Core governance target-tool combination identification method based on reference association

Abstract

The invention discloses a reference association-based core governance target-tool combination identification method, which comprises the following steps of: (1) selecting a target field, collecting governance files to construct a data pool, and preprocessing the data pool; (2) annotating and converting file data; (3) constructing a W2NER model; (4) training a W2NER model; (5) extracting a W2NER model; (6) constructing a governance target-tool reference network; (7) identifying a core governance target-tool; according to the method, a core governance target-tool combination and interaction logic thereof can be disclosed, comprehensive understanding of relationships and effects among governance file elements is facilitated, scientific data support and decision basis are provided for governance effect evaluation and optimization, and the method has important theoretical significance and practical value.

Description

Technical Field

[0001] The present invention relates to the technical field of data representation and computer system, and in particular to a method for identifying a core governance target-tool combination based on reference association. Background Art

[0002] With the continuous improvement of governance capabilities, the complexity and diversity of the governance system have become increasingly apparent. A single governance tool is difficult to cope with complex governance goals. How to effectively analyze the complex relationship between governance goals and governance tools, track the evolution path of governance combinations, and evaluate governance effects has become an important topic in the field of governance research. Governance goals and tools are semantic structures extracted from texts, which can reflect the main goals, implementation plans, and expected results of governance. The development and evolution of governance is often reflected in the dynamic combination of governance goals and governance tools. This governance combination design is not only affected by the content of the governance documents themselves, but also closely related to the citation and reference relationship between documents. Therefore, constructing a network structure that can reveal the interaction pattern between governance goals and tools is of great value to improving the scientificity and systematicity of governance research.

[0003] At present, the identification of governance objectives and tools has the following shortcomings: (1) The identification methods rely on manual annotation or shallow semantic features, which makes it difficult to capture the complex relationship between governance objectives and tools; (2) The citation network construction technology is simple and fails to reveal the deep interaction patterns between documents; (3) There is a lack of comprehensive analysis methods for the evolution of governance objective-tool combination patterns and the dynamic changes of core combinations. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for identifying core governance goal-tool combinations based on citation associations. Through semantically driven goal and tool identification, combined with citation network analysis and model identification of governance goal-tool combinations based on the self-sphere unit algorithm, the core goal-tool combinations and their evolutionary trajectories are revealed, providing scientific data support and decision-making reference for governance decisions.

[0005] Technical solution: The core governance objective-tool combination identification method based on reference association described in the present invention comprises the following steps: (1) Select the target area, collect governance documents, build a data pool and perform preprocessing; (2) Sample data annotation and conversion: Collect sample data, and use Doccano software to annotate discontinuous governance goals, tools, and entity relationships according to the prompt words of the goals and tools, and divide them into training, validation, and test sets; (3) Constructing the W2NER model: Combining the bidirectional transformer model BERT, bidirectional long short-term memory LSTM, convolutional layer and biaffine mechanism to build a model for sequence labeling and relationship extraction tasks, extracting the semantic features of the text, and extracting governance objectives, tool entities and relationships; (4) Training the W2NER model: a training model for the joint extraction of text entities and relations based on adversarial training and dynamic learning rate adjustment; (5) Extraction W2NER model: A text entity and relationship joint extraction method based on deep learning is used to extract governance objectives, tool entities, and entity relationships in the text; (6) Constructing a governance goal-tool citation network: Extracting the citation and cited relationships between texts based on the book title marks and citation prompts in the text content; using the W2NER model to identify, clean, and integrate the governance goals and tools in the text, and constructing a governance goal-tool citation network based on this information; (7) Identify core governance goals and tools: Use the self-sphere unit algorithm (ESU) to perform directed motif analysis on the governance goal-tool reference network. Recursively identify the induced subgraph in the network, that is, the multiple relationships and interaction patterns between the goal nodes and the tool nodes. Based on this, identify the core governance goal-tool combination and analyze its evolution process.

[0006] Furthermore, in step (1), the preprocessing is: performing a cleaning operation on the data.

[0007] Furthermore, step (1) includes the following steps: (11) Construct a keyword search strategy for the target field and search the database for target document data whose titles contain keywords; (12) Identify and summarize a list of relevant keywords in the target field by reading document data; (13) Remove duplicate files by combining and matching the two fields of the document number and title in the data pool, and only keep one file; (14) Automatically parse the text structure based on natural language processing technology to identify the basic elements such as document title, publishing agency, document type, and date of writing and effective date.

