Non-performing asset cross-scene question and answer framework based on knowledge graph

Abstract

The invention provides a non-performing asset cross-scene question and answer framework based on a knowledge graph, and relates to the field of artificial intelligence and legal science and technology, and the framework comprises a framework, a knowledge fusion layer is provided with a multi-source data access module which is used for obtaining heterogeneous data of a law and regulation database, a judicial case library, an enterprise financial statement system and an asset transaction platform; configuring a legal entity analysis engine, and converting the unstructured legal text into an entity-relationship-attribute triple which can be represented by a knowledge graph based on an ontology mapping technology; setting a financial data standardization component, and converting the financial data in different formats into unified semantic representation through a predefined financial index mapping rule; the knowledge fusion layer of the framework can integrate multi-source heterogeneous data such as laws and regulations, financial data, business processes and the like through a multi-source data access module, a legal entity analysis engine, a financial data standardization component and the like.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and legal technology, and specifically to a cross-scenario question-answering framework for non-performing assets based on a knowledge graph. Background Art

[0002] With the continuous progress of artificial intelligence and big data technologies, the demand for intelligent and efficient solutions in the field of non-performing asset disposal has become increasingly urgent. In the process of non-performing asset disposal, accurately understanding complex business scenarios and quickly conducting risk assessment and decision-making have become the key to improving disposal efficiency and reducing risks.

[0003] According to a legal question-answering system based on intention recognition and knowledge graph disclosed in Chinese Patent Application No. 202410113798.4, it includes a data storage module, a data processing module, a dialogue management module, a natural language understanding module, a model API interface service module, a knowledge calculation module, and a front-end interaction module. The data storage module is responsible for managing Elasticsearch and Neo4j databases; the data processing module collects data through web crawlers and transfers it to the data storage module after processing; the dialogue management module configures the operation mode of the question-answering system and distributes the processing tasks of the question asked; the natural language understanding module is responsible for analyzing the intention of the question statement; the model API interface service module is responsible for docking the dialogue module and the natural language understanding module; the knowledge calculation module is responsible for selecting an appropriate question retrieval method according to the configuration, generating a query statement for retrieval; the front-end interaction module is responsible for receiving user question input and displaying answer feedback; this invention provides a convenient way of legal consultation for the public.

[0004] Although the above technologies have made certain progress in legal question-answering and text understanding, in the practical application of cross-scenario question-answering and disposal of non-performing assets, the existing technologies still have the following problems:

[0005] Problem 1: Insufficient knowledge integration and management. Existing systems mostly focus on single-domain knowledge processing, and it is difficult to integrate multi-source heterogeneous data such as laws and regulations, financial data, and business processes. Knowledge representation mostly uses simple matching or shallow analysis, and it is difficult to accurately understand the relationship between complex legal concepts and multi-source knowledge. Moreover, the knowledge update mechanism is imperfect and cannot adapt to the dynamic requirements of regulatory changes and business scenarios in a timely manner, making it difficult to build a comprehensive and accurate knowledge system.

[0006] Problem 2: Limited reasoning and decision-making capabilities. Traditional reasoning methods lack multi-step reasoning and interpretability support. When facing complex legal, financial, market and other comprehensive scenarios in non-performing asset disposal, it is difficult to conduct in-depth analysis and effective decision-making. At the same time, the reasoning efficiency and scenario adaptability are insufficient, and it cannot quickly respond to the diverse needs in different scenarios, making it difficult to meet the timeliness and accuracy requirements for risk assessment and disposal plan formulation in actual business.

[0007] Problem 3: Insufficient system application performance. The existing system has poor scalability in architecture design and is difficult to cope with the growing data scale and business complexity in the field of non-performing assets; its concurrency performance is low and cannot meet the needs of multiple users such as financial institutions and asset management companies using it simultaneously; the security protection mechanism is imperfect, and it is difficult to ensure the security of sensitive information related to non-performing assets during data transmission, storage, and processing, and it is difficult to meet the actual application requirements of large scale and high reliability.

[0008] Therefore, a cross-scenario question-answering framework for non-performing assets based on knowledge graph is needed to solve the above problems. Summary of the Invention

[0009] Technical Problems to be Solved

[0010] Aiming at the deficiencies of the prior art, the present invention provides a cross-scenario question-answering framework for non-performing assets based on knowledge graph, which solves the problems in the following background technology.

