A policy knowledge graph construction method for the infrastructure field

Abstract

The application discloses a policy knowledge graph construction method for infrastructure field, and relates to the technical field of knowledge graph, which comprises the following steps: S1, knowledge modeling is performed on the policy knowledge graph of the infrastructure field, and entity types, entity attributes and entity relationships are determined; S2, provincial and ministerial infrastructure policy provisions are collected and preprocessed to obtain an initial data set; S3, an initial keyword library of the infrastructure field is constructed, the initial keyword set and historical infrastructure policy texts are input into a large model, the initial keyword library is expanded, and then the infrastructure policy texts are classified; S4, the classified infrastructure policy texts are jointly extracted by using a pretraining language model to obtain structured knowledge; and S5, a Neo4j graph database is used to construct an infrastructure policy knowledge storage system, and the structured knowledge is stored in association. The method realizes automatic collection, structured analysis, associated mining and dynamic updating of multi-source policies, and provides accurate and timely policy knowledge services for enterprises.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and more specifically to a method for constructing a policy knowledge graph for the infrastructure sector. Background Technology

[0002] Currently, my country's infrastructure construction has expanded from traditional railways and highways to emerging areas such as clean energy and digital infrastructure, forming a diversified and cross-sectoral modern infrastructure system. Related policies play a crucial guiding role in enterprises' R&D layout, project decision-making, and resource allocation. However, enterprises generally face a "supply-demand mismatch" problem when actually applying these policies: on the one hand, they need to promptly obtain and understand new policies to support business decisions; on the other hand, they still rely on manual retrieval and general tools for analysis, which suffers from bottlenecks such as low efficiency, weak relevance, and slow knowledge transformation, making it difficult for policies to effectively support enterprise R&D and project implementation.

[0003] Existing policy information processing technologies are mostly based on general natural language processing tools and general knowledge graph construction methods. In the highly specialized field of infrastructure, these technologies have the following obvious shortcomings: Incomplete policy collection coverage and lack of field screening mechanisms: Infrastructure-related policies are scattered across multiple platforms. Existing technologies lack automatic collection and screening mechanisms for the infrastructure field, relying on manual platform-by-platform searches. Furthermore, the lack of unified industry classification standards leads to the omission of important policies, affecting enterprises' comprehensive understanding of the policy system.

[0004] The ontology model lacks characteristics specific to the infrastructure sector and has weak expressive capabilities: existing policy knowledge graphs mostly use general entities and relationships such as "policy name, issuing agency, release date, and applicable region" and do not establish exclusive entity types for the infrastructure industry (such as "infrastructure project", "investment and financing model", "responsible entity" etc.). As a result, the knowledge graph cannot reflect the inherent business logic and relational network of infrastructure policies and is difficult to support complex queries in actual infrastructure applications.

[0005] Insufficient capabilities in identifying technical terms and extracting deep relationships: Infrastructure policy texts contain a large number of engineering terms, investment and financing models, construction standards, and other professional expressions. General-purpose NLP tools have low accuracy in entity identification and relationship extraction. In addition, existing methods are mostly limited to the analysis of single policies and are unable to automatically discover implicit relationships across policies and clauses, resulting in low efficiency in policy knowledge transformation and affecting project decision-making.

[0006] Therefore, there is an urgent need for a policy knowledge graph construction method for the infrastructure sector, which can realize the automated collection, structured analysis, correlation mining and dynamic updating of multi-source policies, and provide enterprises with accurate and timely policy knowledge services. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing a policy knowledge graph for the infrastructure field.

[0008] The objective of this invention is achieved through the following technical solution: This application discloses a method for constructing a policy knowledge graph in the field of infrastructure, including the following steps: S1. Based on ontology editing tools, using the resource description framework RDF and the OWL language, knowledge modeling is performed on the infrastructure-related policy knowledge graph to determine entity types, entity attributes, and entity relationships; the entity types include general policy entities and infrastructure-related entities. S2. Collect provincial and ministerial-level infrastructure policy documents from authoritative sources through web crawlers, API interfaces, or direct database connections, and preprocess them to obtain an initial dataset; S3. Based on the industry attributes and project type characteristics of infrastructure policy texts, and referring to the core industry categories in the infrastructure field, an initial keyword library for the infrastructure field is constructed with the help of infrastructure experts, and the keyword descriptions for each category are clarified; the initial keyword set for industry classification within the infrastructure field is summarized by experts, and the initial keyword set and historical infrastructure policy texts are input into the large model. The large model is used to expand the initial keyword library, and then the infrastructure policy texts are classified. S4. Jointly extract structured knowledge from the infrastructure policy text classified in step S3 using a pre-trained language model. S5. Use Neo4j graph database to build a knowledge storage system for infrastructure policies, and store the structured knowledge obtained in step S4 in association.