[0008] Furthermore, step (2) includes the following steps: (21) Label the governance objectives. Based on the governance objective entity type and the verb phrase “realize / promote”, use the Doccano software to point the relationship pointer from one end of the discontinuous entity to the other end. If it is a continuous entity, directly label it as the entity type of governance objective. (22) Labeling governance tools. Based on the entity type of governance tools and common verbs and nouns, including the semantic structure of "setting standards and establishing systems", Doccano software is used to use the relationship pointer from one end of the discontinuous entity to the other end. If it is a continuous entity, it is directly labeled as the entity type of governance tool; (23) Mark the relationship between governance objectives and tools, combine the relationship between adopting governance tools to “achieve” governance objectives, and mark the relationship between governance objectives and tool models; (24) Read the annotated data and export the annotated data set as a json file, which contains the annotated entity set and the annotated entity relationship set; (25) Preprocessing of entity-relationship data: defining the target data format, removing isolated entities that are not involved in any relationship, and extracting the text index range of the starting entity and the target entity to generate relationship data; (26) Dataset division: divide the dataset into a training set for model training, a validation set for model parameter tuning, and a test set for model performance testing.

[0009] Furthermore, step (3) includes the following steps: (31) The model receives preprocessed data, including sentence word ID sequence, subword to word mapping matrix, two-dimensional mask matrix, entity distance matrix and sentence length, and verifies and masks the input data to make the data format consistent and mask invalid positions; (32) Generate contextual representations based on multi-layer feature extraction. Use the pre-trained BERT model to contextually encode the input sequence and extract word-level or sub-word-level embeddings. Use the pre-trained BERT model to contextually encode the input sequence and extract word-level or sub-word-level embeddings. Input the word-level embeddings into a bidirectional LSTM to capture the global context information of the sentence. (33) Conditional normalization processing, using the conditional layer normalization module to enhance the features of LSTM output; (34) Fusion of distance and relationship features, mapping the distance matrix between entities into a high-dimensional embedding representation, generating an embedding of the relationship type based on the relationship matrix, and concatenating the distance embedding, relationship embedding, and context representation, i.e., the LSTM output processed by CLN, to form a feature input; (35) Use a multi-expansion rate convolutional layer to locally model the feature input and use the convolutional layer output as the feature input; (36) Bi-affine interaction modeling projects entity representations into a specific space, calculates the interaction scores of entity pairs, and predicts the relationship categories between entities based on the interaction scores.

[0010] Furthermore, step (4) includes the following steps: (41) Data loading and preprocessing: loading training sets, validation sets, and test sets. The data formats include text, entity annotations, and relationship annotations. The distribution of the loaded data is checked and counted. (42) Model initialization: construct a joint extraction model according to the configuration file. The model structure includes using BERT for context feature extraction, bidirectional LSTM to further capture global context, convolutional layer for local feature modeling, and dual affine mechanism for joint prediction of entities and relationships; setting the loss function and grouping parameters, and setting the learning rate scheduler to dynamically adjust the learning rate; (43) Model training: Set the model to training mode, perform forward propagation, loss calculation, adversarial training, gradient clipping and parameter update, and learning rate update steps for each batch of data; after each training round, calculate the average loss and record the predicted accuracy, recall rate, and F1 score; (44) Model verification: set the model to evaluation mode, verify batch by batch, perform forward propagation and prediction, and statistically analyze the performance, saving the current model parameters whose F1 score on the verification set is higher than the historical best; (45) Early stopping mechanism: when the performance of the validation set fails to improve after several consecutive rounds, the number of stagnations is recorded; when the number of stagnations reaches the set threshold, the training is terminated early to avoid overfitting; (46) Model saving and loading: save the model parameter file in the specified path for subsequent loading or inference; load the saved model parameters to support continued training or direct evaluation.

[0011] Furthermore, step (5) includes the following steps: (51) Text segmentation: perform character-level segmentation on the input text, replace consecutive spaces with single spaces, and remove redundant characters; (52) Data loading and construction: Use the segmented text to build a prediction dataset, including the text segmentation results and auxiliary features. After encapsulating the text into the model input format, load the data in batches; (53) Model loading and initialization, loading the joint extraction model, and loading the trained model parameters; (54) Model reasoning: input batch-loaded data into the model, perform forward propagation batch by batch, calculate the category score of each position, and classify and output the entity category, i.e., the target or tool and entity relationship category; (55) Decoding and result generation: decode the index and category of the model output, map the decoded entity and relationship results into a readable format, and finally generate a JSON file of the extraction results; (56) Result storage: output the extracted governance goals, tools and relationships, and store the governance combination in the local file according to the file ID; the governance combination includes: governance goals, implementation, and governance tools.

