A method for intelligently generating due diligence reports on non-performing financial assets

Abstract

The present invention provides a method for intelligently generating due diligence reports on financial non-performing assets, which relates to the field of management systems. It deeply integrates multi-source heterogeneous data such as legal documents and financial statements, and forms a comprehensive context-aware model of the target assets through multimodal feature extraction, semantic alignment and knowledge graph construction technology; uses evolutionary computing methods such as genetic algorithms to automatically mine potential and non-explicit risk factor combinations that are deeply coupled with asset characteristics, and constructs an adaptive risk assessment network to dynamically quantitatively assess and predict the comprehensive risk level; uses an explainable artificial intelligence model to perform attribution analysis on the risk assessment results, clearly revealing key impact paths and core data evidence; and automatically outputs a due diligence report on financial non-performing assets that includes in-depth analysis, risk warnings, diversified disposal suggestions and compliance review points based on a report logic framework and narrative template that can be dynamically adjusted by users.

Description

Technical Field

[0001] The present invention relates to the field of management systems, and in particular to a method for intelligently generating due diligence reports on non-performing financial assets. Background Art

[0002] Traditional due diligence reports are mostly static descriptions and analyses of the historical and current status of assets. They are unable to dynamically track real-time changes in key factors such as the debtor's operating conditions, collateral value, legal environment, and market environment. They also lack the ability to effectively predict and provide early warnings of risk evolution trends. When current public and practical applications combine AI or related intelligent systems for processing, their system iterations rely on the cognition and work of back-end developers, and the iteration timeliness is far weaker than real-time iteration and monitoring. At the same time, due to the single AI, its data processing is concentrated on the processing of single or formatted files, making it difficult to process unstructured and non-standardized data. Summary of the Invention

[0003] Problems Solved

[0004] In response to the existing deficiencies, the present invention provides a method for intelligently generating due diligence reports on financial non-performing assets, which solves the problems of the prior art.

[0005] Technical Solution

[0006] To achieve the above objectives, the present invention is implemented through the following scheme: a method for intelligently generating a due diligence report on non-performing financial assets, the method comprising:

[0007] Sp1. Deep fusion of multimodal data and context-aware modeling:

[0008] Collect structured, semi-structured and unstructured data related to target financial non-performing assets to obtain heterogeneous data;

[0009] Mapping the heterogeneous data into a unified semantic space, identifying and establishing entities, attributes, and complex relationships related to non-performing financial assets, and forming a context-aware model of the non-performing financial assets;

[0010] Sp2. Risk factor mining and adaptive evaluation based on evolutionary computing:

[0011] Sp2.1. Initializing a candidate set of risk factors: Based on the context-aware model, using a genetic algorithm evolutionary computation method, iteratively search for optimality within a preset risk dimension space to discover potential and non-explicit risk factor combinations that are coupled with the characteristics of the financial non-performing assets;

[0012] Sp2.2. Constructing a risk assessment network: The risk assessment network dynamically adjusts the weights and interactions of various risk factors based on newly input data and historical assessment feedback, and combines it with an integrated learning strategy to quantitatively assess the comprehensive risk level of non-performing financial assets, outputting a risk profile and confidence intervals.

[0013] Sp3, Insight generation and report output based on interpretable models:

[0014] Sp3.1. Use an interpretable intelligent model to conduct attribution analysis on the output results of the risk assessment network to identify the key impact paths and core data evidence that lead to specific risk assessment conclusions;

[0015] Sp3.2. Based on the preset report logic framework and narrative template that can be dynamically adjusted by the user, the key information in the context-aware model is generated using natural language to output a due diligence report on non-performing financial assets.

[0016] Preferably, the deep fusion of multimodal data and context-aware modeling further includes: using a pre-trained language model enhanced by an attention mechanism to parse long-distance dependencies and complex clause structures in legal texts, and encoding the extracted key elements of legal texts and their confidence scores as nodes and weighted edges in the knowledge graph.

[0017] Preferably, the risk factor mining based on evolutionary calculation further includes:

[0018] The explanatory power of the decision rules generated by the risk factors was quantified by calculating the minimum description length;

[0019] The prediction accuracy of the risk factor combination is measured by evaluating its F1 score for historical default events on the backtest dataset;

[0020] The correlation is evaluated by calculating the vector space cosine similarity between the risk factor combination and the patterns in a preset typical non-performing asset risk pattern library verified by domain experts.

[0021] Preferably, the risk assessment network adopts a reinforcement learning mechanism, which further includes: constructing a deep deterministic policy gradient agent, the state space of the agent represents the risk factors and assessment results of the current assets, and the action space corresponds to the adjustment strategy for the weights or activation functions of specific risk factors in the risk assessment network; quantifying the user's confirmation, correction or rejection of the risk assessment into a scalar reward signal to guide the agent's policy learning.

[0022] Preferably, the insight generation of the explainable model also includes: generating a visual map of the risk transmission path, and the generating a visual map of the risk transmission path further includes: using an attention-weighted graph neural network inference algorithm on the knowledge graph of the context-aware model to identify and quantify the influence intensity and transmission probability between different risk entities and risk factors to form a transmission path.

[0023] Preferably, in the report output, the natural language generation further includes: utilizing a conditional text generation model, taking a preset portrait of the target audience as a conditional input, combining the structured semantic representation extracted from the interpretability model insights, dynamically selecting a narrative template, adjusting the professional level of terminology, and controlling the detailed level of argumentation to generate customized report text that meets the needs of specific audiences.

[0024] Preferably, the method further includes a continuous learning and model iteration module, and the continuous learning and model iteration module further includes:

[0025] Data and concept drift detection unit, which monitors the statistical characteristics of input data streams and changes in user feedback patterns in real time, and automatically triggers the model update process when significant drift is detected;

[0026] The differential knowledge graph update unit incrementally merges new or changed entities, relationships, and their confidence levels into the existing knowledge graph, and uses transfer learning to apply the knowledge learned from historical data to the iterative upgrade of the new model.

[0027] Preferably, a system for intelligently generating a due diligence report on non-performing financial assets comprises a processor and a memory coupled to the processor, wherein the memory stores computer program instructions, and the computer program instructions are executed by the processor. The system further comprises:

[0028] Multimodal data deep fusion and context-aware modeling engine:

[0029] Collect data related to target financial non-performing assets from heterogeneous data sources;

[0030] Build a context-aware model of financial non-performing assets;

[0031] Sp2, risk factor mining and adaptive assessment engine based on evolutionary computing:

[0032] Perform iterative optimization within the preset risk dimension space;

[0033] Build and operate a risk assessment network;

[0034] Sp3, insight generation and report output engine based on explainable models:

[0035] Use interpretable intelligent models to conduct attribution analysis on risk assessment results;

[0036] Output due diligence report on financial non-performing assets.

[0037] Preferably, the pre-trained language model processing unit in the multimodal data deep fusion and context-aware modeling engine further includes:

[0038] Complex long sentence segmentation and dependency parsing module, used to accurately identify the master-slave structure and restrictive conditions in contract terms;

[0039] The semantic role labeling module, based on domain knowledge enhancement, labels the specific roles of participants and core legal behaviors in financial transactions.

