AI-assisted reading summary generation method and system based on double-mode semantic analysis of government affair knowledge graph

Abstract

The application discloses an AI-assisted reading summary generation method and system based on a government affair knowledge graph double-mode semantic analysis, a structured knowledge graph in the government affair field is constructed in advance, contains a government affair term library, a policy level graph, a department right and responsibility graph, a double-mode fusion and knowledge-guided attention reinforcement, a text semantic feature vector and a knowledge graph coding vector are fused, an attention weight is calculated, a summary is generated and a multidimensional compliance verification is carried out, and three verifications of semantic feature verification, graph entity verification and policy logic verification are carried out. Technical effects: improve the accuracy of the summary, avoid understanding deviation. The compliance of the summary is guaranteed, the three verification mechanisms ensure that the summary information is consistent with the government affair knowledge graph. The adaptability of the AI summary in the government affair scene is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence natural language processing technology, and in particular to a method and system for generating intelligent text summaries based on the fusion of knowledge graph and NLP in government affairs scenarios. Background Technology

[0002] With the continuous advancement of e-government, the demand for technologies that can accurately and efficiently process government documents and generate high-quality summaries is becoming increasingly urgent. Therefore, developing an AI-assisted reading method and system specifically designed for e-government scenarios, based on intelligent semantic analysis and natural language processing technologies, is of significant practical importance.

[0003] In today's information-saturated world, the internet is flooded with massive amounts of text information, especially in the government sector, where news releases, notices, and announcements are abundant. For users, quickly accessing the core content of these texts has become crucial.

[0004] Traditional text reading methods mainly rely on users manually reading word by word, which is extremely inefficient, consumes a lot of time and energy, and is difficult to meet the needs of modern fast-paced work and life.

[0005] Most existing automatic summarization technologies are based on general-purpose Natural Language Processing (NLP) models. While these general-purpose models have achieved certain results in general text summarization, they have many shortcomings in government affairs scenarios. Government affairs texts are characterized by high professionalism and complex implicit logic, containing a large amount of government terminology, policy hierarchies, and departmental responsibilities. General-purpose NLP models often fail to accurately understand these implicit logics and professional terms specific to the government affairs field, leading to comprehension biases in the generated summaries and failing to accurately reflect the core content of the text. For example, when processing news articles involving complex policy interpretations, general-purpose NLP models may ignore the hierarchical relationships and internal connections between policies, resulting in one-sided or inaccurate summaries.

[0006] Furthermore, existing technologies lack effective verification mechanisms when generating summaries, failing to guarantee the consistency and compliance of the summary information with the government knowledge system. This may result in generated summaries containing incorrect or inappropriate information, misleading users.

[0007] With the continuous advancement of e-government, the demand for technical solutions that can accurately and efficiently process government documents and generate high-quality summaries is becoming increasingly urgent. Therefore, developing a method specifically designed for e-government scenarios, capable of fully executing a computer program and achieving end-to-end intelligent summarization of government documents, along with a supporting AI-assisted reading system, is of significant practical importance. Summary of the Invention

[0008] The core objective of this invention is to provide an AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs.

