Enterprise risk analysis method and system based on news co-occurrence

Abstract

The invention relates to the technical field of enterprise risk assessment, in particular to a news co-occurrence-based enterprise risk analysis method and system, which comprises the following steps: extracting co-occurrence enterprise name pairs, recording first occurrence time and frequency to construct a network, matching policy keywords to generate a standard set, extracting default emotion words, calculating offset values, and distributing the offset values to enterprise pairs. Analyzing the fracture word and the amount, screening the conduction chain according to the time sequence, integrating the network, the emotion value and the difference value, and calculating the intensity to generate a risk topology result. According to the method, a dynamic network is constructed based on first co-occurrence time and frequency, homonymous enterprise disambiguation classification is realized through keyword coincidence, a directional emotion value is calculated in combination with part-of-speech weight and emotion word co-occurrence, subject-predicate-object structures are extracted, and high-coupling events are screened according to a time sequence to eliminate interference; and fusing the offset and the difference weighted diffusion intensity to generate a risk-oriented topological graph.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise risk assessment, and in particular to an enterprise risk analysis method and system based on news co-occurrence. Background Art

[0002] The technical field of enterprise risk assessment encompasses the application of technical means and theoretical models to identify and predict potential risks in enterprise operations. The core focus is on improving risk identification accuracy and enhancing the timeliness of predictions, encompassing four key technical aspects: data collection and cleaning, risk indicator modeling, risk transmission path tracing, and dynamic risk monitoring and early warning. From a technical perspective, enterprise risk assessment requires five key technical steps: unstructured data conversion, multi-source heterogeneous data fusion, risk feature vector extraction, risk probability calculation model construction, and visualization of risk assessment results. Traditional methods primarily employ four implementation paths: financial indicator trend analysis, supply chain relationship mapping, industry policy text mining, and expert experience quantitative evaluation.

[0003] Among them, the enterprise risk analysis method based on news co-occurrence refers to a technical solution that identifies risk correlations by mining the co-occurrence relationships of enterprise entities in news events. To address the low utilization of news data and the difficulty in identifying cross-industry risk transmission pathways in traditional risk assessment, this method constructs an enterprise entity co-occurrence matrix from a news media corpus, uses text similarity calculations to determine the correlation strength threshold, and establishes an enterprise risk correlation network map. The specific implementation process includes four technical steps: time series segmentation of news data, standardization of enterprise entity names, encoding of semantic features of co-occurrence events, and dynamic correction of correlation weights. It integrates an industry classification code library and an enterprise registration information database for multi-dimensional verification, and uses a sliding time window mechanism to ensure timely updates of risk correlation relationships.

[0004] The risk identification process often relies on static indicator maps or co-occurrence frequency aggregation, ignoring the time sequence and content relationship between events, resulting in the inability to effectively judge the leading nodes and actual causes of risk relationships, and the risk transmission chain is broken; text paragraphs of companies with the same name in different business scenarios are directly classified without context screening, which can easily cause confusion in subject attribution and significantly reduce the accuracy of subsequent emotion or event recognition; word frequency-based emotion assessment methods do not consider part-of-speech differences and semantic directionality, and the direction of emotional changes cannot be accurately expressed, and the assessment results have significant noise; the event behavior extraction process lacks a clear analysis of the subject-predicate-object structure, and descriptions without direct logical relationships are easily included in the path, resulting in a significant increase in semantic misconnections between risk events; the risk diffusion results are manifested at the graph level as an imbalance of central nodes, a lack of continuity in edge weight settings, and the topological structure is difficult to truly reflect the fluctuations in risk impact, resulting in blurred diffusion trends and distorted aggregation patterns. Summary of the Invention

[0005] In order to solve the technical problems existing in the prior art, the present invention provides a method and system for enterprise risk analysis based on news co-occurrence. The technical solution is as follows: The enterprise risk analysis method based on news co-occurrence includes the following steps: S1: Obtain the co-occurring company name pairs in the news paragraph, record the first co-occurrence timestamp and co-occurrence frequency, and generate a company co-occurrence association network with the name pairs as nodes and the frequency as edge weights; S2: Based on the enterprise co-occurrence network, extract the enterprise name context, match financial regulatory policy keywords, calculate the keyword overlap ratio between paragraphs of companies with the same name, classify those exceeding the threshold, and generate a standardized set of financial entities; S3: Traverse the standardized set of financial entities, extract credit default trigger words and market panic sentiment words in the news paragraphs, count the word frequencies and part-of-speech weights, calculate the absolute value of the paragraph sentiment shift, assign them to company name pairs based on co-occurrence relationships, and generate corporate debt risk sentiment values; S4: Extract supply chain disruption action words and debt default amounts from event sentences, sort triplets by time, calculate the difference in amounts between adjacent events and the semantic similarity of actions, filter out sequences exceeding the threshold, and generate a debt risk transmission chain; S5: Call the enterprise co-occurrence relationship network, enterprise debt risk sentiment value and debt risk transmission chain, superimpose the sentiment offset and amount difference, calculate the diffusion intensity in combination with the edge weight, and generate the debt risk diffusion topology result.

[0006] As a further solution of the present invention, the enterprise co-occurrence association network includes a set of enterprise name pairs, a first co-occurrence timestamp, and a co-occurrence frequency distribution. The standardized set of financial entities is specifically a policy keyword matching set, an enterprise alias mapping table, and a paragraph overlap threshold parameter. The enterprise debt risk sentiment value includes default trigger word frequency statistics, sentiment part-of-speech weight distribution, and offset allocation rules. The debt risk transmission chain includes an event amount difference sequence, a semantic similarity score, and a transmission path identifier. The debt risk diffusion topology result covers a diffusion intensity coefficient, a topological node hierarchical distribution, and a path weight attenuation parameter.