[0040] Preferably, the risk factor mining and adaptive assessment engine based on evolutionary computing further includes:

[0041] The feedback-driven model parameter adjustment module further includes:

[0042] User feedback real-time capture and structured processing interface, used to receive and analyze user annotation information on risk factors and assessment results;

[0043] An incremental model training and version control unit, wherein the incremental model training and version control unit has an embedded evolutionary algorithm and supports online fine-tuning of the population initialization strategy of the evolutionary algorithm and the local connection weights of the evaluation network;

[0044] Offline batch retraining scheduler, used to trigger global optimization of the entire model system after accumulating enough new data or feedback.

[0045] Beneficial effects

[0046] The present invention provides a method for intelligently generating due diligence reports on non-performing financial assets. It has the following beneficial effects:

[0047] This invention deeply integrates multimodal data, constructs a panoramic contextual knowledge graph, and uses evolutionary computing to mine hidden risk factors. Combined with adaptive advanced evaluation networks and explainable AI insights, it can quickly extract accurate insights from heterogeneous data, reveal complex correlations and potential risks, and improve the efficiency of information collection and preliminary analysis to a new level. It can not only accurately quantify known risks, but also actively discover and evaluate non-explicit and combined risks, greatly improving the sensitivity and accuracy of risk identification.

[0048] The present invention has the ability of continuous learning and self-evolution, can adapt to the dynamically changing market and risk environment, realize the digital precipitation and intelligent iteration of organizational knowledge and experience, and ensure the effectiveness and value of long-term application. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a system composition diagram of the present invention;

[0050] Figure 2 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the accompanying drawings to clearly and completely describe the solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without making creative efforts are within the scope of protection of the present invention. Specific embodiment:

[0053] like Figures 1 to 2 As shown, a method for intelligently generating a due diligence report on non-performing financial assets includes obtaining data related to the non-performing financial assets and generating a report structure; the method further includes the following steps:

[0054] Sp1. Deep fusion of multimodal data and context-aware modeling lays the foundation for a comprehensive understanding of target financial non-performing assets, involving the entire process from data collection and preprocessing to multimodal feature extraction and alignment, and finally forming a context-aware model.

[0055] Data collection and preprocessing:

[0056] Data sources and types: including but not limited to: (1) Legal documents: including loan contracts, guarantee contracts, mortgage agreements, litigation / arbitration documents (complaints, judgments, mediation documents), bankruptcy reorganization documents, etc., usually in PDF, Word or scanned image format; (2) Financial statements: balance sheets, income statements, cash flow statements, statements of changes in owners' equity and their notes of the debtor and related parties, usually in Excel, PDF or image format; (3) Market transaction data: secondary market transaction prices, trading volumes, relevant macroeconomic indicators, industry indexes, interest rates, exchange rates, etc. of similar non-performing assets, usually in structured time series data; (4) Public opinion information: news portals, financial media, social platforms, industry forums, etc. about negative news, default records, litigation information, business anomalies, etc. about debtors, guarantors, related parties, and collateral, usually in unstructured text or semi-structured data; (5) Description of the physical status of assets: including appraisal reports, on-site inspection photos, videos, geographic location information, ownership certificates, etc. of collateral such as real estate and machinery and equipment, involving text, images, and geospatial data.

[0057] Data processing solution preprocessing: legal documents and financial statements (text / image type):

[0058] OCR and Layout Analysis: For documents in scanned or image format, high-precision OCR engines (including Tesseract-OCR, optimized with the deep learning model CRNN) are used for text recognition. Layout analysis (including models based on Faster RCNN or Layout LM) is used to identify document structure, including titles, paragraphs, tables, and seals.

[0059] Table extraction: For tables in financial statements or contracts, we use OpenCV image processing combined with heuristic rules, or specialized table recognition models (including TableNet and Tab Transformer) to extract structured table data.

[0060] Text cleaning: remove headers, footers, watermarks, irrelevant symbols, and garbled characters; convert between traditional and simplified Chinese, and unify full-width and half-width text; identify and process fill-in-the-blank items and handwritten supplements in contracts.

[0061] Preliminary element extraction: Based on regular expressions and keyword libraries, core elements such as contract number, party name, amount, date, and collateral name are preliminarily extracted as auxiliary information or verification basis for subsequent model processing.

[0062] Financial statements (structured data):

[0063] Account standardization: Map financial accounts disclosed by different accounting standards and different enterprises to a unified standardized account system (including those based on XBRL taxonomy or customized standard chart of accounts).

[0064] Data verification and cleaning: Check the cross-references between financial statements (including assets = liabilities + owner's equity), identify and handle outliers (including using box plots or Z-score methods), and missing values ​​(including using mean / median filling, regression interpolation, or multiple interpolation).

[0065] Financial indicator calculation: Automatically calculate key financial indicators such as solvency ratio (including current ratio, quick ratio, debt-to-asset ratio), profitability ratio (including net profit margin, return on equity), operating capacity ratio (including accounts receivable turnover rate, inventory turnover rate), etc.

[0066] Market Trading Data: API Access and Data Parsing: Obtain data through APIs (including Bloomberg, Reuters, and Wind) or web crawlers (subject to robots.txt compliance). Parse data in formats such as JSON, XML, or CSV.

[0067] Data cleaning: processing missing points in time series data (linear interpolation, spline interpolation) and abnormal fluctuations (moving average filtering, exponential smoothing).

[0068] Time alignment and frequency conversion: Align data from different sources and frequencies to a unified timeline (including daily and weekly).

[0069] Public opinion information: Targeted crawlers and content extraction: Based on keywords such as debtors and related parties, use frameworks such as Scrapy to develop targeted crawlers to extract news text, release time, source, etc.

[0070] Data deduplication: Remove duplicate or highly similar public opinion information based on text similarity (including SimHash, MinHashLSH).

[0071] Initial judgment of sentiment tendency: Use a dictionary (including Hownet sentiment dictionary) or a simple classification model (including Naive Bayes) to perform a preliminary positive or negative sentiment score on the public opinion text.

[0072] Description of the physical status of the asset: Image data processing: unified format conversion, size normalization, and image enhancement (including histogram equalization and denoising) of photos and videos.

[0073] Text description structuring: Extract key attributes from the text description in the appraisal report (including the area, location, and purpose of the property; the model, purchase year, and depreciation of the equipment).

[0074] Multimodal feature extraction and alignment: Mapping data from different sources and modalities into a unified shared semantic space that can capture cross-modal semantic relationships.

[0075] Text feature extraction:

[0076] Model Building: Select a Transformer model pre-trained and fine-tuned for the financial or legal domain, including FinBERT (pre-trained on financial news and research reports), LawBERT / LegalBERT (pre-trained on legal documents and precedents), or general-purpose BERT, RoBERTa, or ERNIE models. Perform domain-adaptive pre-training and fine-tuning for downstream tasks on a specific NPL-related corpus (including contracts, litigation documents, financial notes, market analysis reports, etc.). Downstream tasks may include named entity recognition, relation extraction, text classification (including risk type determination), and semantic similarity calculation. During fine-tuning, use the cross-entropy loss function for parameter optimization.

[0077] Model Application: The model takes preprocessed text snippets (including contract clauses, financial notes, and news summaries) as input and outputs fixed-dimensional semantic vector representations (token-level embeddings, sentence / clause-level embeddings, and document-level embeddings). For long texts, a holistic representation can be obtained using segmented processing followed by pooling (including mean / max pooling) or a hierarchical Transformer (including HiBERT).

