A project declaration guide intelligent generation method based on atlas enhancement and reflection reconstruction

Abstract

This invention discloses an intelligent method for generating project application guidelines through graph enhancement and reflective reconstruction. It addresses the problems of homogenized project application guide generation and the lack of prior constraints and closed-loop verification in large models. This invention proposes an algorithmic architecture that integrates hard graph constraints, multi-model reflection, and reinforcement learning. Specifically, this invention first parses cutting-edge multi-source data to construct a structured prior knowledge base; then, it combines joint states and demonstration samples to drive a decision engine to map execution plans, thereby retrieving graph feature constraints and generating preliminary project application guidelines; next, it calls the Actor-Critic dual model to perform compliance verification, and if the verification fails, it forces reconstruction based on mapping error feedback; finally, it relies on reinforcement learning to update parameters and establish a long-term feedback loop. This invention significantly improves the compliance and forward-looking nature of generated project application guidelines, effectively overcoming the knowledge illusion of large models.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing, specifically to a method for intelligently generating project application guidelines using graph enhancement and reflective reconstruction. Background Technology

[0002] With the digital transformation of government services and scientific research management, the accurate compilation of project application guidelines has become a crucial step. Traditional, manually-driven compilation methods suffer from low information retrieval efficiency and weak semantic connections across documents when faced with massive amounts of heterogeneous macro-policies and historical archives. Existing template-matching-based auxiliary systems often rely excessively on historical archives, resulting in rigid rules, a lack of deep reasoning about complex policy logic, and a high risk of generating guidelines that become homogenized. This static generation model lacks the keen ability to capture cutting-edge developments in top journals like *Nature* and *Science*, as well as macro-level constraints, making it difficult for guidelines to play a forward-looking and guiding role in scientific and technological innovation.

[0003] In recent years, while large language models have demonstrated powerful text generation capabilities, they have revealed significant limitations in applications within specialized vertical fields. Due to the lack of rigid constraints from structured prior knowledge, generative models are prone to knowledge illusions, leading to factual discrepancies in policy alignment and terminology in the output application guidelines. Furthermore, existing generative large model frameworks generally adopt a unidirectional open-loop generation mode, lacking a rigorous post-validation closed loop based on discriminative rules. Simultaneously, the underlying tool scheduling strategies are often static and fixed, unable to self-evolve and optimize parameters based on historical errors. These shortcomings make it difficult for existing technologies to meet the stringent requirements of high accuracy, absolute compliance, and forward-looking guidance in high-standard government application scenarios. Therefore, there is an urgent need for an automatic guide generation method that integrates knowledge graph constraints, multi-model reflective validation, and reinforcement learning closed-loop optimization to address these technical bottlenecks. Summary of the Invention

[0004] This invention provides an intelligent generation method for project application guidelines through graph enhancement and reflective reconstruction. The method includes: parsing multi-source data to construct a domain knowledge graph and performing semantic encoding using a graph attention mechanism; configuring the output entity embedding feature set as a structured prior knowledge base; fusing graph entity features from real-time interactive text, historical archives, and context matching to perform cross-modal cascaded encoding and projection transformation, mapping the output environment joint state vector; calculating the matching score between the state vector and intent features based on a dynamic demonstration sample library and retrieving the optimal execution strategy, driving the decision engine to map and output a global execution plan sequence according to the feature space distribution; parsing the plan and retrieving local subgraph nodes within the knowledge base to aggregate them into constraint feature vectors, using these as prior constraints to control the decoding engine to iteratively generate a preliminary guide text sequence; calculating the compliance score of the text and rule features using an Actor-Critic dual model, and mapping error feedback features to update the state vector and call back to the decision engine when the score is below the target, forcibly triggering plan reconstruction until compliance is achieved; parsing the scheduling trajectory log and quantifying immediate rewards based on the compliance score and graph similarity measure, calculating the value distribution based on a reinforcement learning architecture and synchronously updating weight parameters, and establishing a long-term policy gradient feedback link through parameter write-back.

