Public text data-based enterprise credit risk assessment method and system

Abstract

The invention provides an enterprise credit risk assessment method and system based on public text data, and the method comprises the steps: obtaining public text data corresponding to a plurality of data sources, carrying out the text preprocessing of the public text data, and generating a structured enterprise text data set; performing enterprise feature extraction according to the preprocessed public text data to obtain a feature set associated with the target enterprise; inputting the feature set into a pre-trained enterprise credit risk assessment model to obtain a credit classification result of the target enterprise; determining a credit risk level of the target enterprise according to a classification label corresponding to the credit classification result; and based on the credit risk level and the classification weight parameter, generating a credit risk assessment parameter of the target enterprise, and outputting a visual assessment report according to the credit risk assessment parameter. According to the invention, the accuracy of enterprise credit risk assessment can be improved.

Description

Technical Field

[0001] The present invention relates to the field of text processing, and in particular to a method and system for enterprise credit risk assessment based on public text data. Background Art

[0002] Enterprise credit risk assessment aims to determine the credit risk level of an enterprise by analyzing its relevant data. Existing assessment methods usually rely on corporate financial statements or structured credit records, and generate risk assessment results based on preset rules or models. However, in existing methods, due to excessive reliance on structured data and neglect of industry dynamics, public opinion correlation and time-series change information in unstructured public texts, feature extraction is one-sided and unable to capture implicit risk signals of corporate credit. At the same time, traditional models lack an adaptive weight allocation mechanism for features of different dimensions, making it difficult to integrate industry rules, semantic correlation and dynamic fluctuation characteristics, resulting in significant deviations in classification results. In addition, the separation of static assessment parameters and visual presentation methods makes it difficult for assessment results to reflect the evolution trend of corporate credit within the time window, and there is a lack of support for secure transmission of multi-source data and real-time alarms, further limiting the accuracy and practical application value of risk assessment. Summary of the Invention

[0003] The present invention provides a method and system for assessing enterprise credit risk based on public text data.

[0004] In a first aspect, an embodiment of the present invention provides a method for assessing corporate credit risk based on public text data, the method comprising: obtaining public text data corresponding to multiple data sources, and performing text preprocessing on the public text data to generate a structured corporate text data set; performing corporate feature extraction based on the preprocessed public text data to obtain a feature set associated with a target enterprise; the feature set comprises feature items based on industry rules, feature items based on semantic associations, and feature items based on time series changes; inputting the feature set into a pre-trained corporate credit risk assessment model to obtain a credit classification result of the target enterprise; the corporate credit risk assessment model comprises multiple sub-assessment models, each sub-assessment model corresponding to classification weight parameters of different dimensions; determining the credit risk level of the target enterprise based on the classification label corresponding to the credit classification result; generating credit risk assessment parameters of the target enterprise based on the credit risk level and the classification weight parameters, and outputting a visual assessment report based on the credit risk assessment parameters.

[0005] In a second aspect, an embodiment of the present invention provides an enterprise credit risk assessment system, comprising: a memory storing a computer program; A processor is used to load the computer program to implement the above-mentioned enterprise credit risk assessment method based on public text data.

[0006] The enterprise credit risk assessment method based on public text data provided by the present invention solves the problems of insufficient coverage and delayed timeliness caused by a single data source by integrating multi-source heterogeneous public text data and generating a structured enterprise text data set; by extracting a multidimensional feature set of industry rule feature items, semantic association feature items and time series change feature items, it comprehensively covers the static attributes and dynamic evolution laws of enterprise credit, thereby breaking through the limitation of one-sided reliance on structured data in traditional methods; using a hierarchical sub-assessment model to assign classification weights to different dimensional features, and generating a comprehensive classification result by dynamically fusing the classification contribution values of each sub-model, effectively adapting to the differences in multidimensional features and improving classification accuracy; dynamically correcting and normalizing the classification weight parameters based on the risk level interval to form a self-calibrated assessment logic to ensure the consistency of assessment results in different industries and risk scenarios; finally, through a visual assessment report, the risk level, feature contribution and time series fluctuation are intuitively presented, significantly improving the interpretability and decision support value of the assessment results, while enhancing data security and response timeliness through encrypted transmission and dynamic alarm mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0008] Figure 1 This is a flowchart of a method for assessing corporate credit risk based on public text data provided by an embodiment of the present invention.

[0009] Figure 2 The figure is a schematic diagram of the composition of an enterprise credit risk assessment system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0010] 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. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0011] See also Figure 1 , Figure 1A flowchart of a method for assessing corporate credit risk based on public text data provided by an embodiment of the present invention. The method for assessing corporate credit risk based on public text data can be executed by an enterprise credit risk assessment system. The method for assessing corporate credit risk based on public text data may include the following steps: Step S100: obtaining common text data corresponding to a plurality of data sources, and performing text preprocessing on the common text data to generate a structured enterprise text data set.

[0012] In the embodiments of the present invention, public text data refers to various types of enterprise-related text information from multiple different data sources, including news media, industry reports, and information released by the enterprise itself. Text preprocessing involves performing a series of operations on the acquired raw public text data to make it easier to process and analyze. Structured enterprise text datasets are datasets obtained by organizing and arranging the text data according to preset structures and rules after preprocessing, facilitating subsequent operations such as feature extraction and model training.

[0013] Specifically, when obtaining public text data corresponding to multiple data sources, web crawler technology can be used to capture relevant corporate text information from major news websites, industry forums, and other data sources. It is understood that the data involved above is obtained from public information in public channels.

[0014] Performing text preprocessing on public text data to generate a structured enterprise text dataset can specifically include the following steps: Step S110: extracting original text data associated with the target enterprise from multiple data sources, and performing format standardization processing on the original text data to generate uniformly coded initial text data.

[0015] Raw text data is unprocessed text information directly obtained from various data sources. Because text formats and encodings may vary across data sources, format standardization is necessary. Format standardization involves converting different text formats (such as HTML, PDF, and TXT) into a unified format and also standardizing the text encoding, such as converting to UTF-8. Initial text data is the text data obtained after format standardization, with a unified encoding and format to facilitate subsequent processing.

[0016] For example, reports about a target company captured from different news websites may be in HTML format or TXT format, and may use different encoding methods. By using specialized text format conversion tools and encoding conversion functions, these text data in different formats and encodings are converted into a unified UTF-8 encoded TXT format to obtain the initial text data.

[0017] Step S120: performing word segmentation processing on the initial text data to obtain a set of text segments after word segmentation.

[0018] Word segmentation is the process of breaking down continuous text into individual words or phrases according to pre-defined rules. A text segment collection is a collection of individual words or phrases obtained after word segmentation. Common word segmentation methods used in Chinese text processing include rule-based, statistical, and deep learning-based methods.

[0019] For example, for the initial text data "The company recently launched a new product", a statistical-based word segmentation method is used to segment it into "the", "company", "recently", "launched", "a", and "new product". These words constitute the set of text fragments after word segmentation.

[0020] Step S130: Filtering target text segments containing preset keywords from the text segment set, and generating semantically associated text blocks based on the contextual semantic relationship of the target text segments.

[0021] Preset keywords are keywords pre-defined based on the needs of corporate credit risk assessments, such as "debt default," "major lawsuit," and "declining market share." Target text segments are text segments containing pre-set keywords, selected from the set of segmented text segments. Contextual semantic relationships refer to the semantic connections between a target text segment and its preceding and following text segments. Semantically associated text blocks are formed by grouping related text segments based on the contextual semantic relationships of the target text segment, facilitating holistic understanding and analysis of the text's semantics.

[0022] For example, the preset keyword is "debt default", and the target text segment containing "debt default" "The company has defaulted on its debt" is filtered out from the text segmentation set. Then, based on the contextual semantic relationship, the previous and subsequent related text segments "recent tight capital chain" and "unable to repay due debts" are combined together to generate a semantically related text block "recent tight capital chain, the company has defaulted on its debt and is unable to repay due debts".

[0023] Step S140: performing label matching on the semantic association blocks according to the industry attributes of the target enterprise, and generating industry labels corresponding to the industry attributes.

[0024] The target enterprise's industry attributes refer to its industry category, such as manufacturing, finance, or services. Label matching compares semantically related blocks with pre-defined industry labels to identify matching industry labels. Industry labels are defined based on the characteristics of different industries and the needs of credit risk assessment, such as "overcapacity" and "lack of technological innovation" in the manufacturing industry, and "high non-performing loan ratio" and "liquidity risk" in the financial industry.

[0025] For example, the target enterprise belongs to the manufacturing industry. For the semantic association block "the enterprise's production equipment is aging and the production efficiency is low", it is matched with the "equipment aging" industry label of the manufacturing industry through label matching to generate an industry label corresponding to the industry attribute.

[0026] Step S150: performing time alignment processing on the text blocks containing the industry tags to generate a structured enterprise text dataset; wherein the structured enterprise text dataset includes a sequence of text blocks arranged in time order, and each text block sequence corresponds to at least one industry tag.

[0027] Time alignment is the process of arranging and organizing text blocks containing industry tags in chronological order. A text block sequence is a chronological sequence of text blocks obtained after time alignment. Each text block sequence corresponds to at least one industry tag. This clearly displays information related to companies and industries at different points in time, facilitating subsequent feature extraction and analysis.

[0028] For example, the text blocks containing industry labels obtained at different time points are arranged in chronological order, such as "In January 2023, the enterprise had overcapacity (overcapacity)" and "In March 2023, the enterprise lacked technological innovation (lack of technological innovation)", etc., to form a sequence of text blocks arranged in chronological order, forming a structured enterprise text dataset.

[0029] Step S200: extracting enterprise features based on the preprocessed public text data to obtain a feature set associated with the target enterprise; the feature set includes feature items based on industry rules, feature items based on semantic associations, and feature items based on temporal changes.

[0030] Enterprise feature extraction is the process of extracting characteristic information that can reflect the credit risk of an enterprise from preprocessed public text data. A feature set is a collection of various extracted feature items. Feature items based on industry rules are extracted based on the rules and characteristics of different industries, such as the depreciation rate and capacity utilization rate of equipment in the manufacturing industry. Feature items based on semantic associations are extracted by analyzing the semantic association relationships of text, such as the semantic association strength and semantic association path between keywords. Feature items based on temporal changes are extracted based on the changes in text data at different time points, such as the change rate of the number of text blocks within a time window and the frequency of keyword repetition.

[0031] For example, for a manufacturing company, the feature item based on industry rules can be the depreciation rate of its production equipment, which is calculated by analyzing information in the text about the equipment purchase time, service life, etc.; the feature item based on semantic association can be the semantic association strength between the two keywords "product quality" and "customer complaints", which is calculated by the semantic association analysis algorithm; the feature item based on time series changes can be the rate of change of the number of text blocks containing the keyword "market share" in the past month, which is obtained by counting the number of text blocks at different time points and calculating the rate of change.

