Declarative multi-AI agent architecture-based nationality knowledge base construction method

Abstract

The invention relates to the technical field of artificial intelligence and knowledge base construction, and discloses a nationality knowledge base construction method based on a declarative multi-AI agent architecture. The method comprises the steps of obtaining declarative configuration template data and analyzing to generate a structured task demand description, performing vectorization processing on analysis dimensions to generate a dimension semantic dependency graph, generating an agent task execution scheduling plan by utilizing a topological sorting algorithm, identifying a country group by utilizing a graph clustering algorithm and generating a sharing degree matrix, and performing task execution scheduling on the agent task execution scheduling plan. The method comprises the steps of extracting a source country argumentation structure chart and generating an argumentation topology template, evaluating the similarity between a target country and a source country to generate a template applicable dimension list, instantiating the argumentation topology template into a specific argumentation structure of the target country, and cooperatively generating target country knowledge base content and an analysis report.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge base construction technology, and more specifically, to a method for constructing country-specific knowledge bases based on a declarative multi-AI agent architecture. Background Technology

[0002] In the field of international relations research and policymaking, institutions need to build in-depth situational analysis knowledge bases for multiple countries to support transnational policy decisions and risk assessments. Taking the "Global South" countries as an example, institutions need to build knowledge bases covering 16 analytical dimensions for more than 20 countries, including political systems, economic structures, class power, foreign trade dependence, and the degree of financialization. Each country's knowledge base needs to generate 40 to 80 pages of analytical reports and cite more than 3,300 source data points. In this process, non-developers (such as policy researchers) need to configure agent behavior to achieve the large-scale production of multi-country knowledge bases.

[0003] The construction of existing country knowledge bases mainly relies on manual analysis and general knowledge management systems, which typically use fixed analysis frameworks and manual configuration methods to organize the collaborative work of multiple AI agents.

[0004] However, existing methods for constructing country-specific knowledge bases face the following technical challenges: First, the task dependencies among multiple AI agents are complex, with implicit semantic dependencies existing in each analytical dimension. For example, the analysis of the economic structure dimension depends on the analysis results of the foreign trade dependence dimension, but existing scheduling methods cannot automatically identify these implicit dependencies and optimize the execution order, resulting in low agent task execution efficiency. Second, non-developers lack programming skills and find it difficult to configure agent behavior programmatically, while existing methods lack an effective conversion mechanism from declarative configuration to executable constraints. Third, when constructing multi-country knowledge bases in parallel, related country groups often exhibit similar political and economic structures, and their analytical argumentation structures are reusable. However, existing methods fail to utilize this reusability, leading to redundant construction, wasted resources, and difficulty in ensuring the consistency of quality across multiple country reports. Summary of the Invention

[0005] This invention provides a method for constructing a country knowledge base based on a declarative multi-AI agent architecture, which solves the technical problems of difficulty in identifying multi-AI agent task dependencies, difficulty in configuration by non-programmers, and repetitive waste in knowledge base construction in related technologies.

[0006] This invention provides a method for constructing a country knowledge base based on a declarative multi-AI agent architecture, including:

[0007] Obtain declarative configuration template data, parse the analysis dimensions, data source scope, and agent behavior constraints in the template, and generate a structured task requirement description;

[0008] The task descriptions of each analysis dimension are vectorized, the semantic similarity and causal relationship strength between dimensions are calculated, and a dimensional semantic dependency graph is generated. Based on the dimensional semantic dependency graph, the tasks of multiple AI agents are hierarchically divided using a topological sorting algorithm, and an agent task execution scheduling plan is generated.

[0009] Obtain a list of countries to be analyzed and their geographical, political and economic data; use graph clustering algorithms to identify country groups; and generate a country association map and a sharing degree matrix of each analysis dimension within the group.

[0010] Obtain analysis reports of source countries in the country group, extract the arguments, evidence, and reasoning chains in the reports, and generate a source country argumentation structure diagram;

[0011] The source country argumentation structure diagram is abstracted, and specific entities are replaced with type placeholders to generate an argumentation topology template;

[0012] Based on the sharing degree matrix and the basic characteristics of the target country, the similarity between the target country and the source country in each dimension is evaluated, the dimensions with similarity exceeding the threshold are filtered, and a list of applicable dimensions for the template is generated.

[0013] For each dimension in the template applicable dimension list, replace the placeholders in the argument topology template with the specific data and entities of the target country to generate the instantiated argument structure;

[0014] Based on the instantiated argumentation structure, each intelligent agent collaboratively generates the target country's knowledge base content and analysis report.

[0015] The process of generating a dimensional semantic dependency graph includes: inputting the task description text of each dimension into a pre-trained language encoder to obtain vector representations of each dimension; calculating the ratio of the inner product of any two dimensional vectors to their norm product to obtain semantic similarity; analyzing keyword co-occurrence and semantic dependency patterns in the dimensional task descriptions, calculating citation strength, and weighted summing of semantic similarity and citation strength to obtain causal association strength; and establishing directed edges between dimensional pairs with causal association strength exceeding a preset threshold to generate a dimensional semantic dependency graph.

[0016] Furthermore, the generation of country association maps and sharing matrices for each analysis dimension within the group includes:

[0017] By comprehensively calculating geographical, political, and economic data among countries, and then weighting and summing these data, the strength of inter-country connections is obtained.

[0018] Establish edges between country pairs whose association strength exceeds a preset threshold to construct a country association graph;

[0019] A graph clustering algorithm is applied to the country association graph to divide the nodes into multiple country groups based on the tightness of the connections between nodes, thereby generating a country association graph.

[0020] For each country group, calculate the feature similarity score of each analysis dimension for each country within the group, average the feature similarity scores of each country within the group, obtain the sharing degree of each analysis dimension within the group, and generate a sharing degree matrix.

[0021] Furthermore, the generated source country argumentation structure diagram includes:

[0022] Sentence segmentation and semantic analysis are performed on the source country's analysis report text to identify declarative sentences expressing arguments, factual sentences expressing evidence, conjunctions expressing inference relationships, and logical markers.

[0023] For each argument, identify the set of evidence supporting the argument, as well as the reasoning chain from the evidence to the argument; organize the arguments, evidence, and reasoning chain into a directed graph structure, where the nodes include argument nodes and evidence nodes, and the directed edges represent supporting or reasoning relationships, generating the source country argument structure graph.

[0024] Furthermore, generating the argument topology template includes:

[0025] Identify specific entities in the text of each node in the source country argumentation structure diagram; determine the type of each entity using named entity recognition technology;

[0026] Each specific entity is replaced with a placeholder of the corresponding type. The logical structure of the reasoning steps in the reasoning chain is preserved, but the parts involving specific entities are replaced with placeholders to generate an argument topology template.

[0027] To demonstrate the type constraints associated with each placeholder in the topology template, we define the type requirements for the data when filling the placeholder and record the semantic relationships between each placeholder.

[0028] Furthermore, the list of dimensions applicable to generating the template includes:

[0029] For each analysis dimension, extract the feature vectors of the target country and the source country on the dimension-related features;

[0030] The similarity between the target country and the source country in this dimension is obtained by calculating the ratio of the inner product of the feature vectors of the target country and the source country to their norm product.

[0031] Filter dimensions with similarity exceeding a preset threshold and add those dimensions to the template's applicable dimension list;

[0032] Based on the sharing degree of this dimension in the target country's group in the sharing degree matrix, when the sharing degree is high and the similarity is high, this dimension is preferentially added to the template applicable dimension list.

[0033] Furthermore, the generated instantiated argument structure includes:

[0034] For each dimension in the template applicable dimension list, obtain the target country's dimension-related data;

[0035] For each placeholder in the argument topology template, retrieve the matching specific entity or value from the target country data according to its type constraint;

[0036] The retrieved data is filled into the corresponding placeholder positions, and the filled content is verified to meet the semantic association constraints between the placeholders, thus generating the instantiated argument structure.

[0037] Furthermore, after generating the instantiated argument structure, the following steps are also included:

[0038] Perform multi-dimensional defect detection on the instantiated argument structure, including evidence sufficiency assessment, logical validity test, over-inference identification, and evidence bias detection, and generate an argument defect diagnosis report;

[0039] Based on the deficiencies in the argument defect diagnosis report, the argument repair algorithm is used to supplement missing reasoning links, replace inapplicable arguments, and add necessary qualifiers to generate an enhanced argument structure.

[0040] Use the enhanced argumentation structure to generate content for the target country's knowledge base.

[0041] Furthermore, performing multi-dimensional defect detection includes:

[0042] For each argument node in the argument structure, check the quantity and quality of its supporting arguments. When the number of supporting arguments for an argument is less than a preset minimum threshold, or the reliability score of the arguments is lower than the threshold, the argument is judged to have the defect of insufficient evidence.

[0043] For each inference relation edge in the argument structure, analyze the logical rationality of the inference chain, detect whether there are logical jumps, premise contradictions or unsupported conclusions in the reasoning process, and determine that the inference chain has a logical invalidity defect when a logical error is detected.

[0044] For arguments in the argument structure, analyze whether the strength of the conclusion matches the strength of the evidence. By calculating the difference between the evidence support and the certainty of the conclusion, when the difference exceeds a threshold, the argument is judged to have the defect of over-inference.

[0045] For the argument nodes in the argument structure, the diversity and representativeness of the argument sources are detected. By calculating the entropy value and the uniformity of the time distribution of the argument sources, it is determined that there is a defect in evidence bias when the entropy value is lower than the threshold or the distribution is uneven.

[0046] Furthermore, in generating the target country's knowledge base content and analysis report, it also includes:

[0047] Establish version association records for cross-border argumentation templates, linking the argumentation structure diagram of the source country, argumentation topology template, target country, and identifiers for each applicable dimension;

[0048] When the argument structure of the source country is updated, compare the argument structure diagrams before and after the update, identify the changed nodes and edges, and determine whether the changes affect the structure of the argument topology template.

[0049] When the template structure changes, a synchronization update check for related countries is triggered, an update notification is sent to each related country, and the slot filling and instantiation process is re-executed.