[0007] As a further solution of the present invention, the steps for obtaining the enterprise co-occurrence association network are: S101: Traverse news text paragraphs, detect company names, extract all company name pairs within the same paragraph, record the paragraph publishing timestamp of the first appearance of each name pair, count the number of co-occurrences of each name pair, and generate a co-occurrence record of the company name pair, including the company name pair, the first co-occurrence timestamp, and the cumulative number of co-occurrences; S102: Retrieving the co-occurrence records of the enterprise name pairs, merging duplicate name pair records, retaining the earliest timestamp of each name pair as a time reference value, accumulating all co-occurrence times as edge weight values, and generating a co-occurrence frequency of the enterprise name pairs including unique name pairs, edge weight values, and time reference values; S103: Based on the co-occurrence frequency of the enterprise name pair, the two enterprises in the enterprise name pair are set as network nodes, the edge weight value is called as the edge weight parameter between the nodes, the first co-occurrence time of each edge is marked according to the time reference value, the node and edge data are integrated, and an enterprise co-occurrence association network is generated with the enterprise name as the node and the co-occurrence frequency as the edge weight.

[0008] As a further solution of the present invention, the steps for obtaining the standardized set of financial entities are: S201: Retrieving enterprise names in the enterprise co-occurrence association network, extracting a fixed range of text segments before and after the paragraph containing the name, storing the text segments of all paragraph instances, and generating a context segment set containing the enterprise name and the corresponding text segment; S202: Based on the context segment set, a financial regulatory policy keyword library containing terms such as antitrust review and listing compliance is preset, and the number of matching keywords in each company name-related paragraph is counted using the formula: ; Calculate the keyword overlap value KSR of company name i in paragraph j i,j , generating cross-paragraph keyword matching including company name, paragraph identifier and keyword overlap value; Among them, F k is the standard frequency of keywords in policy documents, C i is the total frequency of co-occurrence of enterprises, L j is the length of the paragraph character, ∈ is a very small constant, W n is the regulatory level weight of the keyword, O n is the number of times the keyword appears in the paragraph; S203: Calling the cross-paragraph keyword matching degree, calculating the variance-mean ratio of the keyword overlap values ​​of the same-named enterprises, screening the name classifications below the fluctuation threshold and above the merging threshold, and generating a standardized set of financial entities of the standardized name and alias set.

[0009] As a further solution of the present invention, the steps for obtaining the corporate debt risk sentiment value are: S301: Traverse the company names in the standardized set of financial entities, collect related news paragraph texts, match them with the credit default trigger vocabulary and the market panic sentiment vocabulary, count the co-occurrence times of the two types of words in the paragraphs, and generate a risk co-occurrence record containing the company name, paragraph identifier, and the co-occurrence frequency of trigger words and sentiment words; S302: Call the risk co-occurrence record, based on the trigger word part-of-speech weight table and the emotional word intensity level table, using the formula: ; Calculate the absolute value of sentiment shift EOS of enterprise i in paragraph j i,j , generating a cross-paragraph sentiment offset value including the company name, paragraph identifier and the absolute value of sentiment offset; Among them, F t is the co-occurrence frequency of trigger words, W t is the part-of-speech weight, λ is the time decay factor, C s is the co-occurrence frequency of emotional words, I s is the emotion intensity weight, D j is the dispersion of emotional tone of the paragraph; S303: Call the cross-paragraph sentiment offset value, assign the sentiment offset value according to the edge weight ratio of the enterprise name pair in the enterprise co-occurrence association network, accumulate all paragraph assignment values, and generate the enterprise debt risk sentiment value identified by the enterprise name pair.

[0010] As a further solution of the present invention, the steps for obtaining the debt risk transmission chain are: S401: Traverse the news event sentences, extract the subject-verb-object triples containing the supply chain disruption action and the debt default amount value, arrange the triples in ascending order by timestamp, and generate a time-series triple sequence containing the event subject, action description, amount value and timestamp; S402: Call the time-series triple sequence to calculate the absolute value of the difference between the default amounts of adjacent events using the formula: ; Calculate the transmission correlation R between the rth event and the r-1th event r , generating cross-event correlation indicators including event pair identification and transmission correlation degree; Among them, M r is the default amount of the rth event, T r is the timestamp difference, S r is the semantic similarity of supply chain actions, λ is the industry impact factor, I r is the industry risk level index; S403: Call the cross-event correlation indicator to screen event pairs whose transmission correlation is higher than the transmission threshold and whose absolute value of the amount difference exceeds the amount fluctuation threshold, connect the event pairs that meet the conditions in chronological order, and generate a debt risk transmission chain with the event subject as the node.

[0011] As a further solution of the present invention, the steps for obtaining the debt risk diffusion topology result are: S501: Calling the edge weight values ​​in the enterprise co-occurrence association network, the sentiment offset in the enterprise debt risk sentiment value, and the default amount difference in the debt risk transmission chain, aligning the identifiers of the same enterprise node in different data sources in chronological order, and superimposing the sentiment offset and the default amount difference item by item according to the timestamp to generate a risk parameter superposition table including the enterprise name, timestamp, superposition value, and edge weight value; S502: Based on the risk parameter overlay table, extract the edge weight values ​​of adjacent enterprise nodes, multiply the overlay values ​​by the edge weight values ​​item by item, and accumulate the product results of all related enterprise nodes in the same time window to generate an inter-node diffusion intensity set with enterprise name pairs as identifiers and numerical intensity as content; S503: Call the inter-node diffusion intensity set, filter the enterprise name pairs whose intensity values ​​exceed the preset diffusion threshold in timestamp order, and construct a directed connection path with the enterprise as the node and the diffusion intensity as the edge weight based on the connection relationship of the enterprise nodes in the co-occurrence network, and generate a debt risk diffusion topology result including the risk transmission direction and intensity weight.

[0012] As a further embodiment of the present invention,.

[0013] As a further embodiment of the present invention,.