[0032] Enterprise features are extracted based on the pre-processed public text data to obtain a feature set associated with the target enterprise. Specifically, the following steps may be included: Step S210: extracting a text block sequence corresponding to the industry label from the structured enterprise text data set, and calculating the weight value of each keyword under the industry label based on the keyword distribution in the text block sequence.

[0033] The text block sequence corresponding to an industry tag is a chronological sequence of text blocks related to the industry tag, selected from a structured enterprise text dataset. Keyword distribution refers to the position and frequency of occurrence of keywords within the text block sequence. The keyword weight for an industry tag is a numerical value calculated based on the keyword's distribution within the text block sequence corresponding to the industry tag, indicating the keyword's importance within that industry tag.

[0034] For example, for the manufacturing industry label "overcapacity," we extracted text blocks related to "overcapacity" from a structured enterprise text dataset and analyzed the distribution of keywords such as "output," "inventory," and "market demand." We calculated the weight of each keyword within the "overcapacity" industry label by counting the frequency and location of these keywords in the text blocks.

[0035] Based on the keyword distribution in the text block sequence, the weight value of each keyword under the industry tag is calculated. Specifically, the following steps may be included: Step S211: extract all keywords that appear from the text block sequence, and generate initial frequency distribution parameters based on the appearance position and context relevance of each keyword under the industry tag.

[0036] Initial frequency distribution parameters are generated based on the keyword's position within the industry tag and its contextual relevance. Keywords can appear at the beginning, middle, or end of a text block, and contextual relevance refers to the degree of semantic connection between a keyword and surrounding keywords.

[0037] For example, in the text block sequence corresponding to the "overcapacity" industry label, keywords such as "output," "inventory," and "market demand" were extracted. The occurrence location of the keyword "output" in different text blocks and its contextual relevance to surrounding keywords such as "production line" and "order" were analyzed. Based on this information, the initial frequency distribution parameters for the keyword "output" were generated.

[0038] Step S212: Determine the time window distribution data of each keyword under the industry label based on the initial frequency distribution parameters; the time window distribution data includes the frequency change rate of the keyword in different time periods and the frequency difference coefficient of adjacent time windows.

[0039] Time window distribution data describes changes in keyword frequency within different time periods. A time window divides the entire time range into several time periods, each of which is a time window. The frequency change rate refers to the percentage change in the keyword's frequency within different time windows, and the frequency difference coefficient between adjacent time windows refers to the degree of difference in the keyword's frequency between two adjacent time windows.

[0040] For example, the time range is divided into three time windows: January-March 2023, April-June 2023, and July-September 2023. For the keyword "yield," calculate its frequency of occurrence in each time window based on the initial frequency distribution parameters, and calculate the frequency change rate and frequency difference coefficient between adjacent time windows. For example, if the frequency of occurrence in the first time window is 10 times and the frequency of occurrence in the second time window is 15 times, then the frequency change rate is (15-10) / 10=50%. The frequency difference coefficient of adjacent time windows can be obtained using a preset calculation formula. For example, the difference in the frequency of keyword occurrence in adjacent time windows can be directly calculated and then standardized.

[0041] Let the frequency of occurrence of keywords in the i-th time window be , the frequency of occurrence of keywords in the i+1th time window is , then the frequency difference coefficient of adjacent time windows is It can be calculated by the following formula: Among them, the molecule It represents the absolute difference in the frequency of keyword occurrences in adjacent time windows, reflecting the magnitude of the frequency change. It is used to standardize the difference so that the coefficient of difference ranges from 0 to 1. =0, it means that the frequency of keywords in adjacent time windows has not changed; when =1, it means that the keyword’s frequency of occurrence in one time window is 0, while it has a frequency of occurrence in the other time window, that is, the maximum difference occurs.

[0042] Step S213: Based on the time window distribution data, the dynamic stability parameter of each keyword under the industry label is calculated; the dynamic stability parameter is used to characterize the degree of influence of the frequency fluctuation of the keyword on the semantic consistency of the industry label.

[0043] The dynamic stability parameter measures the impact of keyword frequency fluctuations on the semantic consistency of industry tags. Keyword frequency fluctuations can affect the stability of the semantics expressed by industry tags. Large fluctuations in keyword frequency can blur the semantics of industry tags, affecting the assessment of corporate credit risk.

[0044] Based on the time window distribution data, the dynamic stability parameter of each keyword under the industry label is calculated. Specifically, the following steps may be included: Step S2131: Generate an initial fluctuation sequence of the keyword in each time window according to the occurrence frequency change rate and frequency difference coefficient corresponding to each time window in the time window distribution data.

[0045] The initial fluctuation sequence is a sequence composed of the frequency change rate and frequency difference coefficient of the keyword in each time window, which is used to describe the fluctuation of the keyword in different time windows.

[0046] For example, for the keyword "output", the initial fluctuation sequence [50%, 30%, 20%] is generated based on its frequency change rate and frequency difference coefficient in three time windows (assuming that the frequency change rates of the three time windows are 50%, 30%, and 20% respectively).

[0047] Step S2132: assigning a time window weight to each time window according to the distribution position of the time window in the overall time series; wherein the weight of the time window close to the current time point is higher than the weight of the time window far from the current time point.

[0048] A time window weight is assigned to each time window based on its position in the overall time series. It represents the importance of the data within that window in calculating the dynamic stability parameter. Time windows closer to the current time point receive higher weights because the data at that point in time better reflects the current state of the enterprise and has a greater impact on the enterprise's credit risk assessment. For example, for the three time windows described above, a weight of 0.2 is assigned to January-March 2023, 0.3 to April-June 2023, and 0.5 to July-September 2023.

[0049] Step S2133: Based on the initial fluctuation sequence and the time window weight, calculate the fluctuation trend consistency parameter of the keyword between adjacent time windows; the fluctuation trend consistency parameter is used to characterize whether the frequency change direction of adjacent time windows is continuous.

[0050] The Fluctuation Trend Consistency Parameter measures the continuity of frequency changes between adjacent time windows. If the frequency changes between adjacent time windows are continuous, this indicates that the keyword frequency fluctuations are regular and have a relatively small impact on the semantic consistency of industry tags. Conversely, if the frequency changes between adjacent time windows are discontinuous, this indicates that the keyword frequency fluctuations are more random, potentially having a significant impact on the semantic consistency of industry tags.

[0051] For example, based on the initial fluctuation sequence [50%, 30%, 20%] and the time window weights [0.2, 0.3, 0.5], a fluctuation trend consistency parameter between adjacent time windows is calculated using a preset calculation method, such as the symbol matching method, the correlation coefficient method, or the trend slope comparison method. Assuming that the calculated parameter value is high, it indicates that the frequency change direction of the keyword "yield" between adjacent time windows is relatively continuous.

[0052] Step S2134: Perform weighted correction on the initial fluctuation sequence according to the fluctuation trend consistency parameter and the time window weight to generate a corrected fluctuation indicator set.

[0053] Weighted correction refers to the process of adjusting and modifying the initial fluctuation sequence based on the fluctuation trend consistency parameter and time window weight. The modified fluctuation index set is a new fluctuation index sequence obtained after weighted correction, which more accurately reflects the impact of keyword fluctuations on the semantic consistency of industry tags.

[0054] For example, based on the fluctuation trend consistency parameter and the time window weight, the initial fluctuation sequence [50%, 30%, 20%] is weightedly corrected to obtain the corrected fluctuation indicator set [40%, 25%, 15%].

[0055] Step S2135: Generate a dynamic stability parameter based on the degree of discreteness and trend continuity of each indicator in the revised volatility indicator set; wherein, the lower the degree of discreteness and the higher the trend continuity, the larger the value of the dynamic stability parameter.

[0056] The degree of dispersion refers to the degree of difference between the indicators in the revised set of fluctuation indicators, and the trend continuity refers to whether the changing trends of the indicators in the revised set of fluctuation indicators are continuous. The larger the value of the dynamic stability parameter, the smaller the impact of the frequency fluctuation of the keyword on the semantic consistency of the industry label, and the more stable the semantics of the industry label. For example, for the revised set of fluctuation indicators [40%, 25%, 15%], its degree of dispersion and trend continuity are calculated, and the dynamic stability parameter is generated using a preset calculation method, such as calculating the degree of dispersion using the standard deviation method and the coefficient of variation method, calculating the trend continuity using the adjacent difference sign consistency method and the linear fit goodness of fit method, and then obtaining the dynamic stability parameter through comprehensive weighting. If the degree of dispersion is low and the trend continuity is high, the value of the generated dynamic stability parameter will be larger.

[0057] Step S214: extracting keyword distribution data of other industries associated with the industry tag from the structured enterprise text dataset, and determining the industry difference coefficient of each keyword under the current industry tag based on the keyword distribution data of other industries.

[0058] The keyword distribution data for other industries associated with an industry tag refers to the distribution of keywords from other industries related to the current industry tag within a text block sequence, extracted from a structured enterprise text dataset. The industry variance coefficient is a parameter determined based on the keyword distribution data for other industries, indicating the degree of difference between each keyword under the current industry tag and keywords in other industries.

[0059] For example, for the "overcapacity" industry label in the manufacturing industry, we extracted keyword distribution data related to "overcapacity" in other industries, such as the financial industry and the service industry, from a structured enterprise text dataset. We then analyzed the distribution differences between the keyword "output" in the manufacturing industry and other industries, and based on these differences, we determined the industry variation coefficient for "output" under the "overcapacity" industry label.

[0060] Step S215: Based on the dynamic stability parameter and the industry difference coefficient, the initial frequency distribution parameters are weighted and fused to generate a weight value for each keyword under the industry label; wherein the weight value is positively correlated with the dynamic stability parameter and negatively correlated with the industry difference coefficient.

[0061] Weighted fusion refers to the process of weighting and merging the initial frequency distribution parameters based on the dynamic stability parameter and the industry variation coefficient. The weight value is positively correlated with the dynamic stability parameter, indicating that a higher dynamic stability parameter means a higher keyword weight. The weight value is negatively correlated with the industry variation coefficient, indicating that a higher industry variation coefficient means a lower keyword weight.

[0062] For example, for the keyword "output," a weighted fusion calculation method is used to generate its weight under the "overcapacity" industry tag based on its dynamic stability parameter, industry variation coefficient, and initial frequency distribution parameter. Assuming a high dynamic stability parameter and a low industry variation coefficient, the weight of the keyword "output" will be higher.

[0063] Step S220: Sort the keywords according to the weight values, and filter out a target keyword set whose weight values are greater than or equal to a preset threshold.

[0064] The preset threshold is a weighted value limit set in advance based on the needs of corporate credit risk assessment. The target keyword set is a set of keywords selected from all keywords with a weight greater than or equal to the preset threshold. These keywords are highly important under the industry label and can better reflect the company's credit risk.

[0065] For example, the preset threshold is 0.5, and the weight values of all keywords are sorted to filter out keywords with weight values greater than or equal to 0.5, such as "output", "inventory", "market demand", etc., to form the target keyword set.

[0066] Step S230: Based on the target keyword set, a semantic association analysis is performed on the text block sequence to determine the semantic association path of each target keyword in different text blocks.