[0050] This invention provides a country knowledge base construction system based on a declarative multi-AI agent architecture, comprising:

[0051] The configuration parsing module is used to obtain declarative configuration template data, parse the analysis dimensions, data source scope and agent behavior constraints in the template, and generate a structured task requirement description.

[0052] The dimension dependency analysis module is used to vectorize the task descriptions of each analysis dimension, calculate the semantic similarity and causal relationship strength between dimensions, and generate a dimension semantic dependency graph.

[0053] The task scheduling module is used to hierarchically divide the tasks of multiple AI agents based on the dimensional semantic dependency graph and use the topological sorting algorithm to generate an agent task execution scheduling plan.

[0054] The country association analysis module is used to obtain a list of countries to be analyzed and their geographical, political and economic association data, use graph clustering algorithms to identify country groups, and generate a country association map and a sharing degree matrix of each analysis dimension within the group.

[0055] The argument structure extraction module is used to obtain the analysis reports of source countries in the country group, extract the arguments, evidence and reasoning chains in the reports, and generate the argument structure diagram of the source countries.

[0056] The argument template generation module is used to abstract the source country argument structure diagram, replace specific entities with type placeholders, and generate an argument topology template.

[0057] The template applicability assessment module is used to evaluate the similarity between the target country and the source country in each dimension based on the sharing degree matrix and the basic characteristics of the target country, filter the dimensions with similarity exceeding the threshold, and generate a list of applicable dimensions for the template.

[0058] The argument instantiation module is used to replace the placeholders in the argument topology template with the specific data and entities of the target country for each dimension in the template applicable dimension list, and generate the instantiated argument structure.

[0059] The knowledge base generation module is used to collaboratively generate the target country's knowledge base content and analysis report based on the instantiated argument structure by various intelligent agents.

[0060] The beneficial effects of this invention are as follows:

[0061] This invention lowers the barrier to entry for non-developers by adopting a declarative configuration approach, enabling policy researchers to configure agent behavior without programming, thus solving the technical problem of difficult configuration in existing methods. By automatically identifying implicit semantic dependencies between dimensions and generating dimensional semantic dependency graphs, it optimizes the task scheduling efficiency of multi-AI agents, solving the problem of low execution efficiency caused by the inability of existing scheduling methods to handle implicit dependencies. Through national group identification and argument structure reuse mechanisms, it achieves large-scale construction of multi-country knowledge bases, significantly reducing the workload of repetitive analysis, ensuring the consistency of quality of multi-country reports, and achieving significant technical effects in improving the efficiency and quality of knowledge base construction. Attached Figure Description

[0062] Figure 1 This is a flowchart of the method for constructing a country knowledge base based on a declarative multi-AI agent architecture, as described in this invention.

[0063] Figure 2 This is the causal dependency graph between dimensions of the present invention;

[0064] Figure 3 This invention provides a heatmap of the clustering results and association strength of national groups.

[0065] Figure 4 This is a bar chart illustrating the hierarchical scheduling plan for the intelligent agent task execution of this invention.

[0066] Figure 5 This is a bar chart showing the distribution of the sharing degree of each dimension of the present invention in the South American group;

[0067] Figure 6 This is a line graph illustrating the improved argumentation quality scoring process of this invention.

[0068] Figure 7 This is a scatter plot of the multi-country knowledge base version synchronization status tracking of the present invention. Detailed Implementation

[0069] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0070] At least one embodiment of the present invention discloses a method for constructing a country knowledge base based on a declarative multi-AI agent architecture, such as... Figure 1 As shown, it includes the following steps:

[0071] Step 100: Obtain the declarative configuration template data filled in by non-developers, parse the slot content such as analysis dimensions, data source range, and agent behavior constraints in the template, and generate a structured task requirement description.

[0072] In step 100, the declarative configuration template data contains multiple slots, each slot corresponding to different types of configuration information.

[0073] The Analysis Dimension slot defines the names and descriptions of the 16 dimensions to be analyzed, such as political system, economic structure, and foreign trade dependence. The Data Source Scope slot specifies the types of data sources and access permissions that can be used for each dimension. The Agent Behavior Constraint slot defines the rules that each agent must follow when performing tasks, such as the minimum number of sources to cite and the depth requirements for argumentation.

[0074] Furthermore, the behavioral constraints of the intelligent agent are defined in a structured format, including numerical constraints and Boolean constraints. The minimum number of cited sources is defined as an integer value, typically ranging from 3 to 10; the depth requirement for reasoning is defined as the minimum number of reasoning steps in the reasoning chain, typically ranging from 2 to 4 steps; and the data timeliness constraint is defined as the maximum time elapsed between the data's publication date and the current time, in years, typically ranging from 3 to 5 years. Each constraint rule is recorded in the configuration template as a key-value pair, such as "min_sources: 3, reasoning_depth: 2, data_timeliness: 5", facilitating parsing into structured constraints.

[0075] After parsing the content of the above slots, the text content of each slot is converted into a structured data format to generate a structured task requirement description, which includes a dimension list, a data source configuration list, and a constraint rule list.

[0076] Furthermore, the parsing process for slot content employs a combination of rule matching and regular expressions. For analysis dimension slots, commas or newlines are used to identify boundaries between dimensions, and keywords are extracted from each dimension's text to separate the dimension name from its description. For data source range slots, key information is located by identifying markers such as "source," "type," and "permissions." For agent behavior constraint slots, constraint parameters are extracted by identifying key-value pairs in the format "key:value" or "key=value." For slot content that does not conform to the specified format, the system issues a warning and requires re-entry to ensure data validity.

[0077] Step 200: Use a pre-trained language encoder to vectorize the task descriptions of each analysis dimension, calculate the semantic similarity and causal relationship strength between dimensions, and generate a dimensional semantic dependency graph.

[0078] In step 200, the analysis dimensions and their task description texts from the structured task requirement description generated in step 100 are obtained. The task description texts for each dimension are then input into a pre-trained language encoder, which maps the text into high-dimensional vector representations, yielding vectors for each dimension. ,in , To analyze the number of dimensions, For vector dimensions.

[0079] The aforementioned pre-trained language encoder employs a Transformer architecture. Its input layer receives a sequence of task description text after word segmentation, and its output layer outputs a vector representation of the text sequence. The pre-trained language encoder is obtained through pre-training on a large-scale corpus. Pre-training utilizes a masked language model task, randomly masking some words in the input text sequence. The encoder is trained to predict the masked words based on the context. The loss function is cross-entropy loss, and the optimization algorithm is the Adam optimization algorithm. In application to this invention, the task description text for each dimension is input into the pre-trained language encoder, and the average pooling result of the last hidden state is extracted as the vector representation for that dimension.

[0080] Furthermore, word segmentation employs a combination of dictionary-based and statistical model-based methods. Chinese text is segmented using word segmentation tools (such as jieba or BERT's built-in segmentation), while English text is segmented based on spaces and punctuation marks. Each segmented word is then mapped to the encoder's vocabulary to obtain word vectors. The hidden layer dimension of the pre-trained encoder is... Typically, the value is 768 or 1024, which can be adjusted according to the specific application scenario and computing resources. A value of 768 is recommended to balance representational power and computational efficiency. The final hidden state contains a context-aware representation of each word, and average pooling is used to aggregate the word representations of the entire text sequence into a single vector. The average pooling calculation formula is: ,in The length of the text sequence. For word index, For the first The hidden state of each word element.

[0081] For any two dimensions and Calculate the cosine similarity of their vector representations to obtain the semantic similarity. :

[0082]

[0083] in, Represents the dot product of vectors. Representing vectors The Euclidean norm.

[0084] Furthermore, cosine similarity The value of is in the range [-1, 1], representing the degree of directional similarity between two vectors. In this invention, since the vectors of the task description text processed by the pre-trained language encoder are usually located in the same quadrant of the high-dimensional space, semantic similarity is... The actual values ​​are mainly distributed in the range of [0, 1]. The closer the value is to 1, the higher the semantic similarity between the two dimensions.

[0085] It should be noted that the strength of causal relationship Representing dimensions The analysis results for the dimensions The degree of influence of the analysis. For any dimension... This study utilizes a causal association identification algorithm to analyze keyword co-occurrence and semantic dependency patterns in dimensional task descriptions, and calculates the strength of causal associations. .

[0086] The aforementioned causal association identification algorithm includes the following steps:

[0087] The first step is to obtain the dimensions. and dimensions The task description text is used to extract a set of keywords from each description. The keywords include core concept words and verbs.

[0088] Furthermore, keyword extraction employs the Term Frequency-Inverse Document Frequency (TF-IDF) method. First, the task description text is segmented and part-of-speech tagging is performed, selecting nouns and verbs as candidate keywords. The TF-IDF value of each candidate word is calculated, and the top K words with the highest TF-IDF values ​​are selected as keywords (K is typically 5-10). Part-of-speech tagging is achieved using a rule-based and statistical model-based tool, labeling word elements as nouns (n), verbs (v), adjectives (a), etc. Core concept words are nouns with high TF-IDF values, while verbs represent actions or relationships. The second step involves constructing a keyword co-occurrence matrix, with statistical dimensions... Keywords and dimensions appearing in the description The co-occurrence frequency of keywords in the description.

[0089] Furthermore, the co-occurrence frequency is calculated using a sliding window method. The window size is set to... (Usually takes 5-10 tokens), in dimension In the task description text, for each dimension Keywords, before and after their appearance Detection dimensions within the range of each word element The co-occurrence matrix is ​​obtained by checking if the keyword appears; if it does, its frequency is incremented by 1. After iterating through all keyword combinations, a co-occurrence matrix is ​​obtained. , of which elements Representing dimensions The Key words and dimensions The The third step is to identify the co-occurrence frequency of each keyword. Does the description include information about dimensions? The citation of the analysis results includes: dimension-based citations. Results and basis dimensions The analysis of language patterns, such as co-occurrence frequency and the number of quoted expressions, is performed to count the occurrences of quoted expressions. The fourth step involves calculating the original citation strength based on the co-occurrence frequency and the number of quoted expressions; the original citation strength is directly proportional to both. The original citation strength is then normalized by scaling it to the [0, 1] interval to obtain the normalized citation strength. Step 5: Obtain the semantic similarity calculated in step 200. The causal association strength is obtained by weighted summation of semantic similarity and normalized citation strength.