[0078] Structured data feature extraction:

[0079] Processing Solution: Numerical financial indicators and market data, after cleaning and normalization / standardization (including min-max scaling and Z-score standardization), can be directly used as feature vectors. For categorical features (including industry and region classifications), one-hot encoding can be used or they can be mapped into a low-dimensional dense embedding layer (obtained through end-to-end learning).

[0080] Image feature extraction:

[0081] Model Building: Use deep convolutional neural networks (CNNs) pre-trained on large image datasets (including ImageNet), such as ResNet-50 / 101, Efficient-Net (B0-B7 series), or Vision Transformer (ViT). Based on the characteristics of NPL collateral images (including real estate appearance, equipment details, and title deed scans), fine-tune the model on domain-specific image data to better capture visual features relevant to asset valuation.

[0082] Model application: Input the preprocessed image, and the CNN / ViT model outputs a high-dimensional feature vector (including ResNet output 2048 dimensions, and Efficient-Net has different dimensions depending on the series).

[0083] Multimodal alignment model construction:

[0084] Joint Embedding-Based Models: A multi-input neural network architecture is designed to map semantically related samples from different modalities to nearby locations in a shared semantic space. A clause description in a legal document and the corresponding collateral photo should have similar vector representations. During training, either a contrastive loss or a triplet loss is used. The contrastive loss aims to bring positive pairs (the vector describing the contract clause for Property A and the vector for the photo of Property A) closer together and further distance negative pairs. The triplet loss considers anchor points, positive samples, and negative samples simultaneously.

[0085] Methods based on co-training or cross-modal generation: train a model to generate pseudo representations of image features from text descriptions, and require them to be similar to the real image features; vice versa.

[0086] Training data construction: A large number of cross-modal aligned samples need to be constructed, including legal document fragments, related financial data fragments, collateral description text, collateral images, market news, and changes in corresponding company financial indicators.

[0087] Model Application: Raw data from different modalities is transformed into vector representations in the same high-dimensional semantic space through their respective feature extractors and alignment models. These vectors can be directly used in downstream tasks, including similarity calculation, clustering, and risk assessment.

[0088] Knowledge graph construction: Automatically identify entities, attributes and complex relationships related to financial non-performing assets from multimodal data to form a structured knowledge network.

[0089] Entity recognition model construction: Using a sequence labeling model, the mainstream architecture is a pre-trained language model, using BERT + linear layer + conditional random field. The CRF layer can learn the constraints between labels, improving the accuracy of entity boundary recognition.

[0090] Data processing and application: Define entity types in the non-performing asset field, including debtor (DEBTOR), creditor (CREDITOR), guarantor (GUARANTOR), collateral (COLLATERAL), contract (CONTRACT), court (COURT), amount (AMOUNT), date (DATE), risk event (RISK_EVENT), etc.

[0091] Training data: A large amount of annotated corpus is required (including BIO annotation format). Active learning and semi-supervised learning can be used to reduce annotation costs.

[0092] Model application: Input text (including contracts and judgments) and output the identified entities, their types, and their locations within the text. For example, in the sentence "Zhang San borrowed 1 million yuan from Li Si, with Wang Wu providing a joint and several guarantee," identify Zhang San (DEBTOR), Li Si (CREDITOR), 1 million yuan (AMOUNT), and Wang Wu (GUARANTOR).

[0093] Relation extraction model construction: A pipeline-based approach involves performing NER first, followed by relationship classification of the identified entities. The relationship classification model can be an attention-based BiLSTM, a path-based graph convolutional network (GCN) that leverages dependency syntactic paths between entities, or a classifier based on a pre-trained language model (including adding a classification head to a [CLS] representation or entity pair representation).

[0094] Joint NER and RE: A model is designed to simultaneously perform entity recognition and relation extraction tasks, making better use of the dependency information between them. This approach uses a Transformer model based on multi-task learning.

[0095] Relationship type definition: predefined relationship types between entities, including "borrows_from", "guarantees_for", "mortgages_to", "involved_in_lawsuit", etc.

[0096] Training data: triples of entity pairs and their relations need to be labeled.

[0097] Model application: Input contains text and entity pairs of recognized entities, and output the type of relationship between them and their confidence. (Zhang San, borrows_from, Li Si), (Wang Wu, guarantees_for, Zhang San).

[0098] Event-Extraction (EE):

[0099] Model building: Identify specific events and their participants (arguments) in a text. This model typically includes two stages: event trigger word identification and argument role labeling. Methods based on sequence labeling or machine reading comprehension can be used.

[0100] Data processing and application: Define event types (including "signing a contract", "filing a lawsuit", and "declaring bankruptcy") and the argument roles of each event type (including "signatory", "contract subject", and "signing date" of the contract signing event).

[0101] Entity alignment and disambiguation solutions: These solutions address the unification of entities represented differently across data sources (including "Company A" and "A Co., Ltd."), as well as the problem of the same name referring to different entities. Methods include those based on string similarity, entity attribute similarity, network structure similarity (including leveraging link prediction in knowledge graphs), or entity linking methods based on pre-trained models.

[0102] Knowledge storage and reasoning: The graph database uses Neo4j's graph database to store the constructed knowledge graph (entities as nodes, relationships as edges, and attributes as node / edge characteristics).

[0103] Reasoning applications: Use the graph query language CypherQL for complex queries. Graph-based rule reasoning uses SWRL or graph embedding-based link prediction, path discovery, etc.

[0104] Context-aware model construction and application development: Using the constructed knowledge graph as the core framework, the multimodal feature vectors extracted from each step of Sp1 (including text semantic vectors, image feature vectors, and structured data features) are used as attributes of corresponding entity nodes or relationship edges, or are associated with them. In addition to basic information, debtor nodes are also associated with vectors containing financial statement analysis results, the latest public sentiment score, and semantic vectors of key clauses in relevant legal documents. This model provides a comprehensive, multi-dimensional, and structured view of non-performing assets, their related parties, related events, and potential risks.

[0105] Information Retrieval and Aggregation: Quickly retrieve all relevant information about a specific asset, regardless of its original modality or source. Uncover hidden relationships, including shared guarantee circles, indirect control relationships, and complex chains of creditor and debtor relationships.

[0106] Risk transmission analysis: Simulate the transmission path and potential impact range of specific risk events (including the default of a core enterprise) in the knowledge graph network.

[0107] Feature Engineering: Provide high-quality, context-rich input features (including graph embedding representations, path features, neighborhood aggregation features, etc.) for downstream risk assessment models.

[0108] Optimized options for extracting key elements from unstructured text data such as legal documents:

[0109] Application of pre-trained language models enhanced by attention mechanisms: Model selection and construction: Using the Transformer model to process long sequences of text, the sparse attention mechanism is used to capture long-range dependencies in legal texts (including a clause in a contract that cites a definition from dozens of pages ago, or a comprehensive identification of multiple pieces of evidence in a judgment).

[0110] Complex clause structure analysis: During pre-training and fine-tuning, these models learn from a large amount of legal text, enabling them to better understand complex structures such as parallel sentences, compound sentences, conditional clauses, and restrictive clauses, thereby accurately identifying core rights and obligations, prerequisites, and exclusions.