[0067] Semantic association analysis is the process of analyzing the semantic associations between keywords within a sequence of text blocks. Semantic association paths are paths formed by the semantic associations between a target keyword and other keywords within different text blocks. These paths provide a clear understanding of the semantic evolution and associations of the target keyword within the text.

[0068] Based on the target keyword set, a semantic association analysis is performed on the text block sequence to determine the semantic association path of each target keyword in different text blocks. Specifically, the following steps may be included: Step S231: extract the contextual relationship of each text block containing the target keyword from the text block sequence, and generate a contextual relationship set corresponding to each target keyword; the contextual relationship set includes the semantic type and association direction of the adjacent words of the target keyword in the corresponding text block.

[0069] Contextual relationships refer to the semantic connections between a target keyword and surrounding words within a text block. The contextual relationship set is a collection of contextual relationships for each text block containing the target keyword. These relationships include the semantic type (e.g., noun, verb, adjective) and direction (e.g., positive or negative) of the target keyword's adjacent words within the corresponding text block.

[0070] For example, for the target keyword "output", the text block containing "output" is extracted from the text block sequence, such as "The enterprise has increased its output", and the semantic types and association directions of the adjacent words "increase" and "enterprise" in the text block are analyzed to generate a contextual relationship set corresponding to "output".

[0071] Step S232: Determine the co-occurrence pattern of the target keyword in different text blocks based on the semantic type and association direction in the context relationship set; the co-occurrence pattern is used to describe the regularity of the combined appearance of the target keyword and other keywords in the same text block.

[0072] Co-occurrence patterns refer to the patterns and regularities in which the target keyword and other keywords appear together in the same text block. By analyzing the semantic types and association directions in the contextual relationship set, we can identify co-occurrence patterns between the target keyword and other keywords, such as which keywords frequently appear with the target keyword and the direction of their semantic association.

[0073] For example, for the target keyword "output", by analyzing the context relationship set, it is found that "output" often co-occurs with keywords such as "increase", "reduce", and "market demand", and is positively correlated with "increase" and negatively correlated with "reduce", thereby determining the co-occurrence pattern of "output" in different text blocks.

[0074] Step S233: Based on the co-occurrence pattern, the position distribution and semantic jump relationship of the target keyword in different text blocks are analyzed to generate an intermediate path set of the target keyword in the text block sequence.

[0075] Position distribution refers to the occurrence of target keywords in different text blocks, while semantic jump relationships refer to the changes in the semantic associations between target keywords in different text blocks. The intermediate path set is a set of paths generated based on the co-occurrence pattern, position distribution, and semantic jump relationships of the target keywords. These paths reflect the semantic association process of the target keywords in the text block sequence.

[0076] For example, for the target keyword "yield," we analyze its co-occurrence pattern across different text blocks, such as its position at the beginning of some blocks and its position in the middle of others. We also analyze its semantic jumps between different text blocks, such as its association with "market demand" and its association with "inventory." Based on this information, we generate a set of intermediate paths for "yield" within the text block sequence.

[0077] Step S234: performing conflict detection on the intermediate paths according to the semantic jump continuity of each intermediate path in the intermediate path set, filtering intermediate paths with semantic conflicts or inconsistent jump directions, and obtaining a candidate path set.

[0078] Semantic jump continuity refers to the coherence and rationality of semantic jumps within intermediate paths. Conflict detection involves checking each intermediate path in the intermediate path set to determine whether there are semantic conflicts or inconsistent jump directions. The candidate path set is the set of paths obtained after conflict detection and filtering out problematic intermediate paths. These paths exhibit good semantic jump continuity.

[0079] For example, for the set of intermediate paths of "output", the semantic jump continuity of each intermediate path is checked. For example, if some paths suddenly jump from "output increase" to "output drop sharply" without a reasonable transition, there is a semantic conflict in such paths. They are filtered out to obtain the candidate path set.

[0080] Step S235: Based on the context coverage and semantic density parameters of each candidate path in the candidate path set, determine the semantic association path of the target keyword in different text blocks; wherein the semantic association path is composed of the candidate paths with the widest coverage and the highest semantic density parameters, and each semantic association path contains the semantic evolution link of the target keyword in the text block sequence.

[0081] Context coverage refers to the size of the text blocks covered by a candidate path, while semantic density refers to the closeness of the semantic associations between keywords within a candidate path. Semantic association paths are composed of candidate paths selected from a set of candidate paths with the widest coverage and highest semantic density. These paths most accurately reflect the semantic associations of target keywords across different text blocks and encompass the semantic evolution of the target keyword within the text block sequence, specifically, how the semantics of the target keyword evolves as the text blocks change.

[0082] For example, for the set of candidate paths for "output", the context coverage and semantic density parameters of each candidate path are calculated, and the candidate path with the widest coverage and the highest semantic density parameters is selected, such as the path "output-market demand-inventory", as the semantic association path of "output" in different text blocks. This path contains the semantic evolution link of "output" from being associated with "market demand" to being associated with "inventory".

[0083] Step S240: constructing industry rule feature items of the target enterprise according to the semantic association path. The industry rule feature items include the weight value of each keyword in the target keyword set and the semantic density parameter of its association path.

[0084] Industry rule feature items are constructed based on semantic association paths to reflect the credit risk of target enterprises. The weight of each keyword in the target keyword set reflects its importance within the industry label, while the semantic density parameter of the association path reflects the closeness of the semantic association between keywords. By combining keyword weights and semantic density parameters of association paths, industry rule feature items can be constructed that more accurately reflect the credit risk of enterprises.

[0085] For example, for the target keywords "output", "inventory" and "market demand", based on their semantic association path "output-market demand-inventory", combined with the weight value of each keyword and the semantic density parameter of the association path, the industry rule feature items of the target enterprise are constructed, such as "output (weight value 0.6, association path semantic density parameter 0.8) - market demand (weight value 0.5, association path semantic density parameter 0.7) - inventory (weight value 0.4, association path semantic density parameter 0.6)".

[0086] Step S250: Extract the temporal change data associated with the target enterprise from the text block sequence, and generate feature items based on temporal change according to the distribution differences of the temporal change data in different time periods; wherein the feature items of temporal change include the rate of change of the number of text blocks within the time window, the frequency of keyword repetition, and the semantic conflict detection parameters.

[0087] Time-series variation data refers to data related to a sequence of text blocks at different points in time, such as the number of text blocks and the frequency of keyword occurrence. Time-series variation-based feature items are generated based on the distribution differences of time-series variation data across different time periods and are used to reflect the evolution of a company's credit risk over time. The rate of change in the number of text blocks within a time window reflects the change in the number of text blocks within the preset time window. The keyword repetition frequency reflects the change in the frequency of keyword occurrence at different points in time. The semantic conflict detection parameter is used to detect semantic conflicts within text at different points in time.

[0088] For example, extract temporal variation data from a text block sequence, such as the number of text blocks per month and the frequency of occurrence of the keyword "debt default." Calculate the monthly rate of change in the number of text blocks. For example, if the number of text blocks increases from 10 to 15 from January to February, the rate of change is (15-10) / 10 = 50%. Count the frequency of repetition of the keyword "debt default" in different months, for example, 2 times in January and 5 times in February. Simultaneously, use a semantic analysis algorithm to detect semantic conflicts between texts from different months and generate semantic conflict detection parameters. Based on this data, generate feature items based on temporal variation, such as "rate of change in the number of text blocks within the time window (50%), keyword repetition frequency ('debt default': 2 times in January, 5 times in February), and semantic conflict detection parameter (low)."

[0089] Step S300: Input the feature set into a pre-trained enterprise credit risk assessment model to obtain the credit classification result of the target enterprise; the enterprise credit risk assessment model includes multiple sub-assessment models, each sub-assessment model corresponds to a classification weight parameter of a different dimension.

[0090] The feature set is a collection of feature items based on industry rules, feature items based on semantic associations, and feature items based on time series changes, obtained through enterprise feature extraction. The pre-trained enterprise credit risk assessment model is a model used to assess enterprise credit risk, obtained by training with historical data. The credit classification result is the result obtained after the model classifies the credit risk of the target enterprise, such as high risk, medium risk, low risk, etc. The enterprise credit risk assessment model includes multiple sub-assessment models, each of which is responsible for assessing the enterprise credit risk from different dimensions and corresponds to classification weight parameters for different dimensions. These classification weight parameters are used to adjust the importance of each sub-assessment model in the final assessment result.

[0091] For example, a corporate credit risk assessment model consists of three sub-assessment models: an industry rule-based sub-assessment model, a semantic association-based sub-assessment model, and a temporal change-based sub-assessment model. The industry rule-based feature items in the feature set are input into the industry rule-based sub-assessment model, the semantic association-based feature items are input into the semantic association-based sub-assessment model, and the temporal change-based feature items are input into the temporal change-based sub-assessment model. Each sub-assessment model evaluates the input feature items according to its corresponding classification weight parameters, ultimately obtaining a credit classification result for the target enterprise.

[0092] The enterprise credit risk assessment model utilizes a multi-assessment model fusion architecture to comprehensively assess enterprise credit risk from various dimensions. The model primarily consists of multiple sub-assessment models, including those based on industry rules, semantic associations, and temporal changes. Each sub-assessment model processes feature data from a specific dimension and performs calculations based on corresponding classification weight parameters. Ultimately, the outputs of each sub-assessment model are weighted and fused to produce a comprehensive credit classification result.

[0093] For example, a sub-assessment model based on industry rules consists of a feature input layer, a rule matching layer, a weight calculation layer, and an output layer. The feature input layer receives industry rule-related feature items extracted from a structured enterprise text dataset. For example, in the manufacturing industry, these might include industry rule-based feature items such as equipment depreciation rate, capacity utilization rate, and product qualification rate. These feature items exist as text information in the structured enterprise text dataset and are first quantified. The rule matching layer stores a series of rules and standards set for different industries. These rules and standards are developed based on industry characteristics, historical data, and expert experience. The weight calculation layer assigns a weight to each feature item based on the matching results of the rule matching layer. This weighting takes into account the importance of the feature item within the industry rules and its impact on the enterprise's credit risk. The output layer performs a weighted calculation on the input feature items based on the feature item weights obtained by the weight calculation layer. Each feature item value is multiplied by its corresponding weight, and these products are summed to produce the output of the industry rule-based sub-assessment model. This calculation results in a specific numerical value that reflects the enterprise's credit risk assessment under the industry rule dimension.