[0014] An enterprise risk analysis system based on news co-occurrence, the system comprising: The enterprise co-occurrence network construction module obtains all enterprise names in a news paragraph, detects the first co-occurrence timestamp of a name pair, counts the co-occurrence frequency of the name pair within the paragraph, and uses the timestamp and frequency to construct a network structure with name pairs as nodes and frequencies as edge weights, thus generating an enterprise co-occurrence association network. The entity standardization classification module extracts the paragraph text containing the enterprise name based on the enterprise co-occurrence weight network, matches it with the financial regulatory policy keyword set, calculates the overlap ratio of keywords across paragraphs for enterprises with the same name, and classifies entities whose ratio exceeds a preset classification threshold to generate a standardized set of financial entities; A risk sentiment calculation module calls the standardized set of financial entities, extracts credit default trigger words and market panic sentiment words in the paragraph, counts word frequencies and verb and adjective weight coefficients, calculates the absolute value of the paragraph sentiment offset, and assigns the edge weights of the enterprise co-occurrence weight network to the name pairs to generate the enterprise debt risk sentiment value; The risk transmission chain extraction module obtains the supply chain disruption action words and debt default amounts in the event sentences, arranges them into triplets in chronological order, calculates the absolute value of the difference between the amounts of adjacent events and the semantic similarity of the action words, and selects sequences whose difference exceeds the amount fluctuation threshold and whose similarity exceeds the action similarity threshold to generate the debt risk transmission chain. The risk diffusion topology generation module calls the edge weights of the enterprise co-occurrence weight network, the sentiment offset of the enterprise debt risk sentiment value, and the amount difference of the debt risk transmission chain, superimposes the sentiment offset and the amount difference, calculates the diffusion intensity coefficient based on the edge weight, constructs the node diffusion path topology, and generates the debt risk diffusion topology result.

[0015] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least: In the present invention, a dynamic relationship graph is constructed through the first co-occurrence time and frequency of enterprise name pairs, and the companies with the same name are disambiguated and classified in combination with the overlap ratio of contextual keywords, thereby improving the accuracy of entity recognition; the offset value is calculated by the part-of-speech weight and the co-occurrence frequency of emotional words, so that the emotional expression has directionality and difference; the subject-verb-object triples are extracted to filter high-coupling events according to the time series and exclude pseudo-connections; the emotional offset and the amount difference are integrated, and the diffusion intensity is weighted according to the co-occurrence frequency to generate a topological map with continuous structure and clear risk orientation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method of the present invention; Figure 2 A flowchart for obtaining a co-occurrence association network of enterprises according to the present invention; Figure 3 A flowchart for obtaining a standardized set of financial entities according to the present invention; Figure 4 This is a flow chart for obtaining the enterprise debt risk sentiment value of the present invention; Figure 5 A flowchart for obtaining the debt risk transmission chain of the present invention; Figure 6 This is a flow chart for obtaining the debt risk diffusion topology results of the present invention. DETAILED DESCRIPTION

[0017] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0018] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as an "exemplary" in the present invention should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of the word "exemplary" is intended to present concepts in a concrete manner. Furthermore, in the embodiments of the present invention, "and / or" can mean both or either of the two.

[0019] In the embodiments of the present invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same. The terms "of," "corresponding," and "corresponding" may sometimes be used interchangeably. It should be noted that, when the distinction is not emphasized, the meanings they convey are the same.

[0020] In the embodiments of the present invention, sometimes a subscript such as W1 may be written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are the same.

[0021] In order to make the technical problems, technical solutions and advantages to be solved by the present invention clearer, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0022] See also Figure 1 The present invention provides a technical solution: a method for analyzing enterprise risk based on news co-occurrence, comprising the following steps: S1: Obtain the co-occurring company name pairs within the news text paragraphs, record the first co-occurrence timestamp and co-occurrence frequency of each company name pair, and generate a company co-occurrence association network using the company name pairs as nodes and the co-occurrence frequency as edge weights; S2: Based on the company names in the enterprise co-occurrence network, extract the contextual character segments of the paragraphs containing the company names. By matching the keywords related to financial regulatory policies in the paragraphs, calculate the keyword overlap ratio of companies with the same name in different paragraphs. Classify company names with an overlap ratio exceeding the set merging threshold to generate a standardized set of financial entities. S3: Traverse the news paragraphs of enterprises in the standardized set of financial entities, extract trigger words related to credit default and market panic sentiment words from the news paragraphs, count the co-occurrence frequencies of trigger words and sentiment words in the paragraphs, calculate the absolute value of sentiment offset within the paragraphs based on part-of-speech weights, and assign the offset values ​​to corresponding enterprise name pairs based on co-occurrence relationships to generate enterprise debt risk sentiment values; S4: Extract subject-verb-object triples containing supply chain disruption actions and debt default amount values ​​from news event sentences, arrange the triples in timestamp order, calculate the absolute value of the difference in default amounts between adjacent events and the semantic similarity of the supply chain actions, and select event sequences whose difference exceeds the amount fluctuation threshold and whose semantic similarity is higher than the transmission threshold to generate the debt risk transmission chain; S5: Call the edge weight values ​​in the enterprise co-occurrence relationship network, the sentiment offset in the enterprise debt risk sentiment value, and the default amount difference in the debt risk transmission chain, superimpose the sentiment offset and amount difference of the enterprise nodes in the transmission chain in chronological order, and calculate the risk diffusion intensity between nodes in combination with the co-occurrence network edge weight to generate the debt risk diffusion topology result.

[0023] The enterprise co-occurrence association network includes a set of enterprise name pairs, the first co-occurrence timestamp, and the co-occurrence frequency distribution. The standardized set of financial entities specifically includes a policy keyword matching set, an enterprise alias mapping table, and a paragraph overlap threshold parameter. The enterprise debt risk sentiment value includes the default trigger word frequency statistics, the sentiment part-of-speech weight distribution, and the offset allocation rules. The debt risk transmission chain includes the event amount difference sequence, the semantic similarity score, and the transmission path identifier. The debt risk diffusion topology results cover the diffusion intensity coefficient, the topological node hierarchical distribution, and the path weight attenuation parameter.

