is introduced according to the industry to which the enterprise belongs, including: assigning adjustable attention weights to different industries so that sensitive industries can obtain higher expression intensity in semantic embedding, and performing element-wise multiplication of the attention weights with the enterprise semantic embedding vector.

[0023] By inputting enterprise feature vectors into a deep neural network, a non-linear mapping from the original attribute space to a higher-order semantic space is achieved. This allows enterprise representation to move beyond surface indicators and become a deep expression that integrates multi-dimensional feature interactions. A contrastive learning objective is introduced, using enterprises with similar industries, sizes, and innovation levels as positive sample pairs and enterprises with significant differences as negative sample pairs. This drives the network to narrow the semantic distance between similar enterprises and widen the semantic distance between dissimilar enterprises during training, thereby optimizing the distribution structure of semantic embeddings, enhancing clustering characteristics in the space, and allowing enterprises with common development characteristics to naturally cluster in the vector space, thus improving the discriminative power of semantic expression.

[0024] Building upon this foundation, attention-weighted approaches are introduced based on the industry to which the enterprise belongs. By assigning adjustable weights to different industries, those more sensitive to policy responses (such as high-tech and green environmental protection sectors) achieve stronger expressive intensity in semantic embedding. After element-wise multiplication, the features of the corresponding dimensions are dynamically amplified, thus highlighting their policy relevance during the matching process. This approach ensures that enterprise semantic embedding not only reflects its own attributes but also embodies industry-level differences in policy sensitivity, enhancing the targeting and responsiveness of subsequent policy semantic alignment.

[0025] The policy text is segmented, semantic units are extracted and text vectors are generated, and the importance weight of the unit is determined based on the number of keywords and conditions in the semantic unit. By combining the text vectors and the unit importance weights, a policy original semantic representation is generated. The group centrality trend of enterprise semantic embedding is integrated, and the policy original semantic representation is adjusted to obtain the policy semantic center. A weighted average method is used when generating the policy original semantic representation. The weight of the weighted average comes from the unit importance weight of the semantic unit. The unit importance weight is determined by the number of conditions and keyword density contained in the semantic unit, so that semantic units containing more core policy content occupy a higher proportion in the aggregation process.

[0026] Furthermore, when adjusting the original semantic representation of the policy, a group mean vector of enterprise semantic embedding is introduced. The group mean vector is obtained by the arithmetic mean of all enterprise semantic embedding vectors. The adjustment process is achieved by linearly fusing the original semantic representation of the policy with the group mean vector, so that the policy semantic center can adapt to the current overall distribution characteristics of enterprises while maintaining the original theme.

[0027] By segmenting policy texts, extracting semantic units, and generating text vectors, fine-grained semantic deconstruction of lengthy policy documents is achieved. This avoids the dilution of key information caused by overall encoding and allows for the precise capture of local policy intentions. The importance weight of each semantic unit is determined by combining the number of keywords and conditions within it, giving higher weight to paragraphs containing more substantive requirements and core support directions, ensuring that key policy content dominates subsequent aggregation. A weighted average method is used to generate the original semantic representation of the policy, ensuring that the contribution of each semantic unit is proportional to its information density, thus improving the representativeness and accuracy of the overall representation.

[0028] Building upon this foundation, the original semantic representation of the policy is adjusted by integrating the group mean vector of enterprise semantic embedding. Through linear fusion, the overall characteristic trends of current enterprises are introduced, ensuring that the policy semantic center not only remains faithful to the original text but also adapts to the general structural features of the current enterprise database. This enhances the applicability and responsiveness of policy expression in real-world matching scenarios. This approach transforms policy semantics from static text to a dynamic, alignable representation, improving the semantic synergy and matching feasibility between policies and enterprises at the distribution level.

[0029] The policy semantic center is projected onto multiple independent subspaces to form a multi-perspective policy representation. The semantic distance between the enterprise semantic embedding vector and the policy representation of each perspective is calculated and converted into a matching confidence score. When generating the matching confidence score, a negative exponential decay method is adopted. The negative exponential decay takes the semantic distance as input. The smaller the semantic distance, the higher the matching confidence score. The larger the semantic distance, the closer the matching confidence score is to zero, ensuring that enterprises with high semantic fit obtain a confidence output close to the maximum value.