[0090]

[0091] in, and For the weighting coefficients, satisfying .

[0092] Furthermore, the strength of causal association The calculation is based on normalized semantic similarity and reference strength, both of which take values ​​in the range [0, 1]. The value range is also [0, 1]. The constraint condition is... and Ensure that the weight coefficients are non-negative and sum to 1.

[0093] Furthermore, weighting coefficients and The method was determined by performing a grid search on a small set of candidate parameters. Different parameter combinations are selected to calculate the causal relationship strength between sample dimension pairs, and the relationship is compared with the dependency of manually labeled samples. The parameter combination with the highest accuracy is selected as the final weight coefficient.

[0094] The objective function of grid search is to maximize classification accuracy. ,in , To determine the correct number of dimension pairs for dependencies, Let the sample dimension be the total number. The optimization constraint is... and Since the candidate parameter set is finite (5 candidate values), the time complexity of grid search is O(n log n). ,in This represents the number of candidate combinations, typically selected from a small sample size (10-20 dimension pairs). A typical value is... This indicates that semantic similarity and citation strength contribute equally to causal association; if more emphasis is placed on explicit citation relationships, it can be set to... .

[0095] Treat all dimensions as nodes, and consider causal relationships whose strength exceeds a preset threshold. Dimensions Establish directed edges between them The weight of the edge is Generate a dimensional semantic dependency graph ,in For a set of dimension nodes, It is a set of directed edges.

[0096] Furthermore, the strength of causal association The value range is [0, 1], and the preset threshold is... Used to filter weak associations; only causal associations with a strength exceeding [a certain threshold] are considered. Only pairs of dimensions are considered to have a significant causal dependency, thus establishing directed edges. Threshold The selection method is: initial settings After generating the directed graph, the system checks for cyclic dependencies. If a cycle is found, the threshold is gradually increased (in increments of 0.05) until the cycle is eliminated. Simultaneously, the average out-degree of the generated dependency graph is checked. If it is excessively sparse (mean out-degree < 0.5), the threshold is appropriately decreased to increase the number of dependencies. After generating the directed graph, it needs to be verified that the resulting graph is a directed acyclic graph (DAG). If cyclic dependencies exist, the threshold needs to be adjusted. The size of the edge, or manual intervention to exclude the edge with the smallest weight in the loop path, is determined to ensure that topological sorting can be performed.

[0097] Figure 2 The results of constructing the dimensional semantic dependency graph in step 200 are shown.

[0098] Step 300: Based on the dimensional semantic dependency graph, use the topological sorting algorithm to hierarchically divide the tasks of multiple AI agents and generate an agent task execution scheduling plan.

[0099] In step 300, the dimensional semantic dependency graph generated in step 200 is obtained. Perform a topological sorting algorithm on the dimensional semantic dependency graph. The input of the topological sorting algorithm is a directed acyclic graph. The output is the hierarchical partitioning result of the nodes in the graph. The topological sorting algorithm first calculates the in-degree of each node, i.e., the number of edges pointing to that node. Nodes with an in-degree of 0 are assigned to the first level. After removing these nodes from the graph, the in-degree of the remaining nodes is recalculated, and nodes with a new in-degree of 0 are assigned to the second level, and so on, until all nodes are partitioned. The agent tasks corresponding to the dimensions of each level can be executed in parallel, and tasks at different levels are executed sequentially according to the hierarchical order. Based on the above hierarchical partitioning results, an agent task execution scheduling plan is generated. The agent task execution scheduling plan includes the task list of each level, the agent identifier corresponding to each task, and the definition of the data transfer interface between tasks.

[0100] Figure 4 The execution scheduling plan of the agent task generated by the topology sorting in step 300 is shown.

[0101] Step 400: Obtain the list of countries to be analyzed and their geographical, political and economic correlation data, use graph clustering algorithm to identify country groups, and generate country correlation graph and sharing degree matrix of each analysis dimension within the group.

[0102] In step 400, the list of countries to be analyzed contains the identifiers and basic information of all countries for which the knowledge base needs to be built. Geographical association data includes geographical distance between countries and whether they belong to the same regional organization. Political association data includes diplomatic relations between countries and similarity of political systems. Economic association data includes trade volume between countries and similarity of economic development levels. The above-mentioned original association data are normalized, scaling data of different dimensions such as geographical distance and trade volume to the [0, 1] interval. The normalized association data are then comprehensively calculated to obtain the data for any two countries. and The strength of the correlation between :

[0103]

[0104] in, , , These are the normalized strength values ​​of geographical, political, and economic connections, respectively. , , The corresponding weight coefficients and satisfying .

[0105] Furthermore, weighting coefficients , , The weighting is determined using the Analytic Hierarchy Process (AHP). First, pairwise comparison matrices are constructed for the three dimensions of geography, politics, and economics. Domain experts rate the relative importance of each dimension using a 1-9 scale, where 1 indicates that both dimensions are equally important, and 9 indicates that one dimension is extremely important compared to the other. Based on the comparison matrices, eigenvectors for each dimension are calculated and normalized to obtain the weight coefficients. Typical values ​​are... This indicates that in country-specific correlation analysis, economic correlation is the most important, followed by political correlation, and geographical correlation is the least important. The weighting coefficients can be adjusted according to actual needs for different application scenarios.

[0106] Treat all countries as nodes, and determine the correlation strength if it exceeds a preset threshold. The country to Establish an edge between them, with the edge weight being... Constructing a national relationship graph ,in For a set of national nodes, Let it be the set of edges.

[0107] Furthermore, the strength of the association The value range is [0, 1], and the preset threshold is... Used to filter weak associations; only associations with a strength exceeding [a certain threshold] are considered. Only countries establish borders with each other. Threshold The selection method is as follows: based on the statistical distribution of national connectivity strength, a threshold is typically chosen that makes the average degree of the graph (i.e., the average number of other countries connected to each country) reach a preset target. A typical value is [value missing]. At this point, each country in the graph is connected to an average of 3-5 related countries; to identify stronger country connections, the threshold can be increased. If more weak associations are needed for group discovery, the threshold can be lowered to [value missing]. .

[0108] National Relationship Diagram The graph clustering algorithm is applied, with the input being a weighted undirected graph. The output is the grouping result of the nodes.

[0109] Furthermore, graph clustering employs the Louvain algorithm based on modularity optimization. The Louvain algorithm determines the optimal group partitioning by iteratively optimizing the modularity index. Modularity Defined as the difference between the edge weights within the group and the expected edge weights according to the random connection assumption, the calculation formula is: ,in The total edge weight of the graph. and For node indexing, For nodes and Edge weights between them For nodes The degree, For nodes Group to which it belongs This is an indicator function (1 when two nodes are in the same group, 0 otherwise).

[0110] Furthermore, the parameter calculation method in the modularity formula is as follows: total edge weight That is, half the sum of the weights of all edges in the graph, where and Indexing nodes; nodes degree That is, with nodes The sum of the weights of all connected edges, where For node indexing. The optimization objective is to maximize modularity. ,Right now The constraint is that each node must belong to one and only one group, i.e. ,in For group indexing, For node indexing. The algorithm convergence condition is the module degree increment between two consecutive iterations. (generally ), or reach the maximum number of iterations (usually set to 100).

[0111] The Louvain algorithm consists of two phases: The first phase moves nodes one by one to adjacent groups and selects the group with the largest modularity gain until no further improvement is possible; the second phase merges the groups generated in the first phase into new nodes, repeating the first phase until the algorithm converges, ultimately obtaining the optimal group partitioning. A country association graph is generated, containing the member list of each country group and the relationships between groups.

[0112] For each country group, calculate the share of each analysis dimension within that group. Each dimension, in the group Internal sharing Defined as:

[0113]

[0114] in, For groups The number of countries in China For the country In the Feature similarity scoring across multiple dimensions. The feature similarity score is calculated as follows: for countries... Extract its in the first Feature vectors related to each dimension Calculate the feature vector and the group The average cosine similarity of the feature vectors of all other countries in this dimension, i.e. ,in For the country In the A feature vector with 3 dimensions.

[0115] Furthermore, feature similarity scoring Calculated using cosine similarity, the original value range is [-1, 1], which is mapped to the interval [0, 1] in subsequent applications. Sharedness It is the average score of the similarity of features among the countries in the group, so its value range is [0, 1]. The closer the value is to 1, the higher the degree of sharing of this dimension in the group, and the stronger the reusability of the argument structure.

[0116] The sharing degree of each dimension in each group is organized into a sharing degree matrix. , of which Line number Column elements Indicates the first Each dimension in the group Internal sharing degree.

[0117] Figure 3 The results of the country association graph clustering in step 400 are displayed.

[0118] Figure 5 This demonstrates the sharing of each analysis dimension calculated in step 400 within the South American group. .

[0119] Step 500: Obtain the completed high-quality analysis reports of source countries in the country group, and use the argument structure parsing algorithm to extract the arguments, evidence and reasoning chains in the reports to generate the source country argument structure diagram.

[0120] In step 500, the source country is a country in the country group that has completed the construction of the knowledge base and has a high-quality analysis report. Its analysis report can be used as a reference for the analysis of other countries.

[0121] Furthermore, the quality of the analysis report is assessed through a comprehensive scoring system. The scoring is based on three dimensions: completeness of the report, rigor of argumentation, and reliability of data sources, with a value range of [0, 1]. When the comprehensive score... At that time, the report was considered to be of high quality and could be used as a source country. The completeness was calculated based on the proportion of analytical dimensions covered in the report; the rigor of the argumentation was calculated based on the logical defect rate of the argumentation structure diagram; and the reliability of the data sources was calculated based on the average authority score of the cited sources.