[0111] Identifying Implied Guarantees and Contingent Liabilities: Combining semantic understanding with knowledge graph reasoning. The model identifies typical sentences such as "If Party B fails to repay the principal debt on time, Party C agrees to assume joint and several liability." By combining the relationships between Party C and Party B (including parent companies and actual controllers) in the knowledge graph, it infers Party C's contingent liabilities. For more subtle clauses (including "Under certain market conditions, Party A has the right to request additional guarantees from Party B"), the model requires stronger contextual understanding and logical reasoning capabilities, requiring the integration of specific subtask models (including reading comprehension-based models to answer questions such as "Does Party C have a guarantee liability?") or rule-based post-processing.

[0112] Knowledge graph encoding of key elements and their confidence scores:

[0113] Confidence score calculation: When a model identifies entities, relationships, or extracts features, it typically outputs a probability value or softmax score. This score can be used as a confidence level after calibration (including Platt Scaling or Isotonic Regression).

[0114] Constructing weighted edges: Extracting key elements of legal text (including "guaranteed amount," "guarantee scope," and "cause of action" in contracts) as node attributes or independent nodes in the knowledge graph. The weight of these elements in relation to core entities (including the debtor and the contract) can be determined based on the following factors:

[0115] Confidence score: The accuracy of the extraction.

[0116] The importance of this element in the document is calculated using TF-IDF, Text Rank, or model attention weights. Preset a priori weights for the impact of different types of elements on risk assessment (for example, giving "unlimited joint and several liability" a higher weight than "general guarantee"). If the XAI model subsequently determines that this element contributes significantly to the risk assessment, its weight can be increased. These weighted edges can be used in subsequent graph algorithms (including community discovery, centrality calculation, and risk transmission analysis) and risk assessment models to more accurately reflect the contribution of different pieces of information to the overall risk assessment.

[0117] Sp2. Risk factor mining and adaptive assessment based on evolutionary computing automatically discovers potential risk factor combinations related to non-performing asset risks that are difficult to intuitively perceive from high-dimensional and complex feature spaces, and builds a risk assessment model that can dynamically adapt to new situations.

[0118] Sp2.1. Initialize the risk factor candidate set:

[0119] Detailed explanation of the evolutionary computing method - genetic algorithm: within the preset risk dimension space (including credit risk, market risk, operational risk, legal risk and other dimensions, each dimension contains many candidate atomic risk factors), iterative optimization is carried out to automatically discover potential and non-explicit risk factor combinations coupled with the characteristics of financial non-performing assets.

[0120] Risk factor sources: Context-aware model outputs: node attributes in the knowledge graph (including debtor financial ratios and public opinion scores), the presence or weight of edges (including the existence of specific guarantee relationships and litigation relationships), and certain dimensions of graph embedding vectors. Multimodal features: key dimensions of text semantic features and clustering results of image features. Expert-defined feature library: known important risk points pre-defined by financial experts.

[0121] GA construction details: Chromosome encoding (individual representation): Each individual (chromosome) represents a candidate risk factor combination.

[0122] A random set of individuals is generated as the initial population. Domain knowledge can also be incorporated to include known effective risk factors or combinations as part of the initial population (heuristic initialization) to accelerate convergence. The population size ranges from 50 to 200 individuals.

[0123] The selection operator selects excellent individuals to enter the next generation based on their fitness values.

[0124] The crossover operator simulates genetic recombination in biological evolution, operating on selected parent individuals with a certain crossover probability (0.6-0.9) to produce new offspring individuals.

[0125] The mutation operator randomly changes certain gene positions of offspring individuals with a certain mutation probability (0.01-0.1) to maintain population diversity and avoid falling into local optimality.

[0126] Iteration stopping conditions include reaching a preset maximum number of evolutionary generations, no significant improvement in the fitness function value over multiple generations, or finding a solution that meets specific conditions. The GA output is a set (or an optimal) combination of risk factors, which are used to construct subsequent risk assessment networks or directly form interpretable risk rules.

[0127] The fitness function of the evolutionary computation method further includes: The fitness function balances the predictive accuracy, interpretability, simplicity, robustness, and domain relevance of the risk factor combination. The explanatory power of the decision rules generated by the risk factors is quantified by calculating the minimum description length: the risk factor combinations selected by the GA are converted into a set of if-then decision rules. For example, the causal risk factor combination is {financial indicator A < 0.5, legal risk B = True}, generating the rule "IF financial indicator A < 0.5 AND legal risk B = True THEN high risk." The encoding length of the model itself (L(H)) and the encoding length of the data given model D (L(D|H)) are calculated. For decision rules, L(H) is quantified as the number of rules and the number of conditions in each rule. L(D|H) is the encoding length required to misclassify a sample under the rule. The goal is to minimize L(H) + L(D|H), where H represents the encoding length.

[0128] The smaller the MDL value, the simpler the risk factor combination's rules are, the better they fit the data, and the more explanatory they are. The fitness value can be set to the inverse of the MDL or a constant minus the MDL. The predictive accuracy of the risk factor combination is measured by evaluating its F1 score for historical default events on a backtest dataset:

[0129] Prepare a dataset containing historical NPL cases. Each case includes the values ​​of various GA candidate risk factors (at a certain point in time before the event) and the final actual outcome (including default, loss level, recovery rate, etc.). The dataset needs to be divided into training, validation, and test sets. For each GA-generated risk factor combination, use these factors as features to train a simple classification model (including logistic regression, support vector machine, or decision tree) on the training set, or directly construct a scorecard model using these factors. Then, predict historical default events using the validation and test sets.

[0130] The F1 score is the harmonic mean of precision and recall, which can better balance the two and is particularly suitable for default prediction problems with imbalanced categories.

[0131] F1 score calculation formula:

[0132] ;

[0133] The F1 score is used to evaluate the accuracy of a classification model built based on a specific risk factor combination in predicting adverse events on a historical backtesting dataset. It combines the model's precision and recall.

[0134] F1: F1 score value, ranging from 0 to 1, the higher the value, the better the prediction performance of the model.

[0135] Precision: The proportion of samples predicted by the model as positive that are actually positive. The calculation formula is:

[0136] ;

[0137] Recall (or recall rate): The proportion of samples that are actually positive examples that are successfully predicted as positive examples by the model. The calculation formula is:

[0138] ;

[0139] TP: True positive, the number of positive examples that are actually predicted as positive by the model; FP: False positive (Type I error), the number of negative examples that are actually predicted as positive by the model; FN: False negative (Type II error), the number of positive examples that are actually predicted as negative by the model.

[0140] The higher the F1 score, the stronger the predictive power of the risk factor combination. The correlation is evaluated by calculating the vector space cosine similarity between the risk factor combination and the patterns in the preset typical non-performing asset risk pattern library verified by domain experts:

[0141] Data processing is performed by financial experts (including senior credit reviewers and risk managers) who, based on experience or historical case studies, identify typical risk patterns (including "overexpansion risk," "linked guarantee chain risk," and "obsolescence risk"). Each pattern is described by a set of key risk factors and their typical manifestations. Each expert-defined risk pattern and the risk factor combinations generated by the GA are represented as vectors. If the set of atomic risk factors is fixed, each combination / pattern can be represented as a high-dimensional sparse vector, where the dimension corresponding to the factor is 1 or its weight, and 0 if it is absent.