[0012] Furthermore, step (6) includes the following steps: (61) Identification of identifiers: identifying book title marks embedded in text content based on the string matching algorithm KMP; (62) Complete the file name, obtain the issuing agency of the file, and construct the complete file name by combining the issuing agency and the specific file in the quotation marks; (63) Supplement the citation relationship, identify the prompt words "based on", "implement", "based on", "regulations", "combination", and "comply" before and after the quotation marks, and further identify and supplement the citation relationship; (64) Citation relationship storage: according to the logic of the files cited in this article and the files citing this article, the file citation relationship, i.e., the citing file ID and the cited file ID, is stored; (65) Node construction and attribute mapping. Each governance goal or tool is defined as a node in the reference network. The attributes of the node include: node name, type (i.e., governance goal or tool), source text ID, and frequency of occurrence. The size of the node is proportional to the frequency of occurrence of the governance goal or tool in all texts. Nodes with higher frequencies will appear as larger nodes in the network. (66) The governance goal-tool combination citation network is constructed by combining the governance goals and tools in the citation network and according to the existing governance goal-tool combination patterns. Each text data will contain multiple governance goal and tool combination patterns, indicating a many-to-many relationship between governance goals and tools.

[0013] Furthermore, step (7) includes the following steps: (71) Recursive mining using the ESU algorithm: The ESU (Ego-Sphere Unit) algorithm is applied to recursively mine the subgraphs in the governance goal-tool reference network. Each recursion will expand the SUB set, which contains nodes directly adjacent to the current node, and the EXT set contains nodes that are adjacent to at least one node in the SUB set and have a larger numerical label. The ESU algorithm identifies all eligible induced subgraphs through this recursive process; (72) Motif analysis and subgraph classification: Motif analysis is performed on each branch of the ESU tree, and a recursive method is used to identify subgraph structures that frequently appear in the network. During this process, the ESU algorithm performs an isomorphism test on each subgraph using McKay's nauty algorithm to ensure that subgraphs of different structures are classified and sorted according to their frequency and concentration. (73) Identify the core governance goal-tool combination: Use the ESU algorithm to identify the frequent subgraphs in the network, that is, the core governance goal-tool combination, and calculate the frequency F of each subgraph in the random network. R(G′); at the same time, calculate the actual frequency F of the subgraph in the target network G (G′); use the P value to perform statistical hypothesis testing to evaluate whether a subgraph appears significantly in the target network; at the same time, set the N value to generate multiple random network samples to make the evaluation of subgraph frequency more stable and accurate; the formula is as follows: in, N is the number of random network samples, is the P value, The probability of (as its null hypothesis) is given by In a random network frequency; when P When the value is less than the threshold value of 0.05, the sub-graph can be called a significant pattern.

[0014] (74) Visualization of the core governance goal-tool motif. Based on the identification of the core governance goal-tool combination, the network structure of the motif is displayed using visualization technology. The governance goals and tools are used as network nodes, and the size and color of the nodes represent the frequency and type of the nodes. The thickness of the edges reflects the strength of the citation relationship between the governance goals and tools. (75) Based on the citation network of the core governance objective-tool combination motif, analyze its evolution process at different times or document implementation stages; by tracking the trend of core motifs over time, identify changes in the alternation and combination patterns between governance objectives and tools.

[0015] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, it implements any one of the methods for identifying a core governance objective-tool combination based on reference association.

[0016] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The present invention combines the unified named entity recognition model (W2NER) as word-to-word classification and the ego sphere unit (ESU) algorithm, which can deeply explore the interactive relationship between governance goals and tools, and overcome the problem that the interaction between elements is not fully considered in the prior art. Through the semantically driven citation network and efficient motif analysis method, the goals and tools in the text can be identified more accurately, and the accuracy and efficiency of goal and tool identification can be improved; through the construction of citation relationships and motif analysis, the present invention can reveal the interaction patterns between governance goals and tools at multiple levels, identify the induced subgraphs in the network in a recursive manner, and analyze the multiple relationships and interaction patterns between target nodes and tool nodes, which is helpful to deeply understand the complex relationship between governance goals and tools and provide more powerful support for governance optimization; based on the semantic structure and citation network evolution perspective, the present invention extracts the combination pattern of governance goals and tools to identify the key nodes and constraint relationships in the governance goal-tool network; further, by analyzing the dynamic evolution path of the governance document citation network, the interaction and combination patterns of governance goals and tools at different stages are revealed; by constructing a semantic-based citation network and efficient motif analysis, the present invention can reveal the core goal-tool combination and its interaction logic in the governance system, comprehensively understand the relationship and role between the elements of the governance document, and provide scientific data support and decision-making basis for governance effect evaluation and optimization, which has important theoretical significance and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the present invention; Figure 2 is a two-dimensional matrix of word-word relations of the present invention; Figure 3 A two-dimensional matrix of THW and NHW relationships constructed for the relative position information of the grid characters of the present invention and the annotation data entity; Figure 4 A flow chart of the method for extracting governance objectives and tools based on semantic representation based on the W2NER model of the present invention; Figure 5 Pseudo code of the ESU algorithm of the present invention; Figure 6 is the ESU subgraph search algorithm of the present invention; wherein, Figure 6 (a) in the figure is a target image of size 5; Figure 6 (b) in the figure is an ESU tree with a depth of k. DETAILED DESCRIPTION