[0005] The present invention provides an intelligent generation method for project application guidelines based on graph enhancement and reflective reconstruction, comprising the following steps: S1: Parse multi-source project application data to extract entities and semantic relationships to construct a domain knowledge graph; perform semantic vectorization representation on entity nodes in the domain knowledge graph, and combine the local topological structure of the graph to perform joint feature aggregation to construct a set of entity embedding feature vectors for representing a structured prior knowledge base; S2: Parse the real-time interactive text and long and short-term declaration files, combine the entity feature vectors in the structured prior knowledge base to perform feature concatenation encoding, and construct a multi-dimensional joint state vector to represent the current declaration environment; S3: Using the multidimensional joint state vector as a query variable, the optimal demonstration prompt is matched in the dynamic demonstration sample library, and the decision engine is driven to output a global execution plan sequence containing the tool name and calling parameters through autoregressive mapping. S4: Based on the global execution plan sequence, retrieve the graph constraint features in the structured prior knowledge base, and use the graph constraint features as the prior condition constraint decoding engine to generate a preliminary project application guide text sequence; S5: Call the dual-model verification mechanism to calculate the compliance matching degree between the preliminary project application guide text sequence and the graph rules in the structured prior knowledge base; if the verification fails, generate error feedback features and update the multi-dimensional joint state vector, and call back to the decision engine to trigger the iterative reconstruction of the global execution plan sequence; S6: Extract the scheduling trajectory of the decision engine, quantify the instant reward value based on the compliance verification score and the semantic similarity between text and graph, update the policy parameters of the decision engine based on the deep reinforcement learning network, and establish a long-term policy gradient feedback link. According to a specific implementation of an embodiment of the present invention, the specific steps of S2 are as follows: S21. Analyze the current time. Real-time interactive text input sequences are mapped to generate text semantic feature vectors. Load the long-term and short-term application files stored in the memory layer to obtain the long-term behavioral feature vector. With short-term demand feature vector Tensor splicing and nonlinear spatial mapping operations are performed on cross-temporal features, combined with the archive projection matrix. With bias vector Calculate and output the feature vector of the archive. :

[0006] in, It is a non-linear activation function. This is a tensor concatenation operation; S22. Based on the aforementioned structured prior knowledge base, call the entity feature vectors in the entity embedding feature vector set output by S1 that match the current context. The text semantic feature vector Entity feature vector With archive feature vector Concatenated encoding is performed, and the encoding result is subjected to projection transformation and layer normalization. The state projection matrix is ​​then set. With bias vector The calculation outputs a multidimensional joint state vector characterizing the current reporting environment. :

[0007] in, For a moment The multidimensional joint state vector, For layer normalization function; It should be noted that the “cascaded encoding” mentioned repeatedly in this invention is a conventional feature fusion method in the field of deep learning. Its specific implementation is to physically concatenate multiple independent feature vectors along a unified feature dimension to form a joint feature vector of a higher dimension. According to a specific implementation of an embodiment of the present invention, the specific steps of S3 are as follows: S31. Analyze historical standard project application interaction data and tool scheduling logs to map intent feature vectors. And bind the corresponding tool execution strategy. In this way, a dynamic demonstration sample library is pre-built. Based on the multidimensional joint state vector output by S2 Compared with the intent feature vector in the sample library Set the query projection matrix Key-value feature projection matrix Temperature hyperparameter For the joint state vector and each intention feature vector, latent space inner product similarity calculation and temperature scaling processing are performed to calculate the semantic matching score. :

[0008] in, The normalized semantic matching score. The traversal index parameter in the sample database represents any historical intent feature vector within the sample database. ; S32. Based on the semantic matching score From the dynamic demo sample library Target tool execution strategy for extracting the top-K scores Retrieve and construct the optimal set of demonstration prompts At the same time, load the predefined tool parameter description matrix. ; Set the final output global execution plan sequence as follows In the decision engine's first Each decoding step is based on the context feature vector at the current time. Configure the output transformation weight matrix With bias vector Perform linear mapping and normalization exponent operation on the feature space of the current decoding step to calculate candidate actions. Conditional probability distribution:

[0009] in, The predicted probability of the candidate action. This is the sequence of historical actions generated up to the current decoding step; autoregressive sampling decoding is performed based on the conditional probability distribution, progressively mapping and outputting a global execution plan sequence containing the tool name and calling parameters. ; According to a specific implementation of an embodiment of the present invention, the specific steps of S4 are as follows: S41. Parse the global execution plan sequence output by S3. The mapping generates the corresponding tool parameter feature vector. Based on the aforementioned structured prior knowledge base, and according to Retrieve and construct a set of nodes for local relational subgraphs from the structured prior knowledge base. ; Call the set of nodes in S1 Corresponding entity embedding feature vector Set the feature matching transformation matrix Calculate the parameter feature vector Weights of associations between nodes within the subgraph :

[0010] in, The normalized association distribution weights, The set of nodes in the local associated subgraph The traversal index parameter in the subgraph represents any entity node within that subgraph. S42. Utilize the aforementioned correlation distribution weights Perform a weighted summation operation on the node features of the local associated subgraph to calculate the output graph constraint feature vector. :

[0011] S43, constrain the feature vector of the graph. The multidimensional joint state vector output by S2 Perform concatenated encoding to construct the generated cue vector. ; generate the prompt vector The prior constraints configured for the decoding engine are used in the time step of autoregressive decoding. Based on the current hidden layer feature vectors of the decoding engine Set the vocabulary space mapping matrix With bias vector Calculate target text tags Conditional probability distribution:

[0012] in, The predicted probability of the target text tag. Given the previously generated guide text sequence, perform sampling decoding based on the conditional probability distribution to iteratively output a preliminary project application guide text sequence; According to a specific implementation of an embodiment of the present invention, the specific steps of S5 are as follows: S51. Parse the preliminary project application guide text sequence output by S4 and map it to generate text feature vectors. Based on the aforementioned structured prior knowledge base, the feature vectors corresponding to the entities in the review rules are invoked. Set the transpose weight vector With bias scalar The text feature vector With the feature vector Cascaded encoding is performed, and linear dimensionality reduction projection and nonlinear activation normalization are applied to the concatenated high-dimensional feature space to calculate the compliance verification score. :

[0013] in, The normalized compliance verification score. It is a non-linear activation function; S52. Compare the compliance verification scores. With the set preset threshold ,like If the text compliance check fails, the Actor-Critic dual model is invoked, and the Critic component of the dual model processes the text feature vector. With the feature vector of the audit rules Feature alignment and element-level difference operations are performed to extract text deviation feature distributions that violate the rules of the structured prior knowledge base; and multi-level nonlinear spatial mapping processing is performed on the text deviation feature distributions via the Actor in the dual model to generate error feedback feature vectors. The multidimensional joint state vector constructed by calling S2 ,Will and Perform concatenated encoding and set the mapping transformation matrix. With bias vector For the joint state vector containing dynamic error features, perform affine transformation and hyperbolic tangent space mapping operations to calculate the reconstructed state feature vector. :

[0014] in, To reconstruct the state vector, It is the hyperbolic tangent activation function; S53, reconstruct the state vector The callback input is sent to S3, replacing the original multidimensional joint state vector. The latent space similarity calculation of the intent features is re-executed, and the decision engine is guided to recalculate the conditional probability distribution of candidate actions to force real-time iterative reconstruction of the global execution plan sequence until the verification score of the output text meets the requirements. ; According to a specific implementation of an embodiment of the present invention, the specific steps of S6 are as follows: S61. Analyze the decision scheduling trajectory log consisting of S3 to S5, and extract the input state vector actually adopted by the decision engine. With the corresponding scheduling action Set the state value function parameters With the dominant function parameters Forward evaluation of the input state vector and the scheduled action using a two-stream architecture with a competing deep Q-network, the state value and action advantage features are linearly superimposed and mean-eliminating normalized to calculate the output action value function. :

[0015] in, For inclusion and The overall parameters of the deep reinforcement learning network, For the state value function, For the dominant function, For the set of candidate actions, The dimension of the candidate action set. For the candidate action set The traversal index parameter in the decision space represents any candidate scheduling action within the decision space. S62, Call the final output of S4: Graph constraint feature vector The final text feature vector mapped by S5 And the compliance verification score for those that pass the verification. Set a reward balance coefficient Joint quantization and linear weighted calculation are performed on the global text generation quality and compliance constraints to calculate and output an instant reward value. :

[0016] in, For quantified instant reward value, For vectors, the L2 norm constraint; S63. Based on the state vector of the next moment after the system executes the action. Set discount factor With target network parameters Offline policy objective mapping calculations are performed using the Bellman equation for the long-term expected returns in Markov decision-making processes, and the time-difference objective value is calculated as the output. :

[0017] in, For the time-difference objective value, This represents finding the maximum action value among all possible actions within the action space; S64. Utilize the aforementioned time-difference target value With the current action value function Calculate the network loss function :

[0018] S65. By minimizing the network loss function The gradient is used to update the overall parameters of the deep reinforcement learning network. The updated action value function is used as the evaluation criterion to guide the calculation of the policy gradient, and the transformation weight matrix in S3 is updated accordingly. With bias vector Set soft update coefficient Perform synchronous iterative updates based on a moving average mechanism on the target network parameters:

[0019] in, The target network soft update coefficients are used to establish a long-term policy gradient feedback link that reduces the probability of S5 iteration reconstruction by alternating updates of the above network parameters and policy write-back.

[0020] The beneficial effects of this invention are as follows: First, it effectively solves the problems of knowledge illusion and lack of compliance in the generation process. By constructing a structured prior knowledge base based on multi-source data and using graph constraint features as prior conditions for the decoding engine, combined with the compliance verification and forced reconstruction mechanism of the Actor-Critic dual model, it ensures the absolute accuracy of the generated project application guidelines in terms of policy alignment and professional terminology, and overcomes the technical defects of large language models in terms of unstable output in vertical fields.

[0021] Secondly, it breaks through the homogenization bottleneck of traditional guideline compilation models. By introducing high-impact academic literature and macro-level guiding constraints to construct a domain knowledge graph, and capturing technological evolution trends and relationships through graph attention mechanisms, the system can achieve a keen perception of cutting-edge dynamics from the underlying logic, resulting in guidelines with stronger technological foresight and innovative leadership value.