[0142] If a risk factor combination is highly similar to a typical risk pattern recognized by one or more experts, it is considered to have good domain relevance and interpretability. The fitness function can consider the similarity with the most similar pattern or a weighted average of the similarities with multiple patterns. The final fitness function is usually a weighted combination of the above indicators, with the weights adjusted according to business needs. This includes the formula for the comprehensive fitness function of the genetic algorithm:

[0143] ;

[0144] It is used to evaluate the quality of each risk factor combination in the genetic algorithm. It has good interpretability, high prediction accuracy, and correlation with domain expert knowledge.

[0145] : The comprehensive fitness value of an individual (risk factor combination). The higher the value, the better the risk factor combination. , , : are the weight coefficients of interpretability, prediction accuracy, and domain relevance. The weights can be adjusted according to business needs to focus on different optimization goals. If prediction accuracy is more important, then The weights can be set higher. These weights usually take values ​​0-1, and their sum is 1. : The value of the minimum description length. It quantifies the simplicity and explanatory power of the decision rule generated by the current risk factor combination. The smaller the MDL value, the simpler the rule and the stronger the explanatory power. Therefore, it is used in the fitness function. , so that the smaller the MDL value, the greater the contribution of this item. F1 score. This refers to a metric that measures the accuracy of a prediction model based on the current risk factor combination. It is particularly useful for class-imbalanced datasets (focused on default prediction). It is the harmonic mean of precision and recall. A higher F1 score indicates better prediction accuracy. Maximum cosine similarity. This measures the maximum vector space cosine similarity between the current risk factor combination and each pattern in a pre-defined library of typical NPL risk patterns verified by domain experts. A higher value indicates a closer correlation between the current risk factor combination and the expert-approved risk pattern, and a better domain relevance.

[0146] Sp2.2. Construct a risk assessment network: Based on the risk factor combination mined by GA (or other more comprehensive feature sets), construct a network model that can quantitatively assess the comprehensive risk level of financial non-performing assets and can be dynamically adjusted.

[0147] Model architecture selection and construction:

[0148] Multilayer Perceptron:

[0149] Input layer: risk factors mined by GA (numerical or categorical factors after encoding / embedding) and other important supplementary features (including macroeconomic indicators).

[0150] Hidden layers: One or more fully connected layers. The number of neurons in each layer is determined by the complexity of the problem and the amount of data (including 32, 64, 128, etc.). Activation functions include Leaky ReLU, Tanh, etc. To prevent overfitting, a dropout layer or L1 / L2 regularization can be added.

[0151] Output layer: regression task (including prediction loss rate and recovery rate): single neuron, linear activation function, loss function is mean square error and mean absolute error.

[0152] Classification tasks (including predicting risk level: low / medium / high, or whether there is a default): the number of neurons is equal to the number of categories, the activation function is Softmax, and the loss function is cross entropy loss.

[0153] Optimizers: Adam, RMSprop, SGD-with-momentum. Graph neural networks are suitable for risk assessment using knowledge graph information.

[0154] Context-aware models (knowledge graphs) in the form of entire or subgraphs. Node features include entity attribute vectors and multimodal feature vectors extracted from Sp1; edge features can include relationship types, weights, etc.

[0155] GNN layer: Uses GCN, GAT, and Graph-SAGE. GCN updates node representations by aggregating information about neighboring nodes. GAT introduces an attention mechanism to assign different learning weights to different neighboring nodes. Graph-SAGE designs multiple aggregation functions and supports inductive learning of unknown nodes.

[0156] GNN nodes collect information from their neighborhood and update their own representations through multiple iterations. The pooling layer performs graph-level pooling on node representations to produce a representation vector for the entire graph or a subgraph of the target asset. The output layer connects to the fully connected layer for final risk scoring or classification. GNNs can automatically learn the complex interactions and dependencies between entities, capturing the propagation patterns of risk within the network, thereby more accurately assessing systemic and correlated risks.

[0157] Dynamically adjust the weights and interactions of each risk factor:

[0158] Online learning: When new data (including asset performance updates, market changes) or user feedback becomes available, the model can be updated incrementally rather than completely retrained. For neural networks, mini-batch gradient descent can be used to continuously update model parameters. For Bayesian networks, the CPT can be updated using the Bayesian update method. Adaptive learning rate algorithms (including Adam variants) or learning rate decay strategies can be used.

[0159] Adjustments based on feedback:

[0160] Direct feedback: Users (including risk analysts) can directly adjust the weights of certain risk factors or assessment results. These adjustments can be quantified and used to modify model parameters (including by modifying the loss function, rewarding predictions that are consistent with user feedback and penalizing those that are inconsistent).

[0161] Indirect feedback: Users’ usage behavior of the report (including which parts are read carefully and which suggestions are adopted) can also serve as indirect feedback signals.

[0162] Output risk profile and confidence interval:

[0163] Risk profiling: The comprehensive risk score and sub-scores in different risk dimensions (including credit, market, legal, and operational) are displayed through visualization (including radar charts, dashboards, and heat maps) to form an intuitive depiction of the risk status of non-performing assets.

[0164] Confidence interval calculation: quantify the uncertainty of the assessment results.

[0165] Monte Carlo Dropout also keeps the Dropout layer activated during the prediction phase of the neural network, performs multiple forward propagations, and obtains a set of prediction results. The mean and variance (or quantile) of the distribution of this set of results are calculated as the confidence interval.

[0166] Narrower confidence intervals indicate more reliable assessments. Wider confidence intervals indicate greater uncertainty, requiring more cautious decision-making or further investigation.

[0167] The risk assessment network uses a reinforcement learning mechanism: it enables the risk assessment network to learn an optimal adjustment strategy through interaction with the environment (including user feedback), thereby continuously optimizing its risk sensitivity and generalization ability. Deep deterministic policy gradient agent construction and application;

[0168] Reinforcement learning, target network parameter soft update formula:

[0169] ;

[0170] Used to update the parameters of the target actor network and target critic network in reinforcement learning algorithms. Software updates are used to stabilize the learning process and prevent rapid changes in the target network parameters that can cause training instability.

[0171] : Parameters of the target network (target Actor network or target Critic network); : Parameters corresponding to the online network (online Actor network or online Critic network). :Soft update coefficient, is a small positive value ( , the value range is 0.001-0.01), which controls the speed of "transferring" the online network parameters to the target network parameters. The smaller it is, the slower the target network is updated and the more stable the learning process is.

[0172] State space representation: This includes the current asset's risk factor vector (derived from a combination of GA mining or a more comprehensive feature set), the current assessment results of the risk assessment network (including risk scores and scores for each dimension), and even some parameters or confidence metrics of the model itself. This representation needs to be carefully designed to include sufficient information to guide decision-making.

[0173] Data processing: digitization and normalization.

[0174] The action space corresponds to strategies for adjusting the weights or activation functions of specific risk factors in the risk assessment network. Actions are continuous. Each dimension of the action vector can correspond to: an adjustment to the weight of a key risk factor in the assessment model (percent increase / decrease); an adjustment to the slope or threshold of a hidden layer activation function; or adjustments to the weights of different sub-models in a model ensemble.