[0018] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0019] like Figure 1As shown, an embodiment of the present invention provides a method for identifying a core governance objective-tool combination based on reference association, comprising the following steps: S1. Select the target field, collect governance documents to build a data pool, and perform data cleaning operations; including the following steps: S1.1. Construct a keyword search strategy for the target field and search for documents in the database that contain the keyword in the title; S1.2. By reading the above documents, identify and summarize the list of relevant keywords in the target field; based on the selected target field keywords, the relevant texts are sourced from the Peking University Law Database (https: / / www.pkulaw.com / ), and the search formula is constructed as “title=((keyword 1) OR (keyword 2) OR … OR (keyword 𝑛 )), PY = (Start Time - Now); S1.3, remove duplicate texts by combining and matching the two fields of document number and title in the data pool, and only keep one of the texts; S1.4. Automatically parse the text genre structure based on natural language processing technology to identify the basic elements such as document title, issuing agency, document type, and drafting and effective date.

[0020] S2. Document sample data annotation and conversion: Collect a certain proportion of document samples, combine entity type definitions and prompt words, use Doccano software (open source multi-functional text annotation platform) to annotate discontinuous governance goals, tools and entity relationships, and divide the training, validation and test sets according to the ratio of [0.7, 0.2, 0.1]. It includes the following steps: S2.1. Governance goal annotation: Based on the governance goal entity type and verb phrases such as "realize / promote", use Doccano software to use the relationship pointer from one end of the discontinuous entity to the other end. If it is a continuous entity, it is directly annotated as the entity type of governance goal; S2.2, Governance tool annotation, based on the entity type of governance tool combined with common verbs and nouns including semantic structures such as "setting standards, establishing systems, organizing and carrying out", use Doccano software to use the relationship pointer from one end of the discontinuous entity to the other end. If it is a continuous entity, it is directly marked as the entity type of governance tool; S2.3. Annotation of the relationship between governance objectives and tools. Combined with the relationship between adopting governance tools to achieve governance objectives, annotate the relationship between governance objectives and tool models. S2.4, label data reading, assuming that the input sentence X consists of N words, , the model for each word pair (word pair( )) to predict the two word relationship categories R, where Among them, None means that there is no relationship between the two words and they do not belong to the same entity; NNW, namely Next-Neighboring-Word, means that the two words are adjacent to each other in the same entity; THW-* means that the two words are in the same entity and are the end and the beginning of the entity respectively, which is used to determine the category and boundary of the entity, where * indicates the entity type. The annotation data is read as a word-word two-dimensional relationship matrix. like Figure 2 The labeled dataset is exported as a json file. The labeled entity set includes entity ID, category, starting position, and end point. The entity relationship set includes the starting entity ID, target entity ID, and relationship category. S2.5, entity-relationship data preprocessing, define the target data format, limit the length to the maximum sentence length, store the index range and category information of entities and relationships as entities; if the entity is an isolated entity, treat it as a self-loop relationship, generate a virtual relationship of "starting entity equals target entity", traverse the entities and remove isolated entities that do not participate in any relationship, extract the text index range of the starting entity and the target entity to generate relationship data; S2.6. Dataset division: divide the data into a training set for model training, a validation set for model parameter tuning, and a test set for model performance testing in proportion.

[0021] S3, W2NER model construction: Combine the bidirectional encoder representation from transformers (BERT), bidirectional long short-term memory (LSTM), convolutional neural network and biaffine mechanism to build a model for sequence labeling and relationship extraction tasks, extract the semantic features of text, and extract governance goals, tool entities and relationships, such as Figure 4 As shown, the following steps are included: S3.1, the model receives data preprocessing, including sentence word ID sequence, subword to word mapping matrix, two-dimensional mask matrix, entity distance matrix and sentence length, and verifies and masks the input data to ensure the consistency of data format and mask invalid positions; S3.2. Contextual representation generation based on multi-layer feature extraction. Use the pre-trained BERT model to contextually encode the input sequence and extract word-level or sub-word-level embeddings; Use the pre-trained BERT model to contextually encode the input sequence and extract word-level or sub-word-level embeddings; Input the word-level embeddings into the bidirectional LSTM to capture the global contextual information of the sentence; S3.3, conditional normalization processing, using the conditional layer normalization (CLN) module to enhance the features of LSTM output; S3.4, integrate distance and relationship features, map the distance matrix between entities into a high-dimensional embedding representation, enhance the relationship modeling capability, generate embeddings of relationship types based on the relationship matrix, further enrich the feature expression of sentences, and concatenate distance embedding, relationship embedding, and context representation (LSTM output processed by CLN) to form feature input, such as Figure 3 As shown, dist_inputs refers to the relative position information of grid characters, dimension [B, L, L]; grid_labels refers to the two-dimensional matrix of THW and NHW relationship constructed by the annotation data entity, dimension [B, L, L]; S3.5. Use a multi-dilation rate convolution layer to locally model the feature input, and use the convolution layer output as the feature input for subsequent prediction modules. S3.6, Bi-affine interaction modeling, projects entity representations into a specific space, calculates the interaction scores of entity pairs, and predicts the relationship categories between entities based on the interaction scores.

