Improved double-tower model system and method for policy accurate recommendation

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an improved double-tower model system and method for policy accurate recommendation, and the method comprises the steps: carrying out the hierarchical coding of a policy text, and generating a hierarchical policy vector containing context semantics; constructing a user tower and a policy tower, calculating a feature level weight matrix of user vectors and policy vectors through an attention mechanism, and forming an adaptive weighted matching score; adopting a mixed negative sampling strategy and a multi-task learning framework to optimize the negative sample set construction of the model and the matching ability of a multi-dimensional policy objective; an online learning and offline training collaborative dynamic optimization system is designed, and the real-time performance and accuracy of the model are balanced; according to the method, the multi-dimensional enterprise portrait is constructed, the enterprise portrait features are adjusted in real time through the dynamic updating technology, and the effects of improving policy recommendation accuracy, enhancing system dynamic adaptability and optimizing cold start scene performance are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to an improved twin-tower model system and method for accurate policy recommendations. Background Art

[0002] Traditional policy recommendation systems often employ a basic two-tower model architecture, making recommendations through simple vector matching between the user tower and the policy tower. However, these approaches have significant limitations: They lack semantic understanding. Traditional methods rely on techniques like TF-IDF and simple neural networks to encode policy text. These methods are incapable of parsing specialized terminology (e.g., "additional deductions for R&D expenses"), complex logical relationships (e.g., nested conditions within policy clauses), and implicit conditions (e.g., the criteria for identifying "small and medium-sized enterprises"). Consequently, policy vectors only capture surface features and fail to reflect deeper semantics. For example, traditional models often misjudge the scope of "taxable income reductions" in tax policies by ignoring contextual relationships. They also suffer from poor dynamic adaptability, relying on static enterprise information (e.g., registered industry, founding date) and historical interaction data, lacking the ability to track real-time enterprise dynamics. When enterprises undergo key events such as business expansion (e.g., transitioning from traditional manufacturing to intelligent manufacturing) or technological upgrades (e.g., obtaining invention patents), demand models cannot be updated in a timely manner, resulting in a disconnect between recommended policies and actual demand. The cold start problem is prominent. For new enterprises (with no historical policy interaction records) or new policies (with no implementation feedback data), matching can only rely on limited basic attributes, and the recommendation effect is significantly reduced. For example, due to the lack of policy application records, newly established technology companies have difficulty in identifying their potential demand for "innovation fund subsidies" through traditional models. Feature interaction is simple: matching is only calculated through cosine similarity, ignoring the complex relationship between the multi-dimensional characteristics of enterprises and policies. For example, the nonlinear relationship between the "R&D investment intensity" of an enterprise and the "technological innovation subsidy ratio" of the policy cannot be effectively modeled by traditional methods.

[0003] Existing technologies suffer from numerous flaws. For data processing, policy text parsing relies solely on bag-of-words models or shallow neural networks, which are unable to handle long-range semantic dependencies (such as the "only if... and..." logic in policy clauses). Enterprise profiles are constructed based on static data, with long update cycles (typically quarterly or annually), lagging behind dynamic enterprise changes (such as monthly financing events and temporary policy response needs). At the model architecture level, feature interactions between user towers and policy towers are performed solely through vector concatenation or dot products, failing to account for differences in feature importance (for example, "industry classification" has a much greater impact on industrial policy than "registered address"). Negative sampling strategies are limited (mostly random sampling), resulting in a high proportion of simple negative samples in the training data. This makes it difficult for the model to distinguish between difficult negative samples with high similarity, limiting training effectiveness. Regarding application effectiveness, recommendation accuracy is insufficient, with enterprise click-through rates for policies below 30% and application conversion rates at a mere 15%-20%. Timeliness is poor, with recommendation updates taking an average of 2-3 days after the release of new policies, making it impossible to meet the demand for immediate push notifications of urgent policies (such as tax exemptions during the pandemic). In the cold start scenario, the policy matching accuracy rate for new enterprises is less than 40%, and the exposure rate of new policies is less than 20%, resulting in a waste of policy resources.

[0004] In response to the above technical defects, a solution of a multi-dimensional graph data recall strategy system and method for policy recommendation is proposed. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides the following technical solutions:

[0006] An improved twin-tower model approach for precise policy recommendations, including:

[0007] Hierarchical encoding of policy text to generate hierarchical policy vectors containing contextual semantics;

[0008] Build user towers and policy towers, calculate the feature-level weight matrix of user vectors and policy vectors through the attention mechanism, and form an adaptively weighted matching score;

[0009] A hybrid negative sampling strategy and a multi-task learning framework are used to optimize the model's ability to construct negative sample sets and match multi-dimensional policy objectives.

[0010] Design a dynamic optimization system that coordinates online learning and offline training to balance the real-time performance and accuracy of the model;

[0011] Build a multi-dimensional corporate portrait and adjust the corporate portrait features in real time through dynamic update technology.