[0175] Constraints: The action space needs to have reasonable boundaries to avoid excessive adjustments that may cause model instability.

[0176] Reward signal design: Quantify the user's confirmation, revision, or rejection of risk assessment into a scalar reward signal:

[0177] Evaluation confirmation: If the user approves the evaluation result, a positive reward (including +1) will be given.

[0178] Minor corrections: Users make small adjustments to the results, giving small positive rewards or zero rewards (including +0.1, 0).

[0179] Significant Revision / Rejection: If the user significantly adjusts the result or completely rejects it, a negative reward (including -1, -0.5) will be given.

[0180] Correction direction consistency: If the direction of the user's correction is consistent with the direction the RL agent is trying to adjust, even if it is a correction, a certain positive incentive can be given.

[0181] Delayed rewards: Sometimes the true effect of user feedback is not apparent until some time has passed (including the comparison between actual recovery and forecast after the disposal of non-performing assets), so the distribution of delayed rewards needs to be considered.

[0182] Sparse reward problem: User feedback is not available in every evaluation, so it is necessary to deal with the sparse reward problem, including using reward shaping or hierarchical reinforcement learning.

[0183] The loss function formula of the Critic network for reinforcement learning is:

[0184] ;

[0185] The loss function of the Critic network in the reinforcement learning algorithm is defined. The purpose of the Critic network is to learn the state action value function , that is, evaluate in state Next action The purpose of training the Critic network is to make it predict value: , as close to the target as possible value The loss function takes the form of mean squared error.

[0186] : The loss value of the Critic network. The training goal is to minimize this loss value. : The number of experience samples in a mini batch. : Sum all samples in the minibatch. :No. The target of the sample The value is also called the TD target. Its calculation method is shown in the next formula. :Critic network The status of the samples and actions Predicted value. : Parameters of the Critic network.

[0187] Reinforcement Learning Critic Network Objective Value calculation formula:

[0188]

[0189] Used to calculate the target of Critic network learning in DDPG algorithm value It is based on the Bellman equation, which combines immediate rewards with estimates of the value of future states.

[0190] :No. The target Q value of each sample; :No. In the samples, in the state Execute an action Immediate rewards afterward; : Discount factor, the value range is It measures the importance of future rewards relative to current rewards, The closer it is to 1, the more the agent values ​​long-term rewards; :Target Critic Network For the next state and by the target Actor network The action selected in this state Evaluated This represents an estimate of the maximum expected reward that can be obtained starting from the next state. :No. The next state in samples. :Target Actor Network In state Uncertain action on the output. : Parameters of the target Actor network. : Parameters of the target critic network.

[0191] Reinforcement learning Actor network policy gradient formula:

[0192] ;

[0193] Describes the sampled policy gradient used to update the Actor network parameters in the reinforcement learning algorithm. The purpose of the Actor network is to learn an optimal policy. , so that in any state Next select action This gradient guides the Actor network parameters. Update in the direction of actions that produce higher Q values.

[0194] :Objective function of Actor network (The default is expected cumulative return) About its parameters This is the direction and magnitude of the Actor Network parameter update. : The number of experience samples in a mini batch. : Sum all samples in the minibatch. :Critic Network The output of the action The gradient, and in the current state and the action chosen by the Actor network in this state This gradient term indicates if the action selected by the Actor Change it a little bit, The changing trend of the value indicates the direction of action improvement. :Actor Network The output of The gradient, and in the current state This gradient term indicates how the parameters of the Actor network affect the action it outputs; Actor network in state According to the current parameters Output action. The parameters of the Actor network; by multiplying these two gradient terms through the chain rule, the objective function can be obtained About Actor Network Parameters The gradient of , thereby updating the parameters of the Actor network through the gradient ascent method to optimize the strategy.

[0195] To allow the agent to explore different adjustment strategies, noise (including Ornstein-Uhlenbeck process noise or Gaussian noise) can be added to the actions output by the Actor network.

[0196] Application Logic: The RL agent continuously observes the state of the risk assessment network, takes adjustment actions, receives user feedback as rewards, and continuously optimizes its adjustment strategy. This enables the risk assessment network to adaptively improve its risk sensitivity and generalization ability, more closely aligned with the judgment criteria of human experts and actual business needs.

[0197] Sp3. Insight generation and report output based on explainable models.

[0198] This step aims to transform complex model assessment results into human-understandable and credible insights, and generate high-quality due diligence reports based on standardized logic and personalized needs.

[0199] Sp3.1. Use the Explainable Intelligence (XAI) model to perform attribution analysis on the output results of the risk assessment network: identify the key influencing paths and core data evidence that lead to specific risk assessment conclusions (including "high risk" and "recommendation not to pass"), and enhance model transparency and user trust.

[0200] XAI model selection and application: By generating perturbation samples in the local neighborhood of the sample to be explained, a decision tree is used to fit the local behavior of the risk assessment network.

[0201] Perturbation sample generation: For tabular data, the feature values ​​can be randomly perturbed; for text data, words can be randomly masked or replaced; for graph data, node features or edges can be perturbed.

[0202] Locally interpretable model training: Use perturbed samples and their corresponding complex model prediction results to train a weighted linear model where the weights are based on the distance between the perturbed samples and the original samples.

[0203] Explain the output: The coefficients of the linear model can be viewed as the contribution of each feature to the local prediction.

[0204] Application Logic: LIME can explain a single prediction result, telling the user “Why is this particular asset rated as high risk?”

[0205] Feature Importance Ranking: All risk factors input into the risk assessment network are ranked according to the LIME coefficient, and the top-K factors with the greatest impact on the current assessment conclusion are identified.

[0206] Impact Path Tracing (Integrated with Knowledge Graph): If the risk assessment network is based on a GNN, or its input features are linked to a knowledge graph, the key risk factors identified by XAI can be localized to nodes or edges in the knowledge graph. Graph algorithms (including attention-based path search and the shortest critical path algorithm) are then used to trace the transmission paths within the knowledge graph, including how these key factors interact through a series of relationships (including guarantee chains, capital flows, and equity control) to ultimately lead to the occurrence of a risk event. Key risk factors and impact paths are linked back to their original data sources. If a "contract clause risk" factor is identified as significant, the system should be able to locate the specific clause in the contract and highlight it to the user. If a financial indicator anomaly is significant, it should be linked to the corresponding item in the financial statements and its context.

[0207] Provide users with a complete chain of explanations from "what is the risk" to "why is this risk" to "where is the evidence".

[0208] Sp3.2, based on a pre-set and user-adjustable report logic framework and narrative template, embed key information, risk profiles, risk factor combinations, key impact paths, and core data evidence from the context-aware model into the corresponding report sections. Utilizing natural language generation (NLG), output is a due diligence report on non-performing financial assets that includes in-depth analysis, risk warnings, disposal recommendations, and compliance review highlights.

[0209] Report logical framework design: Use XML Schema, JSON Schema or domain-specific languages ​​to define the report's hierarchical structure (including cover, table of contents, abstract, main text chapters, appendixes), chapter titles, content modules to be included (including asset overview, debtor analysis, guarantee analysis, risk assessment summary, disposal recommendations, etc.), and the data type and source of each module.

