Financial risk AI dynamic assessment method

Abstract

The invention relates to the technical field of financial risk management and artificial intelligence, and discloses a financial risk AI dynamic assessment method, which comprises the following steps: constructing an enterprise DNA map representing the internal structure of an enterprise; extracting concept atoms from external data, and calculating a semantic speed and a semantic acceleration between the concept atoms; the high-momentum narrative is recognized based on the semantic acceleration; aligning the high momentum narrative with the enterprise DNA map for polarity calibration, and judging risks or opportunities; and quantitative scoring and evolution stage determination are carried out on the risk. According to the invention, by calculating the semantic acceleration, early warning of potential risks is realized, and the hysteresis problem in the prior art is solved; the external narration is aligned with the internal DNA atlas, so that the pertinence and operability of an evaluation result are ensured, and the defect of evaluation generalization is overcome; through generation of quantitative risk scores and evolution stages, dynamic quantitative tracking and severity comparison of risks are realized, and defects of static and qualitative analysis in the prior art are made up.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of financial risk management and artificial intelligence, in particular to a financial risk AI dynamic evaluation method. BACKGROUND

[0002] Existing enterprise financial risk evaluation methods increasingly attempt to utilize public market information and unstructured data in order to gain early insights into potential risks. However, these existing technologies have several limitations in practical application.

[0003] Firstly, existing methods exhibit significant lag in risk identification. They typically rely on statistical analysis of the frequency of occurrence of negative sentiment or specific keywords. This approach can only identify risk narratives after they have already formed and market sentiment has already shifted definitively, missing the optimal opportunity for intervention at the risk's nascent stage.

[0004] Secondly, existing technologies generally lack effective correlation with specific internal circumstances of enterprises when evaluating risks. Their analysis of external public opinion is generalized and fails to align and calibrate the identified market signals with the enterprise's unique core technology, strategic layout, or inherent vulnerabilities. This results in evaluation results that are often vague and unable to provide targeted and executable risk response criteria for the enterprise.

[0005] Furthermore, existing evaluation results are mostly static and qualitative descriptions, making it difficult to achieve quantitative comparison and dynamic tracking of different risks. They cannot provide a unified quantitative score to measure the severity of risks or reveal the evolution stage of risks, such as whether a risk is in the early generation stage or approaching the development stage of an outbreak. This deficiency fundamentally stems from the lack of exploration and utilization of the hidden deep semantic momentum in unstructured data, particularly the rate of change in concept association strength (i.e., acceleration). SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a financial risk AI dynamic evaluation method, which solves the problems of lag in risk identification, lack of correlation between evaluation results and specific internal circumstances of enterprises, lack of targetedness, and difficulty in dynamic quantification and evolution stage determination of risks in the prior art.

[0007] To achieve the above purpose, the present application is implemented by the following technical scheme: a financial risk AI dynamic evaluation method, which specifically comprises the following steps:

[0008] Step one: concept atom extraction

[0009] Unstructured text data from multiple information sources is acquired and parsed using a natural language processing model to extract standardized concept atoms, which are used to refer to entities, events, or attributes.

[0010] Step 2: Semantic Momentum Calculation

[0011] To quantify the dynamic evolution of the relationships between conceptual atoms, perform the following calculations:

[0012] First, calculate any two conceptual atoms A i and A j The co-occurrence intensity C(A) at time point t i A j This calculation is obtained by weighting and summing the logarithmic values ​​of the co-occurrence frequencies of concept atoms and the logarithmic values ​​of the number of independent information sources. The formula is as follows:

[0013] C(A i A j ,t)=α·log(1+f count (A i A j ,t))+β·log(1+d source (A i A j ,t));

[0014] Among them, f count (A i A j ,t) represents A within a preset time window i and A j co-occurrence frequency; d source (A i A j ,t) represents A within the same time window i and A j The number of independent information sources that co-occur; α and β are preset weighting coefficients; log(1+x) is a logarithmic smoothing function used to reduce the disproportionate impact of extremely high frequency or number of sources.

[0015] Subsequently, based on the change in co-occurrence intensity over time, the semantic speed V is calculated. sm (A i A j (t) and semantic acceleration A sm (A i A j The semantic velocity is obtained by performing a first-order difference calculation in the time dimension on the co-occurrence intensity, and the semantic acceleration is obtained by performing a first-order difference calculation in the time dimension on the semantic velocity. The calculation formula is as follows:

[0016]

[0017] Among them, C(A) i A j C(A,t) represents the co-occurrence intensity at the current time point; i A j ,t-Δt) represents the co-occurrence intensity at the previous time point; Δt represents the time step used for calculation.

[0018]

[0019] Among them, V sm (A i A j ,t) represents the semantic velocity at the current time point; V sm (A i A j ,t-Δt) represents the semantic velocity at the previous time point; Δt represents the time step of the calculation.

[0020] Step 3: Construction of the Enterprise DNA Map

[0021] For a specific target company, obtain its financial reports, strategic planning documents, and other exclusive information to construct a unique corporate DNA profile (G) for that target company. dna The graph contains a node set N. dna and edge set E dna Nodes define the core concepts inherent to the target company, such as core technologies, key suppliers, or strategic objectives; edges define the inherent logical relationships between nodes. These inherent logical relationships are determined based on the financial reports and strategic planning documents, and specifically include dependency relationships, support relationships, or competitive relationships.

[0022] Step 4: High-momentum narrative recognition

[0023] Based on the calculated semantic acceleration, a high-momentum narrative is identified. Structurally, the high-momentum narrative is a subgraph composed of multiple conceptual atoms, which are connected by edges with semantic accelerations higher than a preset threshold.

[0024] Step 5: Polarity Calibration

[0025] First, the conceptual atoms within the high-momentum narrative are aligned with the nodes of the corporate DNA map. This alignment step specifically involves calculating the semantic similarity between the conceptual atoms within the high-momentum narrative and each node in the corporate DNA map, and selecting the node with the highest semantic similarity as the alignment node.

[0026] Subsequently, the alignment result and the inherent logical relationship of the aligned nodes are obtained, and polarity calibration is performed on the high-momentum narrative based on the result and relationship to obtain a polarity attribute P. cal The calibration step specifically involves: for each concept atom A within the high-momentum narrative... k According to its corresponding alignment node n k The inherent logical relationship between the node and Rel(n) k Determine a unipolar value Pol(A). k ,n k ,Rel(n k Then, a weighted sum is performed on all unipolar values. The calculation formula is as follows:

[0027]

[0028] Where S represents high-momentum narrative; w k For the concept atom A k The weights; This represents each conceptual atom A in the narrative S. k Perform traversal calculations and sum the results; Pol(...) is a polarity determination function. In a specific setting, the unipolar value is negative when the intrinsic logical relationship is a dependency relationship; the unipolar value is positive when the intrinsic logical relationship is a competition relationship.