[0094] The sub - evaluation model based on semantic association consists of a text input layer, a semantic analysis layer, an association path construction layer, and an output layer. In the text input layer, text data related to semantic association extracted from the structured enterprise text dataset is received. This text data contains various descriptions and information about the target enterprise. After receiving the text data, the text is first cleaned to remove noise information such as punctuation marks and stop words. For Chinese text, stop words may include words without actual semantic meaning such as "的", "了", "和", etc. These noise information are identified and removed through methods such as regular expression matching to make the text more concise and facilitate subsequent processing. The semantic analysis layer conducts in - depth semantic analysis on the input text data, including lexical semantic understanding, sentence semantic parsing, and context semantic association analysis, etc. In terms of lexical semantic understanding, a pre - trained word vector model such as Word2Vec or GloVe is used to convert each word in the text into a vector representation. These word vectors can capture the semantic information of words, and words with similar semantics are closer in the vector space. By querying the word vector model, the words in the text are converted into corresponding vectors, providing a basis for subsequent semantic analysis. In terms of sentence semantic parsing, dependency parsing technology is adopted to analyze the grammatical relationships between various words in the sentence, such as subject - predicate - object relationships, attributive - adverbial - complement relationships, etc. Through dependency parsing, the structure and semantics of the sentence can be understood, and the associations between words can be clarified. In terms of context semantic association analysis, the context information of words in the text is considered, and through window mechanisms or attention mechanisms, the semantic association strength between words and surrounding words is calculated. The association path construction layer will construct the semantic association path between target keywords based on the results of the semantic analysis layer. This layer analyzes the co - occurrence patterns, position distributions, and semantic jump relationships of keywords in different text blocks, filters out paths with semantic conflicts or inconsistent jump directions, and obtains a set of candidate paths with good semantic continuity. When analyzing co - occurrence patterns, the frequency of keywords appearing simultaneously in the same text block is counted. Higher co - occurrence frequencies may indicate stronger semantic associations between keywords. When analyzing position distributions, the positions of keywords in the text block are observed, such as at the beginning, in the middle, or at the end, and the distances between them. If two keywords often appear adjacent to each other in the text, it may indicate that their semantic associations are closer. When analyzing semantic jump relationships, it is judged whether the semantic changes between keywords in different text blocks are reasonable, such as whether there are sudden semantic turns or contradictions. For paths with semantic conflicts or inconsistent jump directions, they are filtered by setting rules or thresholds. For example, if the semantic associations between two keywords show obvious contradictions in a certain path, such as one indicating profit and the other indicating loss, and there is no reasonable explanation, then this path is filtered out. Then, the path with the widest coverage and the highest semantic density parameter is selected from the set of candidate paths as the semantic association path of the target keywords.A wider coverage indicates that the path involves more textual information, while a higher semantic density indicates a closer semantic connection between keywords. The output layer calculates the output of the semantic connection-based sub-assessment model based on the semantic connection paths and related semantic features. Each semantic connection path can be assigned a weight, determined by factors such as the path's coverage and semantic density. The relevant semantic features of each path are then quantified, such as the semantic similarity and strength of the connection between keywords. The quantified semantic features are multiplied by the corresponding path weight, and these products are summed to produce the output of the semantic connection-based sub-assessment model. This output reflects the credit risk assessment of the enterprise based on the semantic connection dimension. For example, a close connection between keywords may indicate a potential risk or advantage for the enterprise.

[0095] The sub-evaluation model based on temporal change consists of a time series data input layer, a change feature extraction layer, a trend analysis layer, and an output layer. The time series data input layer receives data related to temporal change extracted from structured enterprise text datasets, such as the number of text blocks and keyword frequency at different time points. After receiving the data, it undergoes preprocessing, including missing value handling and data normalization. For missing values, if the number of missing values is small, statistics such as the mean and median can be used to fill them. If the number of missing values is large, more complex processing can be performed in conjunction with business logic, such as interpolation estimation based on data from adjacent time points. Data normalization can use the min-max normalization method to scale the data to the [0, 1] range to make data with different features comparable. The change feature extraction layer processes the input time series data and extracts features that reflect temporal change, such as the rate of change in the number of text blocks within a time window, keyword repetition frequency, and semantic conflict detection parameters. When calculating the rate of change in the number of text blocks within a time window, a fixed time window, such as one month or one quarter, is set. The number of text blocks within each time window is counted, and the rate of change in the number of text blocks between adjacent time windows is calculated. The number of text blocks in the previous time window is subtracted from the number of text blocks in the next time window, and then divided by the number of text blocks in the previous time window. For keyword repetition frequency, the number of keyword occurrences within each time window is counted, and the change in keyword repetition frequency between different time windows is calculated. When detecting semantic conflict, a semantic comparison is performed on text within adjacent time windows to determine whether there is a semantic contradiction or conflict. Semantic similarity calculation methods, such as cosine similarity, can be used to calculate the semantic similarity of text within adjacent time windows. If the similarity is low and there is a clear semantic contradiction, a semantic conflict is considered to exist. The trend analysis layer performs trend analysis on the feature items obtained by the change feature extraction layer to determine the changing trend of the enterprise's credit risk over time. This layer uses time series analysis methods, such as moving average and exponential smoothing, to fit and predict historical data on the feature items and analyze their changing trends and patterns. The moving average method calculates the average value of the data within a time window. By continuously moving the time window, a series of moving averages are generated, thereby smoothing data fluctuations and observing data trends. Exponential smoothing assigns different weights to historical data, with recent data receiving a higher weight and later data receiving a lower weight. This weighted average is then used to predict future data. These methods can be used to determine whether keyword repetition frequency is trending upward or downward, and the impact of this trend on a company's credit risk. For example, a sustained increase in keyword repetition frequency may indicate a potential risk event for the company. Based on the results of the trend analysis layer, the output layer calculates the output of the sub-assessment model based on time series changes.Each trend feature can be assigned a weight, determined by its impact on the company's credit risk. The quantified value of the trend feature is then multiplied by the corresponding weight, and these products are summed to produce the output of the sub-assessment model based on time-series variation. This output reflects the company's credit risk assessment under the time-series variation dimension.

[0096] When the outputs of the sub-assessment models are combined, the weight allocation ratio for each sub-assessment model in the final assessment is determined based on the weight allocation rules corresponding to different risk level intervals. These weight allocation ratios are positively correlated with the sub-assessment model's classification contribution value within the risk level interval; that is, the greater the classification contribution value, the higher the corresponding weight allocation ratio for the sub-assessment model. The weight allocation rules are developed based on extensive historical data and experimental results. By analyzing the performance of each sub-assessment model at different risk levels, their importance in different risk level intervals is determined. The classification weight parameters of each sub-assessment model are then normalized and fused to generate a composite weight for each sub-assessment model. This normalization ensures that the sum of the weights of each sub-assessment model is 1, making the final assessment more reasonable. A simple normalization method can be used to divide the weight of each sub-assessment model by the sum of the weights of all sub-assessment models to obtain the normalized weights. The composite weights represent the relative importance of feature items from different dimensions in the final risk assessment. Next, core feature items corresponding to the risk level range are extracted from the feature set. Core feature items are those ranked in the top digits by contribution based on the comprehensive weight. The preset digits are determined based on actual needs and the required accuracy of the assessment; for example, they can be set to the top 3 or 5 feature items. Core feature items are selected by calculating the contribution of each feature item based on the comprehensive weight. Based on the distribution of core feature items in historical credit labels, a stability offset is calculated for the core feature items in the current target enterprise data. The stability offset reflects the difference between the volatility of the core feature items within the time window and the historical benchmark. The stability offset can be calculated by calculating the difference between the statistics (such as mean and variance) of the core feature items in the current data and those in the historical data.

[0097] Based on the stability offset and the comprehensive weight value, the contribution of the core feature items is calibrated to generate a calibrated feature contribution set. The calibration process is to make the contribution of the core feature items more accurately reflect their impact on the corporate credit risk. The original contribution of the core feature items can be adjusted according to the stability offset. For example, when the stability offset is large, the contribution of the feature item can be appropriately reduced; when the stability offset is small, the contribution of the feature item can be appropriately increased. Finally, based on the comparison between the calibration results of each feature item in the calibrated feature contribution set and the preset contribution threshold, credit risk assessment parameters are generated, including risk probability values, stability indicators, and abnormal fluctuation warning parameters. These parameters can comprehensively and accurately reflect the credit risk status of the target enterprise.

[0098] As an implementation method, the feature set is input into a pre-trained enterprise credit risk assessment model to obtain a credit classification result of the target enterprise, which may specifically include the following steps: Step S310: inputting the industry rule feature item, the semantic association feature item and the time series change feature item into the corresponding sub-assessment models in the enterprise credit risk assessment model respectively.

[0099] Industry rule features are extracted based on industry rules and characteristics to reflect a company's credit risk. Semantic association features are extracted by analyzing the semantic associations within text. Time series change features are extracted based on how text data changes over time. Each sub-assessment model within the enterprise credit risk assessment model is responsible for assessing a company's credit risk from different dimensions, such as the industry rule-based sub-assessment model, the semantic association-based sub-assessment model, and the time series change-based sub-assessment model.

[0100] For example, the target enterprise's industry rule feature items "output (weight value 0.6, association path semantic density parameter 0.8) - market demand (weight value 0.5, association path semantic density parameter 0.7) - inventory (weight value 0.4, association path semantic density parameter 0.6)" are input into the sub-evaluation model based on industry rules; semantic association feature items, such as the semantic association path of "output" and "market demand", are input into the sub-evaluation model based on semantic association; and temporal change feature items "the rate of change of the number of text blocks within the time window (50%), the frequency of keyword repetition ('debt default': 2 times in January, 5 times in February), and the semantic conflict detection parameter (low)" are input into the sub-evaluation model based on temporal changes.

[0101] Step S320: weight each input feature item through each sub-evaluation model to generate a classification contribution value corresponding to each feature item.

[0102] Weight allocation refers to the process by which each sub-assessment model performs a weighted calculation on each input feature item based on its corresponding classification weight parameter. The classification contribution value is the numerical value of the contribution of each feature item to the final credit classification result after weight allocation. For example, the sub-assessment model based on industry rules performs a weighted calculation on the weight value of each keyword in the input industry rule feature item and the semantic density parameter of the associated path based on its corresponding classification weight parameter to generate the corresponding classification contribution value for each feature item. Assuming that the classification weight parameter of the "yield" feature item is 0.3, its weight value is 0.6, and the semantic density parameter of the associated path is 0.8, the classification contribution value of the "yield" feature item is 0.3×0.6×0.8=0.144.

[0103] Step S330: Determine the output results of each sub-evaluation model based on the classification contribution value, and perform weighted fusion on the output results to generate a comprehensive classification score.

[0104] The output of each sub-assessment model is a comprehensive calculation of the categorical contribution values of the input feature items, resulting in an assessment of the enterprise's credit risk. Weighted fusion involves weighting and fusing the outputs of each sub-assessment model based on their importance in the final assessment. The comprehensive classification score, derived from this weighted fusion, is used to comprehensively assess the enterprise's credit risk.

[0105] For example, the output of the sub-assessment model based on industry rules is 0.8, the output of the sub-assessment model based on semantic association is 0.7, and the output of the sub-assessment model based on temporal changes is 0.6. Assuming the weights of the three sub-assessment models are 0.4, 0.3, and 0.3, respectively, the comprehensive classification score = 0.8 × 0.4 + 0.7 × 0.3 + 0.6 × 0.3 = 0.71.