[0009] The technical solution of this invention: An AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs, comprising the following steps: Step S1: Pre-construction of a Government Affairs-Specific Knowledge Graph: Collect standardized corpora from multiple sources in the government affairs domain, use the TF-IDF formula to filter key terms and construct a standardized, dynamically updatable government affairs terminology database; extract metadata from all policy documents, calculate the degree of correlation between policies using the citation strength formula, and construct a directed policy hierarchy graph with individual policies as nodes and hierarchical relationships and citation relationships between policies as edges; analyze the government departments' "three-fixed" plans and lists of powers and responsibilities, calculate the degree of collaboration between departments using the collaboration strength formula, and construct a departmental power and responsibility graph with departments as nodes and business relationships and collaboration relationships as edges; integrate the government affairs terminology database, policy hierarchy graph, and departmental power and responsibility graph to form a standardized, model-callable structured knowledge graph for the government affairs domain. Step S2: Preprocessing of government documents to be processed: Input the government documents to be processed, and perform noise reduction, Chinese word segmentation, part-of-speech tagging, and stop word and function word filtering on the input text to obtain a standardized text sequence; Step S3: Text semantic feature extraction: For the standardized text sequence in S2, a BERT pre-trained model based on the Transformer architecture is used. The standardized text sequence is input into the BERT pre-trained model to extract the text semantic feature vector. ; Step S4, Knowledge Graph Entity Matching and Encoding: Based on the government terminology database, professional terms, policy entities, and departmental entities are extracted from the standardized text sequence to obtain a core entity set; the attribute information and relationships corresponding to the core entities in the knowledge graph are retrieved, and the knowledge graph encoding vector is obtained by mapping through knowledge graph embedding technology. ; Step S5, Bimodal Fusion and Knowledge-Guided Attention Enhancement: Through the fusion formula α is the weighting coefficient for the text semantic feature vector. Knowledge graph encoding vector Weighted fusion is performed to obtain the fused feature vector. Through attention function , The dot product operation calculates the attention weights and applies them to the fused feature vector. This strengthens the guidance of knowledge graphs for text semantic extraction and outputs deep semantic feature vectors; Step S6, Initial Summary Generation: Input the deep semantic feature vector into a Transformer-based Seq2Seq model to maximize the conditional probability. Generate an initial summary, where X is the input text sequence and Y is the output summary sequence; Step S7, Triple Compliance Verification and Iterative Optimization: Perform triple verification on the initial digest sequentially: S7.1 Semantic Feature Verification: Calculate the summary vector using the cosine similarity formula. Compared with the original text vector The similarity is set, and a similarity threshold is set. If the similarity threshold is lower than the set value, the summary is iteratively optimized. S7.2 Graph Entity Verification: Extracting the Summary Terminology Set The set of text-related terms in a knowledge graph Through the matching rate formula Calculate the entity matching rate, set a matching rate threshold, and iteratively optimize the summary if the matching rate threshold is lower than the set value. S7.3 Policy Logic Validation: Based on the policy hierarchy graph and departmental power and responsibility graph, a set of policy hierarchy and departmental power and responsibility logic rules R is pre-constructed. The policy relationship and departmental power and responsibility relationship in the summary are verified to conform to R. If they do not conform, the summary is iteratively optimized. If all three checks pass, the final compliant government document summary is output. If any check fails, the process returns to step S2 to reprocess and generate a summary, forming a closed-loop processing flow.

[0010] Furthermore, step S1, the pre-construction step of the knowledge graph specific to the government domain, also includes a knowledge graph optimization sub-step: building a crowdsourcing update platform for the government knowledge graph, granting government staff and domain experts the authority to submit new terms, policy associations, and departmental collaboration relationships, and incrementally updating the knowledge graph after multi-level review; and using graph neural networks to optimize the knowledge graph embedding technology, mapping the entities and relationships in the knowledge graph to a low-dimensional vector space.

[0011] Furthermore, step S5, the bimodal fusion and knowledge-guided attention enhancement step, also includes an adaptive weight optimization sub-step: introducing a reinforcement learning algorithm to dynamically adjust the weight coefficient α based on the terminology density, policy complexity, and text length of the input text; expanding bimodal fusion to multimodal fusion, and simultaneously incorporating the visual features of the text charts and the features of the policy interpretation audio transcription into the feature fusion process.

[0012] Furthermore, step S6, the initial summary generation step, also includes a summary optimization sub-step based on Reinforcement-GAN: constructing a reinforced generative adversarial network consisting of a generator and a discriminator; the generator outputs a summary based on a Seq2Seq model, and the discriminator evaluates the summary from multiple dimensions such as semantic relevance, compliance, conciseness, and coherence, and feeds back reward and penalty signals to the generator; the generator dynamically adjusts its generation strategy based on the reward and penalty signals to optimize the diversity and usability of the summary.

[0013] Preferably, the weight coefficient α in step S5 is 0.6, and the similarity threshold in S7.1 is... The matching rate threshold for S7.2 is .