[0012] Furthermore, the step of hierarchical encoding of the policy text includes using a pre-trained Transformer model to perform multi-layer encoding on the policy text, with the input containing the policy text content and attribute fields; capturing the long-distance dependency between words through the self-attention mechanism to generate a token-level semantic vector; integrating the full-text semantics through CLStoken to form a hierarchical policy vector; the hierarchical feature extraction architecture includes a bottom-level encoder, a middle-level encoder, and a high-level encoder, which respectively process word-level semantics, sentence logic, and paragraph-level logic to form a multi-granularity policy vector.

[0013] Furthermore, the steps of constructing the user tower and the policy tower include introducing an attention mechanism in the output layer of the user tower and the policy tower, inputting the user vector and the policy vector; calculating the feature-level weight matrix, and normalizing it through the softmax function to reflect the correlation between user features and policy features; designing a domain-specific attention weight function, and dynamically adjusting the feature weights based on the industry, development stage and core needs of the enterprise; automatically identifying the key policy features corresponding to the core needs of the enterprise to improve the targeted matching.

[0014] Furthermore, the step of adopting the hybrid negative sampling strategy and the multi-task learning framework includes combining three strategies: global random sampling, intra-batch sampling, and difficult negative sample sampling, dynamically adjusting the ratio according to data distribution, and constructing a balanced negative sample set;

[0015] Build a multi-task learning framework, set multiple output layers corresponding to different policy goals, share underlying parameters, and independently optimize the loss function of each task;

[0016] A negative sample distribution balancing algorithm is proposed. By calculating sample density and similarity, the ratio of three sampling strategies is dynamically adjusted to alleviate data bias problems.

[0017] Through the multi-task parameter sharing mechanism, the model's generalization ability for multi-dimensional policy objectives is improved, allowing a single model to handle matching tasks for more than five policy objectives at the same time.

[0018] Furthermore, the design of a dynamic optimization system that coordinates online learning and offline training to balance the real-time and accuracy of the model includes designing an online incremental learning module and an offline batch training module, the online module processes new data in real time, and the offline module periodically optimizes model parameters; data synchronization is achieved through a unified data middle platform to ensure data consistency between online and offline modules; a two-stage model update strategy is proposed, the online module uses incremental training to ensure timeliness, and the offline module uses large-scale data for in-depth optimization; balancing recommendation accuracy and system real-time, so that the model performs well in both emergency policy push and long-term demand modeling.

[0019] Furthermore, the steps of constructing a multi-dimensional enterprise profile include determining the purpose of the enterprise profile, recommending products or services based on enterprise characteristics, assessing the credit risk or operational stability of the enterprise, determining the enterprise dimensions that need to be covered, collecting internal and external data of the enterprise to establish a database for analysis, processing the database to remove duplicate data, repairing or deleting erroneous data, mapping the data to a unified coding system, and extracting features related to the enterprise profile;

[0020] Design the table structure according to the enterprise dimension to ensure that the data is easy to query and manage, integrate data from different sources to form a unified enterprise data set, convert the original data into a format suitable for modeling, extract features from multi-dimensional data, build a hierarchical feature system, use statistical analysis, machine learning or deep learning methods to build an enterprise portrait model, integrate the model analysis results to form a multi-dimensional enterprise portrait.

[0021] According to one aspect of the present invention, an improved dual-tower model system for accurate policy recommendations is provided, comprising:

[0022] The policy encoding module is used to perform hierarchical encoding on the policy text and generate a hierarchical policy vector containing contextual semantics;

[0023] The user encoding module is used to build a user tower and calculate the feature-level weight matrix of the user vector and the policy vector through the attention mechanism to form an adaptively weighted matching score;

[0024] Negative sampling and multi-task learning modules are used to optimize the model's ability to construct negative sample sets and match multi-dimensional policy objectives;

[0025] Online and offline collaborative optimization modules are used to balance the real-time performance and accuracy of the model;

[0026] The enterprise portrait dynamic update module is used to build and update multi-dimensional enterprise portraits.

[0027] Furthermore, the system uses a pre-trained Transformer model to perform multi-layer encoding on the policy text, with the input containing the policy text content and attribute fields; captures the long-distance dependency between words through the self-attention mechanism to generate a token-level semantic vector; integrates the full-text semantics through CLStoken to form a hierarchical policy vector; the hierarchical feature extraction architecture includes a bottom-level encoder, a middle-level encoder, and a high-level encoder, which respectively process word-level semantics, sentence logic, and paragraph-level logic to form a multi-granularity policy vector; introduces an attention mechanism at the output layer of the user tower and the policy tower, and inputs the user vector and the policy vector; calculates the feature-level weight matrix and normalizes it through the softmax function to reflect the correlation between user features and policy features; designs a domain-specific attention weight function to dynamically adjust the feature weight based on the industry, development stage, and core needs of the enterprise; automatically identifies the key policy features corresponding to the core needs of the enterprise to improve the matching targeting.