[0029] Step Six: Generation of Evaluation Results

[0030] A financial risk assessment result is generated by combining the polarity attribute, the identifier of the high-momentum narrative, the alignment specification, the current value of the semantic velocity, and the current value of the semantic acceleration. Furthermore, the evolutionary stage of the high-momentum narrative is determined based on the current values ​​of the semantic velocity and semantic acceleration, and this evolutionary stage is included in the financial risk assessment result.

[0031] This invention provides an AI-based dynamic assessment method for financial risk. It offers the following advantages:

[0032] 1. This invention, by calculating the semantic speed and semantic acceleration of the correlation between conceptual atoms, can identify and quantify the evolution trend of information narrative in the nascent or rapid fermentation stage, realizing dynamic monitoring and forward-looking assessment of financial risks, and effectively overcoming the technical defects of traditional methods that cannot respond to market dynamics in a timely manner due to reliance on lagging financial data.

[0033] 2. This invention constructs a corporate DNA map for a specific enterprise and performs polarity calibration on this basis, linking the macro-level high-momentum narrative with the specific enterprise's internal business structure, supply chain dependence, and market competition relationship. This achieves personalized risk assessment and can accurately distinguish the differentiated impact of the same external event on different enterprises' risks or opportunities, thus solving the problem of homogenized risk assessment results in the prior art.

[0034] 3. The financial risk assessment results generated by this invention not only provide polarity attributes representing the direction of risk or opportunity, but also include semantic speed and semantic acceleration reflecting narrative dynamics, as well as traceable alignment node descriptions. This provides quantitative and multi-dimensional data support for risk decision-making, making the assessment conclusions have clear technical interpretability. Attached Figure Description

[0035] Figure 1 This is a block diagram of the financial risk AI dynamic assessment system architecture of the present invention;

[0036] Figure 2 This is a schematic diagram of the corporate DNA map structure of the present invention. Detailed Implementation

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] See attached document Figure 1 , Figure 1 This is a block diagram of a financial risk AI dynamic assessment system architecture according to an embodiment of the present invention. The financial risk AI dynamic assessment method provided by the present invention can be executed by an assessment device deployed on a computer system.

[0039] In one specific embodiment, the computer system includes at least one processor, a memory, one or more graphics processing units (GPUs), and a network interface. The processor executes computer program instructions to control the overall flow of the method. The GPU performs parallel computation to accelerate the training and inference processes of the deep learning model involved in subsequent steps. The memory stores program code and intermediate and final data generated during method execution. The network interface connects to an external network to perform data acquisition.

[0040] The method's execution environment is deployed on one or more servers running a Linux distribution, such as Ubuntu Server 20.04 LTS. The method's implementation depends on a specific software environment, which includes libraries to support the various technical steps.

[0041] The software environment includes natural language processing libraries, such as spaCy or NLTK, for performing text processing tasks such as word segmentation, part-of-speech tagging, and named entity recognition in subsequent steps. It also includes deep learning frameworks, such as PyTorch or TensorFlow, for building, training, and deploying language models based on the Transformer architecture. Furthermore, the software environment includes graph computing libraries, such as NetworkX, or graph databases, such as Neo4j, for building, storing, and analyzing the enterprise DNA graph generated in subsequent steps.

[0042] To achieve the complete process of the method of this invention, the evaluation device can be logically divided into multiple functional modules. (Refer to...) Figure 1 The device includes: a data acquisition module 101, a data preprocessing module 102, a concept atom extraction module 103, an enterprise DNA map construction module 104, a semantic momentum calculation module 105, a high momentum narrative recognition module 106, a polarity calibration module 107, and an evaluation result generation module 108.

[0043] The data acquisition module 101 connects to multiple preset external information sources via a network interface and periodically captures unstructured text data according to a preset acquisition strategy. The data preprocessing module 102 receives data from the data acquisition module 101 and performs data cleaning, deduplication, and format standardization operations, transmitting the processed data to the concept atom extraction module 103. Subsequent modules are called sequentially and perform data transmission until the evaluation result generation module 108 outputs the final financial risk assessment result.

[0044] The first step in the AI-driven dynamic assessment method for financial risk is to acquire and preprocess unstructured text data. This step is performed collaboratively by the data acquisition module 101 and the data preprocessing module 102.

[0045] In one specific embodiment, the data acquisition module 101 is responsible for collecting raw data from the Internet. The data acquisition module 101 internally contains a list of information sources, which includes multiple preset Uniform Resource Locators (URLs). These URLs point to different types of information publishing platforms, including but not limited to national-level financial news portals, business news sections of international news agencies, local government policy announcement websites, and social media platforms focusing on business and financial topics.

[0046] The data acquisition module 101 performs a data collection task at a preset time period (e.g., every 5 minutes). In each task, the data acquisition module 101 initiates an HTTP request to a URL in the information source list via a network interface, or calls its provided application programming interface (API), to obtain the latest text content and its metadata, including the publication timestamp and the source URL.

[0047] The data acquisition module 101 transmits the acquired raw data stream to the data preprocessing module 102. The data preprocessing module 102 is responsible for standardizing the data for use by subsequent data preprocessing modules. This process includes several sub-steps.

[0048] First, data cleaning is performed. This step removes non-text content from the original data, such as Hypertext Markup Language (HTML) tags, Cascading Style Sheets (CSS) code, and JavaScript. Simultaneously, all text content is converted to UTF-8 encoding to avoid garbled characters during subsequent processing.

[0049] Secondly, data deduplication is performed. For each cleaned text data, the data preprocessing module 102 calculates a hash value for its core content, such as using the MD5 algorithm or the SimHash algorithm. The data preprocessing module 102 maintains a set of hash values ​​of processed data. When processing new data, if the calculated hash value already exists in this set, the new data is determined to be duplicate data and discarded; otherwise, its hash value is stored in the set, and the data is retained.

[0050] Finally, data structuring is performed. For each cleaned and deduplicated valid data entry, the data preprocessing module 102 organizes it into a structured data record. This record contains the following fields: a unique identifier (ID), the original source URL, a publication timestamp accurate to the second, the article title, and the article content. After processing, the data preprocessing module 102 outputs these structured data records to the concept atom extraction module 103.