[0009] Furthermore, the core industry categories in the infrastructure sector mentioned in step S3 include: highway engineering, urban rail transit, water conservancy and hydropower engineering, port and waterway engineering, and railway engineering.

[0010] Preferably, the step of expanding the initial keyword library through a large model in step S3 specifically includes: automatically identifying the semantic relationship between emerging terms appearing in infrastructure policy texts and the initial keyword set, and adding them to the initial keyword library; having experts conduct random checks on the keywords expanded by the large model, removing keywords with semantic deviations, and forming a dynamically updated authoritative keyword library in the field of infrastructure.

[0011] Preferably, step S3, which involves classifying infrastructure policy texts, specifically includes the following sub-steps: S31. Based on an authoritative keyword database in the infrastructure field, design a rule-based algorithm to perform coarse classification of infrastructure policy texts. Specifically, this includes: matching infrastructure policy texts with the authoritative keyword database in the infrastructure field, and counting the frequency of infrastructure-specific keywords in the infrastructure policy texts; if the frequency of infrastructure-specific keywords in the infrastructure policy texts is zero, then the infrastructure policy texts are determined to be non-infrastructure policies and are removed. S32. Leveraging the semantic understanding capabilities of the large-scale model, multi-industry label recognition and weight allocation are performed on the coarsely classified infrastructure policy text based on keyword frequency ratio and core semantic relevance. The industry labels are sorted from high to low weight, and the corresponding results of policy-multi-industry label-weight are output, achieving precise fine classification of infrastructure policy text. The keyword frequency ratio represents the proportion of the occurrence of a certain industry category keyword in the policy text to the total occurrence of all industry category keywords. The core semantic relevance represents the degree of matching between the core semantics of the industry category and the content of the policy text.

[0012] Preferably, step S4 specifically includes the following steps: S41. The infrastructure policy text after precise fine classification is segmented using the BERT model with error correction masking strategy. The segmented text is then input into the embedding layer, and all features obtained from the embedding layer are summed to obtain the infrastructure policy text feature vector. The embedding layer includes word embedding, segment embedding, and position embedding. Word embedding is used to represent the information features of the word itself. Segment embedding is used to determine the order of sentences to obtain sentence features. Position embedding is used to encode and learn the position information features of words in sentences. S42. Add adversarial training to enable the error-correcting masking strategy BERT model to better understand the data distribution during learning, improve its resistance to noise and interference, and cope with small changes in the input data. The adversarial training is... ,in Indicates input data, This indicates resistance to disturbances. Indicates model parameters, Indicates the true label, This represents the maximum probability that the input prediction is the true label; S43. Process the word vector sequence E through an ordered neuron long short-term memory network (ON-LSTM) to obtain a vector matrix H containing contextual and hierarchical information of the infrastructure policy text; the update formula of the ON-LSTM is as follows: in, This represents the initial state awaiting selection, and tanh represents the activation function with constraints on the range of values. This represents the input weight matrix. Represents the current input vector. Represents the historical state weight matrix. This represents the hidden state vector of vector matrix H at time t-1 in the time series computation. Indicates the bias term. Represents the forget gate vector. This represents the cumulative sum from left to right. Indicates the input weights of the forget gate. Indicates the weights of the historical states of the input gate. Indicates the input gate bias term. This represents the hierarchical weight coefficient vector. This represents the multiplication of corresponding elements in two matrices. Let be the hierarchical weight coefficient vector at time t. This represents the cumulative sum from right to left. This represents the soft-maximization activation function. Here is the input weight matrix of the input gate. The input gate's historical state weight matrix, For the bias term of the input gate, The state at time t-1 Waiting for the current state; S44. Decode and identify entities using the global pointer GP, transform the vector matrix H, and construct query-key-key vector pairs that simultaneously associate entity type and relation type:

[0013] in Indicates entity type is The relation type is The query vector at the i-th position of the text sequence. Indicates entity type is The relation type is The key vector at the j-th position of the text sequence. This represents the trainable weight matrix of the query vector. The trainable weight matrix represents the key vector. This represents the bias term of the query vector. This represents the bias term of the key vector. Let H represent the hidden state vector at the i-th position in the feature vector matrix H output after processing by the ON-LSTM layer. This represents the vector at the j-th position in the vector matrix; S45. Using the scoring formula The scoring mechanism of the scoring formula includes text fragments. Is the type entity α? Is the type entity α? text fragments and type entity are text fragments Is the type relationship... ; This represents the score obtained through joint sampling. This represents vector concatenation. Indicates the relation type is The weight matrix, This represents a semantic bias term of relation type . This represents the query vector at the i-th position of the text sequence when the entity type is α1. Let represent the key vector at the j-th position of the text sequence when the entity type is α1. This represents the query vector at the m-th position of the text sequence when the entity type is α2. This represents the key vector at the nth position of the text sequence when the entity type is α2. Let represent the hidden state vector at the m-th position in the vector matrix H. Let T represent the hidden state vector at the nth position in the vector matrix H, and let T denote the transpose.

[0014] Preferably, the anti-disturbance method described in step S42 The calculation formula is as follows: [Formula omitted for brevity] ,in Indicates the scaling factor. The gradient is represented by the following formula: ,in This represents the loss for a single sample. This represents the gradient of the input data e.

[0015] The beneficial effects of this invention are: 1) In view of the shortcomings of existing technologies that only contain general entities such as "policy, issuing agency, and region" and cannot carry the association logic in the infrastructure field, this invention constructs a two-layer ontology model of "general policy entity + infrastructure-specific entity", which can fully carry the association logic unique to the infrastructure field, effectively support key queries in infrastructure scenarios, and fill the gap in the association capabilities of general knowledge graphs in the infrastructure field.

[0016] 2) Combining the experience of infrastructure experts with the semantic expansion of large models, a dynamic keyword library containing 11 core infrastructure industries is constructed. It can automatically identify emerging terms in the infrastructure field and avoid classification omissions caused by term updates. First, a rough classification is performed through rule algorithms, and then the semantic understanding capabilities of large models are used to allocate the weight of multi-industry labels, which greatly reduces the cost of manual screening and improves the completeness of policy cognition.

[0017] 3) To address the problem that existing general-purpose NLP tools cannot accurately identify infrastructure engineering terms and investment and financing expressions, and have low correlation mining efficiency, this invention optimizes extraction capabilities through a combination of multiple technologies, making the accuracy of infrastructure-specific entity identification and entity relationship identification significantly higher than existing general-purpose tools, accelerating the transformation of policy information into structured knowledge, and supporting subsequent in-depth applications. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of a method for constructing a policy knowledge graph in the infrastructure field according to an embodiment of the present invention; Figure 2 This is a schematic diagram of entities and entity relationships in a method for constructing a policy knowledge graph in the field of infrastructure, according to an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1-2 This application discloses a method for constructing a policy knowledge graph in the field of infrastructure, including the following steps: S1. Based on ontology editing tools (such as Protégé), using the resource description framework RDF and the OWL language, perform knowledge modeling on the infrastructure-related policy knowledge graph to determine entity types, entity attributes, and entity relationships; the entity types include general policy entities and infrastructure-domain entities; the entity and entity relationship diagram is shown below. Figure 2 As shown; S2. Collect provincial and ministerial-level infrastructure policy documents from authoritative sources through web crawling, API interfaces, or direct database connections, and preprocess them to obtain an initial dataset. Authoritative sources include development and reform commissions, transportation commissions, construction commissions of various provinces and cities, and the Peking University Legal Database. In this embodiment, provincial and ministerial-level infrastructure policy documents since 2015 are collected. S3. Based on the industry attributes and project type characteristics of infrastructure policy texts, and referring to the core industry categories in the infrastructure field, an initial keyword library for the infrastructure field is constructed with the help of infrastructure experts, and the keyword descriptions for each category are clarified; the initial keyword set for industry classification within the infrastructure field is summarized by experts, and the initial keyword set and historical infrastructure policy texts are input into the large model. The large model is used to expand the initial keyword library, and then the infrastructure policy texts are classified. S4. Jointly extract structured knowledge from the infrastructure policy text classified in step S3 using a pre-trained language model. S5. Use Neo4j graph database to build a knowledge storage system for infrastructure policies, and store the structured knowledge obtained in step S4 in association.