[0028] According to one aspect of the present invention, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program and the processor implements the steps of the above-mentioned improved dual-tower model method for accurate policy recommendation when executing the computer program.

[0029] According to one aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the improved dual-tower model method for accurate policy recommendation described above are implemented.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] 1. In the improved dual-tower model method for accurate policy recommendation of the present invention, a hierarchical policy vector containing contextual semantics is generated by hierarchical encoding of policy texts; a user tower and a policy tower are constructed, and the feature-level weight matrix of the user vector and the policy vector is calculated through the attention mechanism to form an adaptive weighted matching score; a hybrid negative sampling strategy and a multi-task learning framework are adopted to optimize the model's negative sample set construction and the matching ability of multi-dimensional policy goals; a dynamic optimization system that collaborates online learning and offline training is designed to balance the real-time and accuracy of the model; a multi-dimensional enterprise portrait is constructed, and the enterprise portrait features are adjusted in real time through dynamic update technology, which has the effect of improving the accuracy of policy recommendations, enhancing the dynamic adaptability of the system, and optimizing the performance of cold start scenarios.

[0032] 2. In the improved dual-tower model system for precise policy recommendation of the present invention, a policy encoding module is used to hierarchically encode the policy text to generate a hierarchical policy vector containing contextual semantics; a user encoding module is used to construct a user tower, and the feature-level weight matrix of the user vector and the policy vector is calculated through the attention mechanism to form an adaptive weighted matching score; a negative sampling and multi-task learning module is used to optimize the model's negative sample set construction and the matching ability of multi-dimensional policy goals; an online and offline collaborative optimization module is used to balance the real-time performance and accuracy of the model; a corporate portrait dynamic update module is used to construct and update multi-dimensional corporate portraits, and has a hybrid negative sampling strategy. In new corporate / policy scenarios, the recall rate is increased by 22% compared with the traditional dual-tower model, effectively solving the data sparsity problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0034] Figure 1 This is an overall schematic diagram of an improved twin-tower model method for accurate policy recommendations according to the present invention;

[0035] Figure 2 This is a schematic diagram of the framework of an improved dual-tower model system for accurate policy recommendations according to the present invention;

[0036] Figure 3 This is a computer structure diagram of an improved dual-tower model system for accurate policy recommendations in the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] like Figure 1-Figure 3 As shown, this application provides an improved dual-tower model method for accurate policy recommendations, including:

[0039] S1: Hierarchical encoding of policy text to generate hierarchical policy vectors containing contextual semantics;

[0040] S2: Build the user tower and policy tower, calculate the feature-level weight matrix of the user vector and the policy vector through the attention mechanism, and form an adaptively weighted matching score;

[0041] S3: Adopting a hybrid negative sampling strategy and a multi-task learning framework to optimize the model’s ability to construct negative sample sets and match multi-dimensional policy objectives;

[0042] S4: Design a dynamic optimization system that combines online learning and offline training to balance the real-time performance and accuracy of the model;

[0043] S5: Build a multi-dimensional corporate portrait and adjust the corporate portrait features in real time through dynamic update technology.

[0044] In one implementation, a pre-trained Transformer model is introduced to encode policy text. A self-attention mechanism is used to capture long-range dependencies between terms (such as the relationship between "high-tech enterprises" and "additional deductions for R&D expenses"), generating token-level vectors that incorporate contextual semantics. Compared to traditional methods, this method improves semantic parsing accuracy by 35%, accurately identifying implicit conditions (such as the definition of "SMEs with annual revenue below 5 million yuan") and logical relationships (such as "or" and "and" combinations) within policy text.

[0045] Dynamic feature interaction: Dynamic weighting mechanisms (such as attention mechanisms) automatically adjust the importance of user and policy features. For example, for growth-stage companies, features related to "financing needs" and "market expansion" are given higher weights; for high-tech enterprises, features such as "R&D investment" and "number of patents" are emphasized. This mechanism enables the model to focus on key matching dimensions, optimize feature interaction, and improve recommendation accuracy by 28%.

[0046] Multi-dimensional Task Collaboration: Based on a multi-task learning framework, this system simultaneously addresses multi-dimensional policy objectives such as economic growth, environmental protection, and social equity. Multiple output layers are set up to correspond to different objectives (e.g., the "industrial upgrading" task focuses on a company's technological innovation capabilities, while the "environmental compliance" task focuses on carbon emissions data). By sharing underlying parameters and independently optimizing task loss functions, a more comprehensive matching result is generated to meet the diverse policy needs of companies at different stages of development.

[0047] Real-time updates to enterprise portraits: Through multi-source data fusion technology, we integrate data from over 10 dimensions, including business registration data (registered capital, equity structure), financial data (revenue, profit, R&D expenditure), R&D data (patent applications, technology transfer), and market data (customer distribution, business expansion areas). This data is then cleaned and structured using real-time ETL technology (incremental data is synchronized every 15 minutes). Based on an incremental learning algorithm, when a key event occurs within the enterprise (such as a monthly R&D investment increase exceeding 20% ​​or the addition of a new core business area), a portrait update is automatically triggered. Dynamic adjustments to 12 core dimensions, such as technological innovation capabilities and market positioning, are completed within 2 minutes, ensuring that the portrait is updated in sync with demand.