[0051] The method provided by this invention proceeds to the concept atom extraction step after data preprocessing. This step is executed by the concept atom extraction module 103, which first performs entity recognition. The input for entity recognition is the structured data record output by the data preprocessing module 102.

[0052] In one specific embodiment, the concept atom extraction module 103 internally integrates a specifically trained Named Entity Recognition (NER) model. The NER model is based on a Transformer architecture, such as the BERT model, which is pre-trained on a large-scale general text corpus and subsequently fine-tuned on a manually annotated named entity dataset in the financial and business domains.

[0053] The goal of the NER model is to assign a predefined entity label to each token in the input text. This embodiment uses the BIOES labeling system, where B represents the start of an entity, I represents the interior of an entity, O represents a non-entity portion, E represents the end of an entity, and S represents an entity consisting of a single token. Predefined entity categories include, but are not limited to: ORG (Organization), PER (Person), LOC (Location), PROD (Product or Service), and TECH (Technology).

[0054] When the concept atom extraction module 103 receives a structured data record, it first extracts its title and body content. Then, it uses a lexical analyzer compatible with the NER model to segment the text into a lexical sequence. This lexical sequence is converted into an input format acceptable to the model, namely a tensor containing an input ID, an attention mask, and a lexical type ID.

[0055] The input tensor is fed into the NER model for forward propagation computation. The last layer of the model is a fully connected layer, followed by a softmax activation function. For each word in the sequence, the model outputs a probability distribution vector, the dimension of which is equal to the total number of predefined entity labels. The label with the highest probability is selected as the predicted label for that word, thus generating a label sequence of the same length as the input word sequence.

[0056] Next, the concept atom extraction module 103 decodes the output tag sequence. Based on the BIOES tag combination rules, the concept atom extraction module 103 merges consecutive B, I, and E tag tokens with the same entity category to reconstruct the complete entity text span. For example, tokens corresponding to a tag sequence [B-ORG, I-ORG, E-ORG] will be merged into an entity of type ORG.

[0057] To achieve entity standardization, the concept atom extraction module 103 is also configured with an entity knowledge base. For each identified entity text, the concept atom extraction module 103 queries the entity knowledge base. This knowledge base stores a mapping between standard entity names and their aliases (e.g., International Business Machines Corporation and IBM) and unique canonical entity identifiers (Canonical Entity IDs). If the query is successful, the canonical entity identifier is assigned to the identified entity; if the query fails, a new temporary identifier is assigned to it.

[0058] Finally, the concept atom extraction module 103 outputs a data object annotated with entities. This object adds a list of entities to the original structured data record. Each element in the list represents an identified entity and contains the following information: the entity text, entity category, start and end positions in the text, and a normalized entity identifier. This data object is then used to perform event extraction and attribute extraction.

[0059] After entity recognition is completed, the concept atom extraction module 103 continues to perform event extraction. The input to this step is the data object with completed entity annotation, which is the output of the previous sub-step. The goal of event extraction is to identify predefined event types from the text and extract the arguments that constitute the event.

[0060] In one specific embodiment, the event extraction task is performed by a specially trained event extraction model. This model is also based on the Transformer architecture and employs a two-stage processing flow: the first stage is event trigger word identification, and the second stage is event argument role labeling. The model is fine-tuned on a corpus containing labeled financial events to identify financially relevant event types.

[0061] Predefined event types are organized in an event schema library. Each event type contains an event type identifier, a trigger word thesaurus, and an argument role list. For example, for the corporate merger and acquisition event type, its trigger word thesaurus includes words such as acquisition, merger, and consolidation, and its argument role list includes the acquiring party, the acquired party, the transaction amount, and the transaction time.

[0062] During event extraction, the concept atom extraction module 103 first inputs the entity-annotated text sequence into the first stage of the event extraction model. This stage is responsible for identifying words or phrases in the text that serve as event triggers. The model classifies each lexical unit in the sequence to determine whether it belongs to a trigger word of a predefined event type. If a lexical unit is identified as a trigger word, the lexical unit and its corresponding event type are marked, triggering the second stage of processing.

[0063] In the second stage, the event argument role labeling stage, the model analyzes the relationship between each identified event trigger word and all identified entities in the text. For each trigger word entity pair, the model performs classification judgment, assigning an argument role corresponding to the event type to the entity, or determining that it has no direct relationship with the event.

[0064] For example, given the text "Company A announced yesterday that it will acquire Company B for $500 million," in the previous sub-step, Company A and Company B were identified as ORG (Organizational Group) entities, and $500 million was identified as a MONEY (Currency) entity. In this step, the acquisition is identified as the trigger word for the corporate merger event by the first-stage model. Subsequently, the second-stage model labels Company A as the acquirer, Company B as the acquired party, and $500 million as the transaction amount.

[0065] After processing, the concept atom extraction module 103 constructs a structured event record for each successfully extracted event. This record, as a concept atom of an event class, contains the following fields: a unique event instance ID, an event type identifier, the original text of the trigger word, and a list of arguments. Each element in the argument list contains an argument role and a canonical entity identifier pointing to the entity of that role.

[0066] Finally, the concept atom extraction module 103 passes the data object containing the entity list and event list to the subsequent attribute extraction sub-step.

[0067] After the event extraction is complete, refer to the appendix. Figure 1 The concept atom extraction module 103 then continues to perform the attribute extraction and sentiment analysis tasks. The input to this task is the data object from the previous sub-step, which already contains entity and event annotations. Its goal is to identify the attributes describing entities or events and quantify the sentiment tendency of the relevant text.

[0068] In one specific embodiment, attribute extraction is achieved through dependency parsing. The concept atom extraction module 103 uses a dependency parser that processes sentences containing identified entity or event-triggered words. The parser generates a dependency tree for each word in the sentence, revealing the grammatical relationships between words.

[0069] The concept atom extraction module 103 traverses the dependency tree, searching for specific dependency relations pointing to identified entity or event trigger words. For example, for an entity noun, the concept atom extraction module 103 searches for dependency relations of type adjective modifier or noun modifier. When such a relation is found, the modifier it points to (e.g., innovative, high-risk) is extracted as a concept atom of attribute class. This attribute atom is recorded and associated with the canonicalized identifier of the entity or event it modifies.

[0070] Simultaneously with or after attribute extraction, the concept atom extraction module 103 performs sentiment analysis. The concept atom extraction module 103 integrates a text sentiment classification model based on the Transformer architecture, fine-tuned on a large-scale sentiment-annotated corpus in the financial field. This model is capable of classifying input text fragments into three categories: positive, negative, or neutral.