[0210] Dynamic user adjustment: Provides a graphical configuration interface or parameterized interface, allowing users to select or customize report chapters, modules, level of detail, and presentation style based on different reporting purposes (including internal approval, external transfer, and litigation support), audience types (including executives, sales personnel, and legal affairs), or asset characteristics.

[0211] Narrative template library construction:

[0212] Template type: Design diverse narrative templates based on different analysis scenarios (including “short-term defaults caused by liquidity crisis,” “systemic risks caused by over-collateralization,” and “significant decline in collateral value”), risk levels, asset types, and disposal strategies.

[0213] Template content: Templates contain fixed text and dynamic placeholders. The placeholders are populated by the NLG module based on analysis results. A risk description template might read: "Debtor [Debtor Name] is currently assessed to have a risk rating of [Risk Level] due to the influence of [Key Risk Factor 1] and [Key Risk Factor 2], primarily as shown by [Specific Data Evidence Description]..."

[0214] Template management: supports template creation, editing, version control and on-demand calling.

[0215] Encode the output of upstream modules (including lists of key risk factors, knowledge graph subgraphs, and risk scores) into a format acceptable to the model, including structured inputs (key-value pairs), linearized triple sequences, or text sequences combined with control codes. Generate the target report paragraph or summary.

[0216] A large amount of "input data-output text" parallel corpus is required. It can be extracted from existing written reports or constructed semi-automatically. The "risk description" section is extracted from a large number of due diligence reports, and the corresponding structured risk factors and financial data are used as input. The pre-trained model is fine-tuned on the parallel corpus of the specific field to adapt to the language style, professional terminology and logical structure of the due diligence report. The loss function is usually cross-entropy loss (word-by-word prediction). When generating text, a beam search decoding strategy is adopted to balance the fluency, diversity and accuracy of the generated text. It can generate more natural, flexible and contextually coherent text, which is suitable for writing analytical comments, risk summaries, disposal recommendations and other parts that require complex logic and detailed expression.

[0217] The NLG module populates the various analysis results (in-depth analysis, risk warnings, disposal recommendations, and compliance review points) into the selected report logic framework and narrative template, ultimately generating a complete draft due diligence report (including Word, PDF, and HTML formats).

[0218] Generate a visual map of the risk transmission path:

[0219] Application of attention-weighted graph neural network inference algorithm:

[0220] Using a knowledge graph containing historical risk transmission data or expert-annotated transmission relationships, including GAT, GNNs can be trained to predict transmission probability or impact intensity. Using a pre-trained GNN on the constructed context-aware knowledge graph, we can perform inference and identify potential impact paths from a risk source node to other nodes. Attention weights can be interpreted as a proxy for impact intensity or transmission probability.

[0221] Combined with the Dijkstra algorithm, a high-weight (high-impact / high-probability) transmission path starting from a specific risk event node is searched on a graph with attention weights.

[0222] By inputting one or more initial risk nodes (including "core debtor A experiences a liquidity crisis"), GNN infers and outputs a subgraph, which contains the risk transmission path connected by high-attention-weight edges, as well as the quantitative impact intensity / transmission probability of each node and edge on the path.

[0223] Visual graph and interactive node drilling:

[0224] Front-end implementation: Use a graph visualization library to render a risk transmission path map on a web interface. Visual elements such as node size, color, edge thickness, and arrow direction can be used to indicate risk level, impact intensity, and transmission direction.

[0225] When a user clicks on any node in the graph (including debtors, collateral, and risk events), the system should be able to dynamically request and display detailed information, related attributes, and associated raw data fragments (including contract terms, financial statement screenshots, news links, or text evidence). This requires close coordination between front-end and back-end APIs.

[0226] Predetermine the target audience's profile: focus (including whether they prioritize legal risk or market risk), expertise level (including junior analysts, senior experts, and senior executives), and report purpose (including internal decision-making, external disclosure, and regulatory reporting). Encode these discrete or continuous profile dimensions into numeric vectors. Use an embedding layer to map each dimension to an embedding vector. These vectors are then concatenated or fused through a small network to serve as additional conditional input to the Transformer model.

[0227] Dynamic Adjustment Mechanism: Conditional input guides the model to select or generate different high-level narrative structures. For example, if the audience is senior executives, the model tends to generate a more generalized, conclusion-oriented text structure.

[0228] Adjustments to the professional level of terminology:

[0229] Vocabulary control: Dynamically adjust the vocabulary selection range during decoding based on the professional level profile (including limiting the use of overly professional terms or giving priority to easy-to-understand synonyms).

[0230] Style transfer: Models can be trained to transfer text styles between different levels of expertise, or style control codes can be added during generation.

[0231] Model training: It is necessary to construct training data containing triplets of (conditional profile, structured semantic representation, and target-customized report text). This requires extensive writing and annotation, or the use of weakly supervised learning, multi-task learning, etc.

[0232] The continuous learning and model iteration modules ensure that the system can adapt to changing data distribution, risk patterns and user needs, maintaining and improving its long-term performance.

[0233] Data and Concept Drift Detection Unit: Statistical characteristics of the input data stream: Monitor the distribution (including mean, variance, skewness, kurtosis, and category frequency) of each risk factor (numerical and categorical) for significant changes. For high-dimensional data (including text / image embeddings), changes in its distribution in the low-dimensional projection space can be monitored.

[0234] Model prediction performance: Monitor whether key performance indicators (including precision, recall, F1 score, AUC, and stability of risk scores) decline over time.

[0235] Changes in user feedback patterns: the extent and frequency of user revisions to assessment results, new risk points raised, etc.

[0236] Model relationship drift: The true relationship between features and the target variable changes. This is usually detected indirectly by monitoring a consistent decline in model performance. It can also be detected by comparing models trained on new and old data.

[0237] Triggering the model update process: When significant drift is detected (including when the statistical test value is less than the threshold, or when performance degrades by more than a preset amount), the system automatically takes the following actions:

[0238] 1. Issue an alert: notify operation and maintenance personnel and model maintainers.

[0239] 2. Collect new data: Start collecting new data after drift occurs for model retraining or adjustment.

[0240] 3. Schedule retraining tasks: Automatically or semi-automatically initiate the model retraining, fine-tuning, or structural adjustment process.

[0241] Differential knowledge graph update unit: Data source monitoring: Continuously monitor changes in original data sources (including legal document libraries, financial databases, and public opinion APIs) to capture newly added, modified, or deleted data.

[0242] Change Detection and Extraction: Rerun the information extraction process in Sp1 on the changed data to identify new entities, relationships, attributes, or changes to them. Incrementally merge these changes into the existing knowledge graph. Support efficient addition, deletion, and modification of nodes, edges, and their attributes. Define conflict resolution strategies, including those based on timestamps (latest valid), source trustworthiness, or arbitration. Ensure atomicity, consistency, isolation, and durability (ACID properties) of updates, especially in concurrent update scenarios.

[0243] When the model is iteratively upgraded, transfer learning effectively utilizes the knowledge learned from historical data to accelerate the convergence of the new model, reduce dependence on newly labeled data, and improve the performance of the model on new tasks or new data distributions.

[0244] Parameter transfer uses the parameters of an old model trained on historical data as the initial weights (or weights of some layers) when training a new model. This is suitable for situations where the new and old tasks are similar or when there is less new data. For example, if you need to identify risk factors for a new type of non-performing asset based on the FinBERT model, you can use the general FinBERT weights as a starting point and fine-tune them on new data instead of starting from random initialization.

