Risk portrait generation method based on knowledge graph

Abstract

The invention relates to the technical field of risk portrait generation based on a knowledge graph, and discloses a risk portrait generation method based on the knowledge graph. According to the scheme, the document is segmented through the sentence end symbols, semantic coherence is ensured, and the accuracy and stability of candidate word extraction are improved. Continuous Chinese character sequence segmentation is adopted, dependence of a word segmentation tool is avoided, the occurrence frequency and position of candidate words are recorded, and a detailed statistical basis is provided for risk factor screening. A minimum occurrence number constant is set, noise interference and information coverage are balanced, the screening flexibility is improved, and candidate word screening is more representative. And counting the co-occurrence relationship of candidate words in the sentence, quantifying the co-occurrence frequency through an indicator function, enhancing word association expression, and providing key data for initial graph construction. Finally, the scheme constructs a knowledge graph based on the candidate words and the co-occurrence relation thereof, ensures clear expression of risk information, and improves the accuracy of the system in risk identification and portrait generation.

Description

Technical Field

[0001] The present invention relates to the technical field of risk profile generation based on knowledge graphs, and specifically to a risk profile generation method based on knowledge graphs. Background Art

[0002] In recent years, with the rapid development of big data and artificial intelligence technologies, risk management has become increasingly widely used in various fields such as finance, healthcare, and cybersecurity. Knowledge graphs, as a data representation method capable of expressing multivariate relationships and associative structures, have gradually become an important tool for constructing risk profiles and assisting risk decision-making. Existing risk profile generation methods based on knowledge graphs typically rely on pre-trained language models, deep learning algorithms, and complex semantic analysis frameworks, extracting, associating, and summarizing risk factors through large-scale data mining and expert rules. These methods can capture the implicit information in text data to a certain extent and generate risk profiles through steps such as entity extraction, relationship construction, and knowledge fusion for reference by decision-making systems.

[0003] However, existing technologies have many shortcomings. First, many methods rely on external pre-trained models and rules formulated by experts, which makes the system vulnerable to the limitations of pre-set models and dictionaries when integrating information across domains and contexts, making it difficult to flexibly respond to emerging risk factors in texts from different domains. Second, existing deep learning or natural language processing models mostly use fuzzy semantic embedding and probabilistic calculation methods. The model has many internal parameters and is difficult to interpret. It lacks clear mathematical descriptions and deterministic expressions, resulting in deficiencies in the transparency and traceability of the generated risk profiles. In addition, due to the significant differences in data volume, language style, and structural features of text data from different fields, existing methods often face the problem of insufficient data standardization and normalization when integrating cross-domain information, which affects the accuracy of the final map and the effectiveness of the risk profile. In addition, current risk profile generation technologies often use expert systems and rule-based reasoning methods. The system development cycle is long, and they rely heavily on the rule base, making it difficult to update in real time under rapidly changing risk environments. Some methods attempt to use statistical information and natural language processing techniques to construct an initial graph, but they still suffer from technical ambiguity and imprecise algorithmic expression in areas such as candidate word extraction, frequency statistics, and sentence position-based discreteness calculation. Existing methods often rely on manually set thresholds and rules when screening candidate words, lacking a unified and clear mathematical description. This makes the system susceptible to noise interference when processing large-scale data, affecting the accurate extraction of risk factors. Currently, to address the problem of cross-domain risk knowledge integration, some technologies have attempted to use multimodal data fusion and domain adaptation techniques to achieve risk factor alignment and graph integration. However, these methods generally require large amounts of pre-training data and complex model parameter tuning, and have a strong reliance on external data, making it difficult to achieve the desired results with limited resources. In addition, these methods often use vector space matching or deep semantic embedding in the cross-domain entity alignment process, which has certain black-box characteristics and uncertainties, resulting in a lack of intuitive and clear explanation for the risk profiling results.

[0004] In summary, this case aims to propose a risk profile generation method based on knowledge graphs. It can directly extract risk features from raw text data without relying on existing technical frameworks and pre-trained models, use clear mathematical formulas to describe the frequency and position distribution of candidate words, and achieve cross-domain graph construction through unified standards. This method can not only solve the problem of inconsistent cross-domain data formats, semantic expressions, and statistical features, but also maintain strict and verifiable mathematical expressions during the information extraction and graph fusion process, thereby improving the accuracy and transparency of risk profiles. At the same time, the use of candidate word screening and graph construction based on clear statistical indicators can help reduce the introduction of noise and enhance the stability and adaptability of the system in practical applications. Summary of the Invention

[0005] The present invention provides a risk profile generation method based on knowledge graph, which promotes the solution of the problems mentioned in the above background technology.

[0006] The present invention provides the following technical solution: a risk profile generation method based on a knowledge graph, comprising:

[0007] Set the original text dataset to

[0008] For documents Using the end-of-sentence symbol as a delimiter, split the document d into a set of sentences: in, is the sentence set obtained by segmenting document d according to the symbol; m d is the total number of sentences in document d;

[0009] For each sentence According to the continuous Chinese character sequence, the candidate word set is obtained as follows:

[0010] in, is the set of candidate words obtained by segmenting the sentence s according to the continuous Chinese character sequence; N s is the number of candidate words in sentence s, and each t s,j is a string; j is The jth string in

[0011] For each candidate word t in document d, set: Among them, f1(t) is the number of times the candidate word t appears in the document d; is an indicator function, which takes the value 1 when the condition is met, otherwise it takes the value 0;

[0012] Simultaneous recording Appears in in, is the index set of the sentence in which the candidate word t appears in the document d; is the index of the sentence in which the candidate word t appears in the document d;

[0013] Set f0 = 2, construct in, is the node set of the initial graph; f0 is a constant, which is the minimum number of occurrences of candidate words for screening;

[0014] Each node records the candidate word t and its related information

[0015] For sentence s, let in, is the candidate word set after frequency screening;

[0016] For any sentence s, the indicator function is set as:

[0017] Update edge counts: Among them, c0 is the edge (t i ,t j ) the number of times it co-occurs in all sentences in document d; (t i ,t j ) are any two different candidate words t i and t j The candidate word pairs formed;

[0018] Set the set ε0, ε0 is all different candidate word pairs (t i ,t j ) and their counts;

[0019] Set the initial map to:

[0020] Optionally, it also includes the extraction and structuring of risk factors within the field, specifically:

[0021] For each like Then calculate: Where L(t) is the sentence index dispersion of candidate word t in document d;

[0022] when When , set L(t) = +∞;

[0023] Set L0=3, where L0 is a constant and represents the discreteness screening threshold;

[0024] structure in, is the risk factor candidate set;

[0025] According to the source of document d, document d is classified into a specific field D i , for field D i set up:

[0026] Among them, D i is the identifier of different fields, i is the field number; For field D i The set of risk candidate nodes within;

[0027] Set edge collection For the connection in ε0 The part of the candidate words in the domain constitutes the domain graph:

[0028] in, For field Di Inner initial subgraph;

[0029] For field D i , the node is saved in the form of (t,f1(t),L(t)), and the edge records the co-occurrence count c0(t i ,t j ).