[0011] Technical Solution

[0012] To achieve the above objectives, the present invention is realized through the following technical solutions: A cross-scenario question-answering framework for non-performing assets based on knowledge graph, including a framework, the framework includes a knowledge fusion layer, a dynamic reasoning layer, a scenario adaptation layer, a decision verification layer, a knowledge evolution module, a cross-scenario migration module, and a federated knowledge learning architecture. Each layer works together through a standardized service interface. The specific composition and data flow mode are as follows:

[0013] The knowledge fusion layer is used to integrate heterogeneous data and construct a unified knowledge graph, including a multi-source data access module, a legal entity parsing engine, a financial data standardization component, a business process knowledge module, and a knowledge fusion engine. Among them: The multi-source data access module is configured to obtain heterogeneous data from a legal regulations database, a judicial case database, an enterprise financial statement system, and an asset trading platform; the legal entity parsing engine, based on ontology mapping technology, converts unstructured legal texts into entity-relationship-attribute triples that can be represented by the knowledge graph; the financial data standardization component converts financial data in different formats into a unified semantic representation through predefined financial indicator mapping rules; the business process knowledge module analyzes historical disposal cases based on process mining technology and extracts a standardized non-performing asset disposal process template; the knowledge fusion engine uses graph embedding technology to achieve semantic alignment of legal, financial, and business knowledge subgraphs and generates a fused knowledge graph;

[0014] The dynamic reasoning layer is used to process user queries and generate preliminary answers, including a query understanding module, a multi-strategy reasoning engine, a three-dimensional risk assessment model, and an answer generation module. Specifically: The query understanding module converts the user's natural language query into a semantic query expression executable by the knowledge graph; the multi-strategy reasoning engine performs cross-entity multi-hop reasoning, supporting rule reasoning, case-based reasoning, and graph reasoning; the three-dimensional risk assessment model calculates the legal risk index, market volatility index, and asset liquidity index based on the entity relationship network in the knowledge graph; the answer generation module generates preliminary answers containing evidence chains.

[0015] Preferably, the scenario adaptation layer is used to adjust the answer output according to the user context, including a scenario recognition module, a cross-scenario knowledge transfer module, an answer reconstruction engine, and a multi-modal display module. Specifically: The scenario recognition module automatically matches the corresponding disposal scenario template based on the user context information; the cross-scenario knowledge transfer module identifies reusable knowledge components between different scenarios; the answer reconstruction engine converts the preliminary answer into a structured output according to the requirements of the scenario template; the multi-modal display module supports three output forms: text, charts, and knowledge graph visualization.

[0016] The decision verification layer is used to verify the accuracy and feasibility of the answer, including a multi-dimensional evaluation module, a disposal plan executability verification module, an evidence chain traceability engine, and a feedback learning module. Specifically: The multi-dimensional evaluation module verifies the legal compliance, financial feasibility, and process integrity of the answer based on predefined evaluation metrics; the disposal plan executability verification module predicts the resource consumption and time cost of the disposal process through simulation technology; the evidence chain traceability engine records the complete reasoning path from the query to the answer; the feedback learning module updates the relationship weights and reasoning rules of the knowledge graph according to user feedback.

[0017] The knowledge evolution module is used to dynamically update and optimize the knowledge graph, including a knowledge update detection unit, a knowledge incremental update engine, a knowledge conflict resolution module, and a knowledge graph dynamic evolution mechanism. Specifically: The knowledge update detection unit monitors the update of the data source in real time; the knowledge incremental update engine integrates the new knowledge into the existing knowledge graph; the knowledge conflict resolution module resolves the conflicts in the knowledge graph.

[0018] The cross-scenario migration module is used to support the migration of knowledge between different scenarios, including a scenario feature extraction unit, a meta-learning engine, a solution verification unit, and a cross-modal knowledge representation module. Specifically: The scenario feature extraction unit extracts features from the scenario-related information; the meta-learning engine learns the general knowledge representation between different scenarios; the solution verification unit verifies the effectiveness of the cross-scenario migration solution.

[0019] The described federal knowledge learning architecture is used for distributed knowledge processing, including local knowledge processing nodes, a secure aggregation center, a privacy protection mechanism, and an institutional credit assessment module, where: the local knowledge processing nodes perform preprocessing and feature extraction on local data; the secure aggregation center aggregates the knowledge uploaded by each node; the privacy protection mechanism ensures privacy security during data transmission and aggregation;

[0020] Each functional module of the framework is implemented through a standardized microservice interface, supporting independent deployment and horizontal expansion; the data flow between layers adopts a message queue mechanism, and different topics are set for the message queue to distinguish different types of data, including knowledge fusion data topics, inference request data topics, and answer output data topics; the system as a whole supports high-concurrency query processing;

[0021] The legal entity parsing engine in the knowledge fusion layer includes a legal concept extraction unit, a relationship extraction unit, and an attribute annotation unit, where: the legal concept extraction unit uses a pre-trained deep learning model, combined with fine-tuning of legal domain corpora, to identify legal concepts in legal texts; the relationship extraction unit extracts the relationships between legal concepts based on a neural network combined with a legal ontology library; the attribute annotation unit completes the annotation of legal entity attributes through the combination of rule templates and semi-supervised learning;

[0022] The described financial data standardization component contains various financial indicator mapping rules, covering asset, liability, equity, income, and expense accounts, and supports data format parsing, conversion, and verification;

[0023] The disposal process templates extracted by the business process knowledge module include asset inventory, value assessment, judicial auction, debt restructuring, and bankruptcy liquidation, and each template contains process steps, participating parties, execution standards, and time nodes.