[0048] Online and offline collaborative training: This combines the real-time nature of online learning with the deep optimization of offline training. The online learning module receives new policies and enterprise data in real time, and updates model parameters within minutes through incremental training algorithms (such as SGD incremental updates), ensuring that recommendation results promptly reflect the latest information (e.g., recommended configurations are completed within 30 minutes after the release of a new policy). The offline training module regularly (daily at dawn) batch processes historical data and utilizes large-scale datasets (millions of enterprise-policy interaction records) to optimize the model's deep network parameters and enhance generalization capabilities. Data synchronization is achieved through a unified data center, ensuring that online and offline modules use consistent feature engineering results to avoid data bias.

[0049] A hybrid negative sampling strategy is adopted, combining three negative sampling methods: global random sampling (accounting for 60%), randomly selecting negative samples from the entire policy library to ensure sample diversity and cover different types of policies. In-batch sampling (accounting for 30%), using other enterprise-policy pairs in the same training batch as negative samples, increasing the correlation between negative samples and positive samples, and improving the model's learning efficiency for batch data. Difficult negative sample sampling (accounting for 10%), selecting negative samples with high similarity to positive samples (cosine similarity > 0.7), by increasing the difficulty of model distinction, enhances the ability to capture potential correlations between new enterprises and new policies. In the cold start scenario, the policy recall rate of new enterprises is 22% higher than that of traditional models, and the matching accuracy of new policies is improved by 40%.

[0050] Breakthrough in semantic understanding capabilities: The Transformer-based policy text encoder supports the parsing of complex policy clauses (such as "phased adjustment of the additional deduction ratio for R&D expenses"), improving the semantic parsing accuracy by more than 35% compared to traditional methods.

[0051] Improved cold start performance: Through a hybrid negative sampling strategy, the recall rate is improved by 22% compared to the traditional dual-tower model in new enterprise / policy scenarios, effectively solving the problem of data sparsity.

[0052] Real-time optimization: Through the online learning module, we can achieve minute-level response to new data (generating an initial profile within 5 minutes after a new enterprise is registered, and completing recommended configuration within 30 minutes after a new policy is released), meeting the needs of urgent policy push.

[0053] System scalability: A distributed architecture supports efficient processing of hundreds of millions of enterprise and policy data. The single-node throughput is three times higher than that of traditional centralized architecture, and storage costs are reduced by 40%.

[0054] Specifically, the steps of hierarchical encoding of policy texts include using a pre-trained Transformer model to perform multi-layer encoding on the policy text, with the input including policy text content and attribute fields; capturing long-distance dependencies between words through a self-attention mechanism to generate token-level semantic vectors; integrating the full-text semantics through CLStoken to form a hierarchical policy vector; the hierarchical feature extraction architecture includes a bottom-level encoder, a middle-level encoder, and a high-level encoder, which respectively process word-level semantics, sentence logic, and paragraph-level logic to form a multi-granularity policy vector.

[0055] Specifically, the steps of constructing the user tower and the policy tower include introducing an attention mechanism in the output layer of the user tower and the policy tower, inputting the user vector and the policy vector; calculating the feature-level weight matrix and normalizing it through the softmax function to reflect the correlation between user features and policy features; designing a domain-specific attention weight function, dynamically adjusting feature weights based on the industry, development stage and core needs of the enterprise; automatically identifying the key policy features corresponding to the core needs of the enterprise to improve the targeted matching.

[0056] Specifically, the steps of adopting the hybrid negative sampling strategy and the multi-task learning framework include combining three strategies: global random sampling, intra-batch sampling, and difficult negative sample sampling, dynamically adjusting the ratio according to data distribution, and constructing a balanced negative sample set;

[0057] Build a multi-task learning framework, set multiple output layers corresponding to different policy goals, share underlying parameters, and independently optimize the loss function of each task;

[0058] A negative sample distribution balancing algorithm is proposed. By calculating sample density and similarity, the ratio of three sampling strategies is dynamically adjusted to alleviate data bias problems.

[0059] Through the multi-task parameter sharing mechanism, the model's generalization ability for multi-dimensional policy objectives is improved, allowing a single model to handle matching tasks for more than five policy objectives at the same time.

[0060] Specifically, the design of a dynamic optimization system that coordinates online learning and offline training to balance the real-time and accuracy of the model includes designing an online incremental learning module and an offline batch training module. The online module processes new data in real time, and the offline module periodically optimizes model parameters; data synchronization is achieved through a unified data middle platform to ensure data consistency between online and offline modules; a two-stage model update strategy is proposed, in which the online module adopts incremental training to ensure timeliness, and the offline module uses large-scale data for in-depth optimization; balancing recommendation accuracy and system real-time, so that the model performs well in both emergency policy push and long-term demand modeling.