[0030] Optionally, it also includes graph standardization and risk feature normalization, specifically:

[0031] Set the standard node v to: v=(t,F(t),L n (t)); where F(t) is the normalized frequency of candidate word t in the field to which it belongs; L n (t) is the normalized discreteness of candidate word t;

[0032] For field D i ,calculate Among them, f m,i For field D i The maximum raw frequency of all candidate words in ;

[0033] For each Set the normalized frequencies to: Eliminate the magnitude differences of data in various fields and normalize F(t)∈[0,1];

[0034] Set field D i Inside, And the minimum discrete value of L(t)<+∞ is:

[0035]

[0036] set up

[0037] If L(t)=+∞, then let L n (t) = 0;

[0038] Will Update to the standard node v=(t,F(t),L n (t)), side information c0(t i ,t j ) remains unchanged.

[0039] Optionally, cross-domain semantic mapping and entity alignment are also included, specifically:

[0040] For standard nodes v1=(t1,F(t1),L n (t1)) and v2=(t2,F(t2),L n (t2)) Set the matching function to:

[0041]

[0042] For all satisfying S match (v1, v2) = 1, let the number of nodes involved in the merger be k, then the unified node v' is set to: v' = (t, F'(v'), L' n (v')); where t is a common candidate word, and:

[0043] Among them, F'(v') is the normalized frequency of the unified node after merging; L' n (v') is the normalized discreteness of the unified node after merging; v (j) The original standard node participating in the merger;

[0044] For the merged node pair (v',u'), if there is a corresponding edge (v,u) in the original domain, the edge count after the merge is set to:

[0045] Optionally, it also includes cross-domain unified graph construction and unified representation of original frequency and dispersion, specifically:

[0046] Set the unified node set to

[0047] Set the unified edge set to

[0048] The cross-domain unified graph is recorded as:

[0049] For each unified node v' after merging, let the original standard node involved in the merger be v (1) ,v (1) ,...,v (k) ,set up:

[0050] Among them, f orig (v') is the original frequency of the unified node v'; L orig (v') is the original position discreteness of the unified node v'.

[0051] Optionally, it also includes cross-domain graph integration and self-connection structure generation, specifically:

[0052] The cross-domain unified graph has been constructed as Each Record information (t, F'(v'), L' n (v')) and f orig (v') and L orig (v');

[0053] For each Calculate its G * The sum of all edge counts in , specifically:

[0054] in, is the set of adjacent nodes of v';

[0055] For any And v'≠u', if:

[0056] |f orig (v')-f orig (u')|≤1 and |L orig (v')-L orig (u')|≤1, then set the self-connection edge count c s (v',u')=1.

[0057] Optionally, it also includes updating the final map, specifically:

[0058] Set the self-connection edge set to:

[0059]

[0060] The final edge set is: ε final =ε * ∪ε self ;

[0061] The final cross-domain graph is recorded as:

[0062] Optionally, it also includes user risk profile generation, specifically:

[0063] User historical text data The sentence set obtained by segmenting using the end-of-sentence symbol is:

[0064] Among them, K H is the total number of sentences in H;

[0065] For each Extract candidate words and form

[0066] For each set up Where h(t) is the number of times the candidate word t appears in the user text H; For s (H) The total number of candidate words;

[0067] Set the user risk factor set to: v'.t is the candidate word in the user text;

[0068] For each Set the risk contribution value to: m(v') = h(v'.t) × f orig (v'); where h(v'.t) is the number of occurrences of the candidate word v'.t in the user text; f orig (v') is the original frequency of the node;

[0069] Form a risk profile table, each row contains: risk factor string v'.t; user occurrence count h(v'.t); original frequency f orig (v'); risk contribution m(v');

[0070] Set the user's overall risk score to:

[0071] Output the above risk profile table and overall risk score R total For risk assessment.

[0072] The present invention has the following beneficial effects:

[0073] 1. By setting up a raw text dataset and segmenting each document into a set of sentences using sentence-end symbols, this solution addresses the issues of high data noise and semantic discontinuity in the raw text. This step transforms the document into a well-defined basic processing unit, providing a reliable foundation for subsequent candidate word extraction and effectively improving the accuracy and stability of text processing. This segmentation ensures semantic coherence within each sentence, making subsequent statistics on the frequency and positional distribution of candidate words more accurate, thus laying a solid data foundation for risk factor screening. By segmenting each sentence according to a continuous sequence of Chinese characters and extracting a set of candidate words, this solution overcomes the dependency issue of traditional word segmentation tools and enables independent and interference-free candidate word extraction. This step ensures that candidate words are basic text units within a sentence, ensuring that each candidate word has clear semantic boundaries and reducing errors caused by dictionary dependence or model pre-sets. Furthermore, by recording the number of times a candidate word appears in a document and the index of the sentence in which it appears, this solution further addresses statistical inaccuracies in information extraction and provides a comprehensive statistical basis for subsequent candidate word-based risk factor screening. By setting a minimum number of occurrences, f0, for candidate word screening, this solution balances noise interference with information coverage. Larger values effectively eliminate low-frequency noise, ensuring more representative nodes in the graph, although some low-frequency but important risk warning terms may be omitted. Lower values maximize coverage of a wide range of information, but this may also introduce some noise. Therefore, by appropriately adjusting the f0 value, the candidate word screening process becomes more flexible and adaptable to different text characteristics, thereby improving the accuracy and practicality of risk factor extraction. By constructing edge relationships for the frequency-filtered candidate word set in each sentence and counting the co-occurrences of candidate word pairs within the sentence using an indicator function, this solution addresses the problem of unclear relationships between words in the text. Recording the co-occurrence relationships of candidate word pairs not only provides rich information for constructing the initial graph, but also reflects the strength of semantic associations between different candidate words. Ultimately, this series of steps constructs an initial knowledge graph based on the candidate words and their co-occurrence relationships within the document, effectively improving the system's ability to identify risk information and the accuracy of its judgments during the subsequent risk profile generation process.

[0074] 2. By extracting and structuring risk factors within a domain, the system addresses the issues of inaccurate risk factor screening and difficulty quantifying contextual dispersion in cross-domain texts, achieving significant results. Specifically, the system first calculates the index dispersion of the sentences in which candidate terms appear within a document, measuring the concentration of candidate terms within the document. This addresses the issue of noise and irrelevant information being mixed into the original candidate terms. By setting a dispersion screening threshold, L0, the system can filter candidate terms based on their contextual concentration within the document. When L0 is large, the filtering criteria are relaxed, allowing candidates to be retained even when they appear more dispersed within the document, thereby covering more information. However, this may result in some candidate terms, despite their high frequency of occurrence, not possessing clear risk-indicating characteristics due to their dispersed context, thus affecting the accuracy of the risk profile. Conversely, when L0 is small, the system strictly filters, retaining only those candidate terms with highly concentrated occurrences within the document. This helps refine risk factors and highlight risk signals, but may also miss some important information with a wider context. By calculating the discreteness of candidate words and performing threshold screening, the system reduces noise interference and effectively captures risk information during the risk factor extraction process, making the screened risk factors more contextually consistent and more risk-indicative. At the same time, documents are classified into specific fields based on their source, and a set of risk candidate nodes is constructed within each field. Subgraphs are then formed based on the co-occurrence relationships between candidate words within each field, thereby achieving the grouping and structured preservation of risk factors from different fields. This not only fully extracts and clearly expresses the risk information within each field, but also provides accurate data support for subsequent unified standardization processing and the construction of cross-domain graphs. Overall, through the intra-domain risk factor extraction and structuring steps, the system is able to achieve effective screening, hierarchical management, and structured preservation of risk factors in text, laying a solid data foundation for generating accurate and comprehensive risk profiles, and further improving the integration effect of cross-domain risk graphs and the accuracy of risk assessments.