[0030] By projecting the policy semantic center onto multiple independent subspaces, a multi-perspective policy representation is formed, enabling a single policy to be interpreted simultaneously across different semantic dimensions. Each subspace corresponds to a policy focus area, such as innovation investment, enterprise size, regional distribution, or environmental requirements. This overcomes the limitations of single-vector representation and fully demonstrates the multi-dimensional intent of policy support. Based on this, the semantic distance between the enterprise's semantic embedding vector and the policy representations from each perspective is calculated to measure the degree of alignment between enterprises across different policy dimensions, achieving fine-grained cross-modal alignment.

[0031] By employing a negative exponential decay method to convert semantic distance into a matching confidence score, companies with smaller distances receive higher confidence outputs, and the score decreases rapidly as distance increases. This strengthens the identification advantage of semantically highly compatible individuals and ensures that the matching results have clear distinguishing boundaries in the semantic space. This approach enhances the sensitivity to potentially applicable companies, allowing even companies with deviations in some dimensions to still receive reasonable scores in their advantageous dimensions. It enhances the flexibility and inclusivity of the matching process, effectively supporting the accurate association between diverse companies and complex policies.

[0032] By combining the explicit condition satisfaction levels in the policy semantic units, a comprehensive matching score between enterprises and policies is generated. Based on these matching scores, ranking and threshold filtering are performed, and recommendation results are output. Furthermore, when extracting policy semantic units, the enterprise attribute descriptions are identified and transformed into a structured condition set containing attributes, thresholds, and comparison operations; the comparison operations include numerical comparison and category classification judgment.

[0033] The explicit condition satisfaction score is calculated by determining whether the enterprise's actual attributes meet the conditions listed in the policy, and then using a weighted average based on the unit weights to participate in the final matching score calculation. This includes: For each semantic unit, the proportion of conditions met by enterprises is statistically analyzed, and then a weighted sum is calculated based on the importance weight of the semantic units. The results are then combined with the matching confidence score according to a preset ratio to generate a comprehensive matching score.

[0034] By identifying enterprise attribute descriptions when extracting policy semantic units and transforming them into a structured set of conditions containing attributes, thresholds, and comparison operations, precise modeling of the rigid requirements in the policy is achieved. Numerical comparison is used to determine whether quantitative indicators such as revenue, R&D investment, and number of employees meet the standards, while category classification is used to confirm the compliance of qualitative conditions such as industry type, regional location, and qualification level, ensuring that the core thresholds of the policy are fully expressed.

[0035] Based on this, the explicit condition satisfaction rate is calculated, which involves statistically analyzing the proportion of conditions met by enterprises for each semantic unit and then weighting and summing these proportions according to the unit's importance weight. This allows high-importance policy clauses to play a greater role in the final score, reflecting the actual impact of different clauses on policy implementation. This condition satisfaction rate is then integrated with the matching confidence score generated in the semantic space according to a preset ratio to form a comprehensive matching score. This achieves dual verification of flexible semantic fit and rigid rule compliance, avoiding false positives that might arise from pure semantic matching and compensating for false negatives that might occur from pure rule matching.

[0036] Based on the comprehensive score, the system sorts and filters by thresholds, outputting recommended results. This ensures that the final policies presented are not only relevant in terms of linguistic intent, but also feasible in terms of actual application conditions. This significantly improves the accuracy, credibility, and operability of the recommended results, enabling enterprises to efficiently identify policy content that is truly applicable and beneficial.

[0037] On the other hand, this invention proposes a deep learning-based enterprise and policy information matching and recommendation system, such as... Figure 2 As shown, it includes: The information acquisition unit is used to acquire and structure information from multiple sources within the enterprise. The semantic embedding unit is used to perform the mapping from enterprise feature vectors to semantic embedding vectors; The text parsing unit is used to perform segmentation of policy text and generation of semantic unit vectors; The weight calculation unit is used to determine the importance weights of semantic units; The policy representation unit is used to generate policy semantic centers and multi-perspective representations. The matching calculation unit is used to perform a comprehensive calculation of semantic distance, confidence, and condition satisfaction. The recommendation generation unit is used to perform matching score sorting and recommendation result output; The feedback update unit is used to collect user behavior feedback and adjust the parameters of the semantic embedding unit to optimize subsequent matching.