[0071] For each sentence containing an identified event, or a sentence describing a key entity, the sentence text is fed as input to the sentiment classification model. After forward propagation, the model generates a logit value for each of the three categories (positive, negative, and neutral) at its output layer.

[0072] To obtain a continuous sentiment score, the concept atom extraction module 103 first applies the Softmax function to the three log-probability values, converting them into corresponding probability values:

[0073] (p pos ,p neg ,p neu ) = Softmax(l pos ,l neg ,l neu );

[0074] Among them, l pos ,l neg ,l neu These are the log-odds values ​​for the positive, negative, and neutral categories, respectively, p. pos ,p neg ,p neu These are the corresponding probability values, and the sum of the three is 1; subsequently, a standardized sentiment score S is calculated based on the probability values. sent The calculation formula is as follows:

[0075] S sent =p pos -p neg ;

[0076] The emotional score S sentThe value ranges from [-1, 1]. A value of -1 indicates extreme negativity, a value of 1 indicates extreme positivity, and a value close to 0 indicates neutrality.

[0077] Calculated sentiment score S sent This is appended to the event or entity records contained in the sentence as a quantified attribute. At this point, the concept atom extraction step is complete. The concept atom extraction module 103 ultimately outputs a fully labeled data object containing all three types of concept atoms (entities, events, and attributes) extracted from the original text, their interrelationships, and the sentiment scores associated with these atoms. This data object is then transferred to the enterprise DNA mapping module 104 and the semantic momentum calculation module 105 for further processing.

[0078] After completing the conceptual atom extraction of general unstructured text data, the method proceeds to the enterprise DNA map construction step. This step is performed by the enterprise DNA map construction module 104. The first step of the construction process is to collect its specific deep data source for a given target enterprise.

[0079] In one specific embodiment, the enterprise DNA mapping construction module 104 first receives a unique identifier for a target enterprise, such as the company's stock code or unified social credit code. Based on this identifier, the enterprise DNA mapping construction module 104 initiates a dedicated data collection procedure.

[0080] The program first accesses publicly available databases of securities regulatory agencies, such as the EDGAR database of the U.S. Securities and Exchange Commission or disclosure platforms designated by the China Securities Regulatory Commission. From these databases, the program downloads all periodic reports of the target company for the past few financial reporting periods (e.g., the past five years), including annual reports (such as 10-K filings) and quarterly reports (such as 10-Q filings).

[0081] After acquiring these reports, the program automatically parses their structure and extracts key content from specific sections. These extracted sections include, but are not limited to: business descriptions, risk factors, and management discussions and analyses. These sections contain a detailed explanation of the company's core business, known risks, competitive landscape, and future strategies.

[0082] Simultaneously, the data collection program also accesses the investor relations section of the target company's official website. From this section, the program downloads the company's strategic planning documents, annual shareholder meeting presentations, and the latest investor day materials. These documents provide a direct explanation of the company's medium- and long-term strategic goals and key initiatives.

[0083] In addition, to obtain information on a company's technology portfolio, the data collection program connects to national or regional patent databases, such as the public databases of the United States Patent and Trademark Office or the China National Intellectual Property Administration. The program uses the target company's name as a keyword to search and retrieves a list of all its published patent applications and granted patents. For each patent, the program downloads its complete specification text.

[0084] Finally, the Enterprise DNA Mapping Module 104 integrates and structures all the collected proprietary data, attaching metadata to each data source (e.g., a specific annual report or a patent document), including the publication date and source identifier. This integrated proprietary data set will serve as input for subsequent node and edge definition sub-steps.

[0085] After collecting the target company's proprietary data source, the enterprise DNA mapping module 104 continues with the sub-steps of defining and generating nodes and edges. The input to this step is the proprietary data set collected and integrated in the previous sub-step.

[0086] In one specific embodiment, the enterprise DNA map construction module 104 first performs information extraction on the text content in the dedicated dataset to identify nodes representing the enterprise's core concepts. The enterprise DNA map construction module 104 applies key phrase extraction algorithms, such as the statistical TextRank algorithm or the lexical pattern matching method, to extract frequently occurring noun phrases with clear business meanings from sections such as business descriptions, risk factors, and management discussions and analyses.

[0087] For example, phrases such as dependence on a single supplier or semiconductor supply chain disruptions are extracted from the risk factors section as nodes of the vulnerability type. Phrases such as 5nm process technology or AI chip architecture are extracted from business descriptions and patent documents as nodes of the core technology type. Expanding market share in Europe is extracted from strategic planning documents as a node of the strategic objective type. Each extracted phrase is created as a unique graph node and assigned a label containing its textual description.

[0088] Next, the Enterprise DNA Mapping Module 104 establishes edges representing inherent logical relationships between the identified nodes. This process is achieved by applying predefined syntactic and semantic rules within a sentence or paragraph. The rule base defines how specific words or phrases indicate different types of logical relationships.

[0089] To establish dependencies, the Enterprise DNA Mapping Module 104 searches the text for patterns that connect two nodes, such as [Node A]...depends on...[Node B] or [Node B]...is a key supplier of...[Node A]. When a match is found, the Enterprise DNA Mapping Module 104 creates a directed edge of type dependency from node A to node B.

[0090] To establish support relationships, the Enterprise DNA Mapping Module 104 searches for patterns like [Node A]...supports...[Node B] or [Node A]...improves...[Node B]'s performance. When a match is successful, the Enterprise DNA Mapping Module 104 creates a directed edge of type support from Node A to Node B.

[0091] To establish competitive relationships, the Enterprise DNA Mapping Module 104 searches for patterns like "[Node A]... competes with...[Node B]" or "[Our Product A]... is...[Competitor Product B]". When a match is found, the Enterprise DNA Mapping Module 104 creates an undirected edge or two bidirectional edges connecting Node A and Node B, indicating a competitive relationship.

[0092] See attached document Figure 2 , attached Figure 2 This is a schematic diagram of a corporate DNA map structure according to an embodiment of the present invention. Ultimately, all identified nodes and the edges between them are loaded into a graph database. This graph database physically stores the corporate DNA map of the target company. This map is formally represented as G. dna =(N dna E dna ), where N dna It is the set of all defined nodes, E dna It is a collection of all established edges with type attributes. The completed enterprise DNA map is stored for use in subsequent polarity calibration steps.

[0093] The method provided by this invention, after extracting concept atoms, proceeds to a semantic momentum calculation step. This step is executed by the semantic momentum calculation module 105, whose primary task is to calculate the co-occurrence strength between concept atoms. The input to this step is all the labeled data objects output by the concept atom extraction module 103.