[0022] Third, it achieves autonomous evolution and long-term closed-loop of the system's decision-making logic. Based on the policy gradient feedback link established by deep reinforcement learning, the parameters of the decision engine can be automatically optimized according to the scheduling trajectory log and the verification score. The parameter write-back mechanism reduces the probability of triggering reconstruction in the subsequent generation process, enabling the system to have the evolutionary ability to self-correct based on historical deviations.

[0023] Fourth, it significantly enhances the ability to perform deep semantic representation and adaptation of complex, multi-source, and heterogeneous data. By employing concatenated coding techniques to integrate real-time interactive text, long-term and short-term application archives, and context-matched entity features, a multi-dimensional joint state space is constructed, enabling the decision-making process for generating guidelines to be deeply compatible with application environments and personalized needs across different industries and levels. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method; Figure 2 This is the architecture diagram of this method. Detailed Implementation

[0025] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to examples and accompanying drawings.

[0026] As attached Figure 1 and attached Figure 2 As shown, a method for intelligently generating project application guidelines based on graph enhancement and reflective reconstruction includes the following steps: Step 1: Analyze multi-source project application data, extract entities and semantic relationships, and construct a domain knowledge graph. In practice, data sources are precisely limited to macro-level guiding and constraint outlines, digital frontier technology literature databases of high-impact top-tier academic journals, top-level planning macro-policy documents, provincial science and technology department application guidelines from the past five years, archives of completed historical approved projects, and standardized application form templates. A domain knowledge graph is constructed based on the above data. ;in, It is a set of entity nodes, containing cutting-edge technology nodes extracted from the aforementioned cutting-edge technology literature database. It is a collection of semantic relation types, containing evolutionary relations that represent the trends of technological evolution. Given a set of directed edges, topological units are represented as triples. ,in For the head entity node, To connect the semantic relationships between the head and tail nodes, This is the tail entity node;

[0027] A pre-trained language model encoder based on the Transformer architecture with 12 hidden layers and 768 output dimensions is loaded to perform semantic vectorization representation on the entity nodes, generating initial feature vectors. Based on the feature vector of the first hidden state and The relationship projection matrix is ​​set to have a dimension of 768 by 768. With the transformation weight matrix Adaptive focusing operations are performed on contextual features to calculate feature attention weight coefficients. The specific formula is as follows:

[0028] in, To normalize the weights, For the transpose of the shared weight vector, The LeakyReLU nonlinear activation function is used. For nodes The set of neighboring nodes, For set The traversal index parameter in the code represents the center node. any local neighbor node; Using the weighting coefficients A weighted aggregation operation is performed on the features of neighboring nodes, and combined with the self-loop transitive parameter matrix. Mapping the current node's own features, performing joint representation with the local topological structure data of the graph and the high-dimensional semantic features of neighboring nodes, and calculating the output of the first node. Hidden state feature vector of layer :

[0029] in, This represents the feature representation of the next layer of nodes. Employ the ELU activation function; set the aggregation iteration depth of the graph neural network. After three layers of iterative calculation, the final entity embedding feature vector set is output. and the domain knowledge graph The topological structure data and the final entity embedding feature vector set are jointly configured to construct a structured prior knowledge base for subsequent controlled generation; Step 2: Parse the semantic features of the real-time interactive text and load the long and short-term declaration files in the memory layer. With the above-mentioned structured prior knowledge base as support, perform tensor concatenation and concatenation encoding on the semantic features, declaration files and entity embedding feature vector set to construct a joint state vector to represent the current declaration environment. Analyze the current moment The real-time interactive text input sequence is mapped to generate a fixed-dimensional text semantic feature vector. Load the long and short-term application files stored in the distributed memory layer to obtain the long-term behavioral feature vector. With short-term demand feature vector Tensor splicing and nonlinear spatial mapping operations are performed on cross-temporal features, combined with the archive projection matrix. With bias vector Calculate and output the feature vector of the archive. :

[0030] in, The GELU nonlinear activation function is used. This is a tensor concatenation operation; Based on the aforementioned structured prior knowledge base, the entity feature vectors that match the current context are called from the entity embedding feature vector set output in step 1. The text semantic feature vector Entity feature vector With archive feature vector Concatenated encoding is performed, and the encoding result is subjected to projection transformation and layer normalization. The state projection matrix is ​​then set. With bias vector The calculation outputs a multidimensional joint state vector characterizing the current reporting environment. :