[0021] For example, the core industry categories in the infrastructure field mentioned in step S3 specifically include: highway engineering, urban rail transit, water conservancy and hydropower engineering, port and waterway engineering and railway engineering, etc. (11 categories).

[0022] For example, the step of expanding the initial keyword library through a large model in step S3 specifically includes: automatically identifying the semantic association between emerging terms appearing in infrastructure policy texts and the initial keyword set, and adding them to the initial keyword library; having experts conduct random checks on the keywords expanded by the large model, removing keywords with semantic deviations, and forming a dynamically updated authoritative keyword library in the field of infrastructure.

[0023] For example, the step of classifying infrastructure policy texts in step S3 specifically includes the following sub-steps: S31. Based on an authoritative keyword database in the infrastructure field, design a rule-based algorithm to perform coarse classification of infrastructure policy texts. Specifically, this includes: keyword detection and semantic matching. The infrastructure policy texts are matched with the authoritative keyword database in the infrastructure field, and the frequency of occurrence of infrastructure-specific keywords in the infrastructure policy texts is counted. If the frequency of occurrence of infrastructure-specific keywords in the infrastructure policy texts is zero, the infrastructure policy texts are judged to be non-infrastructure policies and are removed. S32. Leveraging the semantic understanding capabilities of the large-scale model, multi-industry label recognition and weight allocation are performed on the coarsely classified infrastructure policy text based on keyword frequency ratio and core semantic relevance. The industry labels are sorted from high to low weight, and the corresponding results of policy-multi-industry label-weight are output, achieving precise fine classification of infrastructure policy text. The keyword frequency ratio represents the proportion of the occurrence of a certain industry category keyword in the policy text to the total occurrence of all industry category keywords. The core semantic relevance represents the degree of matching between the core semantics of the industry category and the content of the policy text.

[0024] For example, MacBERT is a novel pre-trained language model proposed by the Harbin Institute of Technology-iFlytek Joint Laboratory. It uses a masked language model as correction (MAC) strategy. Based on the full-word masking strategy, it uses an N-gram method to select the marker to be masked, where the probabilities of 1-gram to 4-gram are 40%, 30%, 20%, and 10%, respectively. Then, it uses similar words to replace the mask marker, which effectively solves the problem of inconsistency between upstream and downstream tasks, where there are mask markers in the pre-training stage but no mask markers in downstream tasks, thereby improving the generalization ability of the model.