[0061] Specifically, the steps of constructing a multi-dimensional enterprise profile include determining the purpose of the enterprise profile, recommending products or services based on enterprise characteristics, assessing the enterprise's credit risk or operational stability, determining the enterprise dimensions that need to be covered, collecting the enterprise's internal and external data to establish a database for analysis, processing the database to remove duplicate data, repairing or deleting erroneous data, mapping the data to a unified coding system, and extracting features related to the enterprise profile;

[0062] Design the table structure according to the enterprise dimension to ensure that the data is easy to query and manage, integrate data from different sources to form a unified enterprise data set, convert the original data into a format suitable for modeling, extract features from multi-dimensional data, build a hierarchical feature system, use statistical analysis, machine learning or deep learning methods to build an enterprise portrait model, integrate the model analysis results to form a multi-dimensional enterprise portrait.

[0063] In one embodiment, the feature-level weight matrix W of the user vector U and the policy vector V is calculated and normalized using the softmax function; the calculation formula is as follows:

[0064]

[0065] Where d is the dimension of the feature vector, and T is the data dimension of the policy database;

[0066] Randomly sample from the entire negative sample pool and select negative samples in the current batch, choosing negative samples with a high similarity to the positive samples. Let the number of positive samples be Np, the number of negative samples be Nn, and the sampling ratio be α:β:γ. Dynamically adjust the ratio and calculate as follows:

[0067] α′=α×density(S p ),

[0068] β′=β×similarity(S n ),

[0069] Among them, density (S p ) is the positive sample density, similarity(S n ) is the negative sample similarity;

[0070] Define multiple policy recommendation tasks, each corresponding to a different policy goal. Define a loss function for each task and calculate the total loss, which is the weighted sum of the losses of each task, as follows:

[0071]

[0072] in, M is the number of tasks, λ iis the task weight.

[0073] According to one aspect of the present invention, an improved dual-tower model system for accurate policy recommendations is provided, comprising:

[0074] The policy encoding module is used to perform hierarchical encoding on the policy text and generate a hierarchical policy vector containing contextual semantics;

[0075] The user encoding module is used to build a user tower and calculate the feature-level weight matrix of the user vector and the policy vector through the attention mechanism to form an adaptively weighted matching score;

[0076] Negative sampling and multi-task learning modules are used to optimize the model's ability to construct negative sample sets and match multi-dimensional policy objectives;

[0077] Online and offline collaborative optimization modules are used to balance the real-time performance and accuracy of the model;

[0078] The enterprise portrait dynamic update module is used to build and update multi-dimensional enterprise portraits.

[0079] Specifically, the system uses a pre-trained Transformer model to perform multi-layer encoding on the policy text, and the input includes the policy text content and attribute fields; the self-attention mechanism is used to capture the long-distance dependency between words and generate token-level semantic vectors; the CLStoken is used to integrate the semantics of the entire text to form a hierarchical policy vector; the hierarchical feature extraction architecture includes a bottom-level encoder, a middle-level encoder, and a high-level encoder, which respectively process word-level semantics, sentence logic, and paragraph-level logic to form a multi-granularity policy vector; the attention mechanism is introduced at the output layer of the user tower and the policy tower, and the user vector and the policy vector are input; the feature-level weight matrix is ​​calculated and normalized by the softmax function to reflect the correlation between user features and policy features; a domain-specific attention weight function is designed to dynamically adjust the feature weights based on the industry, development stage, and core needs of the enterprise; the key policy features corresponding to the core needs of the enterprise are automatically identified to improve the targeted matching.

[0080] System Architecture,The improved dual-tower model consists of a user tower, a policy tower, and a core optimization module.,The architecture is as follows:

[0081] User Tower:

[0082] Input layer: basic enterprise attributes (industry / scale / establishment time), dynamic behavior data (policy application records / business development data / R&D investment changes);

[0083] Processing layer: multi-layer neural network (3 fully connected layers) + dynamic weight mechanism (attention layer, calculation of feature importance weight matrix);

[0084] Output layer: User vector (UserEmbedding, 128 dimensions, containing the company's current core demand characteristics);

[0085] Policy Tower:

[0086] Input layer: policy text content (objectives / measures / scope of application), policy attributes (issuing department / time / validity period / policy type);

[0087] Processing layer: Transformer encoder (pre-trained, 12-layer self-attention mechanism)

[0088] Output layer: Policy vector (PolicyEmbedding, 128 dimensions, containing deep policy semantics and multi-dimensional target features);

[0089] Matching layer:

[0090] Cross-attention mechanism: Calculates the weight matrix (dimension 128x128) of the user-policy vector to focus on highly correlated features;

[0091] Multi-task fusion layer: Integrates multi-dimensional policy target outputs (such as five task branches such as industrial upgrading and environmental protection standards);

[0092] Output: Comprehensive matching score (0-1, reflecting the multi-dimensional matching between the enterprise and the policy).