[0024] See also Figure 2 , the steps to obtain the enterprise co-occurrence association network are: S101: Traverse news text paragraphs, detect company names, extract all company name pairs within the same paragraph, record the paragraph publishing timestamp of the first appearance of each name pair, count the number of co-occurrences of each name pair, and generate a co-occurrence record of the company name pair, including the company name pair, the first co-occurrence timestamp, and the cumulative number of co-occurrences; Traverse the news text paragraphs and use entity recognition rules based on regular expression matching to detect company names. For example, extract the content in the text that matches the naming pattern of "XX Company" and "XX Group" as candidate names, and then perform a secondary verification through the company name database. For example, extract the two names "CATL" and "BYD" from the paragraph "CATL and BYD reach cooperation in the field of power batteries". If the release time is March 12, 2023, 14:30, the timestamp is converted to 202303121430. If the latter If the same name pair appears again in the same paragraph, the co-occurrence count increases by 1. For example, when "CATL-BYD" is first detected, the timestamp 202303121430 is recorded and the co-occurrence count is set to 1. If the two names appear three times in the same paragraph, the co-occurrence count is accumulated to 3, and the record item "CATL-BYD, 202303121430, 3" is generated. At the same time, other name combination pairs in the paragraph, such as "CATL-Tesla", are processed, and finally a co-occurrence record of enterprise name pairs covering all name pairs is formed.

[0025] S102: Retrieving the co-occurrence records of enterprise name pairs, merging duplicate name pair records, retaining the earliest timestamp of each name pair as the time reference value, accumulating all co-occurrence times as the edge weight value, and generating the co-occurrence frequency of enterprise name pairs including unique name pairs, edge weight values, and time reference values; When retrieving co-occurrence records of enterprise name pairs, different record entries of the same enterprise name pair are merged. For example, there are two records for "CATL-BYD" with timestamps 202303121430 and 202304150930, and co-occurrence counts of 3 and 2, respectively. When merging, the earlier timestamp, 202303121430, is retained as the time reference value. The co-occurrence counts are added together to obtain a total count of 5, which is used as the edge weight value. If there are spelling differences in the name pairs (such as "CATL" and "CATL New Energy"), the similarity is calculated using a string fuzzy matching algorithm. A similarity threshold of 0.9 is set. When the similarity exceeds the threshold, they are considered to be the same enterprise. For example, the similarity between "CATL" and "CATL New Energy" is 0.85. If the similarity is below the threshold, they are treated as independent names. After merging, the co-occurrence frequency of enterprise name pairs is generated, which includes unique name pairs, edge weight values, and time reference values.

[0026] S103: Based on the co-occurrence frequency of the enterprise name pair, the two enterprises in the enterprise name pair are set as network nodes, the edge weight value is used as the edge weight parameter between the nodes, the first co-occurrence time of each edge is marked according to the time reference value, the node and edge data are integrated, and an enterprise co-occurrence association network is generated with the enterprise name as the node and the co-occurrence frequency as the edge weight; When constructing the network, each company name is mapped to a unique node identifier. For example, "CATL" is mapped to node ID N001, and "BYD" is mapped to N002. The node attribute table stores the correspondence between names and IDs, and the edge attribute table stores node pairs (N001, N002), edge weight values ​​5, and time reference values ​​202303121430. If the same company name pair has different weights in different pre-merger records, for example, the pre-merger weights are 3 and 2, the final weight is 5. During the network generation process, nodes are checked for duplication. For example, if "CATL" already exists in the node table, its ID is directly referenced. Otherwise, a new node is created. After integrating all node and edge data, an enterprise co-occurrence association network containing a node list, an edge list, and attributes is output.

[0027] See also Figure 3 , the steps to obtain the standardized set of financial entities are: S201: Retrieving enterprise names from the enterprise co-occurrence network, extracting a fixed range of text segments before and after the paragraph containing the name, storing the text segments of all paragraph instances, and generating a context segment set containing the enterprise name and the corresponding text segment; S202: Based on the context segment set, a financial regulatory policy keyword library containing terms such as antitrust review and listing compliance is preset, and the number of matching keywords in each company name-related paragraph is counted using the formula: ; Calculate the keyword overlap value KSR of company name i in paragraph j i,j , generating cross-paragraph keyword matching including company name, paragraph identifier and keyword overlap value; Among them, F k is the standard frequency of keywords in policy documents, C i is the total frequency of co-occurrence of enterprises, L j is the length of the paragraph character, ∈ is a very small constant, W n is the regulatory level weight of the keyword, O n is the number of times the keyword appears in the paragraph; S203: Call the cross-paragraph keyword matching degree, calculate the variance-mean ratio of the keyword overlap values ​​of the companies with the same name, filter the name classifications that are lower than the fluctuation threshold and higher than the merging threshold, and generate a standardized set of financial entities of the standardized name and alias set.

[0028] See also Figure 4 , the steps to obtain the corporate debt risk sentiment value are: S301: Traverse the company names in the standardized set of financial entities, collect related news paragraph text, match them with the credit default trigger vocabulary and the market panic sentiment vocabulary, count the co-occurrence times of the two types of words in the paragraph, and generate a risk co-occurrence record containing the company name, paragraph identifier, and the co-occurrence frequency of trigger words and sentiment words; Traverse the company names in the standardized set of financial entities, collect related news paragraph text, match the credit default trigger word library (including noun trigger words such as "debt extension", "repayment difficulty", "credit downgrade", etc., with a weight of 1.2) and the market panic sentiment word library (including high-intensity sentiment words such as "panic selling", "liquidity exhaustion", and "plunge", with a weight of 1.5), scan the text content paragraph by paragraph, record the position and co-occurrence relationship between trigger words and sentiment words, and if the interval between trigger words and sentiment words in the same paragraph is less than 200 characters, it is considered a valid co-occurrence. Count each pair The paragraph-level co-occurrence frequency under the company name, taking the company "A Group" as an example, its news paragraph mentioned "A Group faces debt extension risks, triggering panic selling in the market", the trigger word "debt extension" appeared once, and the sentiment word "panic selling" appeared once, and the co-occurrence frequency was counted as 1. In another paragraph "A Group's bond repayment difficulties caused the stock price to plummet", the trigger word "repayment difficulties" and the sentiment word "plunge" each appeared once, and the co-occurrence frequency accumulated to 2. All paragraph data were integrated to generate a risk co-occurrence record containing the company name, paragraph number and co-occurrence frequency.