[0024] Preferably, the multi-strategy inference engine in the dynamic inference layer includes a rule inference unit, a case inference unit, a graph inference unit, and an inference fusion unit, where: the rule inference unit is based on legal, financial, and business rule libraries and uses forward and backward reasoning; the case inference unit matches historical disposal cases through a similarity algorithm; the graph inference unit uses a graph neural network to model the relationships in the knowledge graph; the inference fusion unit integrates the inference results through a weighting mechanism;

[0025] The three-dimensional risk assessment model calculates the legal risk index, market volatility index, and asset liquidity index through weighted calculation, and adjusts the weights in combination with the entity relationship tightness and frequency; the evidence chain contains at least five inference steps, including query parsing, subgraph retrieval, rule matching, case comparison, and result integration.

[0026] Preferably, the scene recognition module of the scene adaptation layer adopts a multi-modal fusion model, and calculates the matching probabilities of each scene template based on the text information input by the user, historical query records, current time, and geographical location;

[0027] The scene templates of the scene adaptation layer include judicial auction scene templates, debt restructuring scene templates, and bankruptcy liquidation scene templates. Each template contains multiple structured sub-templates, and each sub-template defines at least 10 standardized fields. Among them, the sub-templates of the judicial auction scene template include auction asset information sub-template, auction process stage sub-template, participating entity information sub-template, auction result record sub-template, and risk reminder sub-template. The auction asset information sub-template contains fields such as asset name, asset category, appraised value, starting price, and ownership certificate number;

[0028] When visualizing the knowledge graph, the multi-modal display module uses different graphic layouts and color coding to distinguish entity and relationship types.

[0029] Preferably, the multi-dimensional evaluation module of the decision verification layer includes a legal compliance evaluation unit, a financial feasibility evaluation unit, and a process integrity evaluation unit. Among them: the legal compliance evaluation unit matches legal clauses with disposal plans through a rule engine and a legal knowledge graph to judge compliance; the financial feasibility evaluation unit evaluates the financial risks of the disposal plan based on a simulation method; the process integrity evaluation unit verifies the integrity of the disposal process through process modeling technology;

[0030] The disposability verification module of the disposal plan constructs a simulation model containing resource, activity, and event elements based on discrete event simulation technology, and predicts resource consumption and time costs.

[0031] Preferably, the knowledge update detection unit of the knowledge evolution module adopts an incremental detection algorithm to monitor the updates of data sources such as the laws and regulations database and the enterprise financial statement system in real time;

[0032] The knowledge incremental update engine uses graph embedding technology to integrate new knowledge into the existing knowledge graph;

[0033] The knowledge conflict resolution module resolves conflicts based on a conflict resolution strategy library through priority sorting and negotiation strategies;

[0034] The dynamic evolution mechanism of the knowledge graph is based on a reinforcement learning algorithm, and adjusts the structure and parameters of the knowledge graph according to the disposal effect.

[0035] Preferably, the scene feature extraction unit of the cross-scene migration module uses a neural network to extract features from the scene text information and the knowledge graph structure;

[0036] The meta - learning engine learns the general knowledge representation between different scenarios based on the model - agnostic meta - learning algorithm;

[0037] The solution verification unit verifies the effectiveness of the cross - scenario migration solution by comparing simulation with actual cases;

[0038] The cross - modal knowledge representation module converts unstructured judicial documents and evaluation reports into semantic vectors that can be represented by a knowledge graph.

[0039] Preferably, the local knowledge processing node of the federated knowledge learning architecture pre - processes and extracts features from local non - performing asset data;

[0040] The secure aggregation center aggregates the knowledge uploaded by each node through a secure multi - party computing protocol;

[0041] The privacy protection mechanism uses encryption technology to ensure privacy security during data transmission and aggregation;

[0042] The institutional credit assessment module records the knowledge contribution and data quality of each participating institution based on blockchain technology.

[0043] Preferably, the framework supports cross - jurisdiction reasoning, including multi - jurisdiction knowledge sub - graphs, legal conflict identification engines, conflict resolution decision - making modules, and legal knowledge migration modules, where: the legal conflict identification engine compares legal clauses in different jurisdiction knowledge sub - graphs to identify potential conflicts; the conflict resolution decision - making module migrates the disposal experience of one jurisdiction to another based on legal analogical reasoning technology;

[0044] The legal knowledge migration module supports cross - jurisdiction knowledge migration based on legal analogical reasoning technology.