[0031] in, For a moment The multidimensional joint state vector, The LayerNorm normalization function is used for the LayerNorm layer. The "cascaded encoding" method mentioned in this invention is specifically implemented by physically concatenating multiple independent feature vectors along a unified feature dimension to form a joint feature vector of a higher dimension. Step 3: Use the joint state vector as the retrieval query variable to call the dynamic demonstration sample library, match the optimal demonstration sample based on the feature space distance, and use the sample as a parameter to control the decision engine. Autoregression maps out a global execution plan sequence containing the tool name and calling parameters. Analyze historical standard project application interaction data and tool scheduling logs to map intent feature vectors. And bind the corresponding tool execution strategy. Pre-built dynamic demonstration sample library Based on the multidimensional joint state vector output in step 2 Compared with the intent feature vector in the sample library Set the query projection matrix Key-value feature projection matrix With fixed temperature hyperparameters For the joint state vector and each intention feature vector, latent space inner product similarity calculation and temperature scaling processing are performed to calculate the semantic matching score. :

[0032] in, The normalized semantic matching score. The traversal index parameter in the sample database represents any historical intent feature vector within the sample database. ; Based on the semantic matching score From the dynamic demo sample library Extract the top 5 scoring target tools and execution strategies. Retrieve and construct the optimal set of demonstration prompts At the same time, load the predefined tool parameter description matrix. ; Set the final output global execution plan sequence as follows In the decision engine's first Each decoding step is based on the context feature vector at the current time. Configure the output transformation weight matrix With bias vector Perform linear mapping and normalization exponent operation on the feature space of the current decoding step to calculate candidate actions. Conditional probability distribution:

[0033] in, The predicted probability of the candidate action. This is the sequence of historical actions generated up to the current decoding step; autoregressive sampling decoding is performed based on the conditional probability distribution, progressively mapping and outputting a global execution plan sequence containing the tool name and calling parameters. ; Step 4: Based on the global execution plan sequence, accurately retrieve the corresponding policy constraint features and guideline template features from the structured prior knowledge base, configure the features as prior knowledge input, and strictly constrain the decoding engine to generate a preliminary application guideline text sequence through autoregression; Parse the global execution plan sequence output in step 3. The mapping generates the corresponding tool parameter feature vector. Based on the aforementioned structured prior knowledge base, and according to Retrieve and construct a set of nodes for local relational subgraphs from the structured prior knowledge base. ; Call the node set mentioned in step 1 Corresponding entity embedding feature vector Set the feature matching transformation matrix Calculate the parameter feature vector Weights of associations between nodes within the subgraph :

[0034] in, The normalized association distribution weights, The set of nodes in the local associated subgraph The traversal index parameter in the subgraph represents any entity node within that subgraph. Using the aforementioned associated distribution weights Perform a weighted summation operation on the node features of the local associated subgraph to calculate the output graph constraint feature vector. :

[0035] constrain the feature vector of the map The multidimensional joint state vector output from step 2 Perform concatenated encoding to construct the generated cue vector. ; generate the prompt vector Strong prior constraints configured for a decoding engine incorporating a 128-head attention mechanism; at time steps of autoregressive decoding Based on the current hidden layer feature vectors of the decoding engine Set the vocabulary space mapping matrix With bias vector Calculate target text tags Conditional probability distribution:

[0036] in, The predicted probability of the target text tag. Given the previously generated guide text sequence, perform a bundle search decoding operation based on the conditional probability distribution, set the bundle width to 4, and iteratively output a preliminary project application guide text sequence. Step 5: Call the Actor-Critic dual model, which includes an evaluation network and a generation network, to calculate the compliance matching degree between the initially generated application guide text sequence and the graph rules in the structured prior knowledge base; The preliminary project application guide text sequence output in step 4 is parsed and mapped to generate text feature vectors. Based on the aforementioned structured prior knowledge base, the feature vectors corresponding to the entities in the review rules are invoked. Set the transpose weight vector With bias scalar The text feature vector With the feature vector Cascaded encoding is performed, and linear dimensionality reduction projection and nonlinear activation normalization are applied to the concatenated high-dimensional feature space to calculate the compliance verification score. :

[0037] in, The normalized compliance verification score. The Sigmoid function is used to map the output to the range of 0 to 1; Set extremely strict preset thresholds Compare the compliance verification scores. With the set preset threshold ,like If the text compliance check fails, the Critic network, which is structured as a multilayer perceptron in the dual-model approach, will process the text feature vector. With the feature vector of the audit rules Feature alignment and element-level difference operations are performed to extract text deviation feature distributions that violate the rules of the structured prior knowledge base; and multi-layer nonlinear spatial mapping processing is performed on the text deviation feature distributions via the Actor network in the dual model to generate error feedback feature vectors. ; Call the multidimensional joint state vector constructed in step 2 ,Will and Perform concatenated encoding and set the mapping transformation matrix. With bias vector For the joint state vector containing dynamic error features, perform affine transformation and hyperbolic tangent space mapping operations to calculate the reconstructed state feature vector. :