[0014] An AI-assisted reading summary generation system based on bimodal semantic analysis of government knowledge graphs is provided for executing the aforementioned AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs. The system includes the following functional modules: The knowledge graph pre-construction module is used to execute step S1, the government domain-specific knowledge graph pre-construction step, and output a standardized government structured knowledge graph. The text preprocessing module, connected to the knowledge graph pre-construction module, is used to execute step S2, the government text preprocessing step, and output a standardized text sequence. The text semantic extraction module, connected to the text preprocessing module, is equipped with a BERT pre-trained model fine-tuned from a large-scale corpus in the government affairs field. It is used to perform step S3 to extract text semantic feature vectors. ; The knowledge graph encoding module, connected to the knowledge graph pre-construction module and the text preprocessing module, is used to perform step S4, matching entities and outputting the knowledge graph encoding vector. ; The dual-modal fusion attention enhancement module is connected to the text semantic extraction module and the knowledge graph encoding module, respectively, and is used to perform the dual-modal weighted fusion and attention enhancement in step S5, and output a deep semantic feature vector; The summary generation module, connected to the dual-modal fusion attention enhancement module, is equipped with the Transformer architecture Seq2Seq model and the Reinforcement-GAN optimization framework, and is used to perform step S6 to generate the initial summary. The triple compliance verification module has inputs connected to the summary generation module, the knowledge graph pre-construction module, and the text preprocessing module, respectively. The output is split into two paths: one path connects to the summary output terminal, and the other path sends the text back to the text preprocessing module. The triple compliance verification module performs step S7 semantic feature verification, knowledge graph entity verification, and policy logic verification. If the verification passes, the final compliance summary is pushed to the output terminal. If the verification fails, the text is sent back to the text preprocessing module for reprocessing.

[0015] The beneficial effects of this invention are as follows: 1. Improved information processing efficiency: Through an end-to-end automated summary generation method, government officials can quickly extract key information from a large number of news releases, notices, and announcements, saving significant reading time and significantly improving work efficiency. Ordinary citizens can also easily and quickly understand policy highlights through summaries, improving information acquisition efficiency.

[0016] 2. Improved Summary Quality: Through dedicated knowledge graph pre-construction, bimodal fusion, and attention mechanisms, the system effectively addresses the problem of general NLP models' misunderstanding of implicit logic and technical terminology when processing government documents. The generated summaries accurately extract core points, avoiding information errors or biased interpretations, and providing users with accurate and reliable information. Whether it's policy interpretation or departmental work-related texts, the summaries accurately reflect their core content.

[0017] 3. Ensuring Compliance: Through a triple verification process, the abstract is strictly reviewed from multiple dimensions such as semantics, entities, and logic to ensure that the abstract is consistent with the government knowledge graph, conforms to policy logic and departmental responsibilities, avoids errors or non-compliant information, and ensures the standardization and seriousness of government information.

[0018] 4. Flexible Adaptability: Optional optimization techniques enable the method to better adapt to changes in the government environment and diverse text processing needs. The introduction of a knowledge graph crowdsourcing update mechanism and embedding technology enriches and optimizes the construction of the knowledge graph; adaptive weights and multimodal fusion enhance the model's flexibility; and the combination of reinforcement learning and generative adversarial networks optimizes summary generation, enhancing the method's adaptability in different scenarios.

[0019] 5. Promoting government transparency and communication: Accurate and concise summaries of government documents enhance public understanding of government policies and work, and increase the transparency of government information. At the same time, they improve the efficiency of internal and external government communication, promote the efficient conduct of government work, and increase public participation in policy. Attached Figure Description

[0020] Figure 1 This is a flowchart of the abstract generation method of the present invention. Detailed Implementation

[0021] The method of this invention is mainly applied to the intelligent processing of government documents in government portal websites and government service platforms. The processing objects of the method are various government documents such as news releases, notices and announcements, policy interpretations, official documents, and regulatory documents in the government field. It can be directly embedded into government information release systems, government retrieval systems, and AI-assisted reading systems to realize the automated and standardized generation of government document summaries.