[0106] Step S340: Determine a classification label corresponding to the credit classification result based on the comparison result of the comprehensive classification score and the preset classification threshold; wherein the classification label includes multiple risk level intervals, and each risk level interval corresponds to a different combination of classification weight parameters.

[0107] Pre-set classification thresholds are the scoring boundaries used to categorize different risk levels, based on the needs of corporate credit risk assessment. Classification labels, such as high risk, medium risk, and low risk, are determined based on the comparison of the comprehensive classification score with the pre-set classification thresholds. Each risk level corresponds to a different combination of classification weight parameters, which are used to adjust the importance of each sub-assessment model at different risk levels.

[0108] For example, the preset classification thresholds are 0.6 and 0.8. When the comprehensive classification score is less than 0.6, the corresponding classification label is low risk; when the comprehensive classification score is between 0.6 and 0.8, the corresponding classification label is medium risk; and when the comprehensive classification score is greater than 0.8, the corresponding classification label is high risk. Different risk level intervals correspond to different classification weight parameter combinations. For example, in the low risk level interval, the weight of the sub-assessment model based on industry rules may be 0.3, the weight of the sub-assessment model based on semantic association may be 0.3, and the weight of the sub-assessment model based on time series changes may be 0.4.

[0109] The training process of the enterprise credit risk assessment model provided by the embodiment of the present invention includes the following steps: Step S10: Obtain a historical enterprise text data set and corresponding historical credit labels, and perform feature extraction processing on the historical enterprise text data set to generate a historical feature set.

[0110] A historical enterprise text dataset contains historical text data related to multiple enterprises. This text data can come from news reports, corporate financial statements, and other sources. Historical credit labels are labels for the credit risk ratings of enterprises corresponding to the historical enterprise text dataset, such as high risk, medium risk, or low risk. Feature extraction is performed on the historical enterprise text dataset, similar to step S200, to extract feature information that reflects the credit risk of the enterprises. A historical feature set is a collection of feature items from multiple enterprises, obtained after feature extraction.

[0111] Step S20: Divide the historical feature set into a training subset and a verification subset based on the correlation between each feature item type in the historical feature set and the historical credit label; wherein the training subset contains feature item data whose correlation with the historical credit label is greater than a first correlation threshold, and the verification subset contains feature item data whose correlation with the historical credit label is less than a second correlation threshold.

[0112] Correlation refers to the degree of correlation between each feature type in the historical feature set and the historical credit label. The training subset is a dataset used to train the enterprise credit risk assessment model. It contains feature data with a correlation with the historical credit label greater than a first correlation threshold. This data better reflects the relationship between the enterprise's credit risk and the feature items. The validation subset is a dataset used to verify the performance of the trained model. It contains feature data with a correlation with the historical credit label less than a second correlation threshold. The validation subset can be used to test the model's generalization ability across different data sets. The first correlation threshold is greater than or equal to the second correlation threshold, and the specific values of both can be set based on actual needs or experience. For example, the correlation between each feature type in the historical feature set and the historical credit label is calculated using a correlation analysis algorithm. Feature data with high correlation, such as key feature items based on industry rules and semantically related feature items closely related to credit risk, is allocated to the training subset. Feature data with low correlation, such as minor temporal variation feature items, is allocated to the validation subset.

[0113] Step S30: For each sub-evaluation model, extract feature item data of the dimension corresponding to the sub-evaluation model from the training subset, and dynamically match the feature item data with the historical credit label to generate a sub-model training sample set.

[0114] The feature data for the corresponding dimension refers to the feature data related to the assessment dimension for each sub-assessment model. For example, a sub-assessment model based on industry rules corresponds to feature data based on industry rules, a sub-assessment model based on semantic association corresponds to feature data based on semantic association, and a sub-assessment model based on time series changes corresponds to feature data based on time series changes. Dynamic matching refers to the process of associating and matching the extracted feature data with the corresponding historical credit labels. The sub-model training sample set is the sample set obtained after dynamic matching and used to train each sub-assessment model. For example, for a sub-assessment model based on industry rules, feature data based on industry rules, such as "output (weight 0.6, association path semantic density parameter 0.8) - market demand (weight 0.5, association path semantic density parameter 0.7) - inventory (weight 0.4, association path semantic density parameter 0.6)" are extracted from the training subset. These feature data are then matched with the corresponding historical credit labels to generate the training sample set for the sub-model based on industry rules.

[0115] Step S40: The sub-model training sample set is classified for the first time using the initialized classification weight parameters, and the classification weight parameters are iteratively adjusted layer by layer based on the error distribution between the classification results and the historical credit labels.

[0116] Initialized classification weight parameters are pre-set for each sub-evaluation model at the start of model training. First classification refers to the process of classifying the sub-model training sample set using the initialized classification weight parameters. Error distribution refers to the distribution of differences between classification results and historical credit labels. By analyzing the error distribution, we can understand which samples the model inaccurately classifies. Layer-by-layer iterative adjustment refers to the process of repeatedly adjusting and optimizing the classification weight parameters based on the error distribution until the model's classification accuracy meets the preset requirements.

[0117] The sub-model training sample set is first classified using the initialized classification weight parameters. Based on the error distribution between the classification results and the historical credit labels, the classification weight parameters are iteratively adjusted layer by layer. Specifically, the following steps may be involved: Step S41: Determine the initial error offset for initializing the classification weight parameters according to the feature item data distribution in the sub-model training sample set.

[0118] The initial error offset is a value used to measure the initial error between the initial classification weight parameters and the actual optimal weight parameters, determined based on the distribution of feature item data in the sub-model training sample set. The feature item data distribution includes information such as the value range and distribution frequency of the feature item.

[0119] For example, by statistically analyzing the value range and distribution frequency of characteristic items based on industry rules in the sub-model training sample set, the distribution differences of these characteristic items under different credit risk levels are analyzed, and the initial error offset of the initial classification weight parameters is determined based on these differences.

[0120] Step S42: performing a first classification on the sub-model training sample set based on the initial error offset to generate an initial classification result set.

[0121] The first classification is the process of classifying the sub-model training sample set using the classification weight parameters that take into account the initial error offset. The initial classification result set is the classification result set of the sub-model training sample set obtained after the first classification.

[0122] For example, the initialized classification weight parameters are adjusted according to the initial error offset, and then the adjusted classification weight parameters are used to classify the sub-model training sample set to obtain an initial classification result set, such as classifying certain samples as high risk, medium risk, or low risk.

[0123] Step S43: Calculate the error between the classification result of each sample in the initial classification result set and the corresponding historical credit label, and prioritize the sub-model training sample sets according to the error value.

[0124] The error value refers to the degree of difference between the classification result and the corresponding historical credit label for each sample in the initial classification result set. Prioritization sorting involves sorting the sub-model training sample set based on the error value. Samples with larger error values are given higher priority, allowing for subsequent focus on these inaccurately classified samples. For example, for a sample in the sub-model training sample set, the initial classification result is high risk, while the corresponding historical credit label is medium risk. The error value for this sample is calculated using an error calculation method (such as the cross-entropy loss function). After calculating the error values for all samples, the sub-model training sample set is prioritized in descending order of error value to obtain the priority sorting result.

[0125] Step S44: Based on the priority sorting result, samples with error values greater than a preset error threshold are screened from the sub-model training sample set to generate a high-error sample subset.

[0126] The preset error threshold is a pre-set error limit based on model training requirements. The high-error sample subset is a collection of samples selected from the sub-model training sample set whose error values exceed the preset error threshold. These samples are the primary source of model classification inaccuracies and require focused processing and analysis. For example, if the preset error threshold is 0.3, based on the priority sorting results, samples with an error value greater than 0.3, such as those whose classification results significantly differ from historical credit labels, are selected to form the high-error sample subset.

[0127] Step S45: Based on the feature item data distribution of the high-error sample subset, reversely adjust the classification weight parameter to generate a first adjustment parameter.

[0128] Backward adjustment involves adjusting the current classification weight parameters based on the data distribution of the feature items in the high-error sample subset to reduce the error in these samples. The first adjustment parameter is the parameter used to update the classification weight parameters after the reverse adjustment. For example, analyzing the data distribution of the feature items in the high-error sample subset may reveal significant differences between the values of certain feature items in these samples and those in the overall sample set. Based on these differences, the classification weight parameters are reversely adjusted, such as increasing the weight of some feature items and decreasing the weight of others, to generate the first adjustment parameter.

[0129] Step S46: Perform secondary classification on the sub-model training sample set using the first adjustment parameter, and repeat the high-error sample screening and parameter adjustment steps based on the secondary classification results and the error distribution of historical credit labels until the error value is less than or equal to the preset error threshold.

[0130] Secondary classification refers to the process of reclassifying the sub-model training sample set after the classification weight parameters are updated using the first adjustment parameters. Based on the error distribution between the secondary classification results and the historical credit labels, the high-error sample subset is screened out again, and the classification weight parameters are further adjusted. This process is repeated until the error value is less than or equal to the preset error threshold, indicating that the classification accuracy of the model has met the preset requirements. For example, after the classification weight parameters are updated using the first adjustment parameters, the sub-model training sample set is subjected to secondary classification. The error distribution between the secondary classification results and the historical credit labels is calculated. If there are still samples with error values greater than the preset error threshold, the high-error sample subset is screened out again, the feature item data distribution is re-analyzed, the classification weight parameters are adjusted, new adjustment parameters are generated, and then the next classification is performed until the error value meets the requirements.

[0131] Step S47: After completing the layer-by-layer iterative adjustment, the finally adjusted classification weight parameters are locked as the intermediate parameter configuration of the current sub-evaluation model.

[0132] When, after multiple iterative adjustments, the error value is less than or equal to the preset error threshold, it indicates that the classification weight parameters have been adjusted to a relatively appropriate state. At this point, the final adjusted classification weight parameters are locked as the intermediate parameter configuration of the current sub-evaluation model for subsequent model verification and optimization. For example, after multiple iterative adjustments, the classification weight parameters of the sub-evaluation model based on industry rules are adjusted to an appropriate value. These final adjusted classification weight parameters are locked as the intermediate parameter configuration of the sub-evaluation model based on industry rules.

[0133] Step S50: Classify and verify the adjusted sub-evaluation model and the feature item data in the verification subset, and optimize the classification weight parameters across models based on the verification results until the classification accuracy of all sub-evaluation models reaches the preset threshold simultaneously.

[0134] Classification verification involves using the adjusted sub-assessment model to classify the feature data in the validation subset and comparing the classification results with the actual credit labels corresponding to the validation subset to verify the model's performance on different data. Cross-model optimization involves comprehensively adjusting and optimizing the classification weight parameters of each sub-assessment model based on the verification results, so that the classification accuracy of all sub-assessment models reaches the preset threshold.

[0135] The adjusted sub-evaluation model is classified and verified with the feature item data in the validation subset, and the classification weight parameters are cross-model optimized based on the verification results. Specifically, the following steps may be included: Step S51: input the feature item data in the verification subset into the adjusted sub-evaluation model according to the corresponding dimensions to generate a verification classification result set.