[0075] 3. Through the graph standardization and risk feature normalization steps, this solution solves the problem of inconsistent magnitude, format and indicator scale of text data in different fields, and achieves significant results. Specifically, by first setting the standard node, the original frequency and position discreteness of the candidate words are normalized separately, thus achieving the consistency of the risk factor data structure in various fields. In this step, by calculating the maximum original frequency of all candidate words in each field and using this value to normalize the occurrence frequency of each candidate word, the frequency of the candidate words in each field is mapped to a fixed interval, usually [0,1], thereby eliminating the comparison difficulty caused by the different magnitudes of the field data; at the same time, by calculating the minimum discreteness of the candidate word in each field and using this value to normalize the discreteness of each candidate word, it is mapped to the same numerical interval to ensure that the data of different candidate words in position distribution are comparable. Through this normalization process, the system can ensure that the data structure of each node is unified as (candidate word, normalized frequency, normalized discreteness) when constructing standard nodes, so that the subsequent cross-domain merging and entity alignment processes can be carried out under the same indicator system, avoiding the introduction of noise or distortion due to inconsistent data scale and format. In addition, the normalized data has good stability and interpretability, laying a solid foundation for further construction of a cross-domain unified risk knowledge graph. Ultimately, this step not only improves the robustness of the system when processing multi-domain text data, but also makes the comparison and aggregation of risk factors intuitive and accurate, effectively improving the overall quality of risk portraits. In general, through the graph standardization and risk feature normalization steps, the problem of inconsistent data magnitude, format and indicators of cross-domain risk factors is solved, achieving the effect of eliminating differences between domains, unifying data representation and improving the accuracy of subsequent risk map construction, providing a reliable, objective and traceable data basis for risk management and decision support.

[0076] 4. Through cross-domain semantic mapping and entity alignment, the solution addresses the issue of inconsistent representation of the same risk factor in texts from different domains due to differences in expression, achieving significant results. Specifically, a matching function is first used to compare standard nodes from different domains to determine whether their candidate word strings are completely identical, thereby eliminating mismatches caused by semantic ambiguity or differences in expression. This step ensures that only semantically identical candidate words across different domains are considered the same risk factor, resolving the semantic inconsistency issue inherent in cross-domain entity alignment. Next, for all nodes that meet the matching criteria, the system calculates the number of nodes involved in merging and merges these nodes into a unified node using the arithmetic average. The merged unified node has a uniform normalized frequency and normalized dispersion. The normalized frequency, calculated as the arithmetic average, reflects the occurrence intensity of the candidate word across each domain, while the normalized dispersion, calculated as the arithmetic average, reflects the concentration of the candidate word's occurrence across each domain. This process effectively eliminates inconsistencies in data from different domains due to differences in statistics and scales, resulting in a unified, clear, and comparable representation of risk factors across domains. At the same time, for the merged node pairs, if there are corresponding co-occurrence edges in the original domain, the system merges the information of these edges and sets the merged edge count to the sum of all relevant original edge counts. This step retains the original co-occurrence information of each domain, realizes the unified fusion of edge information, and ensures that the correlation between various risk factors is accurately reflected when constructing a unified cross-domain graph. Overall, through the steps of cross-domain semantic mapping and entity alignment, the solution effectively solves the problem that risk factors in different domains are difficult to unify due to inconsistent language expression, data statistics and formats, so that cross-domain information can be accurately reflected in a unified graph. As a result, not only the same risk factor is effectively integrated in various domains, but also subsequent steps such as graph construction and risk profile generation can be carried out under a unified and objective indicator system, thereby improving the accuracy of risk assessment and the overall robustness of the system.

[0077] 5. By constructing a unified cross-domain graph and uniformly representing raw frequencies and dispersions, this step addresses the inconsistency in statistical indicators and structural information across data from different domains, thereby achieving comprehensive integration and unified representation of cross-domain risk factor information. Specifically, by establishing a unified set of nodes and edges, the standard nodes and edge information from different domains, after semantic mapping and entity alignment, are merged to construct a unified cross-domain graph. This graph encompasses the associated information of the same risk factors across all domains, thus resolving the information silos caused by heterogeneous cross-domain data and inconsistent indicators. Furthermore, for each merged unified node, the raw frequencies and raw position dispersions of the original standard nodes involved in the merger are calculated and their arithmetic mean is computed to obtain the raw frequencies and raw dispersions of the unified node. This approach addresses the difficulty in directly comparing quantitative indicators due to differences in the frequency and contextual distribution of risk factors across different domains, making the statistical characteristics of the unified nodes comparable and interpretable. By using raw frequency as an objective indicator of the intensity of candidate word occurrence in each domain, and raw dispersion as a reflection of the concentration of candidate word locations within the text, the unified graph not only accurately reflects the co-occurrence relationships of cross-domain risk factors but also provides a concrete raw numerical basis for subsequent risk profile generation. In summary, by constructing a cross-domain unified graph and uniformly representing raw frequency and dispersion for unified nodes, the inconsistencies in statistical measurement and structural fusion of cross-domain data are resolved, achieving the goal of information fusion. This not only ensures the complete collection and effective integration of risk information from different domains, but also provides standardized and objective foundational data for subsequent steps (such as self-connection edge generation and risk profile construction). This makes the entire risk profile generation process more rigorous and transparent, and improves the accuracy of risk assessment and the robustness of the system.

[0078] 6. Through cross-domain graph integration and self-linking structure generation, the problems of weak connections between risk factors, information silos, and local structural incoherence in the unified cross-domain graph are resolved, thereby enhancing the overall internal consistency and stability of the graph. Specifically, after the cross-domain unified graph is constructed, each node records detailed information such as candidate terms, normalized metrics, raw frequencies, and raw dispersion. Next, for each node, the sum of all edge counts in the unified graph is calculated, that is, the connection strength between the node and all adjacent nodes is calculated, providing an objective basis for quantifying node associations. This step can identify weakly connected or isolated nodes in the graph, providing data support for subsequent structural optimization. Furthermore, for any two nodes that meet the condition of minimal difference in raw frequency and raw dispersion, a self-linking edge is introduced, with the self-linking edge count fixed to 1. This condition ensures that only nodes with similar raw statistical characteristics are additionally connected, effectively compensating for local information differences caused by data collection and domain characteristics. By introducing self-connecting edges, the system not only strengthens the intrinsic connections between similar risk factors, but also improves the connectivity of the entire graph, reducing the problem of missing or mismatched risk information that may be caused by data differences between fields. Ultimately, after this series of operations, the risk factors in the graph have been further optimized and improved in structure, the connections between nodes have become closer, information transmission has become smoother, and the graph as a whole has shown higher internal consistency and robustness. This not only makes the weights and correlations of risk factors more accurate and reliable during the risk profile generation process, but also provides a more interpretable and operational basic data structure for subsequent risk assessment and decision analysis. Through this step, the system can more accurately reflect the true connections between cross-domain risk factors, realize the refinement and quantification of risk profiles, and thus enhance overall risk management capabilities.