[0038] In addition, the modules mentioned above are also used to implement other steps in the deep learning-based enterprise and policy information matching and recommendation method, as follows: Step 1: Enterprise Multi-Source Information Collection and Structured Processing In the initial stage of matching enterprise and policy information, it is necessary to extensively collect and standardize relevant enterprise information. Sources of enterprise information include business registration data, tax declaration records, industry classification labels, science and technology project application materials, intellectual property registration information, and annual report disclosures. This information is usually in unstructured or semi-structured form and needs to be cleaned, normalized, and integrated using a unified framework to form structured enterprise files that can be used for subsequent analysis.

[0039] Basic enterprise registration information, including unified social credit code, enterprise name, establishment date, registered capital, legal representative, registered address, and business scope, is obtained from multiple public databases and authorized interfaces. The obtained information is mapped to a unified data table according to preset fields, and initial completion is performed for missing or outlier values ​​using a method based on the characteristics of neighboring enterprises. For example, if a company's "industry category" field is empty, the most likely industry affiliation is determined by matching keywords from its business scope to an industry dictionary.

[0040] Text extraction is performed on innovation activity information such as enterprise-submitted technology project documents, patent lists, and software copyright registrations. Natural language processing techniques are used to identify technical field keywords, R&D direction descriptions, and project goal statements in the documents, and these are transformed into a set of tagged features. For example, technology tags such as "artificial intelligence," "edge computing," and "industrial Internet of Things" are extracted from a high-tech enterprise certification document to form a preliminary profile of the enterprise in terms of technological innovation.

[0041] The system integrates corporate financial and operational data, including quantitative indicators such as operating revenue, total tax payment, number of employees, and R&D investment ratio over the past three years. Continuous variables are divided into intervals and labeled with levels; for example, "R&D investment ratio" is divided into four levels: [0-3%), [3%-6%), [6%-10%), and [10%+), and assigned corresponding codes. These codes will be used as weighting adjustment factors in subsequent vector generation.

[0042] Based on the structured data mentioned above, construct enterprise feature vectors. Its form is: ; in, Indicates the first The company in the Values ​​on the dimensional features, This represents the total number of features. The vector covers dimensions such as the company's industry (using one-hot encoding), size level (based on number of employees and revenue), innovation intensity (based on number of patents and R&D ratio), geographical distribution (administrative division code), and frequency of keywords related to its main business. Each component of the vector is normalized to ensure its distribution is within a certain range. Within the interval, to eliminate the impact of dimensional differences: ; For categorical variables, label encoding is used to directly normalize them. Scope; For text-derived features (such as keyword weights), TF-IDF weighting is applied and then normalized.

[0043] This step systematically integrates multi-source heterogeneous information about enterprises, transforming scattered and non-standard data into a unified numerical vector representation. This vector not only preserves the enterprise's static attributes but also integrates dynamic operational and innovation capability indicators, providing rich input for subsequent semantic space mapping. Normalization ensures the fair participation of features from different dimensions in subsequent calculations, preventing high-order variables from dominating similarity judgments. The output at this stage is the enterprise feature vector. This will serve as the input basis for semantic embedding generation in the second step.

[0044] Step Two: Construction of Enterprise Semantic Embedding Space After obtaining the structured enterprise feature vectors, they need to be further mapped to a high-dimensional semantic space, so that semantically similar enterprises are closer together in this space. This space not only reflects the similarity of external attributes of enterprises, but also implies the intrinsic connection between their potential development needs and policy sensitivities. Through nonlinear transformation, the original feature vectors are transformed into a more expressive semantic representation.

[0045] The enterprise feature vector generated in step one The input is a deep neural network structure, which consists of multiple fully connected layers stacked together. The first layer receives... As input, the output dimension gradually decreases, forming a bottleneck structure, and finally outputting a fixed-length semantic embedding vector. The network structure is represented as follows: ; in, The parameter is nonlinear mapping function, For the number of network layers, and The first Layer weight matrix and bias terms, This is an activation unit (such as ReLU). This mapping process learns complex nonlinear relationships from the original features to the semantic space.