[0022] S4. W2NER model training: training model of text entity and relationship joint extraction model based on adversarial training and dynamic learning rate adjustment; including the following steps: S4.1. Data loading and preprocessing: loading training sets, validation sets, and test sets. The data formats include text, entity annotations, and relationship annotations. The distribution of loaded data is checked and counted to ensure the balance of training samples. S4.2. Model initialization: build a joint extraction model according to the configuration file. The model structure includes using BERT for context feature extraction, bidirectional LSTM to further capture global context, convolutional layer for local feature modeling, and dual affine mechanism for joint prediction of entities and relationships; setting the loss function and grouping parameters, and setting the learning rate scheduler to dynamically adjust the learning rate; S4.3, model training, set the model to training mode, perform forward propagation, loss calculation, adversarial training, gradient clipping and parameter update, and learning rate update steps for each batch of data; after each training round, calculate the average loss and record the predicted accuracy, recall rate and F1 score; S4.4, model verification, set the model to evaluation mode, verify batch by batch, perform forward propagation and prediction, and statistically analyze the performance, and save the current model parameters whose F1 score of the verification set is higher than the historical best; S4.5, early stopping mechanism: when the performance of the validation set fails to improve for several consecutive rounds, the number of stagnation times is recorded; when the number of stagnation times reaches the set threshold, the training is terminated early to avoid overfitting; S4.6, model saving and loading, save the model parameter file in the specified path for subsequent loading or reasoning; load the saved model parameters to support continued training or direct evaluation.

[0023] S5. W2NER model extraction: A text entity and relationship joint extraction method based on deep learning is used to extract governance objectives, tool entities, and entity relationships in the text; including the following steps: S5.1, text segmentation, perform character-level segmentation on the input text, replace consecutive spaces with single spaces, and remove redundant characters; S5.2, data loading and construction, using the segmented text to build a prediction data set, including the text segmentation results and auxiliary features, encapsulating the text into the model input format, and loading the data in batches; S5.3, model loading and initialization, loading the joint extraction model, loading the trained model parameters; S5.4, model reasoning, input the batch loaded data into the model, perform forward propagation batch by batch, calculate the category score of each position, and classify and output the entity category (target or tool) and entity relationship category; S5.5, decoding and result generation, decoding the index and category output by the model, mapping the decoded entity and relationship results into a readable format, and finally generating a JSON file of the extraction results; S5.6, Result storage, output the extracted governance goals, tools and relationships, and store the governance combination (governance goals, implementation, governance tools) in local files according to the file ID.

[0024] S6. Construction of the governance goal-tool citation network: Extract the citation and cited relationships between texts based on the book title marks and citation prompts in the text content; identify, clean and integrate the governance goals and tools in the text through the W2NER model, and build a governance goal-tool citation network based on this information; including the following steps: S6.1. Identify identifiers, based on the string matching algorithm (Knuth-Morris-Pratt, KMP) to efficiently identify the book title marks embedded in the text content; S6.2. Complete the file name, obtain the file issuing agency, and construct the complete file name by combining the issuing agency and the specific file in the book title quotation marks; S6.3, supplement the citation relationship, identify the prompt words "based on", "implement", "according to", "regulations", "combination", "comply with" before and after the quotation marks, and further identify and supplement the citation relationship; S6.4, citation relationship storage, according to the logic of the files cited in this article and the files citing this article, store the file citation relationship (citing file ID, cited file ID); S6.5. Node construction and attribute mapping. Each governance goal or tool is defined as a node in the reference network. The attributes of the node include: node name, type (governance goal or tool), source text ID, and frequency of occurrence. The size of the node is proportional to the frequency of occurrence of the governance goal or tool in all texts. Nodes with higher frequencies will be displayed as larger nodes in the network. S6.6, Governance goal-tool combination reference network, combining the governance goals and tools in the reference network, and according to the existing governance goal-tool combination pattern, build a "governance goal-tool" combination reference network. Each text data will contain multiple governance goal and tool combination patterns, indicating a many-to-many relationship between governance goals and tools.