[0025] For example, step S4 specifically includes the following steps: S41. The infrastructure policy text, after precise fine-classification, is segmented using the BERT model with error correction masking strategy. The segmented text is then input into the embedding layer, and all features obtained from the embedding layer are summed to obtain the infrastructure policy text feature vector (used as the embedding for subsequent modules). The embedding layer includes word embedding, segment embedding, and position embedding. Word embedding represents the information features of the words themselves; segment embedding determines the order of sentences to obtain sentence features; and position embedding encodes and learns the positional information features of words within a sentence. In position embedding, CLS represents the end-of-sentence marker, and SEP represents the end-of-sentence marker. S42. Add adversarial training to enable the error-correcting masking strategy BERT model to better understand the data distribution during learning, improve its resistance to noise and interference, and cope with small changes in the input data. The adversarial training is... ,in Indicates input data, This indicates resistance to disturbances. Indicates model parameters, Indicates the true label, This represents the maximum probability that the input prediction is the true label; S43. The word vector sequence E is processed by an ordered neuron long short-term memory network (ON-LSTM) to obtain a vector matrix H containing contextual and hierarchical information of the infrastructure policy text. Unlike the unordered neurons of a traditional LSTM, ON-LSTM orders the neurons in the network and obtains specific structures through this ordering. Therefore, the hierarchical information of the text can be obtained using the ordered structure of the neurons. The update formula of the ON-LSTM is shown below: in, This represents the initial state awaiting filtering (containing the semantic features of the current input text), and tanh represents the activation function constraining the range of values. Represents the input weight matrix (used for the current input vector) (perform feature weighting) Represents the current input vector. Represents the historical state weight matrix (used for...) (perform feature weighting) This represents the hidden state vector of vector matrix H at time t-1 in the time series computation. This represents the bias term (used to adjust the baseline value of the candidate state). Represents the forget gate vector (control) (the percentage of information affected by previous storage units). This represents the cumulative sum from left to right. Indicates the input weights of the forget gate. Indicates the weights of the historical states of the input gate. Indicates the input gate bias term; This represents the hierarchical weight coefficient vector (reflecting the fusion strength of historical and current information). This represents the multiplication of corresponding elements in two matrices. Let be the hierarchical weight coefficient vector at time t. This represents the cumulative sum from right to left. This represents the soft-maximization activation function. Here is the input weight matrix of the input gate. The input gate's historical state weight matrix, For the bias term of the input gate, The state at time t-1 Waiting for the current state (which is the basis for subsequent hidden state calculations); Achieve feature amplification at the hierarchical level. To avoid loss of high-level information, To avoid ignoring key information at lower levels, ONLSTM uses the traditional LSTM network structure, adds hierarchical information to obtain semantic representation, and improves the dependency update rules of each neuron through some special methods. S44. After obtaining the context and hierarchical information of the text, the output needs to be decoded. This application uses a global pointer (GP) to decode the output and perform entity recognition. It utilizes a global normalization approach, treating the beginning and end of an entity as a whole for discrimination, providing a more global perspective. Unlike CRF, GP does not require recursive calculation of the denominator during training, and the prediction process does not require dynamic programming, significantly improving training and prediction efficiency, reducing computational load, and lowering computational costs. The vector matrix H is transformed to construct query-key-key vector pairs that simultaneously associate entity type and relation type:

[0026] in Indicates entity type is The relation type is The query vector at the i-th position of the text sequence (used to capture features related to "entity first boundary + relationship association" at that position). Indicates entity type is The relation type is The key vector at the j-th position of the text sequence (used to capture features related to "entity tail boundary + relation association" at that position). This represents the trainable weight matrix of the query vector. The trainable weight matrix represents the key vector. This represents the bias term of the query vector. The two bias terms represent the bias terms of the key vector, used to adapt the semantic offsets for different entity-relation combinations. This represents the hidden state vector at the i-th position in the feature vector matrix H output after processing by the ON-LSTM layer (containing the context and hierarchical information of that position). This represents the vector at the j-th position in the vector matrix; S45. To reduce the number of parameters and improve generalization ability, in actual implementation, for the same entity type... It shares most of the weights, adjusting the relation-sensitive representation only through relation type embedding; and uses a scoring formula. The scoring mechanism of the scoring formula includes text fragments. Is the type entity α? Is the type entity α? text fragments and type entity are text fragments Is the type relationship... ; This represents the score of the joint sampling (the higher the score, the better). for Entity type for Entity type and both exist The higher the confidence level of the type relationship, the better. This represents vector concatenation. Indicates the relation type is The weight matrix, Indicates the relation type is Semantic bias terms (inherent semantic features used to characterize specific relationships in the infrastructure field). This represents the query vector at the i-th position of the text sequence when the entity type is α1. Let represent the key vector at the j-th position of the text sequence when the entity type is α1. This represents the query vector at the m-th position of the text sequence when the entity type is α2. This represents the key vector at the nth position of the text sequence when the entity type is α2. Let represent the hidden state vector at the m-th position in the vector matrix H. Let T represent the hidden state vector at the nth position in the vector matrix H, and let T denote the transpose.

[0027] For example, in step S42, in order to increase the probability of predicting the true label after perturbation, the adversarial perturbation... The perturbation is performed within a certain perturbation space, and the calculation is performed using the FGM method. The calculation formula is as follows: ,in Indicates the scaling factor. The gradient is represented by the following formula: ,in This represents the loss for a single sample. This represents the gradient of the input data e.

[0028] For example, step S5 uses the structured knowledge (including entity, relation, and attribute triples) obtained by the aforementioned MacBERT-based joint extraction method to construct an infrastructure policy knowledge storage system using the Neo4j graph database, thereby realizing the structured and associated storage of knowledge and providing underlying data support for graph analysis applications.