[0046] To enhance the discriminative power of semantic embeddings, a contrastive learning objective is introduced during training. Several enterprise samples are selected. If two samples are highly similar in industry, scale, and innovation level, they are considered positive sample pairs; otherwise, they are considered negative sample pairs. The Euclidean distance between embedding vectors is defined as: ; The training objective is to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. The loss function is designed as follows: ; in For the positive sample set, For the negative sample set, For preset boundary values, This is the balancing coefficient. By optimizing this loss, the network learns to effectively separate unrelated firms and bring similar firms closer together in their embedding distribution.

[0047] After training is complete, fix the network parameters. Generate semantic embedding vectors for all enterprises in batches. Each vector is located in In 3D space (such as This can be viewed as a "digital twin" of a company at a deeper semantic level. This vector no longer directly corresponds to the original features, but rather captures the interaction patterns and potential trends among the overall attributes of the company.

[0048] To further enhance the contextual adaptability of the embedding, an industry-specific attention weighting mechanism is introduced. Based on the varying sensitivities of different industries to policy responses, a learnable attention weight is assigned to each industry. ( (For industry categories). The final semantic representation is adjusted as follows: ; in Represents element-wise product. Indicates enterprise Industry affiliation. This weight is updated jointly with the loss function during training, enabling highly sensitive industries (such as high-tech and green environmental protection) to obtain higher representation strength in the matching.

[0049] This step transforms the original features of enterprises into high-order semantic representations using deep networks, enhancing the ability to capture potential enterprise attributes. Contrastive learning ensures the rationality of the semantic space, allowing similar enterprises to cluster in the vector space. Introducing an industry attention mechanism allows the embedded vectors to reflect the structural differences in policy responses across different industries, improving the targeting of subsequent matching. The generated semantic embeddings... This will serve as the basis input for cross-modal alignment with the policy text in the third step.

[0050] Step 3: In-depth semantic analysis of policy texts After semantic modeling is completed at the enterprise level, the policy text needs to be semantically parsed at the same depth to achieve two-way alignment. Policy documents are usually lengthy, with complex sentence structures and multi-level structures, requiring layered parsing to extract their core applicable conditions, support directions, and descriptions of beneficiaries.

[0051] Obtain the original policy documents issued by the government, including national, provincial, and municipal level documents on industry support, tax incentives, talent introduction, and R&D funding. Preprocess the entire text, including removing headers and footers, standardizing numbered lists, and segmenting paragraphs, to create a continuous text stream. Then, divide the policy into clauses or paragraph blocks according to semantic units, each block denoted as [block name missing]. The set of local semantic units constituting policy .

[0052] For each semantic unit Keyword extraction and entity recognition are performed to identify relevant enterprise attribute descriptions, such as "annual sales revenue less than 50 million yuan," "belonging to strategic emerging industries," "possessing no fewer than 5 valid invention patents," and "located in the central and western regions." These descriptions are then transformed into structured conditional expressions in the form: ; in For attribute names (such as "revenue scale"), For threshold or category values, For comparison operations (such as "<", "belongs to"). This set constitutes the set of explicit applicable conditions for the policy unit.

[0053] Utilizing pre-trained language understanding structures for each semantic unit Generate context-aware text vectors This structure employs a bidirectional encoding mechanism to capture the dependencies between words within a sentence. The output vector is: ; in For the first The hidden state of each word after encoding This represents the aggregated representation of the entire unit by averaging all word vectors. This vector contains deep semantic information about the semantic unit, transcending the limitations of keyword matching.

[0054] To measure the importance of different semantic units within the overall policy, a unit importance scoring mechanism is introduced. Each unit is defined. Importance score The number of conditions it contains Weighted sum with keyword density: ; in To adjust the parameters, This represents the number of keywords identified in the unit. This represents the total number of words. This weight will be used for weighted aggregation in the subsequent overall policy representation.