[0022] This example uses a draft interpretation of a provincial urban and rural residents' pension insurance policy as the government document to be processed, and fully executes the intelligent summary generation method for government documents of this invention. The specific execution steps are as follows:

[0023] S1: Pre-built knowledge graph specific to the government sector This step provides a standardized knowledge baseline for the entire method. The pre-construction process is completed before the method is executed and supports dynamic iterative updates. The specific sub-steps are as follows: S1.1: Construction of a Government Terminology Database Input: Multi-source standardized corpus of government affairs documents, including government regulations, policy documents, official professional reports, and formal documents. Technical processing procedure: 1. Perform noise reduction, word segmentation, and stop word filtering preprocessing on the input corpus to obtain a candidate term set; 2. Calculate the importance weight of each candidate term using the TF-IDF formula TF-IDF(t,d,D)=TF(t,d)×IDF(t,D), where TF(t,d) is the term frequency of term t in document D, and IDF(t,D) is the inverse document frequency of term t in document set D; 3. Set a threshold for the importance of terms, and select terms with a weight higher than the threshold to be included in the government terminology database. Standardize the definition and attribute labeling of the terms in the database (including the field, scope of application, related policies, etc.). 4. Establish a dynamic update mechanism for the terminology database, and complete the supplementation and iteration of terms in the incremental corpus according to a preset cycle.

[0024] Example: In a set of documents on social security policies, the term "pension replacement rate" has a high TF-IDF value, indicating its high importance in the document set. It is included in the government terminology database and standardized.

[0025] Output: A standardized, dynamically updatable government terminology database, providing a unified terminology benchmark for subsequent entity recognition, semantic matching, and entity verification.

[0026] S1.2: Construction of Policy Hierarchy Map Input: All officially released government policy documents and policy interpretation texts Technical processing procedure: 1. Parse the input government policy documents and policy interpretation texts to extract core metadata such as policy subject, issuing unit, citation relationship, and scope of application; 2. Set policies Policy The number of citations is Calculated by referring to the strength formula (n is the total number of policies) to determine the degree of hierarchical linkage between policies; This represents the frequency with which the i-th policy cites the k-th policy; 3. Using individual policies as nodes and the hierarchical relationships and citations between policies as edges, construct a directed policy hierarchy graph, and label each node with attribute information such as policy release time, issuing unit, scope of application, policy objectives, and superior / subordinate policy relationships.

[0027] Example: If policy P A Frequent citation of policies P B , making If the value is large, then the policy is determined. P B It is a policy P A The higher-level policies establish corresponding directed edges in the graph.

[0028] Output: A structured policy hierarchy map, providing a hierarchical rule benchmark for subsequent semantic understanding and policy logic verification.

[0029] S1.3: Construction of Departmental Authority and Responsibility Map Input: Government department's "Three-Fixed" Plan (defining functions, institutions, and staffing), functional planning documents, officially released list of departmental powers and responsibilities, and business cooperation documents. Technical processing procedure: 1. Parse the input file to clarify core information such as the scope of authority, boundaries of responsibility, and corresponding business of each department; 2. Establish departments With Department The number of collaborations on business O is Through the formula of cooperation strength (m is the total number of departments) is used to measure the closeness of cooperation between departments; This represents the frequency of collaboration between the i-th department and the k-th department; 3. Construct a departmental authority and responsibility map with departments as nodes and business relationships and collaboration relationships as edges, and mark each department node with detailed attribute information such as its main responsibilities, scope of authority, and corresponding business.

[0030] Example: In urban construction, planning departments and construction departments frequently collaborate. A higher value indicates a close collaborative relationship between the two parties. Corresponding edges should be established in the graph and business attributes should be labeled.

[0031] Output: A structured departmental responsibility graph, providing a rule benchmark for subsequent semantic understanding and responsibility logic verification.

[0032] S2: Preprocessing of Government Documents to be Processed Input: The government document to be processed (in this example, it is an interpretation of the urban and rural residents' pension insurance policy of a certain province). Technical processing procedure: 1. Perform noise removal on the input text, including removing HTML tags, irrelevant symbols, and meaningless special characters; 2. Perform Chinese word segmentation on the denoised text and use part-of-speech tagging (POS) to tag the part of speech of each segmentation result; 3. Filter out stop words and meaningless function words, retain the core content words and technical terms of the text, and obtain a standardized text sequence.

[0033] Example: For the text "Notice on Strengthening Environmental Protection", the word segmentation results in "about", "strengthen", "environment", "protection", "of", and "notice", with corresponding parts of speech marked. "About" is a preposition, "strengthen" is a verb, etc.

[0034] Output: The standardized text sequence, which serves as input for subsequent semantic feature extraction.