[0029] In summary, addressing the issue that existing general policy ontologies only contain generic entities such as "policy," "issuing agency," and "region," failing to accommodate the unique relational logic specific to the infrastructure sector, this application, for the first time, integrates infrastructure-related business logic into the ontology model. This overcomes the domain adaptability bottleneck of general ontologies, achieving a structured representation of infrastructure policy relational logic and laying the foundation for subsequent accurate queries. Addressing the lack of specific screening standards for the infrastructure sector and the inefficient and prone-to-omission policy classification in existing technologies, this application combines the experience of infrastructure experts with the semantic expansion of a large-scale model to construct a dynamic keyword library containing 11 core infrastructure industries. This library can automatically identify emerging terms in the infrastructure sector, avoiding classification omissions caused by terminology updates. First, a rule-based algorithm is used for coarse classification, then the semantic understanding capabilities of the large-scale model are used to allocate multi-industry label weights, resolving the issues of policy omissions and low classification efficiency in subdivided infrastructure sectors. To address the issues that existing general-purpose NLP tools cannot accurately identify infrastructure engineering terms and investment and financing statements, and have low efficiency in association mining, this application implements MacBERT pre-trained model adaptation, adversarial training to enhance anti-interference capabilities, ON-LSTM hierarchical information capture, and global pointer joint decoding, breaking through the bottleneck of extraction accuracy of specific entities and relationships, and significantly outperforming the recognition performance of general-purpose NLP tools.

[0030] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for constructing a policy knowledge graph for the infrastructure sector, characterized in that, Includes the following steps: S1. Based on ontology editing tools, using the resource description framework RDF and the OWL language, knowledge modeling is performed on the infrastructure-related policy knowledge graph to determine entity types, entity attributes, and entity relationships; the entity types include general policy entities and infrastructure-related entities. S2. Collect provincial and ministerial-level infrastructure policy documents from authoritative sources through web crawlers, API interfaces, or direct database connections, and preprocess them to obtain an initial dataset; S3. Based on the industry attributes and project type characteristics of infrastructure policy texts, and referring to the core industry categories in the infrastructure field, an initial keyword library for the infrastructure field is constructed with the help of infrastructure experts, and the keyword descriptions for each category are clarified; the initial keyword set for industry classification within the infrastructure field is summarized by experts, and the initial keyword set and historical infrastructure policy texts are input into the large model. The large model is used to expand the initial keyword library, and then the infrastructure policy texts are classified. S4. Jointly extract structured knowledge from the infrastructure policy text classified in step S3 using a pre-trained language model. S5. Use Neo4j graph database to build a knowledge storage system for infrastructure policies, and store the structured knowledge obtained in step S4 in association.

2. The method for constructing a policy knowledge graph in the infrastructure field according to claim 1, characterized in that, The core industry categories in the infrastructure sector mentioned in step S3 include: highway engineering, urban rail transit, water conservancy and hydropower engineering, port and waterway engineering, and railway engineering.

3. The method for constructing a policy knowledge graph in the infrastructure field according to claim 1, characterized in that, The steps in step S3 to expand the initial keyword library through a large model specifically include: automatically identifying the semantic relationship between emerging terms appearing in infrastructure policy texts and the initial keyword set, and adding them to the initial keyword library; having experts conduct random checks on the keywords expanded by the large model, removing keywords with semantic deviations, and forming a dynamically updated authoritative keyword library in the field of infrastructure.

4. The method for constructing a policy knowledge graph in the infrastructure field according to claim 3, characterized in that, Step S3, which involves classifying infrastructure policy texts, specifically includes the following sub-steps: S31. Based on an authoritative keyword database in the infrastructure field, design a rule-based algorithm to perform coarse classification of infrastructure policy texts. Specifically, this includes: matching infrastructure policy texts with the authoritative keyword database in the infrastructure field, and counting the frequency of infrastructure-specific keywords in the infrastructure policy texts; if the frequency of infrastructure-specific keywords in the infrastructure policy texts is zero, then the infrastructure policy texts are determined to be non-infrastructure policies and are removed. S32. Leveraging the semantic understanding capabilities of the large-scale model, multi-industry label recognition and weight allocation are performed on the coarsely classified infrastructure policy text based on keyword frequency ratio and core semantic relevance. The industry labels are sorted from high to low weight, and the corresponding results of policy-multi-industry label-weight are output, achieving precise fine classification of infrastructure policy text. The keyword frequency ratio represents the proportion of the occurrence of a certain industry category keyword in the policy text to the total occurrence of all industry category keywords. The core semantic relevance represents the degree of matching between the core semantics of the industry category and the content of the policy text.