[0122] The analysis report text of the source country is processed using an argument structure parsing algorithm.

[0123] The aforementioned argument structure parsing algorithm includes the following steps:

[0124] The first step is to segment the report text into sentences, dividing it into several sentences, and then performing part-of-speech tagging and dependency parsing on each sentence.

[0125] Furthermore, sentence segmentation employs rule-based and statistical model-based methods, segmenting the text by recognizing punctuation marks such as periods, question marks, and exclamation marks, as well as line breaks. For Chinese text, it also needs to recognize Chinese punctuation marks such as periods "." and semicolons ";". Part-of-speech tagging uses a pre-trained part-of-speech tagging model (such as BERT-POS or CRF models) to label each word as a noun (n), verb (v), adjective (a), adverb (d), etc. Dependency parsing uses a dependency parsing model (such as the BiAffine dependency parser) to output the dependency structure of the sentence, generating headwords and dependency relation types (such as subject SBJ, object OBJ, modifier ATT, etc.).

[0126] The second step is to identify declarative sentences that express arguments. By detecting whether the sentences contain conclusive markers (such as "therefore", "indicates", "explains", etc.) and the tone of the sentences, sentences containing evaluative or conclusive statements are marked as candidate sentences for arguments.

[0127] Furthermore, argument identification employs a hybrid approach. First, it checks for the presence of words from a conclusive marker lexicon containing over 100 common conclusive expressions, such as "therefore," "so," "explain," "prove," and "show." Second, it analyzes the sentence's tone characteristics by identifying the subject-verb-object structure and detecting whether the verb is an evaluative verb (such as "need," "should," or "ought to") to determine whether it is a normative or suggestive statement. Finally, it calculates the argument probability score for each sentence. If the score exceeds the threshold of 0.5 (usually set to 0.4-0.6), the sentence is marked as a candidate argument.

[0128] The third step is to identify factual sentences that express arguments. By detecting whether the sentences contain data, statistical results, historical facts, etc., sentences containing descriptions of objective facts are marked as candidate sentences for arguments.

[0129] Furthermore, the argument identification employs a combination of pattern matching and semantic analysis. Numerical evidence is matched using regular expressions, while named entity recognition detects entities such as time, location, and organization; sentences containing these entities tend to contain factual statements. Simultaneously, an objectivity score is calculated for each sentence. By detecting the frequency of subjective words (such as "maybe," "perhaps," "think"), the fewer subjective words, the higher the objectivity score. (Argument probability score) Based on a comprehensive analysis of numerical features, statistical expressions, and objectivity scores, sentences with scores exceeding a threshold of 0.5 are marked as candidate arguments.

[0130] The fourth step is to identify conjunctions and logical markers that express reasoning relationships, including causal relationship markers (such as "due to" or "cause of"), progressive relationship markers (such as "further" or "in addition"), and adversative relationship markers (such as "however" or "but").

[0131] The fifth step is to analyze the candidate sentences for each argument. By analyzing the inference relationship markers and candidate sentences of arguments in the preceding and following text, the set of arguments supporting the argument can be identified. ,in The first argument supporting this point is... To establish a support relationship between the evidence and the argument for point 2.

[0132] Step 6: Construct a chain of reasoning from evidence to argument. A reasoning chain contains intermediate reasoning steps and logical relationships, describing how an argument is derived from the evidence.

[0133] Step 7: Organize the arguments, evidence, and reasoning chains into a directed graph structure. The nodes in the graph include argument nodes and evidence nodes, and the directed edges represent supporting or reasoning relationships, generating the source country argument structure graph. ,in For the set of argument and evidence nodes, This is a set of supporting relation edges and reasoning relation edges.

[0134] Furthermore, the confidence score of an argument ranges from [0, 1] and is calculated based on the quantity and quality of supporting evidence. The calculation formula is the weighted average of the reliability scores of the supporting evidence. The reliability score of the evidence also ranges from [0, 1] and is calculated based on the authority and timeliness of the evidence source. Authority is assessed based on the academic standing or official nature of the source institution, while timeliness is assessed based on the distance between the data's publication date and the current time; the closer the distance, the higher the score.

[0135] Step 600: Abstract the source country argumentation structure diagram, and use the topological abstraction algorithm to replace specific entities with type placeholders to generate a reusable argumentation topology template.

[0136] In step 600, the source country argumentation structure diagram generated in step 500 is obtained. The argument nodes and evidence nodes in the graph are abstracted using a topological abstraction algorithm.

[0137] The aforementioned topology abstraction algorithm includes the following steps:

[0138] The first step is to traverse the argument and evidence nodes in the argument structure diagram and extract the text content of each node.

[0139] The second step is to use named entity recognition technology to process the text of each node and identify specific entities in the text, including country names, personal names, place names, organization names, numerical values, time, event names, etc.

[0140] The third step is to determine the entity type of each identified entity. For example, "Country A" is a country-type entity, "Soybean export volume" is an economic indicator-type entity, and "35 million tons" is a numerical entity.

[0141] The fourth step is to establish a mapping rule from type to placeholder based on the entity type. For example, the entity type of country is mapped to the placeholder "", the entity type of economic indicator is mapped to the placeholder "", and the entity type of numerical value is mapped to the placeholder "".

[0142] The fifth step is to replace the specific entities in the text of each node with placeholders of the corresponding type. For example, replace "Country A's soybean exports reached 35 million tons" with "The exports of [country name] reached [number]".

[0143] The sixth step is to process the reasoning steps in the reasoning chain, retaining the logical structure and relationship types of the reasoning, but replacing the parts involving specific entities with placeholders.

[0144] Step 7: Generate a reusable argument topology template The template retains the structure and logic of the argument but removes content specific to the source country.

[0145] The aforementioned named entity recognition technology is based on a conditional random field model or a bidirectional LSTM-CRF model. The input layer of the named entity recognition model receives word vector representations of the text sequence, and the output layer outputs a sequence of entity type labels corresponding to each word. The named entity recognition model is trained on annotated corpora using supervised learning, with the annotated text sequence and its corresponding entity type label sequence used as training samples. The loss function is negative log-likelihood loss, and the optimization algorithm is the Adam optimization algorithm. When applied to this invention, the text sequence of each node in the argument structure graph is input into the named entity recognition model, which outputs entity type labels for each word, identifying entities such as country type, numerical type, time type, and economic indicator type.

[0146] Furthermore, the type constraints for each placeholder include not only data type information but also the threshold values ​​for the constraints. All constraints are recorded in a machine-resolvable format, such as numerical range constraints recorded as "min: 15, max: 100, unit: percentage", facilitating automated retrieval and verification during slot filling.

[0147] Step 700: Based on the dimension sharing degree matrix and the basic features of the target country, use the country similarity calculation algorithm to evaluate the similarity between the target country and the source country in each dimension, filter the dimensions with similarity exceeding the threshold, and generate a list of applicable dimensions for the template.

[0148] In step 700, the target country is the country from which the knowledge base is to be built. The sharing degree matrix generated in step 400 is obtained. This includes basic characteristic data of the target country, such as its political system type, economic development level, major industrial structure, and diplomatic relations. The analysis focuses on the 16 analytical dimensions... The target country is evaluated using a country similarity calculation algorithm across several dimensions. With source country The degree of similarity in this dimension. The country similarity calculation algorithm first extracts the similarity between the target country and the source country in the second dimension. Feature vectors of each dimension-related feature and Calculate the similarity between two feature vectors. :

[0149]

[0150] For each dimension If similarity Exceeding the preset threshold Then it is considered that the topological template of the argument is... If this dimension applies to the target country, add this dimension to the template's applicable dimensions list. The template is applicable to a list of dimensions, which includes the identifier and similarity score for each applicable dimension.

[0151] Furthermore, similarity The similarity is calculated using cosine similarity, with a value range of [-1, 1]. To facilitate comprehensive calculation with other indicators, in practical applications, the similarity value is mapped to the interval [0, 1], with a preset threshold. eigenvectors and The specific acquisition method is as follows: For each dimension Relevant features are extracted from the national basic feature database. For example, features such as political system type, democracy index, and regime stability are extracted from the political system dimension, while features such as GDP composition, industrial structure index, and development level are extracted from the economic structure dimension. The values ​​of each feature are normalized, scaling feature values ​​of different dimensions to the [0, 1] interval. Then, the normalized feature values ​​are combined into a feature vector. The vector dimension is equal to the number of features contained in that dimension (usually 3-10 features).

[0152] It should be noted that when generating the list of applicable dimensions for the template, the sharing degree of that dimension in the target country's group in the sharing degree matrix was also taken into account. Sharedness The similarity of this feature across all countries within the group is calculated, with a value ranging from [-1, 1], and then mapped to the interval [0, 1] before use. If the sharing degree... Exceeding the preset sharing threshold And similarity Also exceeded If a dimension is found to be highly shared, it will be prioritized for addition to the template's applicable dimension list because high shareability indicates that the analysis and argumentation for that dimension has strong general applicability within the group. Both conditions must be met simultaneously, and dimensions that meet the shareability condition have higher priority.

[0153] Furthermore, similarity threshold Determined based on the precision requirements for reusing argumentation templates, a typical value is [value to be filled in]. This indicates that the argument template can be reused when the similarity between the target country and the source country in this dimension reaches 60% or more; if higher reuse accuracy is required, the threshold can be increased to [value missing]. If template migration with lower similarity is allowed to expand the scope of reuse, the threshold can be lowered to... Shared degree threshold Determined based on the requirement of generality of argumentation within the group, the typical value is... This indicates that this dimension has a high degree of feature similarity among more than 70% of the countries within the group, and the argument structure has strong universality; when This indicates that the dimension is highly consistent within the group and is the optimal dimension for reuse.

[0154] Step 800: For each dimension in the template applicable dimension list, use the slot filling algorithm to replace the placeholders in the argument topology template with the specific data and entities of the target country to generate the instantiated argument structure diagram.

[0155] In step 800, the argument topology template generated in step 600 is obtained. And the template applicable dimension list generated in step 700 For each dimension in the list The slot filling algorithm is applied to instantiate the argument topology template part corresponding to this dimension.