[0094] In one specific embodiment, the semantic momentum calculation module 105 discretizes the time axis in units of a fixed time window Δt (e.g., 24 hours) to obtain a series of continuous time steps t. For any two concept atoms A extracted from different texts... i and A jIf they appear in the same sentence of the same article, they are counted as one co-occurrence.

[0095] For each time step t, the semantic momentum calculation module 105 iterates through all labeled data objects within that time step and performs a semantic momentum calculation for each pair of concept atoms (A i A j Calculate two basic indicators. The first indicator is the co-occurrence frequency C(A). i A j ,t), that is, within time step t, the concept atom A i and A j The total number of sentences that co-occur. The second indicator is the number of independent information sources, S(A). i A j ,t), that is, within time step t, it contains A i and A j The total number of independent information sources (counted by unique source URL domains) for co-occurring sentences.

[0096] Based on the two fundamental indicators mentioned above, the semantic momentum calculation module 105 calculates the concept atom pair (A) at time step t. i A j The co-occurrence intensity I(A) i A j This calculation is performed by weighting and summing the logarithm of the co-occurrence frequency and the logarithm of the number of independent information sources. The specific formula is as follows:

[0097] I(A i A j ,t)=w c ·log(C(A i A j ,t)+1)+w s ·log(S(A i A j ,t)+1);

[0098] Among them: I(A i A j ,t) is the concept atom A at time step t. i and A j Co-occurrence intensity between them; C(A) i A j ,t) is A at time step t. i and A j Co-occurrence frequency; S(A i A j (,t) is the observation of A at time step t. i and A jThe number of co-occurring independent information sources; log is the natural logarithm function. Adding 1 to the parameter ensures that the function's input is always positive; w c and w s These are preset weighting coefficients, both of which are non-negative and sum to 1. For example, in one embodiment, w can be set... c =0.4 and w s =0.6.

[0099] The semantic momentum calculation module 105 repeats the above calculation for all concept atom pairs that co-occur at least once in any time step. The result is for each pair of concept atoms (A... i A j Generate a time series of co-occurrence intensities {I(A)} i A j ,t1),I(A i A j The time series data is stored and used as input for subsequent semantic velocity and semantic acceleration calculations within the same semantic momentum calculation module 105.

[0100] After calculating the time series of co-occurrence intensities, the semantic momentum calculation module 105 continues to calculate semantic velocity and semantic acceleration. The input to this calculation is the time series generated in the previous sub-step for each pair of concept atoms (A...). i A j The co-occurrence intensity time series of {I(A)} i A j ,t)}.

[0101] In one specific embodiment, semantic velocity is calculated by performing a first-order difference in the time dimension on the co-occurrence intensity time series. For any time step t, the concept atom pair (A i A j The semantic speed V(A) i A j The formula for calculating t is:

[0102] V(A i A j ,t)=I(A i A j ,t)-I(A i A j ,t-1);

[0103] Where: V(A) i A j ,t) is the concept atom pair (A) at time step t. i A j The semantic speed of I(A) i Aj I(A,t) is the co-occurrence intensity at time step t; i A j ,t-1) is the co-occurrence intensity at the previous time step t-1. This calculation is performed when t>1.

[0104] Next, semantic acceleration is calculated by performing a first-order difference in the time dimension on the obtained semantic velocity time series. For any time step t, the concept atom pair (A i A j The semantic acceleration a(A) i A j The formula for calculating t is:

[0105] a(A i A j ,t)=V(A i A j ,t)-V(A i A j ,t-1);

[0106] Among them: a(A i A j ,t) is the concept atom pair (A) at time step t. i A j Semantic acceleration of V(A); i A j V(A,t) is the semantic velocity at time step t; i A j ,t-1) is the semantic velocity at the previous time step t-1. This computation is performed when t>2.

[0107] The semantic momentum calculation module 105 repeats the above calculation for all concept atom pairs, thereby generating a time series of semantic velocity and semantic acceleration for each concept atom pair. At the current latest time step t... current The semantic momentum calculation module 105 obtains each pair of concept atoms (A i A j The current semantic speed value V(A) i A j ,t current ) and the current semantic acceleration value a(A i A j ,t current ).

[0108] After the calculation is completed, the semantic momentum calculation module 105 outputs the current semantic acceleration values ​​of all concept atom pairs, along with the set of all concept atoms, to the high momentum narrative recognition module 106 for further processing.

[0109] After calculating semantic velocity and semantic acceleration, the method proceeds to the high-momentum narrative recognition step. This step is executed by the high-momentum narrative recognition module 106. Its input is the output of the semantic momentum calculation module 105, representing the current time step t. current The semantic acceleration values ​​of all concept atom pairs, and the set of all concept atoms.

[0110] In one specific embodiment, the high-momentum narrative recognition module 106 first constructs an undirected graph based on semantic acceleration values, which is referred to as the acceleration graph G. accel The set of nodes N in the acceleration graph accel That is, the set of all conceptual atoms.

[0111] The set of edges E of the acceleration graph accel The generation is based on a preset acceleration threshold a. threshold For any two concept atomic nodes A i and A j If and only if they are at the current time step t current The semantic acceleration value a(A) i A j ,t current () greater than threshold a threshold When the two nodes are in a certain state, an undirected edge is created between them. This set of edges can be formally defined as:

[0112] E accel ={(A i A j )∣a(A i A j ,t current )>a threshold};

[0113] Among them, a threshold It is a pre-defined positive parameter; for example, the threshold could be set to the 95th percentile of all non-zero semantic acceleration values ​​at the current time step to filter out atom pairs with significant acceleration; a(A i A j ,t current This is the core criterion for screening in this formula. It represents the concept atom pair (A) i A j At the current time step t current The semantic acceleration value; a threshold This symbol represents a pre-set positive acceleration threshold.

[0114] In the acceleration diagram G accelAfter construction is complete, the high-momentum narrative recognition module 106 executes a community detection algorithm on the graph. The purpose of the community detection algorithm is to divide the nodes in the graph into several subsets, such that the nodes within each subset are tightly connected, while the nodes between different subsets are sparsely connected. In a specific embodiment, the algorithm used is the Louvain method, which finds the optimal community partitioning of the graph by iteratively optimizing the modularity index.

[0115] The output of the community detection algorithm is a partition of a set of nodes, P = {C1, C2, ..., C}. k}, where each subset C k It is a community, representing a high-momentum narrative. Each high-momentum narrative C k Each is composed of a group of semantically closely related conceptual atoms whose association is growing rapidly.