[0045] Preferably, the framework is configured with a system support module. When using the system support module, each functional module of the framework is interconnected through a standardized microservice interface, supporting independent deployment and horizontal expansion; the data flow between layers uses the Kafka message queue mechanism, and the message queue sets knowledge fusion data topics, inference request data topics, and answer output data topics. The framework is equipped with an intelligent prompt module;

[0046] The intelligent prompt module recommends legal clauses and disposal strategies based on the user's historical query behavior and the current disposal scenario, using collaborative filtering algorithms and association rule mining techniques; the system as a whole supports the processing of 1000 concurrent user query requests.

[0047] Beneficial effects

[0048] The present invention provides a cross - scenario Q&A framework for non - performing assets based on a knowledge graph. It has the following beneficial effects:

[0049] 1. The knowledge fusion layer of this framework can integrate heterogeneous data from multiple sources such as laws and regulations, financial data, and business processes through multi-source data access modules, legal entity parsing engines, financial data standardization components, etc. It uses technologies such as ontology mapping and graph embedding to achieve deep knowledge fusion and semantic alignment, solves the problem of insufficient knowledge fusion and management in the existing technology, and constructs a comprehensive, accurate, and dynamically updated knowledge graph. The knowledge evolution module, based on timestamp detection and reinforcement learning mechanisms, can update the knowledge graph in real time to timely adapt to regulatory changes and the dynamic needs of business scenarios, providing a solid data foundation for non-performing asset disposal.

[0050] 2. The dynamic reasoning layer of this invention is equipped with a multi-strategy reasoning engine and a three-dimensional risk assessment model. It combines rule reasoning, case-based reasoning, and graph reasoning to achieve cross-entity multi-hop reasoning and in-depth analysis, and provides interpretability support through the generation of evidence chains, effectively solving the problem of limited reasoning and decision-making capabilities of traditional reasoning methods in non-performing asset disposal. Facing complex comprehensive scenarios such as law, finance, and market, it can quickly and accurately conduct risk assessment and formulate disposal plans, improving the timeliness and accuracy of decision-making and meeting the actual business needs.

[0051] 3. The scenario adaptation layer of this invention, based on multi-modal fusion scenario recognition technology and meta-learning algorithms, can automatically match disposal scenario templates, achieve rapid reconstruction and multi-modal display of answers, and significantly improve the system's adaptability to different scenarios. At the same time, the framework adopts a federated knowledge learning architecture and a standardized microservice interface design, combines Docker containers and Kubernetes for independent deployment and horizontal scaling, and cooperates with the Kafka message queue mechanism to greatly improve the system's concurrent processing ability and data flow efficiency. In addition, the security aggregation center and privacy protection mechanism use technologies such as homomorphic encryption and blockchain to ensure the security of data during transmission and storage, comprehensively solve the problem of insufficient application performance of the existing system, and can meet the application requirements of large-scale and highly reliable non-performing asset disposal. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is the question-and-answer flow chart of this invention;

[0053] Figure 2 is the composition diagram of the question-and-answer framework of this invention;

[0054] Figure 3 is the evaluation simulation diagram of this invention. DETAILED DESCRIPTION OF THE INVENTION

[0055] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. Specific Embodiment 1:

[0057] As Figures 1-3 shown, a bad asset cross-scenario Q&A framework based on a knowledge graph, the framework includes a knowledge fusion layer, a dynamic reasoning layer, a scenario adaptation layer and a decision verification layer, and each layer works together through a standardized service interface. The specific composition and data flow mode are as follows:

[0058] The knowledge fusion layer sets up a multi-source data access module for obtaining heterogeneous data from legal regulations databases, judicial case libraries, enterprise financial statement systems, and asset trading platforms; configures a legal entity parsing engine to convert unstructured legal texts into entity-relationship-attribute triples that can be represented by the knowledge graph based on ontology mapping technology; sets up a financial data standardization component to convert financial data in different formats into a unified semantic representation through predefined financial indicator mapping rules; constructs a business process knowledge module to perform process mining on historical disposal cases based on Petri net theory and extract a standardized disposal process template; through a knowledge fusion engine, uses graph embedding technology to achieve semantic alignment of legal, financial, and business triple knowledge subgraphs and generate a fused knowledge graph;

[0059] The dynamic reasoning layer sets up a query understanding module to convert the user's natural language query into a semantic query expression executable by the knowledge graph; configures a multi-strategy reasoning engine to perform cross-entity multi-hop reasoning; constructs a three-dimensional risk assessment model to calculate the legal risk index, market volatility index, and asset liquidity index based on the entity relationship network in the knowledge graph; sets up an answer generation module to generate a preliminary answer containing an evidence chain;

[0060] The scenario adaptation layer sets up a scenario recognition module to automatically match the corresponding disposal scenario template based on the user's context information; configures an answer reconstruction engine to convert the preliminary answer into a structured output according to the formatting requirements of the scenario template; constructs a multi-modal display module to support three output forms: text, charts, and knowledge graph visualization;

[0061] The decision verification layer: sets up a multi-dimensional evaluation module to verify the legal compliance, financial feasibility, and process integrity of the preliminary answer based on a predefined evaluation index system; configures an evidence chain traceability engine to record the complete reasoning path from the query to the answer; constructs a feedback learning module to update the relationship weights and reasoning rules in the knowledge graph according to the user feedback data.