[0079] 7. By updating the final graph, the system addresses issues such as incomplete information fusion, insufficient structural connectivity, and missing relationships between some risk factors in the cross-domain knowledge graph. This results in a complete cross-domain risk knowledge graph, encompassing both original edges and self-connecting edges. First, the system combines the calculated self-connecting edges with the cross-domain unified graph constructed in the previous steps to form a self-connecting edge set. This ensures the necessary supplementary connections between the original risk factors, enhancing the overall connectivity of the graph. This step effectively mitigates the issue of isolated nodes caused by data imbalance or insufficient sampling between domains, facilitating smoother cross-domain information transfer. Next, the system merges the edge set of the original cross-domain graph with the self-connecting edge set to form the final complete edge set. This not only preserves the original co-occurrence relationships between risk factors but also adds internal connections created during the cross-domain matching process, making the risk factor relationships in the graph more complete and robust. Finally, based on the updated complete edge set, the final cross-domain graph is constructed. This graph not only captures the underlying risk factor relationships across domains but also reinforces the underlying connections between factors, ensuring a more comprehensive representation of cross-domain risk information. The direct effect of this step is to improve the integrity of the cross-domain risk knowledge graph, allowing the system to more accurately utilize cross-domain data for calculations and reasoning in the subsequent risk analysis and assessment process. By supplementing self-connecting edges, the connection between risk factors with similar characteristics is enhanced, the robustness and connectivity of the graph are improved, and the omission or weakening of important risk factors is avoided. In addition, a complete cross-domain risk knowledge graph can more accurately support risk identification, classification, and early warning, allowing the system to conduct in-depth analysis based on a more reliable knowledge structure, improving the accuracy of risk prediction and management. Ultimately, this step ensures that the cross-domain risk knowledge graph has high stability and adaptability, providing a unified and standardized knowledge expression method for risk factors in different fields, thereby improving the practicality and accuracy of the entire risk assessment system.

[0080] 8. Through the steps of user risk profile generation, the system addresses the issues of unsystematic user risk factor extraction, difficulty in quantifying risk contributions, and a lack of clear assessment of overall risk levels. This results in the construction of a user risk profile based on quantitative indicators. First, the system segments the user's historical text data at the end of sentences to generate a sentence set, which provides the basic unit for subsequent candidate word extraction and ensures the integrity and accuracy of text processing. Subsequently, the system extracts candidate words for each sentence and counts the number of times the candidate words appear in the user's text to quantify their importance within the user's text. This process ensures the rigor of candidate word extraction and statistics, providing data support for the subsequent mapping of risk factors. Next, the system establishes a user risk factor set, mapping candidate words appearing in the user's text to standardized risk factors in the cross-domain risk knowledge graph, ensuring that key risk factors in the user's text are aligned with the structure of the unified knowledge graph. This step effectively resolves the mismatch between user text data and standardized risk factors, enabling scientific assessment of user risk analysis based on the established cross-domain risk knowledge system. Simultaneously, the system calculates a risk contribution value for each candidate word in the user's text. This value is determined by the candidate's occurrence count in the user's text and the raw frequency of the risk factor in the graph, thereby quantifying the impact of different risk factors. This quantification process ensures that the risk factor calculation is not limited to simple word frequency statistics but can incorporate prior knowledge from the graph, improving the accuracy of risk analysis. Subsequently, based on the calculated risk contribution values, the system generates a user risk profile table, in which each risk factor is mapped to a specific risk contribution, providing a clear risk structure analysis. This table not only includes the risk factors in the user's text but also annotates their occurrence counts, raw frequency in the graph, and risk contribution values, making the risk assessment intuitive. Furthermore, the system calculates the user's overall risk score, which comprehensively considers the contribution of all risk factors to assess the user's risk level holistically, ensuring the systematic and objective nature of the assessment. Ultimately, this step enables the presentation of user risk profiles in a standardized and quantifiable manner, effectively improving the accuracy of user risk assessments. Through quantitative analysis based on knowledge graphs, the extraction and calculation of user risk factors becomes more systematic, ensuring the interpretability and reliability of risk assessment results. Furthermore, the generation of user risk profiles provides a solid data foundation for subsequent risk warnings, personalized risk management, and decision support, enhancing the system's risk prediction and response capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0082] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0083] Example, see Figure 1 , a risk profile generation method based on knowledge graph, including:

[0084] Set the original text dataset to

[0085] For documents Using the end-of-sentence symbol as a delimiter, split the document d into a set of sentences: in, is the sentence set obtained by segmenting document d according to the symbol; m d is the total number of sentences in document d; it provides the basic processing unit for subsequent candidate word extraction;

[0086] The sentence end symbols include: ".", "?", "!";

[0087] For example, for d1 (financial field), the text "Today, the company's stock price fell sharply. Investor sentiment is depressed. Market risks are increasing." is split by "." to obtain the following sentence set:

[0088]

[0089] For each sentence According to the continuous Chinese character sequence, the candidate word set is obtained as follows:

[0090] in, is the set of candidate words obtained by segmenting the sentence s according to the continuous Chinese character sequence; N s is the number of candidate words in sentence s, and each t s,j is a string; j is Extract the basic text units that make up the sentence without using external word segmentation tools to ensure independent implementation;

[0091] For example, for d1, s1: "The company's stock price fell sharply today", the candidate word set is:

[0092]

[0093] For d1, s2: "investor sentiment is low", the candidate word set is:

[0094]

[0095] For d1, s3: "market risks continue to intensify", the candidate word set is:

[0096]

[0097] For each candidate word t in document d, set: Among them, f1(t) is the number of times the candidate word t appears in the document d; is an indicator function, which takes the value 1 when the condition is met, otherwise it takes the value 0;

[0098] Simultaneous recording in, is the index set of the sentence in which the candidate word t appears in the document d; l is the index of the sentence in which the candidate word t appears in the document d; the number of occurrences and sentence position of each candidate word are counted for subsequent risk factor screening;

[0099] For example: "risk" only appears in s3 in d1, then f1("risk")=1,

[0100] Set f0 = 2, construct in, is the node set of the initial graph; f0 is a constant, which is the minimum number of occurrences for candidate word screening; when the value of f0 is large, only candidate words with a high number of occurrences in the document will be retained, which can effectively exclude low-frequency noise; words that appear only a few times but have risk warning significance in a specific context may be missed; when the value of f0 is large, the result is that the number of nodes in the graph is reduced, the graph is sparser, and the risk factor filtering is stricter; the extracted risk factors are more concentrated, but the coverage and sensitivity may be reduced; when the value of f0 is small, the impact is that low-frequency candidate words will also be retained, thereby containing more information; a large number of noise words may be introduced, interfering with subsequent risk factor screening and graph construction; when the value of f0 is small, the result is that the number of nodes in the graph increases, the noise is more, and subsequent calculations may require supplementary screening through other indicators, such as position dispersion;

[0101] Each node records the candidate word t and its related information Filter out low-frequency noise and retain meaningful candidate words;

[0102] For sentence s, let in, It is a set of candidate words that have been filtered by frequency, and is used to ensure that the candidate words have sufficient statistical significance when constructing edge relationships;

[0103] For any sentence s, the indicator function is set as:

[0104] Update edge counts: Among them, c0 is the edge (t i ,t j ) the number of times it co-occurs in all sentences in document d; (t i ,t j ) are any two different candidate words t i and t j The candidate word pairs constituted by the ,situation are the candidate word pairs that appear together in the same ,context, and are used to record the co-occurrence relationship between them;

[0105] Set the set ε0, ε0 is all different candidate word pairs (t i ,t j ) and their counts; record the co-occurrence relationship of candidate words in the sentence;

[0106] Set the initial map to: Construct an initial knowledge graph based on candidate words and their co-occurrence relationships in document d.