[0025] S7. Identification of core governance goals and tools: Use the Ego-Sphere Unit (ESU) algorithm to perform directed motif analysis on the governance goal-tool reference network, and recursively identify the induced subgraphs in the network, especially the multiple relationships and interaction patterns between the goal nodes and the tool nodes, and based on this, identify the core governance goal-tool combination and analyze its evolution process. It includes the following steps: S7.1. Recursive mining using the ESU algorithm: Apply the ESU (Ego-Sphere Unit) algorithm to recursively mine the subgraphs in the governance goal-tool reference network. Each recursion will expand the SUB set, which contains nodes directly adjacent to the current node, and the EXT set contains nodes that are adjacent to at least one node in the SUB set and have a larger numerical label. The ESU algorithm identifies all eligible induced subgraphs through this recursive process. The pseudo code of the ESU subgraph search algorithm is as follows: Figure 5 As shown; S7.2, Motif analysis and subgraph classification: Motif analysis is performed on each branch of the ESU tree (ESU-Tree), and subgraph structures (motifs) that frequently appear in the network are identified through a recursive method. In this process, the ESU algorithm performs an isomorphism test on each subgraph through McKay's nauty algorithm to ensure that subgraphs of different structures are classified and sorted according to frequency and concentration; The schematic diagram of the ESU-Tree tree structure is as follows Figure 6 As shown, attached Figure 6 (a) shows the target graph of size 5; Figure 6 (b) in Figure 1 shows an ESU tree of depth k, which is used to extract a subgraph of size 3 from the target graph. The leaf nodes correspond to the target nodes. Figure 6 The set S4 in (a) or all induced subgraphs of size 4. The nodes in the ESU-Tree consist of two adjacent sets, the first set is SUB, which contains the nodes directly adjacent to the current node; the second set is EXT, which contains all nodes adjacent to at least one SUB node and whose numerical labels are greater than the SUB node. The algorithm expands the SUB set by expanding the EXT set until the required subgraph size (here 3) is reached, which is at the bottom of the ESU-Tree (or its leaf nodes).

[0026] S7.3. Identify the core governance goal-tool combination: Use the ESU algorithm to identify frequent subgraphs in the network (core governance goal-tool combination) and calculate the frequency F of each subgraph in the random network. R (G′). At the same time, the actual frequency F of the subgraph in the target network is calculated G (G′). Use P-value to perform statistical hypothesis testing to evaluate whether a subgraph appears significantly in the target network; at the same time, set the N value to generate multiple random network samples to make the evaluation of subgraph frequency more stable and accurate; ; in, N is the number of random network samples, is the P value, The probability of (as its null hypothesis) is given by In a random network frequency; when P When the value is less than the threshold value of 0.05, the sub-graph can be called a significant pattern.

[0027] S7.5. Visualization of the core governance goal-tool motif. Based on the identification of the core governance goal-tool combination, the network structure of the motif is displayed using visualization technology. The governance goals and tools are used as network nodes, and the size and color of the nodes represent the frequency and type of the nodes. The thickness of the edges reflects the strength of the citation relationship between the governance goals and the tools. In this visualization process, the interactive relationship of the core goal-tool combination is highlighted, and their importance and connection pattern in the overall system are displayed through dynamic visualization; S7.6. Based on the citation network of the core governance goal-tool combination motif, analyze its evolution at different times or governance document implementation stages. By tracking the trend of the core motif over time, identify the changes in the alternation and combination patterns between governance goals and tools. This analysis helps to reveal the evolution path of governance goals and tools, predict future governance directions, and adjust the existing governance combination to improve its implementation effect.

[0028] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, the computer program implements any one of the methods for identifying a core governance objective-tool combination based on reference association.

Claims

1. A method for identifying core governance objectives and tool combinations based on reference associations, characterized in that: The following steps are involved: (1) Select the target area, collect governance documents, build a data pool and perform preprocessing; (2) Text data annotation and conversion: Collect sample data, and use Doccano software to annotate discontinuous governance goals, tools, and relationships between entities according to the prompt words of the goals and tools, and divide them into training, validation, and test sets; (3) Constructing the W2NER model: Combining the bidirectional transformer model BERT, bidirectional long short-term memory LSTM, convolutional layer and biaffine mechanism to build a model for sequence labeling and relationship extraction tasks, extracting the semantic features of the text, and extracting governance objectives, tool entities and relationships; (4) Training the W2NER model: a training model for the joint extraction of text entities and relations based on adversarial training and dynamic learning rate adjustment; (5) Extraction W2NER model: A text entity and relationship joint extraction method based on deep learning is used to extract governance objectives, tool entities, and entity relationships in the text; (6) Constructing a governance goal-tool citation network: Extracting the citation and cited relationships between texts based on the book title marks and citation prompts in the document content; using the W2NER model to identify, clean, and integrate the governance goals and tools in the text, and constructing a governance goal-tool citation network based on this information; (7) Identify core governance objectives and tools: The self-sphere unit algorithm (ESU) is used to perform directed motif analysis on the governance objective-tool reference network. The induced subgraphs in the network, i.e., the multiple relationships and interaction patterns between the objective nodes and the tool nodes, are recursively identified. The evolution process of the core governance objective-tool combination is analyzed based on the identification of the combination.