[0156] The aforementioned slot filling algorithm includes the following steps:

[0157] The first step is to obtain information about the target country. Data related to this dimension, including statistical data, historical events, and policy texts, will be used to construct a database for the target country.

[0158] Furthermore, the construction of the target country database is achieved through multi-source data fusion. Data sources include: statistical data from international organization databases (World Bank, IMF, UN, etc.), with structured queries used to obtain values ​​for specific countries and indicators; news databases, where relevant events and policy information are retrieved through keyword searches using the format "country name + dimension keywords"; academic literature databases, where analytical documents are obtained through DOIs or academic search engines; and government websites, where officially released policies and statistical data are regularly collected via web crawlers. Metadata is used to annotate the data from each data source, recording information such as data source, publication time, and data type, facilitating subsequent verification and traceability. The data from each source is categorized and stored according to data type (numerical, text, time series) to establish a structured database for the target countries.

[0159] The second step is to traverse the placeholders in the argument topology template and obtain the type constraints of each placeholder.

[0160] The third step is to retrieve the matching specific entity or value from the target country database for each placeholder based on its type constraint. For example, for the "" placeholder, retrieve the commodity category whose total export value in the target country exceeds 15%, and for the "" placeholder, retrieve the actual export volume value of the commodity.

[0161] Furthermore, data retrieval employs a multi-stage filtering and ranking method. First, search conditions are generated based on the type constraints of placeholders. For example, numerical range constraints "min: 15, max: 100, unit: percentage" are converted into the SQL query "SELECT * FROM table WHERE value BETWEEN 15 AND 100". For text-type placeholders, keyword search conditions are generated. Second, database queries are executed, and data is ranked according to timeliness and source authority, prioritizing the return of the latest data from authoritative institutions. Finally, the search results are scored for relevance; the higher the degree of match between the search results and the constraints, the higher the score. The score calculation considers factors such as character matching degree and semantic similarity (calculated using pre-trained word vectors).

[0162] The fourth step is to filter the retrieved candidate data and select the specific data that best matches the placeholder type constraint.

[0163] Furthermore, the degree of matching is assessed through a matching score. Quantization, with a value range of [0, 1].

[0164] The matching score comprehensively considers the data type matching and the degree of constraint satisfaction. The calculation formula is as follows: ,in For type matching degree, To constrain satisfaction, weights satisfy (Typically, values ​​of 0.4 and 0.6 are used). Select the matching score. The candidate data with the highest score is used as the final filling data. If there is no candidate data that meets the conditions, the placeholder is marked as to be filled.

[0165] Furthermore, weighting coefficients and The determination is based on the characteristics of different placeholder types. For numeric placeholders, constraint satisfaction is more important, and they are typically set to... For entity placeholders, type matching is more important and can be adjusted to... Type matching degree Constraint satisfaction is calculated by checking the consistency between the data type labels and placeholder type labels of the candidate data: 1 for complete consistency, 0.5 for partial consistency, and 0 for inconsistency. The proportion of candidate data that satisfy the constraints is obtained by calculating the proportion of such constraints. The calculation formula is as follows: ,in This represents the number of constraints that the candidate data must satisfy. The total number of constraints defined for the placeholder.

[0166] The fifth step is to verify the semantic association constraints between placeholders. For example, verify whether the value of the "" placeholder is consistent with the actual export data of the product corresponding to the "" placeholder. If they are inconsistent, re-retrieve or mark it as pending supplementation.

[0167] Step 6: Fill the corresponding placeholder positions with the retrieved data and replace the placeholders in the template.

[0168] Step 7: Generate the instantiated argument structure diagram The instantiated argument structure diagram contains specific arguments, evidence, and reasoning chains for the target country.

[0169] Step 900: Based on the instantiated argumentation structure diagrams of each dimension, each agent collaboratively generates the target country's knowledge base content and analysis report, establishes version associations for cross-border argumentation templates, and triggers synchronous update checks of associated countries using change propagation algorithms when the source country's argumentation structure diagram is updated, outputting the country's knowledge base and version synchronization records.

[0170] In step 900, the instantiated argument structure diagrams of each dimension generated in step 800 are obtained. According to the agent task execution schedule plan generated in step 300, each agent sequentially executes the content generation task for its assigned dimension. Each agent generates the analytical text content for that dimension based on the arguments, evidence, and reasoning chains in the instantiated argumentation structure diagram. The content generation process utilizes natural language generation technology to convert structured argumentation elements into coherent text paragraphs, ensuring that the generated text is logically rigorous and fluently expressed.

[0171] The aforementioned natural language generation technology is based on a sequence-to-sequence generation model, which employs an encoder-decoder architecture. The encoder receives the sequence representations of structured argument elements as input, and the decoder outputs a sequence of natural language text. The sequence-to-sequence generation model is trained on paired corpora containing structured data and corresponding text descriptions, using a teacher-forced strategy during training. In the decoding stage, real text sequences are used as decoder input, the loss function is cross-entropy loss, and the optimization algorithm is the Adam optimization algorithm. When applied to this invention, the argument nodes, evidence nodes, and their relationship representations in the instantiated argument structure graph are converted into sequence form and input into the encoder. The decoder generates analysis text through autoregression, and the generation process uses a bundle search strategy to select candidate text sequences.

[0172] The content generated from each dimension is integrated according to the knowledge base's organizational structure to form the target country's knowledge base content. Based on this, a comprehensive analysis report of the target country is generated by summarizing the content from each dimension. The comprehensive analysis report includes the analytical conclusions, supporting evidence, and data source citations for each analytical dimension.

[0173] While generating the target country's knowledge base, a version association record for cross-border argumentation templates is established. This version association record is linked to the source country. Argumentation structure diagram Argumentation Topology Template Target countries And the identifiers for each applicable dimension. When the argument structure diagram of the source country is updated, such as by adding arguments or modifying the reasoning chain, the change propagation algorithm is applied to detect the impact of the update on related countries.

[0174] The aforementioned change propagation algorithm includes the following steps:

[0175] The first step is to obtain the argumentation structure diagram of the source country before the update. and the updated argument structure diagram .

[0176] The second step is to compare the argument structure diagram before and after the update, identify the nodes and edges that have changed, and the types of changes include adding nodes, deleting nodes, modifying node content, adding edges, and deleting edges.

[0177] The third step is to determine whether each change affects the argument topology template. If the change affects the template structure, and the placeholders or relationships corresponding to the nodes or edges involved in the change change, then the change is considered to affect the template structure.

[0178] Fourth, if there are changes affecting the template structure, the argument topology template will be regenerated based on the updated argument structure diagram. And increment the template version number.

[0179] Fifth, query the list of associated countries that use this template from the version association records.

[0180] Step 6: Send update notifications to all relevant countries. The update notifications include the template version number, a summary of the changes, and the processing steps that need to be re-executed.

[0181] Step 7 triggers a synchronization update check for related countries, prompting that the slot filling and instantiation process needs to be re-executed to ensure that the argumentation structure diagram of related countries is consistent with that of the source country.

[0182] Output a country-specific knowledge base, which contains structured knowledge entries and a comprehensive analysis report for the target country across 16 analytical dimensions. Output a version synchronization record, which includes the version number of each cross-border argumentation template, a list of associated countries, an update timestamp, and the synchronization status.

[0183] It should be noted that when establishing version associations, each argument topology template... Each template is assigned a unique template identifier and version number. When the template is updated, the version number is incremented, and a summary of the update is recorded in the version synchronization log. Each target country's knowledge base records its used template identifier and version number, allowing for tracing back to the list of countries requiring synchronization when the source country updates. The change propagation algorithm also assesses the scope of the update's impact.

[0184] Furthermore, the scope of the change's impact is quantified by calculating the ratio of the number of changed nodes / edges to the total number of nodes / edges in the template. Let the number of changed nodes be... The total number of nodes in the template is The formula for calculating the change ratio is: The value range is [0, 1]. The determination rule is as follows: if If it is, it is judged as a local optimization, meaning that the update only involves some minor adjustments and has little impact on the core logic of the argument; therefore, the priority of synchronous update checks is low. If so, it is determined to be a medium-level correction, involving changes to some of the argument structure, triggering a medium-priority synchronous update check; if If the change is deemed a major revision, involving a substantial change in the argument structure, it will immediately trigger a high-priority synchronous update, requiring relevant countries to re-verify and synchronize the content of that dimension. The calculation method is as follows: Count the total number of newly added, deleted, and modified nodes in the difference between the argument structure graph before and after the update. If the change of an edge does not lead to a change in a node, the edge weight is 0.5. The calculation formula is... .

[0185] Figure 7 This demonstrates the effectiveness of the version association and synchronization update mechanism for the cross-border argumentation template in step 900.

[0186] In this embodiment of the application, in order to improve the quality of the instantiated argument structure diagram, the following steps are included after step 800:

[0187] Step 810: Perform multi-dimensional defect detection on the instantiated argument structure diagram, and generate an argument defect diagnosis report using algorithms for evidence sufficiency assessment, logical validity test, over-inference identification, and evidence bias detection.

[0188] In step 810, the instantiated argument structure diagram generated in step 800 is obtained. Perform multi-dimensional defect detection on the instantiated argument structure diagram, including the following four detection dimensions:

[0189] First, assess the sufficiency of evidence. For each argument node in the argument structure diagram, examine the quantity and quality of its supporting arguments. If the number of supporting arguments for an argument is less than a preset minimum threshold... Or the reliability score of the argument is below the threshold. If so, then the argument is deemed to have insufficient evidence.

[0190] Furthermore, minimum threshold Typically, it is set with 2 to 3 arguments, and a reliability score. Quantification is performed based on the authority of the source of the argument, with a value range of [0, 1], and a threshold value. The value is usually set to 0.6, which means that the source of the argument must have a certain degree of credibility.

[0191] Second, logical validity testing. For each inference relation edge in the argument structure diagram, a logical validity testing algorithm is used to analyze the logical rationality of the reasoning chain.