[0116] Finally, the high-momentum narrative recognition module 106 will identify all high-momentum narratives {C k The data is structured and stored as a list. Each element in the list represents a high-momentum narrative, consisting of a set of conceptual atoms that constitute that narrative. This list is then transmitted to the polarity calibration module 107 for subsequent polarity determination and risk correlation analysis.

[0117] After identifying high-momentum narratives, the method proceeds to the polarity calibration step. This step is performed by the polarity calibration module 107. The first step of polarity calibration is to perform an alignment process, which aligns high-momentum narratives formed in the external market with the target company's internal corporate DNA profile. The inputs to this step are the list of high-momentum narratives output by the high-momentum narrative identification module 106, and the corporate DNA profile G generated and stored by the corporate DNA profile construction module 104. dna .

[0118] In one specific embodiment, the polarity calibration module 107 integrates a pre-trained language model, such as Sentence-BERT, to convert text phrases into high-dimensional semantic vectors. The alignment process is tailored to each high-momentum narrative C. k Execute independently.

[0119] For a given high-momentum narrative C k ={A1,A2,…,A m It is composed of a set of conceptual atoms. Meanwhile, the corporate DNA map G... dna Contains a set of nodes N dna ={N1,N2,…,N p}, where each node represents a core concept inherent to an enterprise. The goal of alignment is to provide narrative C k Each concept atom A in i In the node set N of the graph dna Find the node that is semantically most corresponding to it.

[0120] Polarity calibration module 107 first traverses the high-momentum narrative C k Each concept atom A in i For each A i The polarity calibration module 107 uses a pre-trained language model to convert its text content into a semantic vector v(A). i ).

[0121] Subsequently, the polarity calibration module 107 traverses the enterprise DNA map G. dna Each node N in j Similarly, the polarity calibration module 107 uses the same language model to calibrate node N. j The text labels are converted into a semantic vector v(N) j ).

[0122] Next, the polarity calibration module 107 calculates the conceptual atom A. i With each graph node N j The semantic similarity between the two vectors. This similarity is calculated by evaluating the semantic similarity between the two vectors v(A). i ) and v(N j The similarity between the two pairs of objects is quantified by cosine similarity. The calculation formula is as follows:

[0123]

[0124] Where: Sim(A) i N j (A) is a conceptual atom. i and graph node N j The semantic similarity score between them has a range of [-1, 1]; v(A i ) and v(N j ) are A i and N j The semantic vector; · represents the dot product of vectors; ||·|| represents the Euclidean norm of the vector.

[0125] For each concept atom A i The polarity calibration module 107 searches for nodes in all the graphs that have a semantic similarity score of Sim(A). i N j The largest node N j If the maximum similarity score exceeds a preset alignment threshold θalign (For example, 0.8), then the concept atom A is considered to be... i Successfully aligned to the graph node This pair It is recorded as an alignment pair.

[0126] After traversing all concept atoms in a high-momentum narrative, the polarity calibration module 107 outputs an alignment result set. This result set contains all successfully aligned concept atom-graph node pairs. These successfully aligned graph nodes are called anchor nodes of the high-momentum narrative, and they form the basis for subsequent polarity determination analysis. This alignment result will be used to perform the polarity determination sub-step.

[0127] After completing the alignment process, the polarity calibration module 107 continues with the polarity determination calculation. The input to this calculation is the high-momentum narrative C generated in the previous sub-step. k The set of anchor nodes, and the sentiment scores of the conceptual atoms associated with the narrative.

[0128] In one specific embodiment, the polarity calibration module 107 first calibrates for each high-momentum narrative C. k Calculate a comprehensive narrative sentiment mean. For a vector array consisting of m concept atoms {A1, A2, ..., A...} m Narrative C constituted by} k Its narrative emotion mean S narrative (C k The calculation method is to take the arithmetic mean of the sentiment scores of all concept atoms that constitute the narrative and whose sentiment scores have been calculated in the concept atom extraction step. The formula is:

[0129]

[0130] Wherein: S narrative (C k ) is a high-momentum narrative C k The mean of narrative emotion; C k Represents a complete high-momentum narrative; m is the constituent of the narrative C. k The concept of total number of atoms; S sent (A i (A) is a conceptual atom. i The sentiment score, which is calculated by the concept atom extraction module 103, has a value range of [-1, 1]. The complete meaning is to start from index i = 1 and end at index i = m, and assign each concept atom A... i Individual emotional score S sent (A i Add them all up to get a total.

[0131] Subsequently, the polarity calibration module 107 calculates the structured influence score of the narrative. This score is based on the anchor nodes of the narrative in the corporate DNA map G. dna The nature of nodes in the enterprise DNA map. Each type of node is pre-assigned a fixed influence weight. For example, a vulnerability type node has a weight of -1.0, a core technology type node has a weight of +0.8, and a strategic goal type node has a weight of +1.0.

[0132] For a narrative C k Its structured influence score S structural (C k The calculation method for ) is to take the arithmetic mean of the influence weights of all its anchor nodes. The calculation formula is:

[0133]

[0134] Wherein: S structural (C k ) is narrative C k Structured influence scores; Anchors (C k ) is narrative C k The set of anchor nodes in the corporate DNA map is determined by the alignment process; |Anchors(C k | represents the number of anchor nodes; w(N) j ) is the anchor node N j Pre-defined influence weights in the corporate DNA profile; Overall, it describes a clear instruction: "Traverse the set of anchor points Anchors(C k Each node N in ) j For each N j Perform the subsequent operation (i.e., obtain its weight w(N)). j Then add the results of all operations together.

[0135] Finally, the polarity calibration module 107 combines the narrative sentiment mean with the structured influence score through a weighted summation to calculate the ultimate score P of the high-momentum narrative. final (C k The calculation formula is as follows:

[0136] P final (C k )=α·S narrative (C k )+β·S structural (C k );

[0137] Where: P final (Ck ) is narrative C k The ultimate score; S narrative (C k ) is a high-momentum narrative C k The mean of narrative emotion; S structural (C k ) is narrative C k The structured influence score; α and β are preset non-negative weighting coefficients, and α + β = 1. For example, in one embodiment, α = 0.3 and β = 0.7 can be set to indicate that the judgment weight of structured influence is higher than the sentiment tendency of the external market.

[0138] Based on the final polarity score, the polarity calibration module 107 classifies the high-momentum narrative. If P final (C k () greater than a preset positive threshold θ pos (For example, +0.1), then the narrative is classified as chance. If P final (C k () less than a preset negative threshold θ neg If the value is -0.1, the narrative is classified as risky. Otherwise, the narrative is classified as neutral.