2. According to claim 1, a method for identifying core governance objectives and tool combinations based on reference associations is characterized in that: In step (1), the preprocessing is to clean the data.

3. According to claim 2, a method for identifying core governance objectives and tool combinations based on reference associations is characterized in that: Step (1) includes the following steps: (11) Construct a keyword search strategy for the target field and search the database for target document data whose titles contain keywords; (12) Identify and summarize a list of relevant keywords in the target field by reading document data; (13) Remove duplicate files by combining and matching the two fields of the document number and title in the data pool, and only keep one file; (14) Automatically parse the text structure of text data based on natural language processing technology to identify the basic elements such as document title, publishing agency, document type, and date of writing and effective date.

4. According to claim 1, a method for identifying core governance objectives and tool combinations based on reference associations is characterized in that: Step (2) includes the following steps: (21) Label the governance objectives. Based on the governance objective entity type and the verb phrase "realize / promote", use the Doccano software to use the relationship pointer from one end of the discontinuous entity to the other end. If it is a continuous entity, directly label it as the entity type of governance objective. (22) Labeling governance tools. Based on the entity type of governance tools and common verbs and nouns, including the semantic structure of "setting standards and establishing systems", the Doccano software is used to use the relationship pointer from one end of the discontinuous entity to the other end. If it is a continuous entity, it is directly labeled as the entity type of governance tool; (23) Mark the relationship between governance objectives and tools, combine the relationship between adopting governance tools to "achieve" governance objectives, and mark the relationship between governance objectives and tool models; (24) Read the annotated data and export the annotated data set as a json file, which contains the annotated entity set and the annotated entity relationship set; (25) Preprocessing of entity-relationship data: defining the target data format, removing isolated entities that are not involved in any relationship, and extracting the text index range of the starting entity and the target entity to generate relationship data; (26) Dataset division: divide the data into a training set for model training, a validation set for model parameter tuning, and a test set for model performance testing in proportion.

5. According to claim 1, a method for identifying core governance objectives and tool combinations based on reference associations is characterized in that: Step (3) includes the following steps: (31) The model receives preprocessed data, including sentence word ID sequence, subword to word mapping matrix, two-dimensional mask matrix, entity distance matrix and sentence length, and verifies and masks the input data to make the data format consistent and mask invalid positions; (32) Generate contextual representations based on multi-layer feature extraction. Use the pre-trained BERT model to contextually encode the input sequence and extract word-level or sub-word-level embeddings. Use the pre-trained BERT model to contextually encode the input sequence and extract word-level or sub-word-level embeddings. Input the word-level embeddings into a bidirectional LSTM to capture the global context information of the sentence. (33) Conditional normalization processing, using the conditional layer normalization module to enhance the features of LSTM output; (34) Fusion of distance and relation features, mapping the distance matrix between entities into a high-dimensional embedding representation, generating an embedding of the relation type based on the relation matrix, and concatenating the distance embedding, relation embedding, and context representation, i.e., the LSTM output processed by CLN, to form the feature input; (35) Use a multi-expansion rate convolutional layer to locally model the feature input and use the convolutional layer output as the feature input; (36) Bi-affine interaction modeling projects entity representations into a specific space, calculates the interaction scores of entity pairs, and predicts the relationship categories between entities based on the interaction scores.

6. According to claim 1, a method for identifying core governance objectives and tool combinations based on reference associations is characterized in that: Step (4) includes the following steps: (41) Data loading and preprocessing: loading training sets, validation sets, and test sets. The data formats include text, entity annotations, and relationship annotations. The distribution of the loaded data is checked and counted. (42) Model initialization: construct a joint extraction model according to the configuration file. The model structure includes using BERT for context feature extraction, bidirectional LSTM to further capture global context, convolutional layer for local feature modeling, and dual affine mechanism for joint prediction of entities and relationships; setting the loss function and grouping parameters, and setting the learning rate scheduler to dynamically adjust the learning rate; (43) Model training: Set the model to training mode, perform forward propagation, loss calculation, adversarial training, gradient clipping and parameter update, and learning rate update steps for each batch of data; after each training round, calculate the average loss and record the predicted accuracy, recall rate, and F1 score; (44) Model verification: set the model to evaluation mode, verify batch by batch, perform forward propagation and prediction, and statistically analyze the performance, saving the current model parameters whose F1 score on the verification set is higher than the historical best; (45) Early stopping mechanism: when the performance of the validation set fails to improve after several consecutive rounds, the number of stagnations is recorded; when the number of stagnations reaches the set threshold, the training is terminated early to avoid overfitting; (46) Model saving and loading: save the model parameter file in the specified path for subsequent loading or inference; load the saved model parameters to support continued training or direct evaluation.