[0035] S3: Text Semantic Feature Extraction Input: The normalized text sequence output by S2 Technical Process: A BERT pre-trained model based on the Transformer architecture was selected, which had been fine-tuned using a large-scale corpus in the government affairs field. Standardized text sequences were input into the fine-tuned BERT model, and the model's multi-layer Transformer encoder extracted deep semantic features of the text, outputting a text semantic feature vector. .

[0036] Output: Semantic feature vector of the text to be processed .

[0037] S4: Knowledge Graph Entity Matching and Encoding Input: Standardized text sequence output by S2, and a pre-built knowledge graph specific to the government affairs domain from S1.

[0038] Technical processing procedure: 1. Based on the government terminology database, professional terms, policy entities, and departmental entities in standardized text sequences are matched and extracted to obtain the core entity set corresponding to the text; 2. Retrieve attribute information and relationships corresponding to core entities from a knowledge graph specific to the government sector. Using knowledge graph embedding technology, map the entities and their relationships to a low-dimensional vector space to obtain the knowledge graph encoding vector. .

[0039] Output: Knowledge graph encoding vector matching the text .

[0040] S5: Dual-modal feature fusion and enhanced attention mechanism Input: Text semantic feature vector output by S3 The knowledge graph encoding vector output by S4

[0041] Technical processing procedure: 1. Dual-modal fusion: through fusion formula (α is the weighting coefficient, which was determined to be 0.6 through multiple experiments and optimizations) to achieve dual-modal fusion; 2. Strengthening the attention mechanism: Defining the attention function ,in Use dot product operation: . It is the encoding vector of the entire knowledge graph. , ,…, The values ​​are calculated one by one and then summed. Attention weights are calculated using this attention function and applied to the fused feature vectors. This strengthens the guidance of knowledge graphs on text semantics.

[0042] Example: When processing the pension insurance policy text in this embodiment, the model uses an attention mechanism to focus on strengthening features corresponding to social security terminology, pension insurance-related policy nodes, and the rights and responsibilities of the human resources and social security department, thereby improving the accuracy of understanding the core semantics of the text.

[0043] Output: A deep semantic feature vector that integrates textual semantics and government professional knowledge.

[0044] S6: Initial Summary Generation Input: Deep semantic feature vector output by S5

[0045] Technical processing: A Transformer-based Seq2Seq model is used, assuming the input text sequence is... The output summary sequence is The model maximizes the probability To generate an initial summary.

[0046] Example: When processing the press release on education policy reform in this embodiment, the model generates each word or phrase in the summary sequentially based on the features of the input text, forming a complete initial summary.

[0047] Output: Initial digest to be verified.

[0048] S7: Triple Compliance Verification and Final Summary Output Input: The initial summary output by S6, the original text, and the government affairs domain-specific knowledge graph pre-built by S1. Technical processing: Triple validation is performed sequentially. If any validation fails, iterative optimization of the summary is triggered, and the data is sent back to S2 for preprocessing of the government documents to be processed, where a standardized text sequence is regenerated. After all validations pass, the final summary is output. The specific steps are as follows: S7.1 Semantic Feature Verification: Using the Cosine Similarity Formula Calculate the summary vector Compared with the original text vector Similarity, setting a threshold If the similarity is lower than If the calculated similarity is 0.65, then the abstract needs to be adjusted. For example, if the calculated similarity is 0.65, then the abstract needs to be optimized.

[0049] Example: If the calculated similarity is 0.65, which is lower than the threshold of 0.7, the summary needs to be optimized.

[0050] S7.2 Graph Entity Verification: Extract the term set from the abstract as follows It calls upon the set of text-related terms in a knowledge graph specific to the government sector. By calculating the matching rate formula , If the matching rate is lower than If so, then revise the summary.

[0051] Example: If the abstract contains 6 terms, and 4 of them match terms in a government affairs domain-specific knowledge graph, the matching rate is... If the value is below the threshold of 0.8, the summary needs to be modified to better match its terminology with the knowledge graph specific to the government affairs sector.

[0052] S7.3 Policy Logic Validation: Based on the policy hierarchy map and departmental authority and responsibility map, pre-construct a set of logical rules related to policy hierarchy and departmental authority and responsibility. R Let E represent the policy relationships and departmental responsibilities involved in the abstract. We will determine whether E satisfies... R The rules in the document are used to verify the policy logic; if it does not conform to the logic rules, the summary content is adjusted to ensure that the policy logic and the expression of rights and responsibilities are accurate.