[0139] The final output of the polarity calibration module 107 is an updated list of high-momentum narratives, with each narrative appended with its final polarity score and classification result (opportunity, risk, or neutral). This list will then be submitted to the risk quantification and early warning module 108 for final processing.

[0140] During the polarity calibration calculation process, the polarity calibration module 107 internally calls a pre-configured polarity judgment rule base. This rule base is used to calculate the structured influence score S. structural (C k The required fixed influence weight values ​​w(N) are provided. j This rule base will map the enterprise's DNA into G. dna Each node type defined in the code is mapped to a fixed, non-subjective numerical weight.

[0141] In one specific embodiment, the rule base contains the following rules:

[0142] Rule 1: If an anchor node N j If the type is defined as vulnerability or risk factor, then the fixed influence weight w(N) of that node is... j The weight is set to -1.0. For example, a node labeled as dependent on a single vendor and of type Vulnerability has a weight of -1.0.

[0143] Rule 2: If an anchor node Nj If the type is defined as core technology, strategic objective, or competitive advantage, then the fixed influence weight w(N) of that node is... j The weight is set to +1.0. For example, a node labeled as 5nm process technology and of type core technology has a weight of +1.0.

[0144] Rule 3: If an anchor node N j If the type is defined as a product line or partner, then the fixed influence weight w(N) of that node is... j The weight is set to +0.5. For example, a node labeled "flagship smartphone series" and categorized as "product line" has a weight of +0.5.

[0145] Rule 4: If an anchor node N j The type is defined as a geographic market or key figure, which does not inherently possess fixed positive or negative attributes. Therefore, the fixed influence weight w(N) of that node is... j It is set to 0.0.

[0146] When calculating the structured influence score for any high-momentum narrative, the polarity calibration module 107 strictly adheres to this rule base to query the weight values ​​of its anchor nodes. This rule base is stored in the system's non-volatile memory and is loaded and accessed by the polarity calibration module 107 at runtime. This design ensures that the evaluation criteria for the internal structured influence are consistent and reproducible when performing polarity judgments on all narratives.

[0147] After completing the polarity calibration of all high-momentum narratives, refer to the appendix. Figure 1 The method provided by this invention proceeds to the step of generating financial risk assessment results. This step is performed by the risk quantification and early warning module 108. Its input is a high-momentum narrative list output by the polarity calibration module 107, with the final polarity score and classification results appended.

[0148] In one specific embodiment, the risk quantification and early warning module 108 first filters all high-momentum narratives classified as risks from the input list. For each identified risk narrative C k The early warning module 108 calculates a final quantitative risk score R(C) for it. k ).

[0149] The risk score R(C) k The calculation of the risk narrative C combines the magnitude of its momentum and the degree of its negative polarity. First, the early warning module 108 calculates the risk narrative C. k The momentum amplitude M(C) k The momentum amplitude is defined as the sum of the semantic acceleration values ​​of the corresponding edges in the acceleration diagram for all conceptual atom pairs constituting the narrative. Its calculation formula is:

[0150]

[0151] Where: M(C k (C) is a risk narrative. k The momentum amplitude; E(C) k ) is in the acceleration diagram G accel In the middle, both endpoints are in the risk narrative C k The set of internal edges; Its instruction system traverses the narrative subgraph C k All edges inside; a(A i A j ,t current ) is at the current time step t current Concept of atomic pairs (A i A j The semantic acceleration of ), where a represents semantic acceleration.

[0152] Next, the risk quantification and early warning module 108 uses the momentum amplitude M(C) k ) and the final polarity fraction P calculated by the polarity calibration module 107 final (C k The absolute value of ) is used to calculate the final quantitative risk score R(C). k The calculation formula is as follows:

[0153] R(C k )=M(C k )×|P final (C k )|;

[0154] Where: R(C k (C) is a risk narrative. k The final quantitative risk score; M(C k (C) is a risk narrative. k The momentum amplitude; |P final (C k )| is the risk narrative C k The absolute value of the ultimate score.

[0155] The risk quantification and early warning module 108 repeats the above calculations for all selected risk narratives. After completing all calculations, the risk quantification and early warning module 108 generates a structured risk assessment report. This report contains a list of all identified risk narratives ranked according to their final quantified risk score R(C). k Sort in descending order.

[0156] Each entry in the report list represents an independent risk and includes the following data fields: the final quantified risk score R(C) for that risk.k The textual descriptions of all conceptual atoms that constitute the risk narrative, as well as a list of corporate DNA map anchor nodes identified during the alignment process and associated with the risk narrative.

[0157] The generated, sorted risk assessment report is the final output of the method of the present invention, and is transmitted to a user interface risk quantification and early warning module 108 or stored in a database for subsequent display or analysis.

[0158] After generating the quantitative risk score, refer to the appendix. Figure 1 The risk quantification and early warning module 108 continues to perform the task of determining the evolutionary stage of each identified risk. The input to this task is a list of identified risk narratives, as well as time-series data of semantic velocity and semantic acceleration related to these narratives, generated by the semantic momentum calculation module 105.

[0159] In a specific embodiment, the risk quantification and early warning module 108 first performs risk quantification and early warning for each risk narrative C. k Calculate its value at the current time step t. current Aggregate semantic speed and aggregate semantic acceleration.

[0160] Risk Narrative C k Aggregate semantic speed V agg (C k ,t current This is calculated as the arithmetic mean of the semantic velocities of all conceptual atom pairs constituting the narrative at the current time step. The formula is:

[0161]

[0162] Where: V agg (C k ,t current (C) is a risk narrative. k Aggregate semantic speed at the current time step; E(C k ) is in the acceleration diagram G accel In the middle, both endpoints are in narrative C k The set of internal edges; |E(C k | represents the number of edges in the set of edges; Its instruction system traverses the narrative subgraph C k All edges inside; V(A) i A j ,t current ) is a concept atom pair (A i A j The semantic velocity at the current time step; V is the concept atom pair (A i A jThe first derivative of the correlation strength C between the two over time.

[0163] Risk Narrative C k Aggregate semantic acceleration a agg (C k ,t current This is calculated as the arithmetic mean of the semantic accelerations of all conceptual atom pairs constituting the narrative at the current time step. The formula is:

[0164]

[0165] Where: a agg (C k ,t current (C) is a risk narrative. k The aggregate semantic acceleration at the current time step; a(A i A j ,t current ) is a concept atom pair (A i A j Semantic acceleration at the current time step.