[0029] S302: Call the risk co-occurrence record, based on the trigger word part-of-speech weight table and the emotional word intensity level table, using the formula: ; Calculate the absolute value of sentiment shift EOS of enterprise i in paragraph ji,j , generating a cross-paragraph sentiment offset value including the company name, paragraph identifier and the absolute value of sentiment offset; Among them, F t is the co-occurrence frequency of trigger words, W t is the part-of-speech weight, λ is the time decay factor, C s is the co-occurrence frequency of emotional words, I s is the emotion intensity weight, D j is the dispersion of emotional tone of the paragraph; Based on the trigger word part-of-speech weight table (noun weight 1.2, verb weight 1.0) and the emotional word intensity level table (high panic word weight 1.5, medium panic word weight 1.0), taking the enterprise "Company B" paragraph P002 as an example, the trigger word "credit downgrade" (noun, weight 1.2) appears 2 times, and the emotional word "liquidity exhaustion" (high intensity, weight 1.5) appears once. The time decay factor λ is calculated based on the time difference between the news release time and the current time. If the time difference is 90 days, λ = 0.3. The emotional tone dispersion D j By analyzing the sentiment polarity scores (positive +1, neutral 0, negative -1) of non-emotional words (such as "bond" and "lead") in the paragraph, the variance is calculated. Assuming the score sequence is [-1, 0, -1], the variance is 0.222. , substitute into the formula: ; Step-by-step calculation: Trigger word part: 2x1.2=2.4, e -0.3 =0.7408; Sum value: 2.4x0.7408=1.778; Emotional word part: 1x1.5=1.5, 1.5x0.471=0.7065; Absolute value of offset: 1.778-0.7065|=1.0715: "Company B - P002, Sentiment Shift Absolute Value 1.0715" is generated. This result indicates that the market sentiment fluctuation intensity triggered by credit default risk in the current paragraph is high and is less affected by time decay. In subsequent steps, this value needs to be allocated according to the weights of related companies to reflect the intensity of risk transmission. All company and paragraph data must be integrated to form a cross-paragraph sentiment shift value.

[0030] S303: Calling the cross-paragraph sentiment offset value, assigning the sentiment offset value according to the edge weight ratio of the enterprise name pair in the enterprise co-occurrence association network, accumulating all paragraph assigned values, and generating the enterprise debt risk sentiment value identified by the enterprise name pair; The cross-paragraph sentiment offset value is called, and the sentiment value is assigned according to the edge weight ratio of the enterprise name pair in the enterprise co-occurrence association network. Taking the name pair "C Group-D Bank" as an example, its edge weight in the network is 8, the sum of all associated edge weights of the enterprise is 40, and the allocation ratio is 20%. The absolute value of the sentiment offset of paragraph P003 is 0.85, and the proportional allocation value is 0.85×20%=0.17. If the allocation value of another paragraph P004 is 0.23, the total sentiment value after accumulation is 0.40, generating "C Group-D Bank, sentiment value 0.40". This result shows that the debt risk correlation between the two enterprises is high. It is necessary to quantify the overall risk sentiment intensity by accumulating multiple paragraph allocation values. All name pairs and associated paragraphs are traversed, and each allocation value is accumulated to generate an enterprise debt risk sentiment value with the enterprise name pair as the index and the numerical sentiment intensity as the content.

[0031] See also Figure 5 , the steps to obtain the debt risk transmission chain are: S401: Traverse the news event sentences, extract the subject-verb-object triples containing the supply chain disruption action and the debt default amount value, arrange the triples in ascending order by timestamp, and generate a time-series triple sequence containing the event subject, action description, amount value and timestamp; Taking the company "Company A" as an example, its news event sentences are "Company A defaulted on a debt of 500 million yuan due to supplier termination on January 1, 2023" and "Company A defaulted on a debt of 800 million yuan due to raw material shortage on February 5, 2023." The actions "supplier termination" and "raw material shortage" and the amounts "500 million yuan" and "800 million yuan" are extracted. The timestamps are converted to the difference in days (the base date, January 1, 2023, is 0 days, and February 5, 2023, is 35 days). The triple sequence "Company A-Supplier Termination-500 million yuan-0" and "Company A-Raw Material Shortage-800 million yuan-35" are generated and sorted in ascending order by timestamp. If there is an event "Group B Defaulted on a 300 million yuan Default Due to Logistics Interruption on January 15, 2023" with a timestamp of 14 days, the sequence is adjusted to "Company A-Supplier Termination-500 million yuan-0", "Group B Defaulted on a 300 million yuan Default on January 14, 2023" after insertion. The timestamp is 14 days. After insertion, the sequence is adjusted to "Company A-Supplier Termination-500 million yuan-0", "Group B Defaulted on a Logistics Interruption-300 million yuan-14", and "Company A-Raw Material Shortage-800 million yuan-35". The final result is a timed triple sequence containing the event subject, action, amount, and timestamp.

[0032] S402: Call the time-series triple sequence to calculate the absolute value of the difference between the default amounts of adjacent events using the formula: ; Calculate the transmission correlation R between the rth event and the r-1th event r , generating cross-event correlation indicators including event pair identification and transmission correlation degree; Among them, M r is the default amount of the rth event, T ris the timestamp difference, S r is the semantic similarity of supply chain actions, λ is the industry impact factor, I r is the industry risk level index; Taking the adjacent events "Company A - Supplier Termination of Cooperation - 500 million - 0" and "Group B - Logistics Disruption - 300 million - 14" as examples, calculate the absolute value of the difference in the amount of breach of contract |3-5|=2 (unit: 100 million yuan), and the timestamp difference T r -T r-1 =14-0=14 days, semantic similarity S r The cosine similarity between "termination of cooperation with suppliers" and "logistics interruption" calculated by pre-training word vectors is 0.72, the industry impact factor is set to the manufacturing benchmark value of 0.8, and the industry risk level index is I r Take the normalized value of the historical default rate of the manufacturing industry to which Company A belongs, 0.6, and substitute it into the formula: ; Step-by-step calculation: Numerator:√14=3.7417, 2x3.7417=7.4834; Denominator: ln(1.72)=0.5433; Fractional item: 7.4834 / 0.5433=13.774; Industry item: 0.8x0.6=0.48; Final result: 13.774+0.48=14.254: The result shows that the transmission correlation of the current event pair is significantly higher than the threshold, and it needs to be used as a key transmission node in the subsequent steps to construct a risk transmission path by screening high-correlation event pairs.