[0055] This step enables fine-grained decomposition and semantic extraction of policy texts, preserving explicit conditional rules while capturing implicit semantic information. By generating local semantic unit vectors and importance weights, multi-level inputs are provided for the overall policy representation. Keyword extraction and conditional structuring allow for quantifiable comparison of policy requirements, while text vectors support flexible semantic matching. The generated... and This will serve as the basis for constructing the overall semantic center of the policy in the fourth step.

[0056] Step 4: Dynamic Generation of Policy Semantic Center After completing the partial parsing of the policy text, it needs to be integrated into a unified policy semantic center vector for matching with the semantic embedding of enterprises. This center vector should reflect the core support direction of the policy and the overall characteristics of the target enterprise group.

[0057] Based on the semantic unit vectors extracted in step three Corresponding importance weights Calculate the weighted semantic representation of the overall policy. This weighted average operation gives higher importance to semantic units in the final representation, ensuring that the core policy content dominates the overall semantics.

[0058] To enhance the responsiveness of policy vectors to firm characteristics, prior information about firm semantic distribution is introduced. This involves calculating the responsiveness of all firms in the current firm database. mean vector As a central trend for the business community: ; Then, the original policy statement By integrating with this prior center, a context-aware policy semantic center can be generated. : ; in The fusion coefficient controls the relative impact of the policy's own semantics on the overall distribution of enterprises. This design allows the policy representation to maintain its core message while slightly biasing towards the general characteristics of the current enterprise group, thus improving the realistic adaptability of the matching.

[0059] To capture the diversity of policy support directions, the policy semantic center is expanded from multiple perspectives. The projection is applied to multiple independent subspaces, each corresponding to a specific policy focus dimension (such as "innovation support," "scale support," "regional preferential treatment," and "environmental protection requirements"). The projection operation is defined as follows: ; in For the first Projection matrix of each viewpoint, For subspace dimension, Total number of viewpoints. Each This represents the semantic emphasis of a policy in a specific dimension.

[0060] The final policy multi-perspective representation set is In the fifth step, this set will be matched with the enterprise semantic embeddings in multiple dimensions.

[0061] This step generates a semantic center that is both faithful to the original policy text and context-adaptive through weighted aggregation and prior fusion. The multi-perspective expansion mechanism allows a single policy to be interpreted from different dimensions, adapting to diverse feature combinations of enterprises. This representation avoids the information loss caused by simplifying complex policies into a single vector, laying the foundation for refined matching. The generated multi-perspective policy vector... This will serve as the reference benchmark for calculating the matching degree in the fifth step.

[0062] Step 5: Calculate cross-modal semantic matching degree After the semantic embedding of enterprises and the multi-perspective representation of policies are constructed, the core matching stage begins. This stage calculates the degree of fit between each enterprise and each policy across multiple semantic dimensions, forming a comprehensive matching score.

[0063] For each enterprise Its corresponding semantic embedding And the first of each policy View vectors Calculate the semantic distance between the two in this dimension: This squared Euclidean distance measures the degree of deviation of a company from a specific policy focus dimension; the smaller the distance, the higher the degree of alignment.

[0064] Convert semantic distance into matching confidence score Using a negative exponential decay function: ;in The decay rate parameter controls the strength of the distance's influence on the score. This function ensures that the confidence level is close to 1 when the distance approaches zero; as the distance increases, the confidence level decreases rapidly.

[0065] The matching results are then double-validated based on the explicit condition satisfaction scores extracted in step three. (Define the enterprise.) Policy Unit The condition satisfaction level is: ; in For enterprises In attributes The actual value on This is an indicator function; it is 1 if the condition is met, and 0 otherwise. This indicates the proportion of conditions that an enterprise meets under this unit.

[0066] The final matching score is generated by combining semantic confidence and condition satisfaction. First, average the confidence scores from multiple perspectives: ; Then combine it with the weighted condition satisfaction: ; in This is the balance coefficient between semantics and rules, and the denominator is a weighted average of condition satisfaction. This score reflects both semantic fit and rule compliance.