5. A method for constructing a policy knowledge graph for the infrastructure field according to claim 4, characterized in that, Step S4 specifically includes the following steps: S41. The infrastructure policy text after precise fine classification is segmented using the BERT model with error correction masking strategy. The segmented text is then input into the embedding layer, and all features obtained from the embedding layer are summed to obtain the feature vector of the infrastructure policy text. The embedding layer includes word embedding, segment embedding and position embedding. The word embedding is used to represent the information features of the word itself. The segment embedding is used to determine the order of sentences to obtain sentence features. The position embedding is used to encode and learn the position information features of words in sentences. S42. Add adversarial training to enable the error-correcting masking strategy BERT model to better understand the data distribution during learning, improve its resistance to noise and interference, and cope with small changes in the input data. The adversarial training is... ,in Indicates input data, This indicates resistance to disturbances. Indicates model parameters, Indicates the true label, This represents the maximum probability that the input prediction is the true label; S43. Process the word vector sequence E through an ordered neuron long short-term memory network (ON-LSTM) to obtain a vector matrix H containing contextual and hierarchical information of the infrastructure policy text; the update formula of the ON-LSTM is as follows: in, This represents the initial state awaiting selection, and tanh represents the activation function with constraints on the range of values. This represents the input weight matrix. Represents the current input vector. Represents the historical state weight matrix. This represents the hidden state vector of vector matrix H at time t-1 in the time series computation. Indicates the bias term. Represents the forget gate vector. This represents the cumulative sum from left to right. Indicates the input weights of the forget gate. Indicates the weights of the historical states of the input gate. Indicates the input gate bias term. This represents the hierarchical weight coefficient vector. This represents the multiplication of corresponding elements in two matrices. Let be the hierarchical weight coefficient vector at time t. This represents the cumulative sum from right to left. This represents the soft-maximization activation function. Here is the input weight matrix of the input gate. The input gate's historical state weight matrix, For the bias term of the input gate, The state at time t-1 Waiting for the current state; S44. Decode and identify entities using the global pointer GP, transform the vector matrix H, and construct query-key-key vector pairs that simultaneously associate entity type and relation type: in Indicates entity type is The relation type is The query vector at the i-th position of the text sequence. Indicates entity type is The relation type is The key vector at the j-th position of the text sequence. This represents the trainable weight matrix of the query vector. The trainable weight matrix represents the key vector. This represents the bias term of the query vector. This represents the bias term of the key vector. Let H represent the hidden state vector at the i-th position in the feature vector matrix H output after processing by the ON-LSTM layer. This represents the vector at the j-th position in the vector matrix; S45. Using the scoring formula The scoring mechanism of the scoring formula includes text fragments. Is the type entity α? Is the type entity α? text fragments and type entity is Is the type relationship of the text fragments as follows? ; This represents the score obtained through joint sampling. This represents vector concatenation. Indicates the relation type is The weight matrix, Indicates the relation type is semantic bias terms, This represents the query vector at the i-th position of the text sequence when the entity type is α1. Let represent the key vector at the j-th position of the text sequence when the entity type is α1. This represents the query vector at the m-th position of the text sequence when the entity type is α2. This represents the key vector at the nth position of the text sequence when the entity type is α2. Let represent the hidden state vector at the m-th position in the vector matrix H. Let T represent the hidden state vector at the nth position in the vector matrix H, and let T denote the transpose.

6. A method for constructing a policy knowledge graph in the infrastructure field according to claim 5, characterized in that, The counter-disturbance described in step S42 The calculation formula is as follows: [Formula omitted for brevity] ,in Indicates the scaling factor. The gradient is represented by the following formula: ,in This represents the loss for a single sample. This represents the gradient of the input data e.

Images