[0038] in, To reconstruct the state vector, It is the hyperbolic tangent activation function; The reconstructed state vector The callback input is passed to step 3, completely replacing the original multidimensional joint state vector. The latent space similarity calculation of the intent features is re-executed, guiding the decision engine to recalculate the conditional probability distribution of candidate actions, and forcibly triggering the real-time iterative reconstruction of the global execution plan sequence until the verification score of the output text meets the requirements. ; Step 6: Construct a deep reinforcement learning network to calculate the action value distribution of the above tool call operations, measure the instant reward function value based on the feature matching of the final output project application guide text sequence with the entities in the structured prior knowledge base, and iteratively update the weight parameters of the reinforcement learning network based on the temporal difference algorithm; Analyze the decision scheduling trajectory log consisting of steps 3 to 5, and extract the input state vector actually adopted by the decision engine. With the corresponding scheduling action Set the state value function parameters With the dominant function parameters Forward evaluation of the input state vector and the scheduled action using a two-stream architecture with a competing deep Q-network, the state value and action advantage features are linearly superimposed and mean-eliminating normalized to calculate the output action value function. :

[0039] in, For inclusion and The overall parameters of the deep reinforcement learning network, For the state value function, For the dominant function, For the set of candidate actions, The dimension of the candidate action set. For the candidate action set The traversal index parameter in the decision space represents any candidate scheduling action within the decision space. Call the final output of step 4: the graph constraint feature vector Step 5: The final mapped text feature vector And the compliance verification score for those that pass the verification. Set a fixed reward balance coefficient Joint quantization and linear weighted calculation are performed on the global text generation quality and compliance constraints to calculate and output an instant reward value. :

[0040] in, For quantified instant reward value, For vectors, the L2 norm constraint; Based on the state vector of the next moment after the system performs an action. Set the attenuation discount factor With target network parameters Offline policy objective mapping calculations are performed using the Bellman equation for the long-term expected returns in Markov decision-making processes, and the time-difference objective value is calculated as the output. :

[0041] in, For the time-difference objective value, This represents finding the maximum action value among all possible actions within the action space; Using the aforementioned temporal difference target value With the current action value function Calculate the network loss function :

[0042] Minimize the network loss function using the AdamW optimization algorithm. The gradient is used to update the overall parameters of the deep reinforcement learning network. The updated action value function is used as the evaluation criterion to guide the calculation of the policy gradient, and the transformation weight matrix in step 3 is updated accordingly. With bias vector Set the target network soft update coefficient. Perform synchronous iterative updates based on a moving average mechanism on the target network parameters:

[0043] in, To limit the soft update coefficient range of the target network, a long-term policy gradient feedback link is established by alternating updates of the network parameters and policy write-back to reduce the probability of triggering the iterative reconstruction in step 5.

Claims

1. A method for intelligently generating project application guidelines through graph enhancement and reflective reconstruction, characterized in that... Includes the following steps: S1: Parse multi-source project application data to extract entities and semantic relationships to construct a domain knowledge graph; perform semantic vectorization representation on entity nodes in the domain knowledge graph, and combine the local topological structure of the graph to perform joint feature aggregation to construct a set of entity embedding feature vectors for representing a structured prior knowledge base; S2: Parse the real-time interactive text and long and short-term declaration files, combine the entity feature vectors in the structured prior knowledge base to perform feature concatenation encoding, and construct a multi-dimensional joint state vector to represent the current declaration environment; S3: Using the multidimensional joint state vector as a query variable, the optimal demonstration prompt is matched in the dynamic demonstration sample library, and the decision engine is driven to output a global execution plan sequence containing the tool name and calling parameters through autoregressive mapping. S4: Based on the global execution plan sequence, retrieve the graph constraint features in the structured prior knowledge base, and use the graph constraint features as the prior condition constraint decoding engine to generate a preliminary project application guide text sequence; S5: Call the dual-model verification mechanism to calculate the compliance matching degree between the preliminary project application guide text sequence and the graph rules in the structured prior knowledge base; If the verification fails, an error feedback feature is generated and the multidimensional joint state vector is updated, and the decision engine is called back to trigger the iterative reconstruction of the global execution plan sequence. S6: Extract the scheduling trajectory of the decision engine, quantify the instant reward value based on the compliance verification score and the semantic similarity between text and graph, update the policy parameters of the decision engine based on the deep reinforcement learning network, and establish a long-term policy gradient feedback link.