[0107] By setting up a raw text dataset and segmenting each document into a set of sentences using the end-of-sentence symbols ".", "?", and "!", this solution addresses the issues of high data noise and semantic discontinuity in the raw text. This step transforms the document into a well-defined basic processing unit, providing a reliable foundation for subsequent candidate word extraction and effectively improving the accuracy and stability of text processing. This segmentation ensures semantic coherence within each sentence, enabling more accurate statistics on the frequency and positional distribution of candidate words, thus laying a solid data foundation for risk factor screening. By segmenting each sentence according to a continuous sequence of Chinese characters and extracting a set of candidate words, this solution overcomes the dependency issue of traditional word segmentation tools and enables independent and interference-free candidate word extraction. This step ensures that candidate words are basic text units within a sentence, ensuring that each candidate word has clear semantic boundaries and reducing errors caused by dictionary dependence or model pre-sets. Furthermore, by recording the number of candidate word occurrences in a document and the index of the sentence in which it appears, this solution further addresses statistical inaccuracies in information extraction and provides a comprehensive statistical basis for subsequent candidate word-based risk factor screening. By setting a minimum number of occurrences, f0, for candidate word screening, this solution strikes a balance between noise interference and information coverage. Larger values effectively eliminate low-frequency noise, ensuring more representative nodes in the graph, although some low-frequency but important risk warning terms may be omitted. Lower values maximize coverage of a wide range of information, but this may also introduce some noise. Therefore, by appropriately adjusting the f0 value, the candidate word screening process becomes more flexible and adaptable to different text characteristics, thereby improving the accuracy and practicality of risk factor extraction. By constructing edge relationships for the frequency-filtered candidate word set in each sentence and counting the co-occurrences of candidate word pairs within the sentence using an indicator function, this solution addresses the problem of unclear relationships between words in the text. Recording the co-occurrence relationships of candidate word pairs not only provides rich information for constructing the initial graph, but also reflects the strength of semantic associations between different candidate words. Ultimately, this series of steps constructs an initial knowledge graph based on the candidate words and their co-occurrence relationships within the document, effectively improving the system's ability to identify risk information and the accuracy of its judgments during the subsequent risk profile generation process.

[0108] It also includes the extraction and structuring of risk factors within the field, specifically:

[0109] For each like Then calculate: Where L(t) is the sentence index dispersion of candidate word t in document d, which is used to measure the concentration of the position where candidate word t appears;

[0110] when When , set L(t) = +∞;

[0111] Set L0=3, L0 is a constant, which represents the discreteness screening threshold; when L0 is large, the impact is that a higher L0 means looser screening conditions, that is, candidate words are allowed to appear more dispersedly in the document and can also be included in the risk factor set; more candidate words may be retained, but some of them may be more contextually dispersed and do not have obvious risk indication characteristics; when L0 is large, the result is that the number of risk factors in the graph increases and the coverage is wider, but the accuracy and discrimination of the risk profile may be reduced; when L0 is small, the impact is that the screening conditions are stricter, and only those candidate words that appear in very concentrated positions in the document can be retained; some risk information may be missed, such as words that are slightly widely distributed but still have risk warning significance; when L0 is small, the result is that the selected risk factors are more refined, the graph structure is sparse, and the risk signals are clearer, but the coverage may be limited;

[0112] structure in, A candidate set of risk factors is created; candidate words that are likely to indicate risk are screened in a concentrated context;

[0113] According to the source of document d, document d is classified into a specific field D i , for field D i set up:

[0114] Among them, D i is the identifier of different fields, i is the field number; For field D i The set of risk candidate nodes in At the same time, each candidate word t in the domain D i The source within has been identified;

[0115] Set edge collection For connections in ε0 The part of the candidate words in the domain constitutes the domain graph:

[0116] in, For field D i Initial subgraph within; grouping and structured storage of risk factors in different fields;

[0117] For field D i , the node is saved in the form of (t,f1(t),L(t)), and the edge records the co-occurrence count c0(t i ,t j ); provide data for subsequent unified standardization processing.

[0118] By extracting and structuring risk factors within a domain, the system addresses the issues of inaccurate risk factor screening and difficulty quantifying contextual dispersion in cross-domain texts, achieving significant results. Specifically, the system first calculates the index dispersion of the sentences in which candidate words appear in a document, measuring the concentration of candidate word locations within the document. This addresses the issue of noise and irrelevant information being mixed into the original candidate words. By setting a dispersion screening threshold, L0, the system can filter candidate words based on their contextual concentration within the document. When L0 is large, the filtering criteria are relaxed, allowing candidates to be retained even when they appear more dispersed within the document, thereby covering more information. However, this may result in some candidate words, despite their high frequency of occurrence, not possessing clear risk-indicating characteristics due to their dispersed context, thus affecting the accuracy of the risk profile. Conversely, when L0 is small, the system applies strict screening, retaining only those candidate words with highly concentrated locations within the document. This refines risk factors and highlights risk signals, but may also miss some important information with a wider context. By calculating the discreteness of candidate words and performing threshold screening, the system reduces noise interference and effectively captures risk information during the risk factor extraction process, making the screened risk factors more contextually consistent and more risk-indicative. At the same time, documents are classified into specific fields based on their source, and a set of risk candidate nodes is constructed within each field. Subgraphs are then formed based on the co-occurrence relationships between candidate words within each field, thereby achieving the grouping and structured preservation of risk factors from different fields. This not only fully extracts and clearly expresses the risk information within each field, but also provides accurate data support for subsequent unified standardization processing and the construction of cross-domain graphs. Overall, through the intra-domain risk factor extraction and structuring steps, the system is able to achieve effective screening, hierarchical management, and structured preservation of risk factors in text, laying a solid data foundation for generating accurate and comprehensive risk profiles, and further improving the integration effect of cross-domain risk graphs and the accuracy of risk assessments.