[0033] S403: Call the cross-event correlation indicator to screen event pairs whose transmission correlation is higher than the transmission threshold and whose absolute value of the amount difference exceeds the amount fluctuation threshold. Connect the event pairs that meet the conditions in chronological order to generate a debt risk transmission chain with the event subject as the node; Call the cross-event correlation indicator, set the amount fluctuation threshold to 100 million yuan, the transmission threshold to 5.0, take the event pair "Company A - Supplier Termination of Cooperation - 500 million - 0" and "Company A - Raw Materials Shortage - 800 million - 35" as examples, the absolute value of the amount difference is |8-5|=300 million yuan (exceeding the threshold of 100 million yuan), and the transmission correlation degree R r=14.254 (exceeding the threshold of 5.0), connecting the two events as "Company A - supplier termination of cooperation → raw material shortage". If the transmission correlation between the event "Bank C - credit tightening - 1 billion - 50" and the aforementioned event is 4.8 (below the threshold of 5.0), then this connection is eliminated, and finally a debt risk transmission chain with the path of "Company A - supplier termination of cooperation → raw material shortage → Group B - logistics interruption" is generated. This result shows that the transmission correlation value directly determines whether the event pair is included in the transmission chain, and forms a risk diffusion path through chronological connection.

[0034] See also Figure 6 , the steps to obtain the debt risk diffusion topology results are: S501: Call the edge weight values ​​in the enterprise co-occurrence association network, the sentiment offset in the enterprise debt risk sentiment value, and the default amount difference in the debt risk transmission chain, align the identifiers of the same enterprise node in different data sources in chronological order, and superimpose the sentiment offset and default amount difference item by item according to the timestamp to generate a risk parameter superposition table containing the enterprise name, timestamp, superposition value, and edge weight value; The edge weight values ​​in the enterprise co-occurrence relationship network and the sentiment offset in the enterprise debt risk sentiment value are called. Taking the enterprises "Company A" and "Group B" as an example, the edge weight value is taken from the co-occurrence frequency of the two enterprises in the supply chain cooperation (for example, co-occurrence 5 times, weight 0.6), and the sentiment offset is extracted from the weighted result of the co-occurrence frequency of trigger words and sentiment words in the news paragraphs related to Company A (trigger word frequency 2, weight 1.2, sentiment word frequency 3, weight 1.5, time decay factor 0.8, and the calculated sentiment offset is 2×1.2 ×0.8-3×1.5×0.63=1.92-2.835=0.915), the difference in the default amount of Company A in the debt risk transmission chain is 300 million yuan (event 1: 500 million, event 2: 800 million, difference 300 million), and the sentiment offset and the amount difference are superimposed in chronological order (timestamp 0 day: 0.915+3=3.915, time stamp 35 days: 0.915+0=0.915), generating a risk parameter superposition table containing company name, timestamp, superposition value and edge weight 0.6.

[0035] S502: Based on the risk parameter overlay table, extract the edge weight values ​​of adjacent enterprise nodes, multiply the overlay values ​​by the edge weight values ​​one by one, and accumulate the product results of all related enterprise nodes in the same time window to generate an inter-node diffusion intensity set with enterprise name pairs as identifiers and numerical intensity as content; Extract the sum of the value of 3.915 for the enterprise node "Company A - Timestamp 0 days" and the edge weight of 0.6 for the adjacent node "Group B - Timestamp 14 days". Multiply the sum by the edge weight (3.915 × 0.6 = 2.349). Add up the product of all the associated nodes in the same time window (if there is a sum of 2.1 for "Group B - Timestamp 14 days" and an edge weight of 0.5, then 2.1 × 0.5 = 1.05, and the total sum is 2.349 + 1.05 = 3. 399), set the diffusion threshold to 5.0 (according to historical data statistics, it is set to 20% higher than the industry average risk intensity value of 4.2), filter the cumulative value to 3.399 (lower than the threshold of 5.0 and not included in the connection path), if the enterprise node "C Bank-timestamp 50 days" has a superposition value of 8.2 and an edge weight of 0.7, calculate 8.2×0.7=5.74 (exceeding the threshold of 5.0), and generate an inter-node diffusion intensity set including the "Company A-C Bank" intensity value of 5.74.

[0036] S503: Calling the inter-node diffusion intensity set, filtering enterprise name pairs whose intensity values ​​exceed a preset diffusion threshold in timestamp order, constructing directed connection paths with enterprises as nodes and diffusion intensity as edge weights based on the connection relationships of enterprise nodes in the co-occurrence network, and generating a debt risk diffusion topology result that includes risk transmission direction and intensity weights; The inter-node diffusion intensity set is called, and the intensity value of "Company A-Bank C" with a value of 5.74 (exceeding the threshold of 5.0) and the intensity value of "Group B-Supplier D" with a value of 6.3 (exceeding the threshold of 5.0) are screened. According to the co-occurrence network edge weight connection relationship (Company A-Bank C weight 0.7, Group B-Supplier D weight 0.8), directed edges are constructed in timestamp order (0 day: Company A→Bank C, weight 5.74, 50 days: Group B→Supplier D, weight 6.3). If there is a "Bank C-E Logistics" with an intensity value of 4.8 (below the threshold of 5.0, it will be eliminated), the effective connection paths are integrated to generate the debt risk diffusion topology results including "Company A→Bank C (5.74)" and "Group B→Supplier D (6.3)".