[0067] This step integrates semantic matching and rule validation, balancing flexible semantic association with rigid condition constraints. The multi-perspective design enables the matching process to identify a company's strengths across different policy dimensions, avoiding the bias of a single scoring method. Final score. By combining the semantic understanding capabilities of deep learning with the interpretability of structured rules, a comprehensive assessment of the applicability of enterprise-policy analysis is provided. This score will serve as the basis for ranking and recommendation in step six.

[0068] Step Six: Ranking Matching Results and Generating Recommendations After obtaining the matching scores between enterprises and policies, the results need to be organized and presented to generate a personalized recommendation list.

[0069] For each enterprise Collect its information and all policies Match score This forms a score vector. The vector is then normalized to unify its range. Interval: Normalized scores facilitate cross-enterprise comparisons and threshold setting.

[0070] Set dynamic recommendation threshold This is determined based on the matching distribution of the enterprise itself. The upper quartile is used as the recommendation threshold: ; Only retain scores higher than The policy items form a preliminary recommendation set. This dynamic threshold avoids recommending too many irrelevant policies to companies with generally low matching scores.

[0071] Policies within the recommended set are categorized as follows: Sort the data in descending order to generate the final recommendation sequence. Simultaneously, extract the semantic units from the policy that are most relevant to the company. (Right now The highest-ranking unit), as a summary of the reasons for recommendation.

[0072] Generate a visual recommendation report, including a list of recommended policies, matching scores, core applicable conditions, application deadlines, and competent authorities. The report is displayed in layers according to matching degree, with highly matched items placed at the top and their main points of convergence marked (such as "meets the standard due to R&D investment ratio" or "meets the definition of micro and small enterprises").

[0073] This step transforms matching scores into actionable recommendations, ensuring accuracy and usability through dynamic thresholds and ranking mechanisms. Normalization eliminates differences in scoring scales between enterprises, ensuring consistent recommendation standards. Extracting recommendation rationale enhances the understandability of the results, helping enterprises quickly assess policy relevance. The generated recommendation report can be directly used for enterprise decision support.

[0074] Step 7: System Feedback and Semantic Space Iterative Update To maintain the long-term effectiveness of the system, a feedback mechanism is introduced to dynamically adjust the semantic space and matching logic.

[0075] Collect feedback from enterprise users regarding the recommendation results, including clicking to view, marking as "relevant" or "irrelevant," and actual application actions. "Relevant" feedback is recorded as positive examples, and others as negative examples.

[0076] For enterprises and policies that generate positive feedback Check its matching score If the score is lower than the preset benchmark... If the score is low and the correlation is high, it is considered a "low score, high correlation" case, indicating that there is a deviation in the current matching logic.

[0077] For such cases, adjust the parameters of the enterprise semantic embedding network. Define the modified loss term: ; in This is a set of low-scoring, highly relevant samples. Network weights are updated via backpropagation. This will enable similar companies to obtain higher matching scores in the future.

[0078] Perform this update process regularly (e.g., monthly) to ensure that the semantic space continues to evolve with changes in the policy environment and corporate behavior, maintaining the timeliness and adaptability of the matching system.

[0079] This step establishes a closed-loop update path from user feedback to system parameters, enabling the matching logic to self-optimize. By focusing on "low-score, high-relevance" cases, the system can identify and correct blind spots or biases in the semantic space, gradually improving recommendation quality. This mechanism ensures that the system maintains high efficiency and relevance in long-term operation.

[0080] In summary, this embodiment provides a complete process for matching and recommending enterprise and policy information, from information collection, semantic modeling, text parsing, matching calculation to recommendation generation and system updates. Through deep semantic space construction and a multi-dimensional matching mechanism, it achieves precise matching between enterprise needs and policy supply.

[0081] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for matching and recommending enterprise and policy information, characterized in that, include: Collect multi-source information about enterprises and perform structured processing to generate enterprise feature vectors. Input the enterprise feature vectors into a deep neural network to obtain enterprise semantic embedding vectors. The policy text is segmented, semantic units are extracted and text vectors are generated, and the importance weight of the unit is determined based on the number of keywords and conditions in the semantic unit. By combining the text vector with the unit importance weight, a policy original semantic representation is generated. The group center trend of enterprise semantic embedding is then integrated, and the policy original semantic representation is adjusted to obtain the policy semantic center. The policy semantic center is projected onto multiple independent subspaces to form a multi-perspective policy representation. The semantic distance between the enterprise semantic embedding vector and the policy representations of each perspective is calculated and converted into a matching confidence score. By combining the explicit condition satisfaction in the policy semantic unit, a comprehensive matching score between enterprises and policies is generated. Based on the matching score, the enterprises are sorted and threshold-filtered, and the recommendation results are output.

2. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, The structuring process includes: Obtain the company's registration information, financial data, and documents related to its innovation activities; The registration information is mapped to fields and missing values ​​are filled in. Keyword extraction is performed on the innovation activity documents to generate a set of technical tags, and the financial data is divided into intervals and labeled with levels. The processed information is integrated into a numerical vector that includes industry, scale, innovation intensity, region, and main business characteristics.

3. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, Also includes: Introduce a contrastive learning objective when generating enterprise semantic embedding vectors; The comparative learning objective includes selecting companies with similar industries, scales, and innovation levels as positive sample pairs and companies with significant differences as negative sample pairs. During training, the semantic distance of positive sample pairs is minimized while the semantic distance of negative sample pairs is maximized to optimize the embedding distribution.

4. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, Also includes: After generating the semantic embedding vector of the enterprise, attention weighting is introduced based on the industry to which the enterprise belongs, including: Adjustable attention weights are assigned to different industries to enable sensitive industries to obtain higher expression intensity in semantic embedding, and the attention weights are multiplied element-wise with the enterprise semantic embedding vector.

5. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, Also includes: When extracting policy semantic units, identify the enterprise attribute descriptions within them and transform them into a structured set of conditions containing attributes, thresholds, and comparison operations; The comparison operations include numerical comparison and category classification.

6. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, Also includes: When generating the original semantic representation of the policy, a weighted average method is used, and the weights of the weighted average are derived from the unit importance weights of the semantic units; The importance weight of the unit is determined by the number of conditions and keyword density contained in the semantic unit, so that semantic units containing more core policy content occupy a higher proportion in the aggregation process.

7. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, Also includes: When adjusting the original semantic representation of the policy, a group mean vector of enterprise semantic embedding is introduced. The group mean vector is obtained by the arithmetic mean of all enterprise semantic embedding vectors. The adjustment process is achieved by linearly fusing the original semantic representation of the policy with the group mean vector, so that the policy semantic center can adapt to the current overall distribution characteristics of enterprises while maintaining the original theme.

8. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, Also includes: When generating the matching confidence score, a negative exponential decay method is used. The negative exponential decay takes semantic distance as input. The smaller the semantic distance, the higher the matching confidence score. The larger the semantic distance, the closer the matching confidence score is to zero, ensuring that enterprises with high semantic fit obtain a confidence output close to the maximum value.

9. The method for matching and recommending enterprise and policy information based on deep learning according to claim 1, characterized in that, The explicit condition satisfaction level is determined by assessing whether the enterprise's actual attributes meet the conditions listed in the policy, and then weighted by unit weights and averaged before being included in the final matching score calculation, including: For each semantic unit, the proportion of conditions met by enterprises is statistically analyzed, and then a weighted sum is calculated based on the importance weight of the semantic units. The results are then combined with the matching confidence score according to a preset ratio to generate a comprehensive matching score.

10. A deep learning-based enterprise and policy information matching and recommendation system for implementing the method as described in any one of claims 1-9, characterized in that, include: The information acquisition unit is used to acquire and structure information from multiple sources within the enterprise. The semantic embedding unit is used to perform the mapping from enterprise feature vectors to semantic embedding vectors; The text parsing unit is used to perform segmentation of policy text and generation of semantic unit vectors; The weight calculation unit is used to determine the importance weights of semantic units; The policy representation unit is used to generate policy semantic centers and multi-perspective representations. The matching calculation unit is used to perform a comprehensive calculation of semantic distance, confidence, and condition satisfaction. The recommendation generation unit is used to perform matching score sorting and recommendation result output; The feedback update unit is used to collect user behavior feedback and adjust the parameters of the semantic embedding unit to optimize subsequent matching.

Images