[0136] Corresponding dimensions refer to the dimensions of feature data associated with the assessment dimension for each sub-assessment model. For example, a sub-assessment model based on industry rules corresponds to feature data based on industry rules, a sub-assessment model based on semantic associations corresponds to feature data based on semantic associations, and a sub-assessment model based on time series changes corresponds to feature data based on time series changes. The validation classification result set is the result set obtained by classifying the feature data in the validation subset using the adjusted sub-assessment model.

[0137] For example, the feature item data based on industry rules in the verification subset is input into the adjusted sub-evaluation model based on industry rules, the feature item data based on semantic association is input into the adjusted sub-evaluation model based on semantic association, and the feature item data based on time series changes is input into the adjusted sub-evaluation model based on time series changes to obtain the verification classification result set.

[0138] Step S52: Compare the verification classification result set with the true credit labels corresponding to the verification subset, and calculate the current accuracy of each sub-evaluation model.

[0139] The current accuracy rate refers to the classification accuracy of each sub-evaluation model on the validation subset. By comparing the validation classification result set with the true credit labels corresponding to the validation subset, and counting the proportion of correctly classified samples to the total number of samples, the current accuracy rate of each sub-evaluation model is obtained.

[0140] For example, for a sub-evaluation model based on industry rules, there are 100 samples in the validation subset, of which 80 are correctly classified. The current accuracy of the sub-evaluation model is 80%.

[0141] Step S53: If the current accuracy rate does not reach the preset threshold, supplementary sample data with similar distribution characteristics to the feature item data of the validation subset is extracted from the training subset.

[0142] The preset threshold is a pre-set limit on classification accuracy based on model performance requirements. Supplementary sample data is extracted from the training subset and has similar distribution characteristics to the validation subset feature data. This sample data can be used to further train the sub-evaluation model and improve its classification accuracy on the validation subset.

[0143] For example, the preset threshold is 90%. If the current accuracy of the sub-evaluation model based on industry rules is 80%, which does not reach the preset threshold, sample data with similar distribution to the feature item data based on industry rules in the verification subset is extracted from the training subset, such as sample data with similar value ranges and distribution frequencies of certain feature items, as supplementary sample data.

[0144] Step S54: merging the supplementary sample data with the sub-model training sample set to generate an extended training set, and performing supplementary iterative adjustment on the classification weight parameters based on the extended training set.

[0145] The expanded training set is a new training dataset created by merging the supplementary sample data with the sub-model training sample set. This dataset contains more samples with similar feature data to the validation subset, enabling more comprehensive training of the sub-evaluation model. Supplementary iterative adjustment involves further iterative adjustments to the classification weight parameters based on the expanded training set to improve the classification accuracy of the sub-evaluation model.

[0146] For example, the extracted supplementary sample data is combined with the training sample set of the sub-model based on industry rules to generate an extended training set. The classification weight parameters of the sub-evaluation model based on industry rules are iteratively adjusted using the extended training set, and the process of steps S41-S47 is repeated until the classification accuracy reaches a preset threshold.

[0147] Step S55: If the current accuracy reaches the preset threshold, the classification weight parameters are cross-model balanced according to the distribution ratio of different risk levels in the verification classification result set; wherein, the cross-model balanced processing includes proportional calibration of the weight parameters of different sub-evaluation models to ensure that each sub-evaluation model has a consistent risk level coverage in the fused output results.

[0148] Cross-model balancing is used to ensure that each sub-assessment model more evenly reflects the different risk levels in the final fusion output. When a sub-assessment model has high classification accuracy for one risk level but low accuracy for other risk levels, cross-model balancing is used to proportionally calibrate the weight parameters of the different sub-assessment models, ensuring more consistent coverage of the different risk levels in the fused output.

[0149] For example, the sub-assessment model based on industry rules has a higher classification accuracy for high-risk levels, but a lower classification accuracy for low-risk levels; the sub-assessment model based on semantic associations has a higher classification accuracy for medium-risk levels. Based on the distribution ratio of different risk levels in the validation classification result set, the weight parameters of these two sub-assessment models are proportionally calibrated. For example, the weight of the sub-assessment model based on industry rules is appropriately reduced for high-risk levels, while the weight for low-risk levels is increased. This ensures that each sub-assessment model has good coverage of high, medium, and low risk levels in the fused output.

[0150] Step S56: After all sub-assessment models have completed cross-model balancing, the classification weight parameters of each sub-assessment model are synchronized to the enterprise credit risk assessment model to generate a final parameter configuration that matches the credit classification result.

[0151] The final parameter configuration is the parameter configuration obtained by synchronizing the classification weight parameters of each sub-assessment model to the enterprise credit risk assessment model after cross-model balancing processing. This parameter configuration enables the enterprise credit risk assessment model to classify and assess enterprise credit risks more accurately.

[0152] For example, after the sub-assessment model based on industry rules, the sub-assessment model based on semantic associations, and the sub-assessment model based on time series changes have completed cross-model balancing processing, their classification weight parameters are synchronized into the enterprise credit risk assessment model to generate a final parameter configuration that matches the credit classification results.

[0153] Step S60: After all sub-assessment models have been optimized, the classification weight parameters of each sub-assessment model are integrated according to the feature item type of the corresponding dimension to generate the final parameter configuration of the enterprise credit risk assessment model.

[0154] The classification weight parameters of each sub-assessment model are obtained through training and optimization and are used to adjust the importance of each sub-assessment model in the final assessment results. The feature item type of the corresponding dimension refers to the feature item type related to the assessment dimension for which each sub-assessment model is responsible, such as feature items based on industry rules, feature items based on semantic associations, and feature items based on temporal changes. Fusion is the process of integrating and unifying the classification weight parameters of each sub-assessment model according to the feature item type of the corresponding dimension to generate the final parameter configuration of the enterprise credit risk assessment model, which is used to accurately assess the credit risk of the target enterprise.

[0155] For example, the classification weight parameters of the sub-assessment model based on industry rules, the classification weight parameters of the sub-assessment model based on semantic association, and the classification weight parameters of the sub-assessment model based on time series changes are integrated according to the dimensions of feature items based on industry rules, feature items based on semantic association, and feature items based on time series changes to generate the final parameter configuration of the enterprise credit risk assessment model.

[0156] Step S400: Determine the credit risk level of the target enterprise according to the classification label corresponding to the credit classification result.

[0157] The classification label corresponding to the credit classification result is the label obtained in step S300, which represents the enterprise's credit risk level, such as high risk, medium risk, or low risk. The target enterprise's credit risk level is determined based on the classification label and helps enterprise managers, investors, and other stakeholders understand the enterprise's credit status and make reasonable decisions.

[0158] Determine the target enterprise's credit risk level based on the classification label corresponding to the credit classification result. Specifically, the following steps may be included: Step S410: extracting a risk level identifier from the classification tag, and matching the identifier with a preset risk level rule library.

[0159] The risk level identifier is used within the classification label to clearly indicate the risk level, such as "high risk," "medium risk," or "low risk." The pre-set risk level rule base is a pre-established database containing detailed rules and standards corresponding to different risk levels. By matching the risk level identifier with the pre-set risk level rule base, you can obtain the specific rules and requirements corresponding to that risk level.

[0160] For example, if the classification label is "high risk", the risk level identifier "high risk" is extracted from the label, and then the rules corresponding to "high risk" are searched in the preset risk level rule library, such as the company's debt ratio exceeds the preset ratio, there have been major default events in the recent period, etc.

[0161] Step S420: Determine the risk level range to which the target enterprise belongs based on the rule entries in the risk level rule library.

[0162] Rule entries are the specific rules and standards corresponding to each risk level in the pre-set risk level rule library. Based on these rule entries, the target company's relevant information is evaluated and judged to determine the target company's risk level range, such as high risk, medium risk, and low risk.

[0163] For example, based on the rule entry corresponding to "high risk," the target company's debt ratio and whether it has had any major default events are examined. If the target company's debt ratio exceeds the specified ratio in the rule and a major default event has occurred recently, the target company's risk level is determined to be high risk.

[0164] Step S430: Read the assessment parameter adjustment strategy from the configuration file corresponding to the risk level interval, and dynamically modify the credit risk assessment parameters according to the adjustment strategy; wherein the dynamic modification includes normalizing the classification weight parameters and filtering abnormal feature items.

[0165] The configuration files corresponding to risk level intervals are pre-configured files containing the assessment parameter adjustment strategies for each risk level interval. These strategies are the specific methods and rules used to adjust credit risk assessment parameters. Dynamic correction refers to the process of real-time adjustment and optimization of credit risk assessment parameters based on these strategies. This includes normalizing classification weight parameters and filtering out abnormal feature items.

[0166] Normalization of classification weight parameters involves adjusting them to meet preset proportional relationships, ensuring that each sub-assessment model receives a reasonable weight in the final assessment results. Abnormal features are those that may significantly bias the assessment results, such as those whose values significantly deviate from the normal range. Filtering these abnormal features can improve the accuracy of the assessment results.

[0167] For example, for high-risk areas, assessment parameter adjustment strategies are read from the corresponding configuration file. For example, the weight of the sub-assessment model based on industry rules is increased, while some time-varying feature items with large fluctuations are filtered out. Based on these strategies, the classification weight parameters are normalized, such as adjusting the weight of the sub-assessment model based on industry rules from 0.3 to 0.4. At the same time, feature items based on time-varying changes are checked to filter out those with abnormal fluctuations, and credit risk assessment parameters are dynamically adjusted.

[0168] Step S500: Generate credit risk assessment parameters for the target enterprise based on the credit risk level and classification weight parameters, and output a visual assessment report based on the credit risk assessment parameters.

[0169] The credit risk rating represents the target enterprise's credit risk level determined in step S400. The classification weight parameters, derived during the model training and optimization process, are used to adjust the importance of each sub-assessment model in the final assessment results. Credit risk assessment parameters, such as risk probability values, stability indicators, and abnormal fluctuation warning parameters, are generated based on the credit risk rating and classification weight parameters to comprehensively assess the target enterprise's credit risk. The visual assessment report presents the credit risk assessment parameters in intuitive charts and text, enabling relevant personnel to quickly understand the target enterprise's credit risk status.

[0170] As an embodiment, in step S500, generating the credit risk assessment parameters of the target enterprise based on the credit risk level and the classification weight parameter may specifically include the following steps: Step S510: Determine the weight distribution ratio corresponding to each sub-assessment model in the classification weight parameter according to the weight distribution rule corresponding to the risk level interval; wherein the weight distribution ratio is positively correlated with the classification contribution value of the sub-assessment model in the risk level interval.

[0171] The weighting rules for risk level intervals are pre-defined rules that determine the weighting ratios for each sub-assessment model based on the risk level interval. The classification contribution value refers to the contribution of each sub-assessment model to the final assessment result within that risk level interval. The weighting ratio is positively correlated with the classification contribution value of the sub-assessment model within that risk level interval; that is, the greater the classification contribution value, the higher the corresponding weighting ratio for that sub-assessment model.