[0037] Enterprise risk analysis system based on news co-occurrence, the system includes: The enterprise co-occurrence network construction module obtains all enterprise names in a news paragraph, detects the first co-occurrence timestamp of a name pair, counts the co-occurrence frequency of the name pair within the paragraph, and uses the timestamp and frequency to construct a network structure with name pairs as nodes and frequencies as edge weights, thus generating an enterprise co-occurrence association network. The entity standardization classification module, based on the enterprise co-occurrence weight network, extracts the paragraph text containing the enterprise name, matches it with the financial regulatory policy keyword set, calculates the overlap ratio of keywords across paragraphs for enterprises with the same name, and classifies entities whose ratio exceeds the preset classification threshold to generate a standardized set of financial entities; The risk sentiment calculation module uses a standardized set of financial entities to extract credit default trigger words and market panic sentiment words within a paragraph, counts word frequencies and verb and adjective weight coefficients, calculates the absolute value of the paragraph sentiment shift, and assigns edge weights to name pairs based on the enterprise co-occurrence weight network to generate enterprise debt risk sentiment values. The risk transmission chain extraction module obtains the supply chain disruption action words and debt default amounts in the event sentences, arranges them into triplets in chronological order, calculates the absolute value of the difference between the amounts of adjacent events and the semantic similarity of the action words, and selects sequences whose difference exceeds the amount fluctuation threshold and whose similarity exceeds the action similarity threshold to generate the debt risk transmission chain. The risk diffusion topology generation module calls the edge weights of the enterprise co-occurrence weight network, the sentiment offset of the enterprise debt risk sentiment value, and the amount difference of the debt risk transmission chain. It superimposes the sentiment offset and the amount difference, calculates the diffusion intensity coefficient based on the edge weight, constructs the node diffusion path topology, and generates the debt risk diffusion topology result.

[0038] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. The enterprise risk analysis method based on news co-occurrence is characterized by: The following steps are involved: S1: Obtain the co-occurring company name pairs in the news paragraph, record the first co-occurrence timestamp and co-occurrence frequency, and generate a company co-occurrence association network with the name pairs as nodes and the frequency as edge weights; S2: Based on the enterprise co-occurrence network, extract the enterprise name context, match financial regulatory policy keywords, calculate the keyword overlap ratio between paragraphs of companies with the same name, classify those exceeding the threshold, and generate a standardized set of financial entities; S3: Traverse the standardized set of financial entities, extract credit default trigger words and market panic sentiment words in the news paragraphs, count the word frequencies and part-of-speech weights, calculate the absolute value of the paragraph sentiment shift, assign them to company name pairs based on co-occurrence relationships, and generate corporate debt risk sentiment values; S4: Extract supply chain disruption action words and debt default amounts from event sentences, sort triplets by time, calculate the difference in amounts between adjacent events and the semantic similarity of actions, filter out sequences exceeding the threshold, and generate a debt risk transmission chain; S5: Call the enterprise co-occurrence relationship network, enterprise debt risk sentiment value and debt risk transmission chain, superimpose the sentiment offset and amount difference, calculate the diffusion intensity in combination with the edge weight, and generate the debt risk diffusion topology result.

2. The enterprise risk analysis method based on news co-occurrence according to claim 1, characterized in that: The enterprise co-occurrence association network includes a set of enterprise name pairs, a first co-occurrence timestamp, and a co-occurrence frequency distribution. The standardized set of financial entities specifically includes a policy keyword matching set, an enterprise alias mapping table, and a paragraph overlap threshold parameter. The enterprise debt risk sentiment value includes default trigger word frequency statistics, sentiment part-of-speech weight distribution, and offset allocation rules. The debt risk transmission chain includes an event amount difference sequence, a semantic similarity score, and a transmission path identifier. The debt risk diffusion topology result covers a diffusion intensity coefficient, a topological node hierarchical distribution, and a path weight attenuation parameter.

3. The enterprise risk analysis method based on news co-occurrence according to claim 1, characterized in that: The steps for obtaining the enterprise co-occurrence association network are: S101: Traverse news text paragraphs, detect company names, extract all company name pairs within the same paragraph, record the paragraph publishing timestamp of the first appearance of each name pair, count the number of co-occurrences of each name pair, and generate a co-occurrence record of the company name pair, including the company name pair, the first co-occurrence timestamp, and the cumulative number of co-occurrences; S102: Retrieving the co-occurrence records of the enterprise name pairs, merging duplicate name pair records, retaining the earliest timestamp of each name pair as a time reference value, accumulating all co-occurrence times as edge weight values, and generating a co-occurrence frequency of the enterprise name pairs including unique name pairs, edge weight values, and time reference values; S103: Based on the co-occurrence frequency of the enterprise name pair, the two enterprises in the enterprise name pair are set as network nodes, the edge weight value is called as the edge weight parameter between the nodes, the first co-occurrence time of each edge is marked according to the time reference value, the node and edge data are integrated, and an enterprise co-occurrence association network is generated with the enterprise name as the node and the co-occurrence frequency as the edge weight.

4. The enterprise risk analysis method based on news co-occurrence according to claim 1, characterized in that: The steps for obtaining the standardized set of financial entities are: S201: Retrieving enterprise names in the enterprise co-occurrence association network, extracting a fixed range of text segments before and after the paragraph containing the name, storing the text segments of all paragraph instances, and generating a context segment set containing the enterprise name and the corresponding text segment; S202: Based on the context segment set, a financial regulatory policy keyword library containing terms such as antitrust review and listing compliance is preset, and the number of matching keywords in each company name-related paragraph is counted using the formula: ； Calculate the keyword overlap value KSR of company name i in paragraph j i,j , generating cross-paragraph keyword matching including company name, paragraph identifier and keyword overlap value; Among them, F k is the standard frequency of keywords in policy documents, C i is the total frequency of co-occurrence of enterprises, L j is the length of the paragraph character, ∈ is a very small constant, W n is the regulatory level weight of the keyword, O n is the number of times the keyword appears in the paragraph; S203: Calling the cross-paragraph keyword matching degree, calculating the variance-mean ratio of the keyword overlap values ​​of the same-named enterprises, screening the name classifications below the fluctuation threshold and above the merging threshold, and generating a standardized set of financial entities of the standardized name and alias set.