2. The intelligent generation method for project application guidelines based on graph enhancement and reflective reconstruction according to claim 1, characterized in that... The specific method for step S1 is as follows: The analysis of multi-source project application data includes macro-level binding planning texts, cutting-edge technology literature databases (including top academic journals, leading laboratory research reports, and industry white papers), macro-level policy documents, application guidelines from previous years, archives of historically approved projects, and standardized application form templates, in order to construct a domain knowledge graph. ;in, It is a set of entity nodes, containing cutting-edge technology nodes extracted from the aforementioned cutting-edge technology literature database. It is a collection of semantic relation types, containing evolutionary relations that represent the trends of technological evolution. Given a set of directed edges, topological units are represented as triples. ,in For the head entity node, To connect the semantic relationships between the head and tail nodes, For the tail entity node; load the initial feature vector generated by the semantic vectorization representation of the entity node by the pre-trained language model encoder based on the Transformer architecture. Based on the feature vector of the first hidden state and Set the relational projection matrix With the transformation weight matrix Adaptive focusing operations are performed on contextual features to calculate feature attention weight coefficients. The specific formula is as follows: ； in, To normalize the weights, For the transpose of the shared weight vector, It is a non-linear activation function. For nodes The set of neighboring nodes, For set The traversal index parameter in the code represents the center node. any local neighbor node; Using the weighting coefficients A weighted aggregation operation is performed on the features of neighboring nodes, and combined with the self-loop transitive parameter matrix. Mapping the current node's own features, performing joint representation with the local topological structure data of the graph and the high-dimensional semantic features of neighboring nodes, and calculating the output of the first node. Hidden state feature vector of layer : ； in, This represents the feature representation of the next layer of nodes. The activation function; after Layer-by-layer iterative computation outputs the final set of entity embedding feature vectors. and the domain knowledge graph The topological structure data and the final entity embedding feature vector set are jointly configured to construct a structured prior knowledge base for subsequent controlled generation.

3. The intelligent generation method for project application guidelines based on graph enhancement and reflective reconstruction according to claim 1, characterized in that... The specific method for step S2 is as follows: S21. Analyze the current time. Real-time interactive text input sequences are mapped to generate text semantic feature vectors. Load the long-term and short-term application files stored in the memory layer to obtain the long-term behavioral feature vector. With short-term demand feature vector Tensor splicing and nonlinear spatial mapping operations are performed on cross-temporal features, combined with the archive projection matrix. With bias vector Calculate and output the feature vector of the archive. : ； in, It is a non-linear activation function. This is a tensor concatenation operation; S22. Based on the aforementioned structured prior knowledge base, call the entity feature vectors in the entity embedding feature vector set output by S1 that match the current context. The text semantic feature vector Entity feature vector With archive feature vector Concatenated encoding is performed, and the encoding result is subjected to projection transformation and layer normalization. The state projection matrix is ​​then set. With bias vector The calculation outputs a multidimensional joint state vector characterizing the current reporting environment. : ； in, For a moment The multidimensional joint state vector, For layer normalization function; It should be noted that the "cascaded encoding" mentioned repeatedly in this invention is a conventional feature fusion method in the field of deep learning. Its specific implementation is to physically concatenate multiple independent feature vectors along a unified feature dimension to form a joint feature vector of a higher dimension.

4. The intelligent generation method for project application guidelines based on graph enhancement and reflective reconstruction according to claim 1, characterized in that... The specific method in step S3 is as follows: S31. Analyze historical standard project application interaction data and tool scheduling logs to map intent feature vectors. And bind the corresponding tool execution strategy. In this way, a dynamic demonstration sample library is pre-built. ; Based on the multidimensional joint state vector output by S2 Compared with the intent feature vector in the sample library Set the query projection matrix Key-value feature projection matrix Temperature hyperparameter For the joint state vector and each intention feature vector, latent space inner product similarity calculation and temperature scaling processing are performed to calculate the semantic matching score. : ； in, The normalized semantic matching score. The traversal index parameter in the sample database represents any historical intent feature vector within the sample database. ; S32. Based on the semantic matching score From the dynamic demo sample library Target tool execution strategy for extracting the top-K scores Retrieve and construct the optimal set of demonstration prompts At the same time, load the predefined tool parameter description matrix. ; Set the final output global execution plan sequence as follows In the decision engine Each decoding step is based on the context feature vector at the current time. Configure the output transformation weight matrix With bias vector Perform linear mapping and normalization exponent operation on the feature space of the current decoding step to calculate candidate actions. Conditional probability distribution: ； in, The predicted probability of the candidate action. This is the sequence of historical actions generated up to the current decoding step; autoregressive sampling decoding is performed based on the conditional probability distribution, progressively mapping and outputting a global execution plan sequence containing the tool name and calling parameters. .