[0166] After calculating the aggregation semantic velocity and aggregation semantic acceleration, the risk quantification and early warning module 108 divides each risk narrative into a predefined evolutionary stage based on the signs of these two indicators. This division is performed according to the following set of rules:

[0167] Rule 1: If V agg (C k ,t current )>0 and a agg (C k ,t current If )>0, then the risk narrative C k The evolutionary stage was identified as the generative period.

[0168] Rule 2: If V agg (C k ,t current )>0 and a agg (C k ,t current If )≤0, then the risk narrative C k The evolutionary stage was identified as the development period.

[0169] Rule 3: If V agg (C k ,t current )≤0 and a agg (C k ,t current If ) < 0, then the risk narrative C k The evolutionary stage was identified as the decline period.

[0170] Rule 4: If V agg (C k ,t current )≤0 and a agg (C k ,t current If )≥0, then the risk narrative C k The evolutionary stage was identified as the incubation period.

[0171] Finally, the risk quantification and early warning module 108 will append the evolutionary stage (generation, development, decline, or incubation period) determined for each risk narrative as a new data field to the corresponding risk entry in the risk assessment report generated in the previous step. At this point, the complete financial risk assessment report is generated.

[0172] After generating a complete financial risk assessment report, the risk quantification and early warning module 108 transmits the report to the user interface module 109, which is responsible for visualizing the assessment results.

[0173] In one specific embodiment, the user interface module 109 generates a scatter plot in a two-dimensional coordinate system to display all identified risk narratives. In this scatter plot, each identified risk narrative C... k It is represented as a separate graphic symbol, such as a circular bubble.

[0174] The horizontal axis (X-axis) of the two-dimensional coordinate system is defined as the final quantified risk score R(C). k The horizontal position of a risk narrative graphic marker is determined by its corresponding R(C). k The value is uniquely determined.

[0175] The vertical axis (Y-axis) of the two-dimensional coordinate system is defined as the aggregate semantic acceleration a. agg (C k ,t current The vertical position of a risk narrative graphic marker is determined by its corresponding a. agg (C k ,t current The value is uniquely determined.

[0176] In addition, the size and color of each graphic marker are also used to encode additional dimensions of information. The size of the graphic marker is related to the absolute value of the semantic velocity of the risk narrative aggregation |V agg (C k ,t current | Proportional.

[0177] The colors of the graphic markers are assigned according to the evolutionary stages of the risk narrative that have been identified. The user interface module 109 has a preset fixed color mapping rule, for example: the generation stage is mapped to red, the development stage to orange, the decline stage to blue, and the latency stage to gray.

[0178] The user interface module 109 also provides interactive functionality. When a user selects any graphic marker on the graph using an input device (such as a mouse), the system displays an information panel on the interface. This information panel displays full details of the selected risk narrative in plain text format, directly derived from the risk assessment report, including:

[0179] A list of texts comprising all the conceptual atoms that make up this risk narrative.

[0180] A list of textual anchor nodes in the corporate DNA map aligned with this risk narrative.

[0181] The final quantified risk score of this risk narrative is R(C). k ), Aggregate semantic speed V agg (C k ,t current ) and aggregate semantic acceleration a agg (C k ,t current The precise value of ).

[0182] The risk narrative is identified by textual labels representing the evolutionary stages of the risk narrative.

[0183] In this way, the user interface module 109 transforms multi-dimensional, structured risk data into an interactive, quantifiable graphical representation.

Claims

1. A dynamic AI-based method for assessing financial risk, characterized in that, Including the following steps: Acquire unstructured text data and extract concept atoms from it; Calculate the co-occurrence intensity between the concept atoms, and calculate the semantic velocity and semantic acceleration based on the change of the co-occurrence intensity over time; For a target company, construct a company DNA graph containing nodes and edges. The nodes define the core concepts of the target company, and the edges define the logical relationships between the nodes. Based on the semantic acceleration, a high-momentum narrative with a semantic acceleration higher than a preset threshold is identified; Align the conceptual atoms within the high-momentum narrative with the nodes of the corporate DNA map to obtain the alignment results and the inherent logical relationships between the aligned nodes. Based on the alignment results and the inherent logical relationships between the aligned nodes, perform polarity calibration on the high-momentum narrative to obtain polarity attributes. The financial risk assessment result is generated by combining the polarity attribute, the identifier of the high-momentum narrative, the alignment description, the current value of the semantic velocity, and the current value of the semantic acceleration.

2. The AI-based dynamic assessment method for financial risk according to claim 1, characterized in that, The steps for constructing the enterprise DNA map also include: Based on the target company's financial reports and strategic planning documents, the inherent logical relationship is determined; The inherent logical relationships include dependency, support, or competition.

3. The AI-based dynamic assessment method for financial risk according to claim 1, characterized in that, The calculation of the co-occurrence intensity includes a weighted summation of the logarithmic values ​​of the co-occurrence frequencies of the concept atoms and the logarithmic values ​​of the number of independent information sources.

4. The AI-based dynamic assessment method for financial risk according to claim 1, characterized in that, The semantic velocity is obtained by performing a first-order difference calculation in the time dimension on the co-occurrence intensity, and the semantic acceleration is obtained by performing a first-order difference calculation in the time dimension on the semantic velocity.

5. The AI-based dynamic assessment method for financial risk according to claim 1, characterized in that, The high-momentum narrative is structurally a subgraph composed of multiple of the aforementioned conceptual atoms.

6. The AI-based dynamic assessment method for financial risk according to claim 1, characterized in that, The alignment steps include: Calculate the semantic similarity between the conceptual atoms within the high-momentum narrative and each node in the corporate DNA map; And select the node with the highest semantic similarity as the alignment node.

7. The AI-based dynamic assessment method for financial risk according to claim 1, characterized in that, The steps for performing polarity calibration include: For each concept atom within the high-momentum narrative, a unipolar value is determined based on the alignment node corresponding to the concept atom and the inherent logical relationship between the alignment node and the alignment node. Furthermore, a weighted summation is performed on all unipolar values ​​to obtain the final polarity attribute.

8. The AI-based dynamic assessment method for financial risk according to claim 7, characterized in that, When the inherent logical relationship is a dependency relationship, the unipolar value is negative.

9. The AI-based dynamic assessment method for financial risk according to claim 7, characterized in that, When the inherent logical relationship is a competing relationship, the unipolar value is positive.

10. The AI-based dynamic assessment method for financial risk according to claim 1, characterized in that, The evolutionary stage of the high-momentum narrative is determined based on the current values ​​of the semantic velocity and the semantic acceleration; and the evolutionary stage is included in the financial risk assessment results.

Images