[0053] Example: If the policy hierarchy map stipulates that a certain policy must be formulated based on another higher-level policy, but this relationship is not reflected in the summary, then the summary needs to be adjusted.

[0054] The verification results of this embodiment are as follows: the semantic similarity is 0.82, the entity matching rate is 0.92, the policy logic meets the requirements of the rule set, and all three verifications are passed.

[0055] Output: A summary of the final compliant government documents.

[0056] (III) Expanding technical solutions

[0057] 1. Knowledge graph construction optimization steps (a further extension of step S1) In addition to traditional official data sources, a knowledge graph crowdsourcing update mechanism is introduced, establishing a government knowledge graph crowdsourcing platform with a multi-role review process. This platform allows government staff and domain experts to submit information such as new terms, policy connections, and departmental collaboration relationships. After multi-level review, the information is incorporated into the knowledge graph, enabling real-time dynamic updates. Simultaneously, the knowledge graph embedding technology is optimized, employing a more efficient graph neural network model to map entities and relationships in the knowledge graph to a low-dimensional vector space, improving the fusion efficiency of the knowledge graph and NLP model.

[0058] 2. Dual-modal fusion model optimization steps (a further extension of step S5) In the bimodal fusion process, an adaptive weight adjustment strategy is adopted. Through reinforcement learning algorithms, the model dynamically adjusts the value of α based on the characteristics of different texts (terminology density, policy complexity, text length, etc.) and task requirements. For example, for highly specialized texts with dense terminology, the model automatically decreases the α value and increases the weight of knowledge graph information; for texts with clear semantics and simple logic, the α value is appropriately increased to enhance the weight of text semantic features. Furthermore, multimodal information fusion is introduced. In addition to text semantic features and knowledge graph structured information, visual features such as charts and images in the text, as well as the transcribed text features of policy interpretation audio, can be incorporated into the model to further enhance its ability to understand government documents.

[0059] 3. Summary generation optimization step (a further extension of step S6) Building upon the Seq2Seq model for summarization, this paper optimizes the model by incorporating reinforcement learning and Reinforcement Generative Adversarial Networks (Reinforcement-GAN). A generator is responsible for generating summaries, while a discriminator not only assesses the similarity between the generated summaries and real human summaries but also provides reward or penalty signals to the generator based on multi-dimensional quality metrics such as relevance, conciseness, coherence, and compliance. The generator dynamically adjusts its generation strategy based on these feedback signals, continuously optimizing the quality of the generated summaries. This ensures that the generated summaries are not only semantically accurate and compliant but also more diverse and practical.

[0060] Simultaneously, a corresponding AI-assisted reading summary generation system based on bimodal semantic analysis of government knowledge graphs is proposed to execute the aforementioned AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs, including the following functional modules: The knowledge graph pre-construction module is used to execute step S1, the government domain-specific knowledge graph pre-construction step, and output a standardized government structured knowledge graph. The text preprocessing module, connected to the knowledge graph pre-construction module, is used to execute step S2, the government text preprocessing step, and output a standardized text sequence. The text semantic extraction module, connected to the text preprocessing module, is equipped with a BERT pre-trained model fine-tuned from a large-scale corpus in the government affairs field. It is used to perform step S3 to extract text semantic feature vectors. ; The knowledge graph encoding module, connected to the knowledge graph pre-construction module and the text preprocessing module, is used to perform step S4, matching entities and outputting the knowledge graph encoding vector. ; The dual-modal fusion attention enhancement module is connected to the text semantic extraction module and the knowledge graph encoding module, respectively, and is used to perform the dual-modal weighted fusion and attention enhancement in step S5, and output a deep semantic feature vector; The summary generation module, connected to the dual-modal fusion attention enhancement module, is equipped with the Transformer architecture Seq2Seq model and the Reinforcement-GAN optimization framework, and is used to perform step S6 to generate the initial summary. The triple compliance verification module has inputs connected to the summary generation module, the knowledge graph pre-construction module, and the text preprocessing module, respectively. The output is split into two paths: one path connects to the summary output terminal, and the other path sends the text back to the text preprocessing module. The triple compliance verification module performs step S7 semantic feature verification, knowledge graph entity verification, and policy logic verification. If the verification passes, the final compliance summary is pushed to the output terminal. If the verification fails, the text is sent back to the text preprocessing module for reprocessing.