[0192] The aforementioned logical validity test algorithm includes the following steps:

[0193] The first step is to obtain the argument nodes, intermediate reasoning nodes, and thesis nodes in the reasoning chain, and extract the text content of each node.

[0194] The second step is to identify the premises and conclusions for each reasoning step in the reasoning chain. The premises are the input node content of the reasoning step, and the conclusions are the output node content of the reasoning step.

[0195] The third step is to detect whether there are logical jumps in the reasoning process. This is done by judging whether the premises can directly lead to the conclusion. If there is a lack of necessary intermediate reasoning steps between the premises and the conclusion, then a logical jump is determined to exist.

[0196] Furthermore, the detection of logical jumps employs a combination of semantic similarity analysis and reasoning chain integrity testing. The semantic similarity between the premise and the conclusion is calculated, and a pre-trained language model is used to obtain the vector representation of the premise text. Vector representation of the conclusion text Calculate the cosine similarity between the two. If the similarity is below the threshold of 0.3 (indicating a significant semantic jump between the premise and the conclusion), the system also checks whether there are common keywords or inference connectors (such as "therefore" or "lead to") in the two texts. If the number of common words is less than a preset number (such as 1), then a logical jump is determined to exist.

[0197] The fourth step is to check for premise contradictions in the reasoning chain. This is done by comparing the premises of different reasoning steps to see if there are any conflicting statements. If a contradiction exists, then a premise contradiction is determined to exist.

[0198] The fifth step is to check whether the conclusion is supported by the premises. By analyzing the logical relationship between the semantic content of the premises and the semantic content of the conclusion, if the premises cannot support the conclusion, it is determined that there is an error that the conclusion is not supported.

[0199] The sixth step is to determine if logical errors such as logical jumps, contradictory premises, or unsupported conclusions are detected, and to record the type and location of the error.

[0200] Third, over-inference identification. For arguments in the argument structure diagram, analyze whether the strength of their conclusion matches the strength of the supporting evidence. If the conclusion of an argument is too absolute, but the evidence only provides partial support or there are counterexamples, then the argument is deemed to have the flaw of over-inference. Over-inference identification is achieved by calculating the difference between the support of the evidence and the certainty of the conclusion; when the difference exceeds a threshold, an over-inference alarm is triggered.

[0201] Furthermore, the support score is the average of the reliability scores of all supporting arguments, ranging from [0, 1], and is calculated using the following formula: ,in To support the number of arguments, For the index of arguments, For the first The reliability score of each argument; the certainty of the conclusion is quantified based on the occurrence of absolute terms (such as necessary, complete) in the text of the argument, with a value range of [0, 1], and the calculation formula is as follows. ,in The number of times a qualifier (such as possible, perhaps) appears. The total number of words in the argument. The degree of over-inference is defined as... ,in To support the argument, To ensure the certainty of the conclusion. When the inference is excessive. Exceeding the preset threshold When (usually taken as 0.3), overinference is considered to exist; if This would be considered a serious over-inference.

[0202] Fourth, evidence bias detection. For each argument node in the argument structure diagram, the diversity and representativeness of the sources of evidence are examined. If the evidence is excessively concentrated in a specific source or within a specific time period, evidence selection bias may exist. Evidence bias detection is achieved by calculating the entropy value of the sources of evidence and the uniformity of their temporal distribution. When the entropy value is below a threshold or the distribution is uneven, evidence bias is determined to exist.

[0203] Furthermore, the entropy of the sources of the argument is calculated by determining the information entropy of the distribution of the argument across each source, using the following formula: ,in The total number of sources of evidence. For source index, For from the The proportion of arguments from each source to the total arguments. The original entropy value ranges from [0, 1]. The normalized value range is [0, 1], and the normalization formula is: When the normalized entropy value Below the preset threshold A value of 0.5 typically indicates that the concentration of evidence sources is too high, suggesting bias. The uniformity of the time distribution is achieved by calculating the variance of the evidence distribution across different time periods, using the following formula: ,in The number of time periods to be divided (usually 5 time periods). Indexed by time period For the first The number of arguments for each time period This represents the average number of arguments across different time periods. When the variance... Exceeding the preset threshold (usually taken) The condition ) indicates uneven time distribution. If either condition is met, evidence bias is considered to exist.

[0204] The various defects detected in the above four dimensions are summarized to generate an argument defect diagnosis report. The argument defect diagnosis report includes the type of each defect, its location in the argument, its severity score, and a suggested remediation plan.

[0205] Furthermore, the severity score of the defect. The value range is [0, 1], calculated comprehensively based on the type of defect and its scope of impact. The severity of the insufficient evidence defect is directly proportional to the number of missing arguments; the severity of the logical invalidity defect is related to the type of logical error, with the severity of premise contradiction (0.8-1.0) being higher than that of logical jump (0.5-0.7); the severity of the over-inference defect is equal to the difference between the support of the arguments and the certainty of the conclusion; the severity of the evidence bias defect is set at 0.3-0.6 based on the degree of bias. Severity Score The defects are marked as high priority and need to be fixed first.

[0206] Step 820: Based on the defects in the argument defect diagnosis report, use the argument repair algorithm to supplement the missing reasoning links, replace inapplicable arguments, add necessary qualifiers, and generate an enhanced argument structure diagram.

[0207] In step 820, the argument defect diagnosis report generated in step 810 is obtained. For each defect item in the argument defect diagnosis report, the corresponding argument repair algorithm is applied to repair it according to its type.

[0208] The aforementioned proof-fixing algorithm includes the following steps:

[0209] The first step is to go through each defect item in the defect diagnosis report and identify the defect type and location.

[0210] The second step is to address the deficiency of insufficient evidence by retrieving additional evidence related to the argument from data sources in the target country, selecting evidence with high reliability scores, adding it as a supplementary argument node for the argument, and establishing new supporting relationship edges in the argument structure diagram.

[0211] Furthermore, evidence retrieval employs a combination of keyword matching and semantic retrieval. First, core keywords are extracted from the argument text (using the TF-IDF algorithm), for example, for the argument "Country X's economy is highly dependent on exports," keywords such as "economy," "dependence," and "exports" are extracted. Second, Boolean searches are performed on the data source using these keywords (using AND / OR operators), for example, searching for "(economy OR GDP) AND (exports OR trade) AND Country X." Third, the retrieved results are semantically relevance ranked, and the similarity between the argument and candidate evidence is calculated using pre-trained sentence vector representations. The top N similarity results (N is typically 3-5) are selected as supplementary evidence. Finally, the selected evidence is scored for reliability based on factors such as source authority (international organizations score higher than news media, official data higher than academic literature) and timeliness (newer data scores higher), selecting evidence with a score exceeding a threshold of 0.6 as final supplementary evidence.

[0212] The third step is to analyze the logical breaks in the reasoning chain for logically invalid defects, identify the missing reasoning links, retrieve relevant background knowledge or theories, and insert them as intermediate reasoning nodes into the reasoning chain to complete the reasoning process.

[0213] Furthermore, the identification and supplementation of missing reasoning links employs a method based on knowledge graphs and reasoning rules. First, a domain knowledge graph is constructed, containing common concepts, relationships, and reasoning rules within the country analysis domain. Second, the semantic relationships between premises and conclusions in the reasoning chain are analyzed by searching the knowledge graph for paths connecting premise concepts and conclusion concepts; the intermediate nodes and relationships in these paths represent the missing reasoning links. Third, the textual descriptions of these intermediate reasoning links are extracted from the knowledge graph. Finally, the supplemented reasoning links are inserted as intermediate nodes into the reasoning chain, establishing a complete reasoning path of premise → intermediate link → conclusion, and the reasoning type (e.g., causal reasoning, deductive reasoning) is recorded.

[0214] The fourth step is to adjust the argument text to address the flaws of over-inference, adding necessary qualifiers or conditional constraints to reduce the absoluteness of the conclusion, and recording the limiting conditions in the argument nodes.

[0215] The fifth step is to address the shortcomings of the evidence by searching for supplementary evidence from different sources and time periods to increase the diversity of the arguments, or to add explanations of the limitations of the evidence to the arguments to draw the reader's attention to possible biases.

[0216] Step 6: Generate the enhanced argument structure diagram. The enhanced argument structure diagram improves upon the sufficiency of evidence, logical rigor, reasonableness of inference, and representativeness of evidence. The enhanced argument structure diagram is used in step 900. Replacement instantiation argument structure diagram Generate knowledge base content.

[0217] Figure 6 The quality improvement effects of defect detection and repair demonstrated in steps 810-820 are shown.

[0218] An international relations research institution needed to build a country knowledge base for four countries in region D (countries A, B, C, and D), covering 16 analytical dimensions including political system, economic structure, class power, and foreign trade dependence. Each country was to generate a 45-60 page in-depth analysis report. Policy analysts in the research team defined the analytical requirements and agent behavior constraints for each dimension by filling in a declarative configuration template. The system automatically identified country A as the source country for initial analysis and then reused the argumentation structure of country A for target countries B, C, and D, significantly reducing the workload of redundant analysis and ensuring consistency in analytical quality.

[0219] In step 100, policy analysts fill out a declarative configuration template, defining 16 analytical dimensions and their descriptions in the analytical dimension slot, specifying access to international organization databases such as the World Bank, IMF, and UN in the data source scope slot, and setting "min_sources: 5, reasoning_depth: 3, data_timeliness: 5" in the agent behavior constraint slot. After parsing the template, the system generates a structured task requirement description, including detailed definitions of the 16 dimensions, data source access configurations, and constraint rules such as each argument requiring at least 5 supporting sources, a reasoning depth of at least 3 steps, and data timeliness limited to the last 5 years.

[0220] Table 1. Example of slot content for declarative configuration template:

[0221]

[0222] Table 2: Structured task requirements description after parsing:

[0223]

[0224] In step 200, the system uses a pre-trained language encoder to vectorize the task descriptions for the 16 analytical dimensions. Taking the economic structure dimension and the foreign trade dependence dimension as examples, the task description for the economic structure dimension is to analyze GDP composition, industrial structure, and economic development level, while the task description for the foreign trade dependence dimension is to analyze foreign trade dependence, major trading partners, and export commodity structure, which needs to be based on the results of the economic structure analysis. The pre-trained language encoder maps the two descriptions into 768-dimensional vectors. and Calculate cosine similarity .