[0245] Feature representation migration uses the feature representations (including text embedding, image embedding, and graph embedding) learned by the old model directly as input features of the new model, or as part of the new model features. The vector representation of non-performing assets learned by the old risk assessment model can be used as node features by new, more complex assessment models (including GNN).

[0246] Domain adaptation: When the data distributions of the source domain (historical data) and the target domain (new data) differ but the tasks are the same, the model can be adapted to the target domain through instance-based weight adjustment. Historical data mainly comes from non-performing assets in the real estate industry, while new data mainly comes from the manufacturing industry. Domain adaptation can be used to better generalize the risk assessment model to the manufacturing industry.

[0247] The overall operation process is as follows Figure 2 As shown:

[0248] [Start] Task reception and initialization, data collection and preprocessing phase, context-aware modeling phase, risk factor mining phase, risk assessment phase, insight generation phase, automatic report generation phase, user interaction and report review phase, report finalization and output phase, continuous learning and model iteration phase, and finally end, task completion, and resource release.

[0249] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further restrictions, an element defined by the statement "comprising a reference structure" does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0250] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently generating a due diligence report on non-performing financial assets, characterized in that: The method comprises: Sp1. Deep fusion of multimodal data and context-aware modeling: Collect structured, semi-structured and unstructured data related to target financial non-performing assets to obtain heterogeneous data; Mapping the heterogeneous data into a unified semantic space, identifying and establishing entities, attributes, and complex relationships related to non-performing financial assets, and forming a context-aware model of the non-performing financial assets; Sp2. Risk factor mining and adaptive evaluation based on evolutionary computing: Sp2.

1. Initializing a candidate set of risk factors: Based on the context-aware model, using a genetic algorithm evolutionary computation method, iteratively search for optimality within a preset risk dimension space to discover potential and non-explicit risk factor combinations that are coupled with the characteristics of the financial non-performing assets; Sp2.

2. Constructing a risk assessment network: The risk assessment network dynamically adjusts the weights and interactions of various risk factors based on newly input data and historical assessment feedback, and combines it with an integrated learning strategy to quantitatively assess the comprehensive risk level of non-performing financial assets, outputting a risk profile and confidence intervals. Sp3, Insight generation and report output based on interpretable models: Sp3.

1. Use an interpretable intelligent model to conduct attribution analysis on the output results of the risk assessment network to identify the key impact paths and core data evidence that lead to specific risk assessment conclusions; Sp3.

2. Based on the preset report logic framework and narrative template that can be dynamically adjusted by the user, the key information in the context-aware model is generated using natural language to output a due diligence report on non-performing financial assets.

2. The method for intelligently generating a due diligence report on non-performing financial assets according to claim 1, characterized in that: The deep fusion of multimodal data and context-aware modeling further includes: using a pre-trained language model enhanced by an attention mechanism to parse long-distance dependencies and complex clause structures in legal texts, and encoding the extracted key elements of the legal texts and their confidence scores as nodes and weighted edges in the knowledge graph.

3. The method for intelligently generating a due diligence report on non-performing financial assets according to claim 1, characterized in that: The risk factor mining based on evolutionary calculation further includes: The explanatory power of the decision rules generated by the risk factors was quantified by calculating the minimum description length; The prediction accuracy of the risk factor combination is measured by evaluating its F1 score for historical default events on the backtest dataset; The correlation is evaluated by calculating the vector space cosine similarity between the risk factor combination and the patterns in a preset typical non-performing asset risk pattern library verified by domain experts.

4. The method for intelligently generating a due diligence report on non-performing financial assets according to claim 1, characterized in that: The risk assessment network adopts a reinforcement learning mechanism, which further includes: constructing a deep deterministic policy gradient agent, where the agent's state space represents the risk factors and assessment results of the current assets, and the action space corresponds to the adjustment strategy for the weights or activation functions of specific risk factors in the risk assessment network; quantifying the user's confirmation, correction or rejection of the risk assessment into a scalar reward signal to guide the agent's policy learning.

5. The method for intelligently generating a due diligence report on non-performing financial assets according to claim 1, characterized in that: The insight generation of the explainable model also includes: generating a visual map of the risk transmission path, and the generating a visual map of the risk transmission path further includes: using an attention-weighted graph neural network reasoning algorithm on the knowledge graph of the context-aware model to identify and quantify the impact intensity and transmission probability between different risk entities and risk factors to form a transmission path.

6. The method for intelligently generating a due diligence report on non-performing financial assets according to claim 1, characterized in that: In the report output, the natural language generation further includes: utilizing a conditional text generation model, taking a preset profile of the target audience as a conditional input, combining the structured semantic representation extracted from the interpretability model insights, dynamically selecting a narrative template, adjusting the professional level of terminology, and controlling the level of detail of the argument to generate customized report text that meets the needs of specific audiences.

7. The method for intelligently generating a due diligence report on non-performing financial assets according to claim 1, characterized in that: The method further includes a continuous learning and model iteration module, which further includes: Data and concept drift detection unit, which monitors the statistical characteristics of input data streams and changes in user feedback patterns in real time, and automatically triggers the model update process when significant drift is detected; The differential knowledge graph update unit incrementally merges new or changed entities, relationships, and their confidence levels into the existing knowledge graph, and uses transfer learning to apply the knowledge learned from historical data to the iterative upgrade of the new model.

8. A system for intelligently generating due diligence reports on non-performing financial assets, comprising a processor and a memory coupled to the processor, wherein the memory stores computer program instructions, and the computer program instructions are executed by the processor, characterized in that: The system further comprises: Multimodal data deep fusion and context-aware modeling engine: Collect data related to target financial non-performing assets from heterogeneous data sources; Build a context-aware model of financial non-performing assets; Sp2, risk factor mining and adaptive assessment engine based on evolutionary computing: Perform iterative optimization within the preset risk dimension space; Build and operate a risk assessment network; Sp3, Insight Generation and Report Output Engine Based on Explainable Models: Use interpretable intelligent models to conduct attribution analysis on risk assessment results; Output due diligence report on financial non-performing assets.

9. The system for intelligently generating due diligence reports on non-performing financial assets according to claim 8, characterized in that: The pre-trained language model processing unit in the multimodal data deep fusion and context-aware modeling engine further includes: Complex long sentence segmentation and dependency parsing module, used to accurately identify the master-slave structure and restrictive conditions in contract terms; The semantic role labeling module, based on domain knowledge enhancement, labels the specific roles of participants and core legal behaviors in financial transactions.

10. The system for intelligently generating due diligence reports on non-performing financial assets according to claim 8, characterized in that: The risk factor mining and adaptive assessment engine based on evolutionary computing further includes: The feedback-driven model parameter adjustment module further includes: User feedback real-time capture and structured processing interface, used to receive and analyze user annotation information on risk factors and assessment results; An incremental model training and version control unit, wherein the incremental model training and version control unit has an embedded evolutionary algorithm and supports online fine-tuning of the population initialization strategy of the evolutionary algorithm and the local connection weights of the evaluation network; Offline batch retraining scheduler, used to trigger global optimization of the entire model system after accumulating enough new data or feedback.

Images