[0061] Furthermore, corresponding to the method scheme, the dual-modal fusion attention enhancement module also includes an adaptive weight optimization submodule, which adjusts α according to the method scheme description. The summary generation module also includes a summary optimization submodule based on Reinforcement-GAN, used to construct the enhanced generative adversarial network described in the method scheme.

Claims

1. An AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs, characterized in that, Includes the following steps: Step S1: Pre-construction of a Government Affairs-Specific Knowledge Graph: Collect standardized corpora from multiple sources in the government affairs domain, use the TF-IDF formula to filter key terms and construct a standardized, dynamically updatable government affairs terminology database; extract metadata from all policy documents, calculate the degree of correlation between policies using the citation strength formula, and construct a directed policy hierarchy graph with individual policies as nodes and hierarchical relationships and citation relationships between policies as edges; analyze the government departments' "three-fixed" plans and lists of powers and responsibilities, calculate the degree of collaboration between departments using the collaboration strength formula, and construct a departmental power and responsibility graph with departments as nodes and business relationships and collaboration relationships as edges; integrate the government affairs terminology database, policy hierarchy graph, and departmental power and responsibility graph to form a standardized, model-callable structured knowledge graph for the government affairs domain. Step S2: Preprocessing of government documents to be processed: Input the government documents to be processed, and perform noise reduction, Chinese word segmentation, part-of-speech tagging, and stop word and function word filtering on the input text to obtain a standardized text sequence; Step S3: Text semantic feature extraction: For the standardized text sequence in S2, a BERT pre-trained model based on the Transformer architecture is used. The standardized text sequence is input into the BERT pre-trained model to extract the text semantic feature vector. ; Step S4, Knowledge Graph Entity Matching and Encoding: Based on the government terminology database, professional terms, policy entities, and departmental entities are extracted from the standardized text sequence to obtain a core entity set; the attribute information and relationships corresponding to the core entities in the knowledge graph are retrieved, and the knowledge graph encoding vector is obtained by mapping through knowledge graph embedding technology. ; Step S5, Bimodal Fusion and Knowledge-Guided Attention Enhancement: Through the fusion formula α is the weighting coefficient for the text semantic feature vector. Knowledge graph encoding vector Weighted fusion is performed to obtain the fused feature vector. Through attention function , The dot product operation calculates the attention weights and applies them to the fused feature vector. This strengthens the guidance of knowledge graphs for text semantic extraction and outputs deep semantic feature vectors; Step S6, Initial Summary Generation: Input the deep semantic feature vector into a Transformer-based Seq2Seq model to maximize the conditional probability. Generate an initial summary, where X is the input text sequence and Y is the output summary sequence; Step S7, Triple Compliance Verification and Iterative Optimization: Perform triple verification on the initial digest sequentially: S7.1 Semantic Feature Verification: Calculate the summary vector using the cosine similarity formula. Compared with the original text vector The similarity is set, and a similarity threshold is set. If the similarity threshold is lower than the set value, the summary is iteratively optimized. S7.2 Graph Entity Verification: Extracting the Summary Terminology Set The set of text-related terms in a knowledge graph Through the matching rate formula Calculate the entity matching rate, set a matching rate threshold, and iteratively optimize the summary if the matching rate threshold is lower than the set value. S7.3 Policy Logic Validation: Based on the policy hierarchy graph and departmental power and responsibility graph, a set of policy hierarchy and departmental power and responsibility logic rules R is pre-constructed. The policy relationship and departmental power and responsibility relationship in the summary are verified to conform to R. If they do not conform, the summary is iteratively optimized. If all three checks pass, the final compliant government document summary is output. If any check fails, the process returns to step S2 to reprocess and generate a summary, forming a closed-loop processing flow.

2. The AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs according to claim 1, characterized in that, The S1 step of pre-constructing a knowledge graph specific to the government domain also includes a knowledge graph optimization sub-step: building a crowdsourcing update platform for the government knowledge graph, granting government staff and domain experts the authority to submit new terms, policy associations, and departmental collaboration relationships, and incrementally updating the knowledge graph after multi-level review; and using graph neural networks to optimize the knowledge graph embedding technology, mapping entities and relationships in the knowledge graph to a low-dimensional vector space.