[0225] Meanwhile, the causal association identification algorithm detected that the description of foreign trade dependence contained references that needed to be based on the economic structure analysis results. Extracting the keyword co-occurrence matrix revealed that the co-occurrence frequency of industrial structure and export commodity structure was 3. The normalized reference strength was then calculated. According to the formula for the strength of causal association Exceeding the preset threshold Therefore, directed edges from economic structure to foreign trade dependence are established in the dimensional semantic dependency graph.

[0226] Table 3 shows the semantic similarity and causal association strength between some dimensions:

[0227]

[0228] In step 300, the system performs a topological sorting algorithm on the dimensional semantic dependency graph. Initially, the in-degree of each dimension node is calculated. The in-degree of the political system and historical evolution dimensions is 0, thus it is classified as the first level. After removing these two nodes, the in-degree of the class power and international relations dimensions becomes 0, thus it is classified as the second level. This process is repeated for all 16 dimensions, resulting in a total of 5 levels. A task execution scheduling plan for the agent is generated: the first level contains 3 tasks executed in parallel, the second level contains 4 tasks executed in parallel, the third level contains 5 tasks executed in parallel, the fourth level contains 3 tasks executed in parallel, and the fifth level contains 1 task.

[0229] In step 400, the system acquires geographical, political, and economic relationship data for four countries: A, B, C, and D. Regarding geographical relationships, the geographical distance between A and B is 2800 kilometers, after normalization... In terms of political ties, both countries are presidential democracies, and their political systems are similar after normalization. In terms of economic ties, bilateral trade reached US$18 billion in 2022, after normalization. .

[0230] According to the weighting coefficients Calculate the correlation strength Exceeding the preset threshold Edges are established in the country association graph. Louvain's algorithm is applied to all four countries for graph clustering. Countries A, B, C, and D are grouped into the same country group, region D. Modularity... .

[0231] Table 4. Inter-country correlation strength matrix:

[0232]

[0233] Calculate the intra-group sharing degree for each analysis dimension in the D region group. Taking the foreign trade dependence dimension as an example, extract the feature vector of country A. (Representing foreign trade dependence, concentration of major trading partners, and concentration of export commodities, respectively), the feature vector of country B. eigenvectors of country C eigenvectors of country D Calculate the mean cosine similarity between country A and the other three countries. Similarly, the feature similarity scores of other countries are calculated to obtain the degree of sharing. This indicates that the foreign trade dependence dimension has a high degree of sharing within the D region group, and the argument structure is highly reusable.

[0234] Table 5. Share of each analysis dimension in the D region group:

[0235]

[0236] In step 500, the system selects country A as the source country because country A's analysis report has a comprehensive score. It exceeds the threshold of 0.7.

[0237] An argument structure parsing algorithm was applied to the analysis report text on the foreign trade dependence dimension of country A. The report text contains 125 sentences. After sentence segmentation and part-of-speech tagging, 23 candidate sentences for arguments and 47 candidate sentences for evidence were identified. Among them, the probability score of the argument that country A's economy is highly dependent on agricultural exports was calculated. This is marked as argument node P1. The arguments supporting this argument include: In 2022, Country A's soybean exports reached 82 million tons, accounting for 21.3% of total exports (Source: Country A's Statistics Bureau, reliability score 0.92); iron ore exports reached 350 million tons, accounting for 18.7% of total exports (Source: World Bank, reliability score 0.88); and agricultural and mineral products accounted for 63.5% of total exports (Source: IMF, reliability score 0.90). The reasoning chain is constructed as follows: From the three arguments, the intermediate reasoning node "Country A's export structure is highly concentrated in agricultural and mineral products" is derived, which then leads to argument P1. The reasoning type is inductive reasoning. The generated source country argument structure diagram contains 23 argument nodes, 47 argument nodes, and 31 reasoning relation edges.

[0238] In step 600, the system applies a topological abstraction algorithm to the argument structure diagram of Country A. For the argument node "Country A's soybean exports reached 82 million tons in 2022, accounting for 21.3% of total exports," named entity recognition is performed, identifying Country A as a country type entity, soybeans as a commodity type entity, 82 million tons as a numerical type entity, and 21.3% as a percentage type entity. According to the mapping rules, this is replaced with "the export volume reached, accounting for % of total exports." After abstracting all argument and thesis nodes, an argument topology template is generated. The templates include 23 placeholder argument templates, 47 placeholder evidence templates, and 31 inference relation templates.

[0239] In step 700, the system uses country B as the target country and evaluates its similarity to the source country A across various dimensions. In the foreign trade dependence dimension, it extracts the feature vector of country B. and the eigenvectors of country A .

[0240] Calculate the inner product of two vectors Vector norm of country B Vector norm of country A Calculate similarity Exceeding the preset threshold Meanwhile, the degree of sharing of this dimension in the D region group. Exceeding the preset sharing threshold Therefore, the foreign trade dependence dimension was added to the template's applicable dimension list. After evaluating all 16 dimensions, a template applicable dimension list was generated containing 12 dimensions, with similarity scores ranging from 0.63 to 0.94.

[0241] Table 6. Similarity between Country B and Country A across various dimensions and template applicability:

[0242]

[0243] In step 800, the system applies a slot-filling algorithm to country B. For the argument topology template of the foreign trade dependence dimension, it iterates through each placeholder and retrieves matching data. For the placeholder "", whose type constraint is a commodity category whose total export value accounts for more than 15%, candidate data is retrieved from the country B database: soybeans (23.5%), corn (9.2%), wheat (7.8%), and beef (6.5%). The matching score is calculated, and the matching score for soybeans is... (Complete match for product type) (Meets the constraint of a percentage > 15%) The placeholder was selected as the data to fill in. For the placeholder "", data on soybean exports from country B was retrieved, yielding "48.5 million tons in 2022" from the World Bank database (timeliness: 2022, authority: 0.90). The argument topology template "The export volume reached, accounting for a certain percentage of the total export volume" is instantiated as "In 2022, Country B's soybean exports reached 48.5 million tons, accounting for 23.5% of the total export volume". After filling the slots for all 47 argument templates and 23 thesis templates, an instantiated argument structure diagram of Country B is generated. .

[0244] In step 810, the system instantiates the argumentation structure diagram of country B. Perform multi-dimensional defect detection. First, assess the sufficiency of evidence. Examine the argument node P8 "Country B's economy is susceptible to fluctuations in international agricultural product prices" in the argument structure diagram, and find that it has only 2 supporting arguments, which does not meet the minimum threshold. Furthermore, one of the arguments received a reliability score of 0.55, which is below the threshold. First, the argument is deemed to have insufficient evidence. Second, logical validity is tested. The reasoning chain of argument node P12 is examined. The premise is "Country B's soybean exports account for 23.5%", and the conclusion is "Country B's economy is completely dependent on agricultural product exports". The semantic similarity between the premise and the conclusion is calculated to be 0.28, which is below the threshold of 0.3. The lack of an intermediate reasoning link, "total proportion of agricultural products", is detected, indicating a logical jump. Third, over-inference is identified.

[0245] For argument node P12, calculate the support of the arguments. The conclusion text contains the absolute word "complete," so the certainty of the conclusion is calculated. Inferring the degree of excess Although it did not exceed the threshold of 0.3, considering the absoluteness of the word "complete," it was still determined that there was a slight over-inference. Fourth, evidence bias detection. Examining the five arguments at argument node P15, four were found to be from the World Bank database and one from academic literature. The source entropy value was calculated. After normalization Below the threshold The study determined that there was a bias in the concentration of evidence sources. An argumentation defect diagnosis report was generated, identifying a total of 8 defects: 3 defects of insufficient evidence (severity score 0.6-0.75), 2 defects of logical jump (severity score 0.55-0.65), 2 defects of over-inference (severity score 0.25-0.40), and 1 defect of biased evidence (severity score 0.48).

[0246] In step 820, the system repairs the instantiated argument structure diagram of Country B based on the argument defect diagnosis report. For the argument node P8 with insufficient evidence, supplementary evidence related to "the impact of international agricultural product price fluctuations" is retrieved from the Country B database. The search is conducted using the keywords "(Economy OR GDP) AND (Price Fluctuation OR Market Risk) AND Country B". Candidate evidence "The decline in soybean prices from 2019 to 2021 caused Country B's GDP growth rate to drop from 2.9% to -2.1%" (Source: IMF, reliability score 0.88) and "35% of Country B's fiscal revenue depends on agricultural product export taxes" (Source: Ministry of Economy of Country B, reliability score 0.85) are added as supplementary arguments for argument P8, bringing the total number of supporting arguments to 4 and increasing the average reliability score to 0.80. For argument node P12, which involves a logical jump, the path connecting the premise "soybean exports account for 23.5%" and the conclusion "the economy depends on agricultural exports" is searched from the domain knowledge graph. The intermediate link "agricultural products account for 57.8% of total exports" is identified and inserted into the reasoning chain as an intermediate reasoning node, establishing a complete reasoning path: "soybean exports account for 23.5% → agricultural products account for 57.8% → the economy is highly dependent on agricultural exports." The reasoning type is marked as inductive reasoning. For argument node P12, which involves over-inference, the argument text is modified from "Country B's economy is completely dependent on agricultural exports" to "Country B's economy is largely dependent on agricultural exports." The qualifier "to a large extent" is added to replace the absolute word "completely," reducing the absoluteness of the conclusion. For argument node P15, which involves evidence bias, supplementary evidence is retrieved from academic literature databases and news databases. Two arguments from different sources are added, increasing the source entropy value to 0.72 after normalization, exceeding the threshold of 0.5. After the repair is completed, an enhanced argument structure diagram is generated. The diagram has been improved in terms of sufficiency of evidence, logical rigor, reasonableness of inference, and representativeness of evidence. All eight deficiencies have been corrected, and the quality score of the argument structure diagram has increased from 0.68 to 0.86.