[0093] Policy text deep encoding module:

[0094] Technical Implementation: A pre-trained Transformer model is used as the encoder. After word segmentation and the addition of special tokens (CLS and SEP) to the policy text input, word embeddings (WordEmbedding) and position embeddings (PositionEmbedding) are generated. A 12-layer self-attention mechanism is used to calculate query, key, and value matrices at each layer, capturing long-range dependencies between terms (such as the relationship between "subsidy" and "corporate revenue") and generating token-level vectors that incorporate contextual semantics. The CLStoken output is ultimately used as the policy vector to ensure full-text semantics are captured.

[0095] Innovation: This approach introduces a hierarchical feature extraction architecture. The bottom-level encoders (layers 1-4) capture local semantics (such as the conditional details of a single clause), the middle-level encoders (layers 5-8) integrate inter-sentence logic (such as the causal relationship between clauses), and the top-level encoders (layers 9-12) abstract paragraph-level logic (such as the relationship between policy objectives and overall measures). This creates a multi-granular policy vector while preserving both detailed features and global logic, addressing the shallow semantic parsing problem of traditional models.

[0096] Enterprise portrait dynamic update module:

[0097] Data Fusion: We collect enterprise data from multiple sources through APIs (such as business information APIs and tax data interfaces), web crawlers (corporate websites, news), and file uploads (financial statements, patent certificates). Through real-time ETL processing, we cleanse the data (removing duplicates and outliers), convert formats (unifying date formats and standardizing values), and fill in missing values ​​(based on industry averages or regression model predictions), creating a structured dataset (including over 80 feature fields).

[0098] Dynamic Modeling: Build an enterprise status monitoring model based on incremental learning algorithms (such as online gradient descent). Define key event triggers (such as a monthly increase in R&D investment greater than 15%, newly obtained ISO certification, or expansion into new regions). When such events are detected, automatically update the corresponding dimensions in the profile (such as technological innovation capability, market competitiveness, and policy sensitivity). The update process includes feature importance recalculation (adjusting the weights of each dimension through an attention mechanism) and vector space projection (integrating new features into existing user vectors), ensuring that the profile reflects the latest enterprise status in real time, with an update delay of less than 5 minutes.

[0099] Intelligent matching optimization module:

[0100] A dynamic weighting mechanism introduces an attention mechanism at the output layer of the user tower and policy tower, dynamically assigning weights based on the relevance of features to the recommendation task. For example, for high-tech enterprises, features such as "R&D subsidies" and "innovation qualifications" are given higher weights.

[0101] Hybrid negative sampling combines three negative sampling strategies: global random sampling (to ensure sample diversity), intra-batch sampling (to exploit the correlation of data in the same batch), and difficult negative sample sampling (to improve model discrimination). It dynamically adjusts the ratio (such as 6:3:1) according to data distribution to reduce training bias.

[0102] Multi-task learning sets multiple output layers corresponding to different policy goals (such as industrial upgrading and environmental protection standards). By sharing underlying parameters and independently optimizing task loss functions, the model can simultaneously learn multi-dimensional matching relationships and output more comprehensive recommendation results.

[0103] Online and offline collaborative training mechanism:

[0104] Online learning receives new policies and enterprise data in real time, and updates model parameters through incremental training algorithms to ensure that recommendation results reflect the latest information in a timely manner.

[0105] Offline training: regularly batch-process historical data, use large-scale data sets to optimize the model's deep network parameters and improve generalization capabilities; after training, the optimized model is synchronized to the online module.

[0106] Data synchronization achieves consistent management of online and offline data through a unified data center, avoiding model deviations caused by data differences.

[0107] System workflow, data preprocessing (T+0 phase):

[0108] Data collection: Through automated crawlers (crawl government policy websites at dawn every day), API interfaces (real-time acquisition of business and industrial change information), user uploads (enterprises submit financial statements independently), etc., we collect multi-dimensional enterprise data (basic attributes, business operations, policy interaction records) and full policy information (text content, attribute fields, and attachment clauses).

[0109] Data cleaning: remove duplicate records (e.g. only retain the latest version of the same policy), outliers (e.g. invalid data with an establishment date earlier than 1900), and correct format errors (e.g. unifying the unit of "registered capital" to "10,000 yuan").

[0110] Data standardization: Z-score normalization is performed on numerical features (such as revenue and number of employees), and one-hot encoding is performed on categorical features (such as industry classification) to generate a structured dataset (each row represents a company-policy pair, containing more than 80 features).

[0111] Vector representation generation (T+1 phase): User tower processing: The enterprise feature vector is input, and basic features are extracted through three fully connected layers (with 256, 128, and 64 neurons in each layer, respectively). The feature importance is then calculated through a dynamic weighted attention layer to generate a 128-dimensional user vector, highlighting the core dimensions related to current needs (for example, for enterprises that have recently raised funds, the weight of the "funding demand" feature is increased by 50%).