[0062] In the legal entity parsing engine of the knowledge fusion layer, the legal concept extraction unit adopts a model that combines a bidirectional encoder representation and a conditional random field. By performing pre-training and fine-tuning on a large-scale legal text corpus, it identifies legal concepts in legal texts. The relationship extraction unit uses a bidirectional long short-term memory network based on the attention mechanism and combines it with an ontology library in the legal domain to extract relationships between legal concepts. The attribute annotation unit uses a combination of rule-based template matching and semi-supervised learning to complete the annotation of legal entity attributes. In the financial data standardization component of the knowledge fusion layer, 128 financial indicator mapping rules cover financial accounts such as asset class, liability class, owner's equity class, income class, and expense class. Each rule contains specific algorithms for source data format parsing, target data format conversion, and data verification. Through regular expression matching and semantic mapping functions, it realizes the standardized conversion of data from different enterprise financial statement systems. The standardized disposal process templates extracted by the business process knowledge module in the knowledge fusion layer include asset inventory, value assessment, judicial auction, debt restructuring, and bankruptcy liquidation. Each template contains elements such as process steps, participating parties, execution standards, and time nodes. Among them, the asset inventory template includes steps such as asset inventory and ownership verification; the value assessment template includes content such as assessment method selection and value calculation; the judicial auction template includes links such as auction announcement release and bidding process; the debt restructuring template includes processes such as plan design and negotiation; the bankruptcy liquidation template includes stages such as asset liquidation and debt repayment.

[0063] In the multi-strategy inference engine of the dynamic inference layer, the rule inference unit adopts a hybrid inference strategy that combines forward inference and backward inference, and conducts inference based on a preset legal rule library, financial rule library, and business rule library; the case inference unit uses a case retrieval algorithm based on cosine similarity to match similar cases from the historical disposal case library; the graph inference unit adopts a graph attention network model to model and reason about the complex relationships in the knowledge graph; the inference fusion unit fuses the results of the three inference units through a weighted voting mechanism. In the three-dimensional risk assessment model of the dynamic inference layer, when calculating the legal risk index, different weights are assigned to different types of legal violations, and the weighted sum is obtained based on factors such as the number of occurrences of violations and the amount involved; the market volatility index is calculated after predicting and analyzing the asset trading price data based on the autoregressive integrated moving average model of time series analysis; the asset liquidity index is comprehensively calculated by constructing an evaluation model that includes indicators such as asset turnover rate and days to liquidate; the risk level is divided into five levels according to the total score weighted by the three indexes. The evidence chain of the dynamic inference layer includes at least 5 consecutive inference steps and their knowledge sources. The consecutive inference steps include user query intention parsing, knowledge graph subgraph retrieval, inference rule matching, relevant case comparison, and result integration; the knowledge sources are specifically the corresponding specific data records in the laws and regulations database, judicial case library, enterprise financial statement system, and asset trading platform. Moreover, when calculating each index, the three-dimensional risk assessment model adjusts the calculation weights in combination with factors such as the tightness and frequency of the association relationships between entities in the knowledge graph; during the generation process of the evidence chain, the credibility of each inference step is evaluated, and the evaluation results are used as a reference basis for answer generation.

[0064] The scenario templates in the scenario adaptation layer include the judicial auction scenario template, the debt restructuring scenario template, and the bankruptcy liquidation scenario template; each scenario template contains 5 structured sub-templates, and each sub-template defines at least 10 standardized fields. Among them, the structured sub-templates of the judicial auction scenario template include the auction asset information sub-template, the auction process stage sub-template, the participating entity information sub-template, the auction result record sub-template, and the risk warning sub-template. The auction asset information sub-template includes fields such as asset name, asset category, appraised value, starting price, and title certificate number. The scenario recognition module in the scenario adaptation layer automatically matches the corresponding disposal scenario template based on the user's context information. It adopts a model based on multi-modal fusion, fusing context information such as the text information input by the user, historical query records, current time, and geographical location, and performs matching by calculating the matching probability of each scenario template. The cross-scenario knowledge transfer module in the scenario adaptation layer identifies reusable knowledge components between different scenario templates based on the meta-learning algorithm. Specifically, by constructing a meta-learning model and training on the historical data of multiple scenarios, it extracts the common features and knowledge patterns between scenarios to achieve the transfer of knowledge components. In the scenario adaptation layer, when the answer reconstruction engine converts the preliminary answer into a structured output, it filters, classifies, and formats the data according to the characteristics of different scenario templates; when the multi-modal display module visualizes the knowledge graph, it uses different graph layouts and color coding methods to distinguish different types of entities and relationships.