[0172] For example, in the high-risk level range, the classification contribution value of the sub-assessment model based on industry rules is higher. According to the weight allocation rules corresponding to the risk level range, the weight allocation ratio of the sub-assessment model is determined to be 0.5, the weight allocation ratio of the sub-assessment model based on semantic association is 0.3, and the weight allocation ratio of the sub-assessment model based on time series changes is 0.2.

[0173] Step S520: Based on the weight distribution ratio, the classification weight parameters are normalized and fused to generate the comprehensive weight value of each sub-assessment model; the comprehensive weight value is used to characterize the relative importance of feature items of different dimensions in the final risk assessment.

[0174] Normalization and fusion processing is the process of adjusting and integrating classification weight parameters according to the weight distribution ratio, so that the weight parameters of each sub-assessment model meet the set proportional relationship and the sum is 1. The comprehensive weight value is the weight value of each sub-assessment model obtained after normalization and fusion processing, which is used to represent the relative importance of feature items of different dimensions in the final risk assessment.

[0175] For example, according to the weight distribution ratio determined in step S510, the classification weight parameters are normalized and fused. Assume that the original classification weight parameter of the sub-evaluation model based on industry rules is 0.4, and the weight distribution ratio is 0.5; the original classification weight parameter of the sub-evaluation model based on semantic association is 0.3, and the weight distribution ratio is 0.3; the original classification weight parameter of the sub-evaluation model based on time series changes is 0.3, and the weight distribution ratio is 0.2. After normalization and fusion processing, the comprehensive weight value of the sub-evaluation model based on industry rules is 0.5, the comprehensive weight value of the sub-evaluation model based on semantic association is 0.3, and the comprehensive weight value of the sub-evaluation model based on time series changes is 0.2.

[0176] Step S530: extracting core feature items corresponding to the risk level interval from the feature set; the core feature items are feature items with a contribution ranking in the top preset digits under the comprehensive weight value.

[0177] Core feature items are those within the feature set that contribute significantly to the final risk assessment result. The preset number of digits is a pre-set number of core feature items used to screen for specific needs. Core feature items corresponding to the risk level range are extracted from the feature set. The contribution of each feature item is calculated based on its comprehensive weight, and the feature items with the highest contribution digits are selected as core feature items.

[0178] For example, the default digit is 3, and the contribution of each feature item in the feature set is calculated based on the comprehensive weight value. Assuming that the contribution of the feature items "output", "market demand", and "inventory" based on industry rules ranks in the top 3, these three feature items are regarded as the core feature items corresponding to the risk level range.

[0179] Step S540: Calculate the stability offset of the core feature items in the current target enterprise data based on the distribution pattern of the core feature items in the historical credit labels; the stability offset is used to reflect the difference between the degree of fluctuation of the core feature items within the time window and the historical benchmark.

[0180] Historical credit labels are enterprise credit risk rating labels corresponding to historical enterprise text datasets. Distribution patterns refer to the distribution of core feature values under different historical credit labels. Stability offsets are calculated based on the distribution patterns of core feature values in historical credit labels, indicating the degree of fluctuation in the core feature values in the current target enterprise data compared to the historical benchmark.

[0181] For example, for the core feature item "output," we analyze the distribution of its historical values when the credit labels were high, medium, and low risk to obtain a historical benchmark. We then calculate the volatility of "output" in the current target enterprise data, such as the rate of change in output over the past month. By comparing the current volatility with the historical benchmark, we calculate the stability offset of "output."

[0182] Step S550: Based on the stability offset and the comprehensive weight value, the contribution of the core feature items is calibrated to generate a calibrated feature contribution set.

[0183] Calibration is the process of adjusting and optimizing the contributions of core feature items based on stability offsets and comprehensive weights. The calibrated feature contribution set is the resulting set of core feature items, which more accurately reflects the role of core feature items in the final risk assessment.

[0184] For example, for the core feature item "yield," its comprehensive weight is 0.5 and its stability offset is 0.2. Based on this information, the contribution of "yield" is calibrated. For example, the original contribution is multiplied by (1 + 0.2) × 0.5 to obtain the calibrated contribution. After calibrating all core feature items, a calibrated feature contribution set is generated.

[0185] Step S560: Generate credit risk assessment parameters based on the comparison between the calibration results of each feature item in the calibrated feature contribution set and the preset contribution threshold; wherein the credit risk assessment parameters include risk probability value, stability index and abnormal fluctuation warning parameter, and each parameter is associated with at least one feature item in the calibrated feature contribution set.

[0186] The preset contribution threshold is a pre-set limit for determining the contribution of a feature item, based on actual needs. The risk probability value indicates the likelihood of a target enterprise incurring credit risk. The stability index measures the stability of the target enterprise's credit risk. The abnormal fluctuation warning parameter is used to warn of abnormal fluctuations in the target enterprise's credit risk. The value of each parameter is determined by comparing the calibration results of each feature item in the calibrated feature contribution set with the preset contribution threshold, generating credit risk assessment parameters.

[0187] For example, if the calibration result of a core feature item in the calibrated feature contribution set exceeds a preset contribution threshold, the feature item is considered to have a significant impact on the risk probability value, and the risk probability value is increased accordingly. Simultaneously, stability indices and abnormal fluctuation warning parameters are calculated based on information such as the stability offset of the core feature item, ensuring that each parameter is associated with at least one feature item in the calibrated feature contribution set.

[0188] As an embodiment, in step S500, outputting a visual assessment report based on the credit risk assessment parameters may specifically include the following steps: Step S570: Convert the risk probability value, stability index and abnormal fluctuation warning parameter into a multidimensional chart data format according to the preset chart mapping rules; the multidimensional chart data format includes the underlying data format of the time series distribution chart, feature contribution stacking chart and risk level heat map.

[0189] Preset chart mapping rules are pre-defined rules for converting credit risk assessment parameters into a multidimensional chart data format. Multidimensional chart data formats are the underlying data formats used to generate different types of charts, such as time series distribution charts that display data changes over time, feature contribution stacked charts that show the contribution of different feature items to the final result, and risk level heat maps that visually display the risk level of different regions or indicators.

[0190] For example, the risk probability value is converted into the underlying data format of the time series distribution chart according to the preset chart mapping rules, the stability index is converted into the underlying data format of the feature contribution stacking chart, and the abnormal fluctuation warning parameter is converted into the underlying data format of the risk level heat map.

[0191] Step S580: Generate an interactive visualization interface according to the multi-dimensional chart data format, and dynamically render the graphic elements of the time series distribution chart, the feature contribution stacking chart, and the risk level heat map in the interface.

[0192] The interactive visualization interface is a user-friendly interface that allows users to view various charts and data. Dynamic rendering refers to the process of generating and displaying graphical elements in real time based on multidimensional chart data formats. Dynamic rendering of graphical elements such as time series distribution charts, feature contribution stacked charts, and risk level heat maps within the interactive visualization interface allows users to intuitively understand the credit risk profile of target companies.

[0193] For example, using visual programming tools, an interactive visualization interface can be generated based on multidimensional chart data formats. In this interface, graphical elements of a time series distribution chart, such as curves and data points, can be dynamically rendered based on the underlying data format of the time series distribution chart. Graphical elements of a feature contribution stacked chart, such as bar charts and percentage labels, can be dynamically rendered based on the underlying data format of the feature contribution stacked chart. Graphical elements of a risk level heat map, such as color blocks and risk level labels, can also be dynamically rendered based on the underlying data format of the risk level heat map.

[0194] Step S590: embed the industry attribute label of the target enterprise in the interactive visualization interface, and adapt the display style of the graphic elements according to the industry attribute label to generate an industry-adapted visualization chart.

[0195] The target enterprise's industry attribute label is used to identify the industry to which the target enterprise belongs, such as manufacturing, finance, or services. Adapting the display style of graphical elements based on industry attribute labels means adjusting the color, shape, size, and other display styles of graphical elements in time series distribution charts, feature contribution stacked charts, and risk level heat maps based on the characteristics and needs of different industries, making the visualization more consistent with the characteristics of the industry and user habits.

[0196] For example, for manufacturing companies, we can set the curve color of the time series distribution graph to a color associated with the manufacturing industry, such as blue; set the bar chart shape of the feature contribution stacked graph to a more industrial-looking shape; and align the color blocks of the risk level heat map with the risk level standards of the manufacturing industry. Through these adjustments, we can generate industry-specific visualization charts.

[0197] Step S5100: Logically associate the industry-adapted visual chart with the text description field in the preset report template to generate a draft assessment report that includes the chart and text linkage.

[0198] Pre-designed report templates are pre-designed report templates that include text description fields and formats. Logical association involves linking industry-specific visualizations with the text description fields in the pre-designed report template, enabling the text descriptions to accurately interpret and explain the data and information in the charts. A draft assessment report with chart-text linkage is generated by linking industry-specific visualizations with the text description fields in the pre-designed report template, facilitating subsequent editing and improvement.

[0199] For example, a pre-set report template includes a text description field for risk probability values. Logically associate the risk probability data in an industry-specific time series distribution chart with this text description field. When users view the time series distribution chart in the visualization interface, the text description field automatically updates to display the text explanation corresponding to the current chart data. This generates a draft assessment report that integrates charts and text.

[0200] Step S5110: Based on the logical relationship between the charts and text in the draft evaluation report, perform semantic verification on the text description field, correct the text content that is inconsistent with the chart data, and generate the final evaluation report document.

[0201] Semantic verification checks the text description fields in the draft assessment report to ensure semantic consistency between the text and the chart data. Correcting inconsistent text with chart data involves modifying and adjusting the text description to accurately reflect the chart data when discrepancies or inconsistencies are found. The final assessment report document, after semantic verification and text correction, is a complete assessment report containing accurate chart and text information.

[0202] For example, in the draft assessment report, the text description field mentioned that the risk probability value was increasing, while the time series distribution chart showed that the risk probability value was decreasing. After the semantic verification discovered this inconsistency, the text description field was corrected to align with the chart data, and the final assessment report document was generated.

[0203] Step S5120: Send the final assessment report document to the target terminal device via an encrypted transmission protocol, and trigger an alarm prompt signal corresponding to the risk level range in the terminal device; wherein the type and intensity of the alarm prompt signal are dynamically adjusted according to the level of the abnormal fluctuation warning parameter.

[0204] The encrypted transmission protocol is a protocol designed to ensure data security. By using this protocol to send the final assessment report to the target terminal device, it prevents data theft or tampering during transmission. The target terminal device is the device that receives the final assessment report, such as a company manager's computer or an investor's mobile phone. Alarm signals alert the target terminal user of the target enterprise's credit risk. Their type and intensity are dynamically adjusted based on the level of the abnormal fluctuation warning parameter, allowing users to promptly monitor changes in the target enterprise's credit risk.