7. According to claim 1, a method for identifying core governance objectives and tool combinations based on reference associations is characterized in that: Step (5) includes the following steps: (51) Text segmentation: perform character-level segmentation on the input text, replace consecutive spaces with single spaces, and remove redundant characters; (52) Data loading and construction: Use the segmented text to build a prediction dataset, including the text segmentation results and auxiliary features. After encapsulating the text into the model input format, load the data in batches; (53) Model loading and initialization, loading the joint extraction model, and loading the trained model parameters; (54) Model reasoning: input batch-loaded data into the model, perform forward propagation batch by batch, calculate the category score of each position, and classify and output the entity category, i.e., the target or tool and entity relationship category; (55) Decoding and result generation: decode the index and category of the model output, map the decoded entity and relationship results into a readable format, and finally generate a JSON file of the extraction results; (56) Result storage: output the extracted governance goals, tools and relationships, and store the governance combination in a local file according to the text ID; the governance combination includes: governance goals, implementation, and governance tools.

8. According to claim 1, a method for identifying core governance objectives and tool combinations based on reference associations is characterized in that: Step (6) includes the following steps: (61) Identification of identifiers: identifying book title marks embedded in text content based on the string matching algorithm KMP; (62) Complete the file name, obtain the issuing agency of the file, and construct the complete file name by combining the issuing agency and the specific file in the quotation marks; (63) Supplement the citation relationship, identify the prompt words "based on", "implemented", "according to", "regulated", "combined with", and "follow" before and after the quotation marks, and further identify and supplement the citation relationship; (64) Citation relationship storage: according to the logic of the files cited in this article and the files citing this article, the file citation relationship, i.e., the citing file ID and the cited file ID, is stored; (65) Node construction and attribute mapping. Each governance goal or tool is defined as a node in the reference network. The attributes of the node include: node name, type (i.e., governance goal or tool), source text ID, and frequency of occurrence. The size of the node is proportional to the frequency of occurrence of the governance goal or tool in all texts. Nodes with higher frequencies will appear as larger nodes in the network. (66) The governance goal-tool combination citation network is constructed by combining the governance goals and tools in the citation network and according to the existing governance goal-tool combination patterns. Each text data will contain multiple governance goal and tool combination patterns, indicating a many-to-many relationship between governance goals and tools.

9. The method for identifying core governance objectives and tool combinations based on reference association according to claim 1 is characterized in that: Step (7) includes the following steps: (71) Use the ESU algorithm for recursive mining: Apply the ESU algorithm to mine the subgraphs in the governance goal-tool reference network in a recursive manner; each recursion will expand the SUB set, which contains nodes directly adjacent to the current node, and the EXT set contains nodes that are adjacent to at least one node in the SUB set and have a larger numerical label; the ESU algorithm identifies all eligible induced subgraphs through this recursive process; (72) Motif analysis and subgraph classification: Motif analysis is performed on each branch of the ESU tree, and a recursive method is used to identify subgraph structures that frequently appear in the network. During this process, the ESU algorithm performs an isomorphism test on each subgraph using McKay's nauty algorithm to ensure that subgraphs of different structures are classified and sorted according to their frequency and concentration. (73) Identify the core governance goal-tool combination: Use the ESU algorithm to identify the frequent subgraphs in the network, that is, the core governance goal-tool combination, and calculate the frequency F of each subgraph in the random network. R (G′); at the same time, calculate the actual frequency F of the subgraph in the target network G (G′); use the P value to perform statistical hypothesis testing to evaluate whether a subgraph appears significantly in the target network; at the same time, set the N value to generate multiple random network samples to make the evaluation of subgraph frequency more stable and accurate; the formula is as follows: ； in, N is the number of random network samples, is the P value, The probability of is given, In a random network frequency; when P When the value is less than the threshold of 0.05, the sub-image is called a significant pattern; (74) Visualization of the core governance goal-tool motif. Based on the identification of the core governance goal-tool combination, the network structure of the motif is displayed using visualization technology. The governance goals and tools are used as network nodes, and the size and color of the nodes represent the frequency and type of the nodes. The thickness of the edges reflects the strength of the citation relationship between the governance goals and tools. (75) Based on the citation network of the core governance objective-tool combination motif, analyze its evolution process at different times or document implementation stages; by tracking the trend of core motifs over time, identify changes in the alternation and combination patterns between governance objectives and tools.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, it implements a method for identifying a core governance objective-tool combination based on reference association according to any one of claims 1 to 9.

Images