[0065] The multi-dimensional evaluation module in the decision verification layer includes a legal compliance evaluation unit, a financial feasibility evaluation unit, and a process integrity evaluation unit. Among them, the legal compliance evaluation unit combines a rule engine with a legal knowledge graph to judge legal compliance by matching legal clauses with legal elements in the disposal plan; the financial feasibility evaluation unit conducts a risk assessment of the financial indicators in the disposal plan based on the Monte Carlo simulation method; the process integrity evaluation unit models and analyzes the disposal process through a Petri net model to verify the integrity of the process. The disposal plan executability verification module in the decision verification layer simulates the execution process of the disposal process based on discrete event simulation technology. By constructing a simulation model containing elements such as resources, activities, and events, it predicts resource consumption and time costs. The decision verification layer configures an evidence chain traceability engine to record the complete reasoning path from query to answer; constructs a feedback learning module to update the relationship weights and reasoning rules in the knowledge graph according to the user feedback data.

[0066] The framework also includes a knowledge evolution module, which is equipped with a knowledge update detection unit, a knowledge incremental update engine, and a knowledge conflict resolution module. The knowledge update detection unit uses an incremental detection algorithm based on timestamps to monitor the updates of data sources such as legal and regulatory databases and enterprise financial statement systems in real time. The knowledge incremental update engine uses a knowledge fusion algorithm based on graph embedding to integrate new knowledge with the existing knowledge graph. The knowledge conflict resolution module is based on a conflict resolution strategy library and uses a priority ranking and negotiation strategy to resolve knowledge conflicts. The dynamic evolution mechanism of the knowledge graph of the framework automatically adjusts the structure and parameters of the knowledge graph based on a reinforcement learning algorithm according to the disposal effect.

[0067] The framework also includes a cross-scenario migration module, which is equipped with a scenario feature extraction unit, a meta-learning engine, and a solution verification unit. The scenario feature extraction unit uses a convolutional neural network to extract features from scenario text information and combines a graph convolutional network to extract features from the knowledge graph structure information related to the scenario. The meta-learning engine is based on a model-agnostic meta-learning algorithm to learn the general knowledge representation between different scenarios. The solution verification unit verifies the effectiveness of the cross-scenario migration solution by comparing simulation with actual cases. The cross-modal knowledge representation module converts unstructured judicial documents and evaluation reports into semantic vectors that can be represented by the knowledge graph.

[0068] The framework adopts a federated knowledge learning architecture, which is equipped with local knowledge processing nodes, a secure aggregation center, and a privacy protection mechanism. The local knowledge processing nodes preprocess and extract features from local non-performing asset data. The secure aggregation center aggregates the knowledge uploaded by each local knowledge processing node through a secure multi-party computing protocol. The privacy protection mechanism uses homomorphic encryption technology to ensure the privacy security of data during transmission and aggregation. The institutional credit assessment module records the knowledge contribution and data quality of each participating institution based on blockchain technology.

[0069] The framework supports cross-jurisdiction reasoning, which is equipped with a multi-jurisdiction knowledge subgraph, a legal conflict identification engine, and a conflict resolution decision-making module. The legal conflict identification engine identifies potential legal conflicts by comparing legal provisions and rules in knowledge subgraphs of different jurisdictions. The conflict resolution decision-making module is based on legal analogical reasoning technology to transfer the disposal experience of one jurisdiction to another to resolve legal conflicts. The legal knowledge migration module transfers the disposal experience of one jurisdiction to another based on legal analogical reasoning technology.

[0070] The framework enables each functional module to be interconnected through standardized microservice interfaces via the configured system support module, supporting independent deployment and horizontal scaling; the data flow between layers adopts the Kafka message queue mechanism, and different topics are set for the message queue to distinguish different types of data, including the knowledge fusion data topic, the inference request data topic, and the answer output data topic; each functional module implements standardized microservices through RESTful API interfaces, and each microservice can be independently deployed through Docker containers and horizontally scaled through Kubernetes; the system as a whole supports the processing of 1000 concurrent user query requests. The intelligent prompt module, based on the user's historical query behavior and the current handling scenario, uses collaborative filtering algorithms and association rule mining techniques to actively recommend legal provisions and handling strategies that may be needed.

[0071] Through the coordinated operation of the knowledge fusion layer, the dynamic inference layer, the scenario adaptation layer, and the decision verification layer, the bad asset cross-scenario Q&A framework based on the knowledge graph realizes the efficient answering of bad asset-related questions and the scientific evaluation of disposal solutions. The layers are closely connected and interlocked. The specific working methods are as follows:

[0072] Knowledge fusion layer: The multi-source data access module serves as the data entry point. With the help of interface protocols that adapt to the characteristics of different data sources, such as database connection protocols, API interfaces, etc., it regularly or real-time collects heterogeneous data from legal regulation databases, judicial case databases, enterprise financial statement systems, and asset trading platforms, covering rich content such as legal provisions, judicial judgment documents, enterprise balance sheets, and asset trading records. These raw data are then transmitted to the subsequent modules in the knowledge fusion layer for in-depth processing.