5. The intelligent generation method for project application guidelines based on graph enhancement and reflective reconstruction according to claim 1, characterized in that... The specific method in step S4 is as follows: S41. Parse the global execution plan sequence output by S3. The mapping generates the corresponding tool parameter feature vector. ; Based on the aforementioned structured prior knowledge base, and according to Retrieve and construct a set of nodes for local relational subgraphs from the structured prior knowledge base. ; Call the set of nodes in S1 Corresponding entity embedding feature vector Set the feature matching transformation matrix Calculate the parameter feature vector Weights of associations between nodes within the subgraph : ； in, The normalized association distribution weights, For the set of nodes in the local associated subgraph The traversal index parameter in the subgraph represents any entity node within that subgraph. S42. Utilize the aforementioned correlation distribution weights Perform a weighted summation operation on the node features of the local associated subgraph to calculate the output graph constraint feature vector. : ； S43, constrain the feature vector of the graph The multidimensional joint state vector output by S2 Perform concatenated encoding to construct the generated cue vector. ; generate the prompt vector The prior constraints configured for the decoding engine are used in the time step of autoregressive decoding. Based on the current hidden layer feature vectors of the decoding engine Set the vocabulary space mapping matrix With bias vector Calculate target text tags Conditional probability distribution: ； in, The predicted probability of the target text tag. The sequence of guide texts generated previously is used; sampling and decoding are performed according to the conditional probability distribution to iteratively output a preliminary project application guide text sequence.

6. The intelligent generation method for project application guidelines based on graph enhancement and reflective reconstruction according to claim 1, characterized in that... The specific method in step S5 is as follows: S51. Parse the preliminary project application guide text sequence output by S4 and map it to generate text feature vectors. ; Supported by the aforementioned structured prior knowledge base, the feature vectors corresponding to the entities in the review rules are invoked. Set the transpose weight vector With bias scalar The text feature vector With the feature vector Cascaded encoding is performed, and linear dimensionality reduction projection and nonlinear activation normalization are applied to the concatenated high-dimensional feature space to calculate the compliance verification score. : ； in, The normalized compliance verification score. It is a non-linear activation function; S52. Compare the compliance verification scores. With the set preset threshold ,like If the text compliance check fails, the Actor-Critic dual model is invoked, and the Critic component of the dual model processes the text feature vector. With the feature vector of the audit rules Feature alignment and element-level difference operations are performed to extract text deviation feature distributions that violate the rules of the structured prior knowledge base; and multi-level nonlinear spatial mapping processing is performed on the text deviation feature distributions via the Actor in the dual model to generate error feedback feature vectors. The multidimensional joint state vector constructed by calling S2 ,Will and Perform concatenated encoding and set the mapping transformation matrix. With bias vector For the joint state vector containing dynamic error features, perform affine transformation and hyperbolic tangent space mapping operations to calculate the reconstructed state feature vector. : ； in, To reconstruct the state vector, It is the hyperbolic tangent activation function; S53, reconstruct the state vector The callback input is sent to S3, replacing the original multidimensional joint state vector. The latent space similarity calculation of the intent features is re-executed, and the decision engine is guided to recalculate the conditional probability distribution of candidate actions to force real-time iterative reconstruction of the global execution plan sequence until the verification score of the output text meets the requirements. .

7. The intelligent generation method for project application guidelines based on graph enhancement and reflective reconstruction according to claim 1, characterized in that... The specific method in step S6 is as follows: S61. Analyze the decision scheduling trajectory log consisting of S3 to S5, and extract the input state vector actually adopted by the decision engine. With the corresponding scheduling action Set the state value function parameters With the dominant function parameters Forward evaluation of the input state vector and the scheduled action using a two-stream architecture with a competing deep Q-network, the state value and action advantage features are linearly superimposed and mean-eliminating normalized to calculate the output action value function. : ； in, For inclusion and The overall parameters of the deep reinforcement learning network, For the state value function, For the dominant function, For the set of candidate actions, The dimension of the candidate action set. For the candidate action set The traversal index parameter in the decision space represents any candidate scheduling action within the decision space. S62, Call the final output of S4: Graph constraint feature vector The final text feature vector mapped by S5 And the compliance verification score for those that pass the verification. Set a reward balance coefficient Joint quantization and linear weighted calculation are performed on the global text generation quality and compliance constraints to calculate and output an instant reward value. : ； in, For quantified instant reward value, For vectors, the L2 norm constraint; S63. Based on the state vector of the next moment after the system executes the action. Set discount factor With target network parameters Offline policy objective mapping calculations are performed using the Bellman equation for the long-term expected returns in Markov decision-making processes, and the time-difference objective value is calculated as the output. : ； in, For the time-difference target value, This represents finding the maximum action value among all possible actions within the action space; S64. Utilize the aforementioned time-difference target value With the current action value function Calculate the network loss function : ； S65, By minimizing the network loss function The gradient is used to update the overall parameters of the deep reinforcement learning network. The updated action value function is used as the evaluation criterion to guide the calculation of the policy gradient, and the transformation weight matrix in S3 is updated accordingly. With bias vector Set soft update coefficient Perform synchronous iterative updates based on a moving average mechanism on the target network parameters: ； in, The target network soft update coefficients are used to establish a long-term policy gradient feedback link that reduces the probability of S5 iteration reconstruction by alternating updates of the above network parameters and policy write-back.

Images