5. The enterprise risk analysis method based on news co-occurrence according to claim 1, characterized in that: The steps for obtaining the corporate debt risk sentiment value are as follows: S301: Traverse the company names in the standardized set of financial entities, collect related news paragraph texts, match them with the credit default trigger vocabulary and the market panic sentiment vocabulary, count the co-occurrence times of the two types of words in the paragraphs, and generate a risk co-occurrence record containing the company name, paragraph identifier, and the co-occurrence frequency of trigger words and sentiment words; S302: Call the risk co-occurrence record, based on the trigger word part-of-speech weight table and the emotional word intensity level table, using the formula: ； Calculate the absolute value of sentiment shift EOS of enterprise i in paragraph j i,j , generating a cross-paragraph sentiment offset value including the company name, paragraph identifier and the absolute value of sentiment offset; Among them, F t is the co-occurrence frequency of trigger words, W t is the part-of-speech weight, λ is the time decay factor, C s is the co-occurrence frequency of emotional words, I s is the emotion intensity weight, D j is the dispersion of emotional tone of the paragraph; S303: Call the cross-paragraph sentiment offset value, assign the sentiment offset value according to the edge weight ratio of the enterprise name pair in the enterprise co-occurrence association network, accumulate all paragraph assignment values, and generate the enterprise debt risk sentiment value identified by the enterprise name pair.

6. The enterprise risk analysis method based on news co-occurrence according to claim 1, characterized in that: The steps for obtaining the debt risk transmission chain are: S401: Traverse the news event sentences, extract the subject-verb-object triples containing the supply chain disruption action and the debt default amount value, arrange the triples in ascending order by timestamp, and generate a time-series triple sequence containing the event subject, action description, amount value and timestamp; S402: Call the time-series triple sequence to calculate the absolute value of the difference between the default amounts of adjacent events using the formula: ； Calculate the transmission correlation R between the rth event and the r-1th event r , generating cross-event correlation indicators including event pair identification and transmission correlation degree; Among them, M r is the default amount of the rth event, T r is the timestamp difference, S r is the semantic similarity of supply chain actions, λ is the industry impact factor, I r is the industry risk level index; S403: Call the cross-event correlation indicator to screen event pairs whose transmission correlation is higher than the transmission threshold and whose absolute value of the amount difference exceeds the amount fluctuation threshold, connect the event pairs that meet the conditions in chronological order, and generate a debt risk transmission chain with the event subject as the node.

7. The enterprise risk analysis method based on news co-occurrence according to claim 1 is characterized by: The steps for obtaining the debt risk diffusion topology result are: S501: Calling the edge weight values ​​in the enterprise co-occurrence association network, the sentiment offset in the enterprise debt risk sentiment value, and the default amount difference in the debt risk transmission chain, aligning the identifiers of the same enterprise node in different data sources in chronological order, and superimposing the sentiment offset and the default amount difference item by item according to the timestamp to generate a risk parameter superposition table including the enterprise name, timestamp, superposition value, and edge weight value; S502: Based on the risk parameter overlay table, extract the edge weight values ​​of adjacent enterprise nodes, multiply the overlay values ​​by the edge weight values ​​item by item, and accumulate the product results of all related enterprise nodes in the same time window to generate an inter-node diffusion intensity set with enterprise name pairs as identifiers and numerical intensity as content; S503: Call the inter-node diffusion intensity set, filter the enterprise name pairs whose intensity values ​​exceed the preset diffusion threshold in timestamp order, and construct a directed connection path with the enterprise as the node and the diffusion intensity as the edge weight based on the connection relationship of the enterprise nodes in the co-occurrence network, and generate a debt risk diffusion topology result including the risk transmission direction and intensity weight.

8. The enterprise risk analysis system based on news co-occurrence is characterized by: The system is used for the enterprise risk analysis method based on news co-occurrence according to any one of claims 1 to 7, and the system includes: The enterprise co-occurrence network construction module obtains all enterprise names in a news paragraph, detects the first co-occurrence timestamp of a name pair, counts the co-occurrence frequency of the name pair within the paragraph, and uses the timestamp and frequency to construct a network structure with name pairs as nodes and frequencies as edge weights, thus generating an enterprise co-occurrence association network. The entity standardization classification module extracts the paragraph text containing the enterprise name based on the enterprise co-occurrence weight network, matches it with the financial regulatory policy keyword set, calculates the overlap ratio of keywords across paragraphs for enterprises with the same name, and classifies entities whose ratio exceeds a preset classification threshold to generate a standardized set of financial entities; A risk sentiment calculation module calls the standardized set of financial entities, extracts credit default trigger words and market panic sentiment words in the paragraph, counts word frequencies and verb and adjective weight coefficients, calculates the absolute value of the paragraph sentiment offset, and assigns the edge weights of the enterprise co-occurrence weight network to the name pairs to generate the enterprise debt risk sentiment value; The risk transmission chain extraction module obtains the supply chain disruption action words and debt default amounts in the event sentences, arranges them into triplets in chronological order, calculates the absolute value of the difference between the amounts of adjacent events and the semantic similarity of the action words, and selects sequences whose difference exceeds the amount fluctuation threshold and whose similarity exceeds the action similarity threshold to generate the debt risk transmission chain. The risk diffusion topology generation module calls the edge weights of the enterprise co-occurrence weight network, the sentiment offset of the enterprise debt risk sentiment value, and the amount difference of the debt risk transmission chain, superimposes the sentiment offset and the amount difference, calculates the diffusion intensity coefficient based on the edge weight, constructs the node diffusion path topology, and generates the debt risk diffusion topology result.

Images