[0073] The legal entity parsing engine analyzes unstructured legal texts, uses a model that combines bidirectional encoder representations and conditional random fields to identify legal concepts, extracts the relationships between concepts using a bidirectional long short-term memory network based on the attention mechanism, and completes attribute annotation by combining rule-based template matching and semi-supervised learning. Finally, the legal text is transformed into an entity-relationship-attribute triple. The financial data standardization component unifies financial data in various formats into a standard semantic representation according to 128 predefined mapping rules covering multiple financial items through regular expression matching, semantic mapping function conversion, and data verification algorithms. The business process knowledge module mines historical disposal cases based on Petri net theory and extracts standardized disposal process templates such as asset inventory and value assessment, clarifying the process steps, participating entities, and other elements of each template.

[0074] The knowledge fusion engine uses graph embedding technology to semantically align the legal, financial, and business tripartite knowledge subgraphs, eliminate semantic differences, and generate a fused knowledge graph. This fused knowledge graph is not only the final product of the knowledge fusion layer but also the core data foundation for subsequent reasoning and analysis in the dynamic reasoning layer, providing rich and structured knowledge reserves for the operation of the entire framework.

[0075] Dynamic reasoning layer: After receiving the natural language query input by the user, the query understanding module uses natural language processing technology to convert it into a semantic query expression executable by the knowledge graph, and then passes this expression to the multi-strategy reasoning engine. The rule reasoning unit in the multi-strategy reasoning engine follows a forward and backward combined reasoning strategy based on the preset legal, financial, and business rule libraries; the case reasoning unit uses a case retrieval algorithm based on cosine similarity to match similar cases from the historical disposal case library; the graph reasoning unit uses a graph attention network model to model and reason about the complex relationships in the knowledge graph, and the reasoning fusion unit integrates the results of the three through a weighted voting mechanism.

[0076] The three-dimensional risk assessment model is based on the entity relationship network of the knowledge graph. Combining the entity association information in the fused knowledge graph generated by the knowledge fusion layer, it calculates the legal risk index, market volatility index, and asset liquidity index, and divides the risk levels according to the weighted total score. During the calculation process, the weights are dynamically adjusted according to the closeness and frequency of the associations between entities. The answer generation module generates a preliminary answer containing at least 5 consecutive reasoning steps and their knowledge sources based on the reasoning results and risk assessment information. The knowledge source of each reasoning step points to the data sources processed by the knowledge fusion layer, and at the same time, the credibility of the reasoning steps is evaluated and incorporated into the answer. After the preliminary answer is formed, it is transmitted to the scenario adaptation layer for further processing.

[0077] Scenario adaptation layer:

[0078] The scenario recognition module uses a multi-modal fusion-based model to fuse context information such as the user input text information and historical query records, calculates the matching probabilities of each scenario template (judicial auction, debt restructuring, bankruptcy liquidation, etc.), and automatically selects the most matching disposal scenario template from the knowledge such as the standardized disposal process templates provided by the knowledge fusion layer.

[0079] The answer reconstruction engine screens, classifies, and formats the preliminary answers generated by the dynamic inference layer according to the formatting requirements of the selected scenario template, and converts them into structured outputs. The multi-modal display module supports three output forms: text, charts, and knowledge graph visualization, presenting the processed answers to users in an intuitive and diverse manner. The cross-scenario knowledge transfer module, based on the meta-learning algorithm, trains a meta-learning model on multi-scenario historical data, extracts common features and knowledge patterns, and identifies reusable knowledge components. When encountering a new scenario, relevant knowledge can be invoked from the knowledge system of the knowledge fusion layer for transfer and application, enhancing the adaptability of the framework. Among them, the cross-scenario knowledge transfer module in the scenario adaptation layer identifies reusable knowledge components between different scenario templates based on the meta-learning algorithm. Specifically, by constructing a meta-learning model and training it on the historical data of multiple scenarios, the common features and knowledge patterns between scenarios are extracted to achieve the transfer of knowledge components. When encountering a new scenario, relevant knowledge can be invoked from the knowledge system of the knowledge fusion layer for transfer and application, enhancing the adaptability of the framework. The processed answers will be passed to the decision verification layer for verification when necessary.