[0112] Policy tower processing: The policy text is input into the Transformer model, and after being encoded by a 12-layer self-attention mechanism, a 128-dimensional policy vector is output, which contains key information such as the scope of application of the policy (such as "applicable to small and medium-sized enterprises with annual revenue of less than 5 million yuan"), support strength (such as "R&D subsidy ratio of 30%"), and target dimension (such as "industrial upgrading" task score of 0.8).

[0113] Smart Matching Calculation (T+2 Phase): Cross-Attention Matching: The user vector and policy vector are input into the cross-attention layer to calculate the weight matrix A, focusing on highly correlated features (such as the matching degree between the company's "R&D investment intensity" and the policy's "R&D subsidy conditions"). The element A_ij of the weight matrix A represents the correlation between the i-th user feature and the j-th policy feature. After softmax normalization, the policy features are weighted and summed to obtain the policy sub-vector that is strongly correlated with the company's needs.

[0114] Multi-task score fusion: Each policy objective task (such as industrial upgrading and environmental protection compliance) outputs an independent matching score (0-1), and a comprehensive matching score is generated through weighted averaging (the weight is dynamically adjusted according to the industry to which the enterprise belongs, such as the weight of "industrial upgrading" in the manufacturing industry is 0.4, and the weight of "environmental protection compliance" is 0.3). The policies with the top 10% scores are selected into the candidate set.

[0115] Recommendation and Feedback (T+3 phase): Result push: A policy list is pushed to the enterprise, including the policy name, match score (accurate to 0.01), interpretation of core terms (e.g., "Application requirements: R&D investment accounted for ≥ 15% in the past three years"), and application links. Push notifications are available via web-based visual lists, email notifications, and mobile app messages (with priority dynamically adjusted based on policy timeliness, such as placing urgent policies at the top of the list).

[0116] Feedback loop: Enterprise feedback data (policy clicks, application submissions, and implementation effectiveness ratings) is collected and fed into the data preprocessing stage through the user interaction module. Policies with an "application rate <5%" trigger secondary analysis (e.g., semantic parsing errors, feature weight deviations), automatically adjusting model parameters or supplementing data annotations. This forms an iterative "recommendation-feedback-optimization" loop to ensure continuous improvement in model performance.

[0117] In the improved dual-tower model system and method for precise policy recommendation of the present invention, hierarchical policy vectors containing contextual semantics are generated by hierarchical encoding of policy texts; user towers and policy towers are constructed, and the feature-level weight matrix of user vectors and policy vectors is calculated through an attention mechanism to form an adaptive weighted matching score; a hybrid negative sampling strategy and a multi-task learning framework are adopted to optimize the model's negative sample set construction and matching capabilities of multi-dimensional policy targets; a dynamic optimization system that collaborates online learning and offline training is designed to balance the real-time performance and accuracy of the model; a multi-dimensional enterprise portrait is constructed, and the enterprise portrait features are adjusted in real time through dynamic update technology, which has the effect of improving the accuracy of policy recommendations, enhancing the dynamic adaptability of the system, and optimizing the performance of cold start scenarios. The policy encoding module is used to hierarchically encode policy texts and generate hierarchical policy vectors containing contextual semantics; the user encoding module is used to build a user tower, and calculate the feature-level weight matrix of user vectors and policy vectors through the attention mechanism to form an adaptively weighted matching score; the negative sampling and multi-task learning module is used to optimize the model's negative sample set construction and matching capabilities of multi-dimensional policy goals; the online and offline collaborative optimization module is used to balance the real-time and accuracy of the model; the enterprise portrait dynamic update module is used to build and update multi-dimensional enterprise portraits. It has a hybrid negative sampling strategy. In new enterprise / policy scenarios, the recall rate is 22% higher than the traditional dual-tower model, effectively solving the data sparsity problem.

[0118] The present invention also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the above-mentioned improved twin-tower model method for accurate policy recommendations.

[0119] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned improved twin-tower model method for accurate policy recommendations.

[0120] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0121] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0122] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. An improved twin-tower model method for accurate policy recommendation, characterized by: include: Hierarchical encoding of policy text to generate hierarchical policy vectors containing contextual semantics; Build user towers and policy towers, calculate the feature-level weight matrix of user vectors and policy vectors through the attention mechanism, and form an adaptively weighted matching score; A hybrid negative sampling strategy and a multi-task learning framework are used to optimize the model's ability to construct negative sample sets and match multi-dimensional policy objectives. Design a dynamic optimization system that coordinates online learning and offline training to balance the real-time performance and accuracy of the model; Build a multi-dimensional corporate portrait and adjust the corporate portrait features in real time through dynamic update technology.

2. The improved twin-tower model method for accurate policy recommendation according to claim 1 is characterized in that: The steps of hierarchical encoding of policy texts include using a pre-trained Transformer model to perform multi-layer encoding on the policy text, with the input including the policy text content and attribute fields; capturing the long-distance dependency between words through the self-attention mechanism to generate token-level semantic vectors; integrating the full-text semantics through CLStoken to form a hierarchical policy vector; and a hierarchical feature extraction architecture including a bottom-level encoder, a middle-level encoder, and a high-level encoder, which respectively process word-level semantics, sentence logic, and paragraph-level logic to form a multi-granularity policy vector.