3. The AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs according to claim 1, characterized in that, The step S5, bimodal fusion and knowledge-guided attention enhancement step, also includes an adaptive weight optimization sub-step: introducing a reinforcement learning algorithm to dynamically adjust the weight coefficient α based on the terminology density, policy complexity, and text length of the input text; expanding bimodal fusion to multimodal fusion, and simultaneously incorporating the visual features of text charts and the features of audio-to-text transcription of policy interpretations into the feature fusion process.

4. The AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs according to claim 1, characterized in that, The initial summary generation step S6 further includes a summary optimization sub-step based on Reinforcement-GAN: constructing a reinforced generative adversarial network consisting of a generator and a discriminator; the generator outputs a summary based on a Seq2Seq model, and the discriminator evaluates the summary from multiple dimensions such as semantic relevance, compliance, conciseness, and coherence and feeds back reward and penalty signals to the generator; the generator dynamically adjusts its generation strategy based on the reward and penalty signals to optimize the diversity and usability of the summary.

5. The AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs according to claim 1, characterized in that, In step S5, the weight coefficient α is set to 0.6, and the similarity threshold in S7.1 is... The matching rate threshold for S7.2 is .

6. An AI-assisted reading summary generation system based on bimodal semantic analysis of government knowledge graphs, characterized in that, The AI-assisted reading summary generation method based on bimodal semantic analysis of government knowledge graphs as described in any one of claims 1-4 includes the following functional modules: The knowledge graph pre-construction module is used to execute step S1, the government domain-specific knowledge graph pre-construction step, and output a standardized government structured knowledge graph. The text preprocessing module, connected to the knowledge graph pre-construction module, is used to execute step S2, the government text preprocessing step, and output a standardized text sequence. The text semantic extraction module, connected to the text preprocessing module, is equipped with a BERT pre-trained model fine-tuned from a large-scale corpus in the government affairs field. It is used to perform step S3 to extract text semantic feature vectors. ; The knowledge graph encoding module, connected to the knowledge graph pre-construction module and the text preprocessing module, is used to perform step S4, matching entities and outputting the knowledge graph encoding vector. ; The dual-modal fusion attention enhancement module is connected to the text semantic extraction module and the knowledge graph encoding module, respectively, and is used to perform the dual-modal weighted fusion and attention enhancement in step S5, and output a deep semantic feature vector; The summary generation module, connected to the dual-modal fusion attention enhancement module, is equipped with the Transformer architecture Seq2Seq model and the Reinforcement-GAN optimization framework, and is used to perform step S6 to generate the initial summary. The triple compliance verification module has inputs connected to the summary generation module, the knowledge graph pre-construction module, and the text preprocessing module, respectively. The output is split into two paths: one path connects to the summary output terminal, and the other path sends the text back to the text preprocessing module. The triple compliance verification module performs step S7 semantic feature verification, knowledge graph entity verification, and policy logic verification. If the verification passes, the final compliance summary is pushed to the output terminal. If the verification fails, the text is sent back to the text preprocessing module for reprocessing.

7. The AI-assisted reading summary generation system based on bimodal semantic analysis of government knowledge graph as described in claim 6, characterized in that, The dual-modal fusion attention enhancement module also includes an adaptive weight optimization submodule, which is used to introduce reinforcement learning algorithms to dynamically adjust the weight coefficient α based on the terminology density, policy complexity, and text length of the input text.

8. The AI-assisted reading summary generation system based on bimodal semantic analysis of government knowledge graph as described in claim 6, characterized in that, The summary generation module also includes a summary optimization submodule based on Reinforcement-GAN, which is used to construct a reinforced generative adversarial network consisting of a generator and a discriminator. The generator outputs a summary based on the Seq2Seq model, and the discriminator evaluates the summary from multiple dimensions such as semantic relevance, compliance, conciseness, and coherence and feeds back reward and penalty signals to the generator. The generator dynamically adjusts its generation strategy based on the reward and penalty signals to optimize the diversity and usability of the summary.

Images