[0119] It also includes graph standardization and risk feature normalization, specifically:

[0120] Set the standard node v to: v=(t,F(t),L n (t)); where F(t) is the normalized frequency of candidate word t in the field to which it belongs; L n (t) is the normalized discreteness of candidate word t; ensure the consistency of node data structure in different fields;

[0121] For field D i ,calculate Among them, f m,i For field D i The maximum raw frequency of all candidate words in ;

[0122] For each Set the normalized frequencies to: Eliminate the magnitude differences of data in various fields and normalize F(t)∈[0,1];

[0123] Set field D i Inside, And the minimum discrete value of L(t)<+∞ is:

[0124]

[0125] set up

[0126] If L(t)=+∞, then let L n (t) = 0; normalize the discreteness to the interval [0, 1] to make the values comparable;

[0127] Will Update to the standard node v=(t,F(t),L n (t)), side information c0(t i ,t j ) remain unchanged; complete the standardization of data within the field and prepare for cross-domain mergers.

[0128] Through the steps of graph standardization and risk feature normalization, this solution solves the problem of inconsistent magnitude, format, and indicator scale of text data in different fields, and achieves significant results. Specifically, by first setting standard nodes, the original frequency and position dispersion of candidate words are normalized separately, achieving the consistency of risk factor data structure in various fields. In this step, by calculating the maximum original frequency of all candidate words in each field and using this value to normalize the occurrence frequency of each candidate word, the frequency of candidate words in each field is mapped to a fixed interval, usually [0,1], thereby eliminating the comparison difficulty caused by the different magnitudes of field data; at the same time, by calculating the minimum dispersion of candidate words in each field and using this value to normalize the dispersion of each candidate word, it is mapped to the same numerical interval, ensuring that the data of different candidate words in position distribution are comparable. Through this normalization process, the system can ensure that the data structure of each node is unified as (candidate word, normalized frequency, normalized discreteness) when constructing standard nodes, so that the subsequent cross-domain merging and entity alignment processes can be carried out under the same indicator system, avoiding the introduction of noise or distortion due to inconsistent data scale and format. In addition, the normalized data has good stability and interpretability, laying a solid foundation for further construction of a cross-domain unified risk knowledge graph. Ultimately, this step not only improves the robustness of the system when processing multi-domain text data, but also makes the comparison and aggregation of risk factors intuitive and accurate, effectively improving the overall quality of risk portraits. In general, through the graph standardization and risk feature normalization steps, the problem of inconsistent data magnitude, format and indicators of cross-domain risk factors is solved, achieving the effect of eliminating differences between domains, unifying data representation and improving the accuracy of subsequent risk map construction, providing a reliable, objective and traceable data basis for risk management and decision support.

[0129] It also includes cross-domain semantic mapping and entity alignment, specifically:

[0130] For standard nodes v1=(t1,F(t1),L n (t1)) and v2=(t2,F(t2),L n (t2)) Set the matching function to:

[0131] Determine whether candidate words from different fields are semantically consistent;

[0132] For all satisfying S match (v1, v2) = 1 nodes (from different fields), let the number of nodes involved in the merger be k, then the unified node v' is set to: v' = (t, F'(v'), L' n (v')); where t is a common candidate word, and:

[0133] Among them, F'(v') is the normalized frequency of the unified node after merging, which is calculated by arithmetic average and reflects the occurrence intensity of the candidate word in each field; L' n (v') is the normalized discreteness of the unified node after merging, which is calculated by arithmetic average and reflects the concentration of the candidate word's position in each field; v (j) The original standard node participating in the merger;

[0134] For the merged node pair (v',u'), if there is a corresponding edge v(,u) in the original domain, the edge count after the merge is set to: The original co-occurrence information of each field is retained to achieve unified fusion of side information.

[0135] Through cross-domain semantic mapping and entity alignment, the solution addresses the problem of inconsistent representation of the same risk factor in texts from different domains due to differences in expression, achieving significant results. Specifically, a matching function is first used to compare standard nodes from different domains to determine whether their candidate word strings are completely identical, thereby eliminating mismatches caused by semantic ambiguity or differences in expression. This step ensures that only semantically identical candidate words across different domains are considered the same risk factor, resolving the semantic inconsistency issue inherent in cross-domain entity alignment. Next, for all nodes that meet the matching criteria, the system calculates the number of nodes involved in merging and merges these nodes into a unified node using the arithmetic average. The merged unified node has a uniform normalized frequency and normalized dispersion. The normalized frequency, calculated as the arithmetic average, reflects the occurrence intensity of the candidate word across each domain, while the normalized dispersion, calculated as the arithmetic average, reflects the concentration of the candidate word's occurrence across each domain. This process effectively eliminates inconsistencies in data from different domains due to differences in statistics and scales, making the representation of cross-domain risk factors uniform, clear, and comparable. At the same time, for the merged node pairs, if there are corresponding co-occurrence edges in the original domain, the system merges the information of these edges and sets the merged edge count to the sum of all relevant original edge counts. This step retains the original co-occurrence information of each domain, realizes the unified fusion of edge information, and ensures that the correlation between various risk factors is accurately reflected when constructing a unified cross-domain graph. Overall, through the steps of cross-domain semantic mapping and entity alignment, the solution effectively solves the problem that risk factors in different domains are difficult to unify due to inconsistent language expression, data statistics and formats, so that cross-domain information can be accurately reflected in a unified graph. As a result, not only the same risk factor is effectively integrated in various domains, but also subsequent steps such as graph construction and risk profile generation can be carried out under a unified and objective indicator system, thereby improving the accuracy of risk assessment and the overall robustness of the system.

[0136] It also includes cross-domain unified graph construction and unified representation of original frequency and dispersion, specifically:

[0137] Set the unified node set to

[0138] Set the unified edge set to

[0139] The cross-domain unified graph is recorded as: Complete semantic mapping and entity alignment of cross-domain information;

[0140] For each unified node v' after merging, let the original standard node involved in the merger be v (1) ,v (1) ,...,v (k) ,set up:

[0141] Among them, f orig (v') is the original frequency of the unified node v'; L orig (v') is the original position discreteness of the unified node v'; it provides a specific original numerical basis for subsequent steps.

[0142] By constructing a unified cross-domain graph and unifying the representation of raw frequencies and dispersions, the inconsistency in statistical indicators and structural information across data from different domains is resolved, thereby achieving comprehensive integration and unified representation of cross-domain risk factor information. Specifically, by establishing a unified set of nodes and edges, the standard nodes and edges from different domains, after semantic mapping and entity alignment, are merged to construct a unified cross-domain graph. This graph encompasses the associated information of the same risk factors across all domains, thus resolving the information silos caused by heterogeneous cross-domain data and inconsistent indicators. Furthermore, for each merged unified node, the raw frequencies and raw position dispersions of the original standard nodes involved in the merge are calculated and their arithmetic mean is computed to obtain the raw frequencies and raw dispersions of the unified node. This approach addresses the difficulty in directly comparing quantitative indicators due to differences in the frequency and contextual distribution of risk factors across different domains, making the statistical characteristics of the unified nodes comparable and interpretable. By using raw frequency as an objective indicator of the intensity of candidate word occurrence in each domain, and raw dispersion as a reflection of the concentration of candidate word locations within the text, the unified graph not only accurately reflects the co-occurrence relationships of cross-domain risk factors but also provides a concrete raw numerical basis for subsequent risk profile generation. In summary, by constructing a cross-domain unified graph and uniformly representing raw frequency and dispersion for unified nodes, the inconsistencies in statistical measurement and structural fusion of cross-domain data are resolved, achieving the goal of information fusion. This not only ensures the complete collection and effective integration of risk information from different domains, but also provides standardized and objective foundational data for subsequent steps (such as self-connection edge generation and risk profile construction). This makes the entire risk profile generation process more rigorous and transparent, and improves the accuracy of risk assessment and the robustness of the system.