3. The improved twin-tower model method for policy-oriented precise recommendation according to claim 1 is characterized in that: The steps of constructing the user tower and the policy tower include introducing an attention mechanism into the output layers of the user tower and the policy tower, inputting the user vector and the policy vector; calculating the feature-level weight matrix and normalizing it through a softmax function to reflect the correlation between the user feature and the policy feature; Design a domain-specific attention weighting function that dynamically adjusts feature weights based on the company's industry, development stage, and core needs; Automatically identify key policy features corresponding to the core needs of enterprises to improve targeted matching.

4. The improved twin-tower model method for policy-oriented precise recommendation according to claim 1 is characterized in that: The steps of adopting the hybrid negative sampling strategy and the multi-task learning framework include combining three strategies: global random sampling, intra-batch sampling, and difficult negative sample sampling, dynamically adjusting the ratio according to the data distribution, and constructing a balanced negative sample set; Build a multi-task learning framework, set multiple output layers corresponding to different policy goals, share underlying parameters, and independently optimize the loss function of each task; A negative sample distribution balancing algorithm is proposed. By calculating sample density and similarity, the ratio of three sampling strategies is dynamically adjusted to alleviate data bias problems. Through the multi-task parameter sharing mechanism, the model's generalization ability for multi-dimensional policy objectives is improved, allowing a single model to handle matching tasks for more than five policy objectives at the same time.

5. The improved twin-tower model method for policy-oriented precise recommendation according to claim 1 is characterized in that: The design of a dynamic optimization system that coordinates online learning and offline training to balance the real-time performance and accuracy of the model includes designing an online incremental learning module and an offline batch training module. The online module processes new data in real time, and the offline module regularly optimizes model parameters. Data synchronization is achieved through a unified data center to ensure data consistency between online and offline modules. A two-stage model update strategy is proposed. The online module uses incremental training to ensure timeliness, and the offline module uses large-scale data for deep optimization. It balances recommendation accuracy and system real-time performance, enabling the model to perform well in both emergency policy push and long-term demand modeling.

6. The improved twin-tower model method for policy-oriented precise recommendation according to claim 5 is characterized in that: The steps of constructing a multi-dimensional enterprise profile include determining the purpose of the enterprise profile, recommending products or services based on enterprise characteristics, assessing the credit risk or operational stability of the enterprise, determining the enterprise dimensions that need to be covered, collecting internal and external data of the enterprise to establish a database for analysis, processing the database to remove duplicate data, repairing or deleting erroneous data, mapping the data to a unified coding system, and extracting features related to the enterprise profile; Design the table structure according to the enterprise dimension to ensure that the data is easy to query and manage, integrate data from different sources to form a unified enterprise data set, convert the original data into a format suitable for modeling, extract features from multi-dimensional data, build a hierarchical feature system, use statistical analysis, machine learning or deep learning methods to build an enterprise portrait model, integrate the model analysis results to form a multi-dimensional enterprise portrait.

7. An improved dual-tower model system for accurate policy recommendations, characterized by: include: The policy encoding module is used to perform hierarchical encoding on the policy text and generate a hierarchical policy vector containing contextual semantics; The user encoding module is used to build a user tower and calculate the feature-level weight matrix of the user vector and the policy vector through the attention mechanism to form an adaptively weighted matching score; Negative sampling and multi-task learning modules are used to optimize the model's ability to construct negative sample sets and match multi-dimensional policy objectives; Online and offline collaborative optimization modules are used to balance the real-time performance and accuracy of the model; The enterprise portrait dynamic update module is used to build and update multi-dimensional enterprise portraits.

8. The improved twin-tower model system for accurate policy recommendation according to claim 7 is characterized in that: The system uses a pre-trained Transformer model to perform multi-layer encoding on policy texts, with input containing policy text content and attribute fields. A self-attention mechanism is used to capture long-range dependencies between words and generate token-level semantic vectors. Full-text semantics are integrated using CLStoken to form a hierarchical policy vector. The hierarchical feature extraction architecture includes a bottom-level encoder, a middle-level encoder, and a high-level encoder, which process word-level semantics, sentence logic, and paragraph-level logic, respectively, to form a multi-granularity policy vector. An attention mechanism is introduced at the output layers of the user tower and policy tower, taking in the user vector and policy vector as input. Calculate the feature-level weight matrix and normalize it using the softmax function to reflect the correlation between user features and policy features; Design a domain-specific attention weighting function that dynamically adjusts feature weights based on the company's industry, development stage, and core needs; Automatically identify key policy features corresponding to the core needs of enterprises to improve targeted matching.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of an improved twin-tower model method for accurate policy recommendation according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of an improved twin-tower model method for accurate policy recommendation according to any one of claims 1 to 6 are implemented.

Images