[0247] In step 900, the system executes the task scheduling plan according to the agents, and each agent sequentially generates analytical text content for country B in 12 applicable dimensions. Taking the foreign trade dependence dimension as an example, the agent Agent_Foreign Trade generates the analytical text content for country B in 12 applicable dimensions based on the enhanced argumentation structure diagram. The system converts 23 argument nodes, 47 evidence nodes, and 31 reasoning relationship edges into sequence form, inputting the sequence into a sequence-to-sequence generation model. The decoder uses autoregression to generate a text paragraph: Country B's economy is highly dependent on agricultural exports. In 2022, Country B exported 48.5 million tons of soybeans, accounting for 23.5% of total exports; and 32 million tons of corn, accounting for 9.2% of total exports. Agricultural products accounted for 57.8% of total exports, indicating a highly concentrated export structure in the agricultural sector. This export structure makes Country B's economy vulnerable to fluctuations in international agricultural product prices… (generating 1850 words). After integrating the content generated from the 12 dimensions, a Country B knowledge base is formed, containing 328 structured knowledge entries, a 48-page comprehensive analysis report, and 3620 cited sources. A version association record is also established, recording Country A's argument topology template ID as "BRA_Foreign Trade_v1.0", with associated target countries including Country B, Country C, and Country D, and applicable dimensions of 12.

[0248] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for constructing a country-specific knowledge base based on a declarative multi-AI agent architecture, characterized in that, Includes the following steps: Obtain declarative configuration template data, parse the analysis dimensions, data source scope, and agent behavior constraints in the template, and generate a structured task requirement description; The task descriptions of each analysis dimension are vectorized, the semantic similarity and causal relationship strength between dimensions are calculated, and a dimensional semantic dependency graph is generated. Based on the dimensional semantic dependency graph, the tasks of multiple AI agents are hierarchically divided using a topological sorting algorithm, and an agent task execution scheduling plan is generated. Obtain a list of countries to be analyzed and their geographical, political and economic data; use graph clustering algorithms to identify country groups; and generate a country association map and a sharing degree matrix of each analysis dimension within the group. Obtain analysis reports of source countries in the country group, extract the arguments, evidence, and reasoning chains in the reports, and generate a source country argumentation structure diagram; The source country argumentation structure diagram is abstracted, and specific entities are replaced with type placeholders to generate an argumentation topology template; Based on the sharing degree matrix and the basic characteristics of the target country, the similarity between the target country and the source country in each dimension is evaluated, the dimensions with similarity exceeding the threshold are filtered, and a list of applicable dimensions for the template is generated. For each dimension in the template applicable dimension list, replace the placeholders in the argument topology template with the specific data and entities of the target country to generate the instantiated argument structure; Based on the instantiated argumentation structure, each intelligent agent collaboratively generates the target country's knowledge base content and analysis report. The process of generating a dimensional semantic dependency graph includes: inputting the task description text of each dimension into a pre-trained language encoder to obtain vector representations of each dimension; calculating the ratio of the inner product of any two dimensional vectors to their norm product to obtain semantic similarity; analyzing keyword co-occurrence and semantic dependency patterns in the dimensional task descriptions, calculating citation strength, and weighted summing of semantic similarity and citation strength to obtain causal association strength; and establishing directed edges between dimensional pairs with causal association strength exceeding a preset threshold to generate a dimensional semantic dependency graph.

2. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 1, characterized in that, The generated country association map and the sharing matrix of each analysis dimension within the group include: By comprehensively calculating geographical, political, and economic data among countries, and then weighting and summing these data, the strength of inter-country connections is obtained. Establish edges between country pairs whose association strength exceeds a preset threshold to construct a country association graph; A graph clustering algorithm is applied to the country association graph to divide the nodes into multiple country groups based on the tightness of the connections between nodes, thereby generating a country association graph. For each country group, calculate the feature similarity score of each analysis dimension for each country within the group, average the feature similarity scores of each country within the group, obtain the sharing degree of each analysis dimension within the group, and generate a sharing degree matrix.

3. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 1, characterized in that, The source country argumentation structure diagram includes: Sentence segmentation and semantic analysis are performed on the source country's analysis report text to identify declarative sentences expressing arguments, factual sentences expressing evidence, conjunctions expressing inference relationships, and logical markers. For each argument, identify the set of evidence supporting the argument, as well as the reasoning chain from the evidence to the argument; organize the arguments, evidence, and reasoning chain into a directed graph structure, where the nodes include argument nodes and evidence nodes, and the directed edges represent supporting or reasoning relationships, generating the source country argument structure graph.

4. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 1, characterized in that, The generated argument topology template includes: Identify specific entities in the text of each node in the source country argumentation structure diagram; determine the type of each entity using named entity recognition technology; Each specific entity is replaced with a placeholder of the corresponding type. The logical structure of the reasoning steps in the reasoning chain is preserved, but the parts involving specific entities are replaced with placeholders to generate an argument topology template. To demonstrate the type constraints associated with each placeholder in the topology template, we define the type requirements for the data when filling the placeholder and record the semantic relationships between each placeholder.

5. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 1, characterized in that, The list of dimensions applicable to the generated template includes: For each analysis dimension, extract the feature vectors of the target country and the source country on the dimension-related features; The similarity between the target country and the source country in this dimension is obtained by calculating the ratio of the inner product of the feature vectors of the target country and the source country to their norm product. Filter dimensions with similarity exceeding a preset threshold and add those dimensions to the template's applicable dimension list; Based on the sharing degree of this dimension in the target country's group in the sharing degree matrix, when the sharing degree is high and the similarity is high, this dimension is preferentially added to the template applicable dimension list.

6. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 1, characterized in that, The generated instantiated argument structure includes: For each dimension in the template applicable dimension list, obtain the target country's dimension-related data; For each placeholder in the argument topology template, retrieve the matching specific entity or value from the target country data according to its type constraint; The retrieved data is filled into the corresponding placeholder positions, and the filled content is verified to meet the semantic association constraints between the placeholders, thus generating the instantiated argument structure.

7. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 1, characterized in that, After generating the instantiated argument structure, the following steps are also included: Perform multi-dimensional defect detection on the instantiated argument structure, including evidence sufficiency assessment, logical validity test, over-inference identification, and evidence bias detection, and generate an argument defect diagnosis report; Based on the deficiencies in the argument defect diagnosis report, the argument repair algorithm is used to supplement missing reasoning links, replace inapplicable arguments, and add necessary qualifiers to generate an enhanced argument structure. Use the enhanced argumentation structure to generate content for the target country's knowledge base.

8. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 7, characterized in that, Performing multi-dimensional defect detection includes: For each argument node in the argument structure, check the quantity and quality of its supporting arguments. When the number of supporting arguments for an argument is less than a preset minimum threshold, or the reliability score of the arguments is lower than the threshold, the argument is judged to have the defect of insufficient evidence. For each inference relation edge in the argument structure, analyze the logical rationality of the inference chain, detect whether there are logical jumps, premise contradictions or unsupported conclusions in the reasoning process, and determine that the inference chain has a logical invalidity defect when a logical error is detected. For arguments in the argument structure, analyze whether the strength of the conclusion matches the strength of the evidence. By calculating the difference between the evidence support and the certainty of the conclusion, when the difference exceeds a threshold, the argument is judged to have the defect of over-inference. For the argument nodes in the argument structure, the diversity and representativeness of the argument sources are detected. By calculating the entropy value and the uniformity of the time distribution of the argument sources, it is determined that there is a defect in evidence bias when the entropy value is lower than the threshold or the distribution is uneven.

9. The method for constructing a country knowledge base based on a declarative multi-AI agent architecture according to claim 1, characterized in that, In addition to generating the target country's knowledge base content and analysis report, it also includes: Establish version association records for cross-border argumentation templates, linking the argumentation structure diagram of the source country, argumentation topology template, target country, and identifiers for each applicable dimension; When the argument structure of the source country is updated, compare the argument structure diagrams before and after the update, identify the changed nodes and edges, and determine whether the changes affect the structure of the argument topology template. When the template structure changes, a synchronization update check is triggered for associated countries, an update notification is sent to each associated country, and the slot filling and instantiation process is re-executed.

10. A country knowledge base construction system based on a declarative multi-AI agent architecture, used to execute the country knowledge base construction method based on a declarative multi-AI agent architecture as described in any one of claims 1 to 9, characterized in that, include: The configuration parsing module is used to obtain declarative configuration template data, parse the analysis dimensions, data source scope and agent behavior constraints in the template, and generate a structured task requirement description. The dimension dependency analysis module is used to vectorize the task descriptions of each analysis dimension, calculate the semantic similarity and causal relationship strength between dimensions, and generate a dimension semantic dependency graph. The task scheduling module is used to hierarchically divide the tasks of multiple AI agents based on the dimensional semantic dependency graph and use the topological sorting algorithm to generate an agent task execution scheduling plan. The country association analysis module is used to obtain a list of countries to be analyzed and their geographical, political and economic association data, use graph clustering algorithms to identify country groups, and generate a country association map and a sharing degree matrix of each analysis dimension within the group. The argument structure extraction module is used to obtain the analysis reports of the source countries in the country group, extract the arguments, evidence and reasoning chains in the reports, and generate the argument structure diagram of the source countries. The argument template generation module is used to abstract the source country argument structure diagram, replace specific entities with type placeholders, and generate an argument topology template. The template applicability assessment module is used to evaluate the similarity between the target country and the source country in each dimension based on the sharing degree matrix and the basic characteristics of the target country, filter the dimensions with similarity exceeding the threshold, and generate a list of applicable dimensions for the template. The argument instantiation module is used to replace the placeholders in the argument topology template with the specific data and entities of the target country for each dimension in the template applicable dimension list, and generate the instantiated argument structure. The knowledge base generation module is used to collaboratively generate the target country's knowledge base content and analysis report based on the instantiated argument structure by various intelligent agents.

Images