[0143] It also includes cross-domain graph integration and self-connection structure generation, specifically:

[0144] The cross-domain unified graph has been constructed as Each Record information (t, F'(v'), L' n (v')) and f orig (v') and L orig (v');

[0145] For each Calculate its G * The sum of all edge counts in , specifically:

[0146] in, is the set of adjacent nodes of v'; quantifies the connection strength of node v';

[0147] For any And v'≠u', if:

[0148] |f orig (v')-f orig (u')|≤1 and |L orig (v')-L orig (u')|≤1, then set the self-connection edge count c s (v',u')=1;

[0149] Introducing additional connections for nodes with similar original features to enhance the internal consistency of the graph.

[0150] Through cross-domain graph integration and self-linking structure generation, the system addresses the issues of weak connections between risk factors, information silos, and local structural incoherence within the unified cross-domain graph, thereby enhancing the overall internal consistency and stability of the graph. Specifically, after the unified cross-domain graph is constructed, each node is recorded with detailed information such as candidate terms, normalized metrics, raw frequencies, and raw dispersion. Next, for each node, the sum of all edge counts within the unified graph is calculated, thereby calculating the strength of the connection between the node and all adjacent nodes. This provides an objective basis for quantifying node associations. This step identifies weakly connected or isolated nodes in the graph, providing data support for subsequent structural optimization. Furthermore, for any two nodes that meet the condition of minimal difference in raw frequency and raw dispersion, a self-linking edge is introduced, with the self-linking edge count fixed to 1. This condition ensures that only nodes with similar raw statistical characteristics are connected, effectively compensating for local information differences caused by data collection and domain characteristics. By introducing self-connecting edges, the system not only strengthens the intrinsic connections between similar risk factors, but also improves the connectivity of the entire graph, reducing the problem of missing or mismatched risk information that may be caused by data differences between fields. Ultimately, after this series of operations, the risk factors in the graph have been further optimized and improved in structure, the connections between nodes have become closer, information transmission has become smoother, and the graph as a whole has shown higher internal consistency and robustness. This not only makes the weights and correlations of risk factors more accurate and reliable during the risk profile generation process, but also provides a more interpretable and operational basic data structure for subsequent risk assessment and decision analysis. Through this step, the system can more accurately reflect the true connections between cross-domain risk factors, realize the refinement and quantification of risk profiles, and thus enhance overall risk management capabilities.

[0151] It also includes updating the final atlas, specifically:

[0152] Set the self-connection edge set to:

[0153]

[0154] The final edge set is: ε final =ε * ∪ε self ;

[0155] The final cross-domain graph is recorded as: Construct a complete cross-domain risk knowledge graph containing original edges and self-connected edges.

[0156] By updating the final graph, the system addresses issues such as incomplete information fusion, insufficient structural connectivity, and missing relationships between some risk factors in the cross-domain knowledge graph. This results in a complete cross-domain risk knowledge graph, comprising both original edges and self-linking edges. First, the system combines the calculated self-linking edges with the cross-domain unified graph constructed in the previous step to form a self-linking edge set. This ensures the necessary supplementary connections between the original risk factors, enhancing the overall connectivity of the graph. This step effectively mitigates the issue of isolated nodes caused by data imbalance or insufficient sampling between domains, facilitating smoother cross-domain information transfer. Next, the system merges the edge set of the original cross-domain graph with the self-linking edge set to form the final complete edge set. This not only preserves the original co-occurrence relationships between risk factors but also adds internal connections created during the cross-domain matching process, making the risk factor relationships in the graph more complete and robust. Finally, based on the updated complete edge set, the final cross-domain graph is constructed. This graph not only captures the underlying risk factor relationships within different domains but also reinforces the underlying connections between factors, ensuring a more comprehensive representation of cross-domain risk information. The direct effect of this step is to improve the integrity of the cross-domain risk knowledge graph, allowing the system to more accurately utilize cross-domain data for calculations and reasoning in the subsequent risk analysis and assessment process. By supplementing self-connecting edges, the connection between risk factors with similar characteristics is enhanced, the robustness and connectivity of the graph are improved, and the omission or weakening of important risk factors is avoided. In addition, a complete cross-domain risk knowledge graph can more accurately support risk identification, classification, and early warning, allowing the system to conduct in-depth analysis based on a more reliable knowledge structure, improving the accuracy of risk prediction and management. Ultimately, this step ensures that the cross-domain risk knowledge graph has high stability and adaptability, providing a unified and standardized knowledge expression method for risk factors in different fields, thereby improving the practicality and accuracy of the entire risk assessment system.

[0157] It also includes user risk profile generation, specifically:

[0158] User historical text data The sentence set obtained by segmenting using the end-of-sentence symbol is:

[0159] Among them, K His the total number of sentences in H; provides a unit for extracting candidate words from user text;

[0160] For each Extract candidate words and form

[0161] For each set up Where h(t) is the number of times the candidate word t appears in the user text H; For s (H) The total number of candidate words;

[0162] Set the user risk factor set to: v'.t is the candidate word in the user text; the candidate words appearing in the user text are matched with the risk factors in the unified graph;

[0163] For each Set the risk contribution value to: m(v') = h(v'.t) × f orig (v'); where h(v'.t) is the number of occurrences of the candidate word v'.t in the user text; f orig (v') is the original frequency of the node; it quantifies the contribution of each risk factor to the user's overall risk;

[0164] Form a risk profile table, each row contains: risk factor string v'.t; user occurrence count h(v'.t); original frequency f orig (v'); risk contribution m(v');

[0165] Set the user's overall risk score to: Generate user risk profiles with quantitative indicators;

[0166] Output the above risk profile table and overall risk score R total For risk assessment.