[0205] For example, the final assessment report is sent to the target terminal device using an encrypted transmission protocol such as SSL / TLS. If the abnormal fluctuation warning parameter is high, a high-intensity alarm signal, such as an audible or vibration alarm, is triggered on the terminal device. If the abnormal fluctuation warning parameter is low, a low-intensity alarm signal, such as a slight vibration or beep, is triggered. In this way, credit risk information about the target enterprise is promptly conveyed to users.

[0206] See also Figure 2 , Figure 2 This is a schematic diagram of the structure of an enterprise credit risk assessment system provided in an embodiment of the present invention. The enterprise credit risk assessment system can be any feasible computer system with data processing capabilities, such as a server, a distributed computing cluster, or a computer device. The enterprise credit risk assessment system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (also known as the central processing unit (CPU)) is the computing and control core of the enterprise credit risk assessment system, capable of parsing various instructions within the enterprise credit risk assessment system and processing various data within the enterprise credit risk assessment system. The communication interface 102 can optionally include a standard wired interface or a wireless interface (such as Wi-Fi, a mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101. The communication interface 102 can also be used for data transmission and interaction within the enterprise credit risk assessment system. The memory 103 is the storage device in the enterprise credit risk assessment system, used to store programs and data. It is understood that the memory 103 herein may include both the built-in memory of the enterprise credit risk assessment system and, of course, the extended memory supported by the enterprise credit risk assessment system. Memory 103 provides storage space for storing the operating system of the enterprise credit risk assessment system, which may include, but is not limited to, Android, iOS, Windows Phone, and the like, although this invention is not limited thereto.

[0207] In one embodiment, the processor 101 executes the enterprise credit risk assessment method based on public text data provided in the above embodiment of the present invention by running the computer program in the memory 103 .

Claims

1. A method for assessing corporate credit risk based on public text data, characterized in that: The method comprises: Acquire public text data corresponding to multiple data sources, and perform text preprocessing on the public text data to generate a structured enterprise text data set; Extract enterprise features based on the preprocessed public text data to obtain a feature set associated with the target enterprise; the feature set includes feature items based on industry rules, feature items based on semantic associations, and feature items based on temporal changes; Inputting the feature set into a pre-trained enterprise credit risk assessment model to obtain a credit classification result of the target enterprise; the enterprise credit risk assessment model includes multiple sub-assessment models, each sub-assessment model corresponds to a classification weight parameter of a different dimension; Determining the credit risk level of the target enterprise based on the classification label corresponding to the credit classification result; Based on the credit risk level and the classification weight parameters, a credit risk assessment parameter of the target enterprise is generated, and a visual assessment report is output according to the credit risk assessment parameter.

2. The method according to claim 1, characterized in that The step of performing text preprocessing on the public text data to generate a structured enterprise text dataset includes: Extracting original text data associated with the target enterprise from the multiple data sources, and performing format standardization processing on the original text data to generate uniformly coded initial text data; Performing word segmentation processing on the initial text data to obtain a set of text fragments after word segmentation; Filtering target text segments containing preset keywords from the text segment set, and generating semantically associated text blocks based on contextual semantic relationships of the target text segments; According to the industry attributes of the target enterprise, label matching is performed on the semantic association block to generate an industry label corresponding to the industry attribute; The text blocks containing the industry labels are subjected to time alignment processing to generate the structured enterprise text dataset; wherein the structured enterprise text dataset includes a sequence of text blocks arranged in time order, and each text block sequence corresponds to at least one industry label.

3. The method according to claim 2, characterized in that Enterprise features are extracted based on the pre-processed public text data to obtain a feature set associated with the target enterprise, including: Extracting a text block sequence corresponding to the industry label from the structured enterprise text dataset, and calculating a weight value of each keyword under the industry label based on keyword distribution in the text block sequence; Sorting the keywords according to the weight values, and screening out a target keyword set whose weight values are greater than or equal to a preset threshold; Based on the target keyword set, performing semantic association analysis on the text block sequence to determine the semantic association path of each target keyword in different text blocks; Constructing industry rule feature items of the target enterprise based on the semantic association path, wherein the industry rule feature items include a weight value of each keyword in the target keyword set and a semantic density parameter of its association path; Extract the time series change data associated with the target enterprise from the text block sequence, and generate the feature items based on time series change according to the distribution differences of the time series change data in different time periods; wherein the feature items based on time series change include the rate of change of the number of text blocks within the time window, the frequency of keyword repetition, and the semantic conflict detection parameters.

4. The method according to claim 3, characterized in that Inputting the feature set into a pre-trained enterprise credit risk assessment model to obtain a credit classification result of the target enterprise includes: Inputting the industry rule feature item, semantic association feature item and time series change feature item into the corresponding sub-assessment models in the enterprise credit risk assessment model respectively; Allocate weights to each input feature item through each of the sub-evaluation models to generate a classification contribution value corresponding to each feature item; Determine the output results of each sub-evaluation model based on the classification contribution value, and perform weighted fusion on the output results to generate a comprehensive classification score; Based on the comparison result of the comprehensive classification score and the preset classification threshold, the classification label corresponding to the credit classification result is determined; wherein, the classification label includes multiple risk level intervals, and each risk level interval corresponds to a different classification weight parameter combination.

5. The method according to claim 4, characterized in that The training process of the enterprise credit risk assessment model includes: Obtaining a historical enterprise text dataset and corresponding historical credit labels, and performing feature extraction processing on the historical enterprise text dataset to generate a historical feature set; Dividing the historical feature set into a training subset and a validation subset based on the degree of association between each feature item type in the historical feature set and the historical credit label; wherein the training subset includes feature item data whose degree of association with the historical credit label is greater than a first correlation threshold, and the validation subset includes feature item data whose degree of association with the historical credit label is less than a second correlation threshold; For each sub-evaluation model, extract feature item data of the dimension corresponding to the sub-evaluation model from the training subset, and dynamically match the feature item data with the historical credit labels to generate a sub-model training sample set; The sub-model training sample set is first classified using the initialized classification weight parameters, and the classification weight parameters are iteratively adjusted layer by layer based on the error distribution between the classification results and the historical credit labels; The adjusted sub-evaluation model is subjected to classification verification with the feature item data in the verification subset, and the classification weight parameters are cross-model optimized according to the verification results until the classification accuracy of all sub-evaluation models reaches a preset threshold simultaneously; When all sub-assessment models have been optimized, the classification weight parameters of each sub-assessment model are integrated according to the feature item type of the corresponding dimension to generate the final parameter configuration of the enterprise credit risk assessment model.

6. The method according to claim 5, characterized in that The sub-model training sample set is first classified using the initialized classification weight parameters, and the classification weight parameters are iteratively adjusted layer by layer based on the error distribution between the classification results and the historical credit labels, including: Determining an initial error offset of the initialized classification weight parameter according to the feature item data distribution in the sub-model training sample set; Performing a first classification on the sub-model training sample set based on the initial error offset to generate an initial classification result set; Calculating the error between the classification result of each sample in the initial classification result set and the corresponding historical credit label, and prioritizing the sub-model training sample sets according to the error values to obtain a priority ranking result; According to the priority sorting result, samples with error values greater than a preset error threshold are screened from the sub-model training sample set to generate a high error sample subset; Based on the feature item data distribution of the high error sample subset, reversely adjust the classification weight parameter to generate a first adjustment parameter; Performing secondary classification on the sub-model training sample set using the first adjustment parameter, and repeating the high-error sample screening and parameter adjustment steps based on the secondary classification result and the error distribution of historical credit labels until the error value is less than or equal to a preset error threshold; After completing the layer-by-layer iterative adjustment, the final adjusted classification weight parameters are locked as the intermediate parameter configuration of the current sub-evaluation model.

7. The method according to claim 6, characterized in that The step of performing classification verification on the adjusted sub-evaluation model and the feature item data in the verification subset, and performing cross-model optimization on the classification weight parameters according to the verification results, includes: Inputting the feature item data in the validation subset into the adjusted sub-evaluation model according to the corresponding dimensions to generate a validation classification result set; Comparing the verification classification result set with the true credit labels corresponding to the verification subset, and calculating the current accuracy of each sub-evaluation model; If the current accuracy rate does not reach the preset threshold, extracting supplementary sample data having similar distribution characteristics to the feature item data of the validation subset from the training subset; Merging the supplementary sample data with the sub-model training sample set to generate an extended training set, and performing supplementary iterative adjustment on the classification weight parameters based on the extended training set; If the current accuracy reaches a preset threshold, the classification weight parameters are subjected to cross-model balancing according to the distribution ratio of different risk levels in the verification classification result set; wherein the cross-model balancing includes proportional calibration of the weight parameters of different sub-assessment models to ensure that each sub-assessment model has a consistent risk level coverage in the fused output result; After all sub-assessment models have completed cross-model balancing, the classification weight parameters of each sub-assessment model are synchronized to the enterprise credit risk assessment model to generate a final parameter configuration that matches the credit classification result.

8. The method according to claim 1, characterized in that Determining the credit risk level of the target enterprise according to the classification label corresponding to the credit classification result includes: Extracting a risk level identifier from the classification label and matching a preset risk level rule library according to the identifier; Determining the risk level interval to which the target enterprise belongs based on the rule entries in the risk level rule library; The assessment parameter adjustment strategy is read from the configuration file corresponding to the risk level interval, and the credit risk assessment parameter is dynamically modified according to the adjustment strategy; wherein the dynamic modification includes normalization of classification weight parameters and filtering of abnormal feature items.

9. The method according to claim 8, characterized in that Generating the credit risk assessment parameters of the target enterprise based on the credit risk level and the classification weight parameter includes: Determine the weight distribution ratio corresponding to each sub-assessment model in the classification weight parameter according to the weight distribution rule corresponding to the risk level interval; wherein the weight distribution ratio is positively correlated with the classification contribution value of the sub-assessment model in the risk level interval; Based on the weight distribution ratio, the classification weight parameters are normalized and fused to generate a comprehensive weight value for each sub-assessment model; the comprehensive weight value is used to characterize the relative importance of feature items of different dimensions in the final risk assessment; Extracting core feature items corresponding to the risk level interval from the feature set; the core feature items are feature items with a preset number of contribution rankings under the comprehensive weight value; Calculate the stability offset of the core feature item in the current target enterprise data based on the distribution pattern of the core feature item in the historical credit labels; the stability offset is used to reflect the difference between the fluctuation degree of the core feature item within the time window and the historical benchmark; Calibrate the contribution of the core feature item based on the stability offset and the comprehensive weight value to generate a calibrated feature contribution set; The credit risk assessment parameters are generated based on the comparison between the calibration results of each feature item in the calibrated feature contribution set and the preset contribution threshold; wherein the credit risk assessment parameters include a risk probability value, a stability index and an abnormal fluctuation warning parameter, and each parameter is associated with at least one feature item in the calibrated feature contribution set.

10. An enterprise credit risk assessment system, characterized in that: include: a memory storing a computer program; A processor, configured to load the computer program to implement the enterprise credit risk assessment method based on public text data as described in any one of claims 1 to 9.

Images