[0167] The steps involved in generating a user risk profile address the issues of unsystematic user risk factor extraction, difficulty quantifying risk contributions, and a lack of clear assessment of overall risk levels. This approach ultimately constructs a user risk profile based on quantitative indicators. First, the system segments historical user text data at the end of sentences to generate a sentence set, which serves as the building block for subsequent candidate word extraction and ensures the integrity and accuracy of text processing. Subsequently, the system extracts candidate words from each sentence and counts the number of occurrences of candidate words in the user text to quantify their importance within the user text. This process ensures the rigor of candidate word extraction and statistics, providing data support for subsequent risk factor mapping. Next, the system establishes a user risk factor set, mapping candidate words appearing in the user text to standardized risk factors in the cross-domain risk knowledge graph. This ensures that key risk factors in the user text align with the structure of the unified knowledge graph. This step effectively resolves the mismatch between user text data and standardized risk factors, enabling scientific user risk analysis based on the established cross-domain risk knowledge system. Simultaneously, the system calculates a risk contribution value for each candidate word in the user's text. This value is determined by the candidate's occurrence count in the user's text and the raw frequency of the risk factor in the graph, thereby quantifying the impact of different risk factors. This quantification process ensures that the risk factor calculation is not limited to simple word frequency statistics but can incorporate prior knowledge from the graph, improving the accuracy of risk analysis. Subsequently, based on the calculated risk contribution values, the system generates a user risk profile table, in which each risk factor is mapped to a specific risk contribution, providing a clear risk structure analysis. This table not only includes the risk factors in the user's text but also annotates their occurrence counts, raw frequency in the graph, and risk contribution values, making the risk assessment intuitive. Furthermore, the system calculates the user's overall risk score, which comprehensively considers the contribution of all risk factors to assess the user's risk level holistically, ensuring the systematic and objective nature of the assessment. Ultimately, this step enables the presentation of user risk profiles in a standardized and quantifiable manner, effectively improving the accuracy of user risk assessments. Through quantitative analysis based on knowledge graphs, the extraction and calculation of user risk factors becomes more systematic, ensuring the interpretability and reliability of risk assessment results. Furthermore, the generation of user risk profiles provides a solid data foundation for subsequent risk warnings, personalized risk management, and decision support, enhancing the system's risk prediction and response capabilities.

[0168] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0169] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A risk profile generation method based on knowledge graph, characterized in that: include: Set the original text dataset to For documents Using the end-of-sentence symbol as a delimiter, split the document d into a set of sentences: in, is the sentence set obtained by segmenting document d according to the symbol; m d is the total number of sentences in document d; For each sentence According to the continuous Chinese character sequence, the candidate word set is obtained as follows: in, is the set of candidate words obtained by segmenting the sentence s according to the continuous Chinese character sequence; N s is the number of candidate words in sentence s, and each t s,j is a string; j is The jth string in For each candidate word t in document d, set: Among them, f1(t) is the number of times the candidate word t appears in the document d; is an indicator function, which takes the value 1 when the condition is met, otherwise it takes the value 0; Simultaneous recording Appears in in, is the index set of the sentence in which the candidate word t appears in the document d; is the index of the sentence in which the candidate word t appears in the document d; Set f0 = 2, construct in, is the node set of the initial graph; f0 is a constant, which is the minimum number of occurrences of candidate words for screening; Each node records the candidate word t and its related information For sentence s, let in, is the candidate word set after frequency screening; For any sentence s, the indicator function is set as: Update edge counts: Among them, c0 is the edge (t i ,t j ) the number of times it co-occurs in all sentences in document d; (t i ,t j ) are any two different candidate words t i and t j The candidate word pairs formed; Set the set ε0, ε0 is all different candidate word pairs (t i ,t j ) and their counts; Set the initial map to:

2. The risk profile generation method based on knowledge graph according to claim 1 is characterized in that: It also includes the extraction and structuring of risk factors within the field, specifically: For each like Then calculate: Where L(t) is the sentence index dispersion of candidate word t in document d; when When , set L(t) = +∞; Set L0=3, where L0 is a constant and represents the discreteness screening threshold; structure in, is the risk factor candidate set; According to the source of document d, document d is classified into a specific field D i , for field D i set up: Among them, D i is the identifier of different fields, i is the field number; For field D i The set of risk candidate nodes within; Set edge collection For connections in ε0 The part of the candidate words in the domain constitutes the domain graph: in, For field D i Inner initial subgraph; For field D i , the node is saved in the form of (t,f1(t),L(t)), and the edge records the co-occurrence count c0(t i ,t j ).

3. The risk profile generation method based on knowledge graph according to claim 2 is characterized in that: It also includes graph standardization and risk feature normalization, specifically: Set the standard node v to: v=(t,F(t),L n (t)); where F(t) is the normalized frequency of candidate word t in the field to which it belongs; L n (t) is the normalized discreteness of candidate word t; For field D i ,calculate Among them, f m,i For field D i The maximum raw frequency of all candidate words in ; For each Set the normalized frequencies to: Eliminate the magnitude differences of data in various fields and normalize F(t)∈[0,1]; Set field D i Inside, And the minimum discrete value of L(t)<+∞ is: set up If L(t)=+∞, then let L n (t) = 0; Will Update to the standard node v=(t,F(t),L n (t)), side information c0(t i ,t j ) remains unchanged.

4. The risk profile generation method based on knowledge graph according to claim 3 is characterized in that: It also includes cross-domain semantic mapping and entity alignment, specifically: For standard nodes v1=(t1,F(t1),L n (t1)) and v2=(t2,F(t2),L n (t2)) Set the matching function to: For all satisfying S match (v1, v2) = 1, let the number of nodes involved in the merger be k, then the unified node v' is set to: v' = (t, F'(v'), L' n (v')); where t is a common candidate word, and: Among them, F'(v') is the normalized frequency of the unified node after merging; L' n (v') is the normalized discreteness of the unified node after merging; v (j) The original standard node participating in the merger; For the merged node pair (v',u'), if there is a corresponding edge (v,y) in the original domain, the edge count after the merge is set to:

5. The risk profile generation method based on knowledge graph according to claim 4 is characterized in that: It also includes cross-domain unified graph construction and unified representation of original frequency and dispersion, specifically: Set the unified node set to Set the unified edge set to And there exists an edge corresponding to the merge}; The cross-domain unified graph is recorded as: For each unified node v' after merging, let the original standard node involved in the merger be v (1) ,v (1) ,...,v (k) ,set up: Among them, f orig (v') is the original frequency of the unified node v'; L orig (v') is the original position discreteness of the unified node v'.

6. The risk profile generation method based on knowledge graph according to claim 5 is characterized in that: It also includes cross-domain graph integration and self-connection structure generation, specifically: The cross-domain unified graph has been constructed as Each Record information (t, F'(v'), L' n (v')) and f orig (v') and L orig (v'); For each Calculate its G * The sum of all edge counts in , specifically: in, is the set of adjacent nodes of v'; For any And v'≠u', if: |f orig (v')-f orig (u')|≤1 and |L orig (v')-L orig (u')|≤1, then set the self-connection edge count c s (v',u')=1.

7. The method for generating risk profiles based on knowledge graph according to claim 6, characterized in that: It also includes updating the final atlas, specifically: Set the self-connected edge set to: The final edge set is: ε final =ε * ∪ε self ; The final cross-domain graph is recorded as:

8. The method for generating risk profiles based on knowledge graph according to claim 7, characterized in that: It also includes user risk profile generation, specifically: User historical text data The sentence set obtained by segmenting using the end-of-sentence symbol is: Among them, K H is the total number of sentences in H; For each Extract candidate words and form For each set up Where h(t) is the number of times the candidate word t appears in the user text H; For s (H) The total number of candidate words; Set the user risk factor set to: v'.t is the candidate word in the user text; For each Set the risk contribution value to: m(v') = h(v'.t) × f orig (v'); where h(v'.t) is the number of occurrences of the candidate word v'.t in the user text; f orig (v') is the original frequency of the node; Form a risk profile table, each row contains: risk factor string v'.t; user occurrence count h(v'.t); original frequency f orig (v'); risk contribution m(v'); Set the user's overall risk score to: Output the above risk profile table and overall risk score R total For risk assessment.

Images