[0080] Decision Verification Layer: The multi-dimensional evaluation module includes a legal compliance evaluation unit, a financial feasibility evaluation unit, and a process integrity evaluation unit. From the perspectives of law, finance, and process respectively, and based on the predefined evaluation index system, it verifies the answers or disposal plans output by the scenario adaptation layer. The legal compliance evaluation unit uses a combination of a rule engine and a legal knowledge graph to match legal clauses with the legal elements of the disposal plan to judge compliance; the financial feasibility evaluation unit evaluates the financial index risks of the disposal plan based on the Monte Carlo simulation method; the process integrity evaluation unit verifies the integrity of the disposal process through a Petri net model. Among them, the executability verification module of the disposal plan in the decision verification layer simulates the execution process of the disposal process based on discrete event simulation technology. By constructing a simulation model containing elements such as resources, activities, and events, it predicts resource consumption and time costs. This module can discover in advance the problems that may be encountered in the actual execution of the disposal plan, provide a basis for the optimization of the plan, and ensure that the finally output disposal plan has high executability. The evidence chain traceability engine records the complete reasoning path from query to answer; a feedback learning module is constructed to update the relationship weights and inference rules in the knowledge graph according to user feedback data.

[0081] Other Modules: The knowledge evolution module works closely with the knowledge fusion layer. The knowledge update detection unit uses an incremental detection algorithm based on timestamps to monitor data source updates in real time. Once an update is detected, the knowledge incremental update engine uses a knowledge fusion algorithm based on graph embedding to integrate the new knowledge into the existing knowledge graph. The knowledge conflict resolution module solves knowledge conflicts based on the conflict resolution strategy library. The knowledge graph dynamic evolution mechanism adjusts the graph structure and parameters according to the disposal effect based on the reinforcement learning algorithm, providing a more accurate knowledge basis for the dynamic inference layer.

[0082] The scene feature extraction unit of the cross-scenario migration module uses a convolutional neural network and a graph convolutional network to extract scene text and knowledge graph structure features. The meta-learning engine learns general knowledge representations based on the model-agnostic meta-learning algorithm. The solution verification unit verifies the effectiveness of the migration solution through simulation and comparison with actual cases, and the process involves the invocation and processing of knowledge in the knowledge fusion layer. The cross-modal knowledge representation module converts unstructured judicial documents and evaluation reports into semantic vectors that can be represented by the knowledge graph, enriching the knowledge content of the knowledge fusion layer.

[0083] The framework adopts a federated knowledge learning architecture. The local knowledge processing node preprocesses and extracts features from local non-performing asset data. The secure aggregation center aggregates the knowledge uploaded by each node through a secure multi-party computing protocol. The privacy protection mechanism uses homomorphic encryption technology to ensure data security. The institutional credit assessment module records the knowledge contribution and data quality of each participating institution based on blockchain technology. This architecture provides a secure and reliable environment for data collection and processing in the knowledge fusion layer.

[0084] The framework supports cross-jurisdictional reasoning. The legal conflict identification engine compares the legal provisions and rules in the knowledge subgraphs of multiple jurisdictions to identify potential legal conflicts. The conflict resolution decision-making module migrates the disposal experience of one jurisdiction to another to resolve conflicts based on legal analogical reasoning technology, relying on the multi-jurisdictional knowledge subgraph constructed in the knowledge fusion layer.

[0085] The functional modules of the framework are all implemented through standardized microservice interfaces. Based on the RESTful API interface, they are independently deployed using Docker containers and horizontally scaled through Kubernetes. The Kafka message queue mechanism is used for data flow between layers, and different topics such as knowledge fusion data, inference request data, and answer output data are set to distinguish data types. The system as a whole supports the processing of 1000 concurrent user query requests. The intelligent prompt module extracts relevant legal provisions and disposal strategies from the knowledge system in the knowledge fusion layer for active recommendation based on the user's historical query behavior and the current disposal scenario, using collaborative filtering algorithms and association rule mining techniques. Specific Embodiment 2:

[0087] As Figures 1-3 shown, the key algorithms mentioned in Embodiment 1 are analyzed in detail below, including their core mathematical formulas and explanations:

[0088] The graph embedding technology in the knowledge fusion layer:

[0089] Graph embedding technology is used to map the entities and relationships of the knowledge graph into a low-dimensional vector space for semantic alignment and fusion. One of the common methods is the TransE model, and its core formula is as follows:

[0090]

[0091] Among them: They are the embedding vectors of the head entity, relation, and tail entity respectively; Denotes the L2 norm; Is the scoring function, used to measure the triple For reasonableness. The loss function is usually defined as:

[0092]

[0093] Among them: Is the set of correct triples, Is the set of negative sample triples; Is the margin hyperparameter.

[0094] Explanation and problem-solving:

[0095] Function of the formula: TransE assumes that the relation Is a translational operation from the head entity To the tail entity That is, . By optimizing the scoring function, entities and relations are embedded into a unified vector space, maintaining semantic consistency. Problem solved: In the knowledge fusion layer, there are significant semantic differences among the legal, financial, and business knowledge subgraphs (for example, "debt restructuring" in legal texts may have a different meaning from "debt restructuring" in financial statements). Through graph embedding technology, the framework aligns heterogeneous knowledge subgraphs to a unified semantic space, generating a fused knowledge graph.

[0096] Multi-strategy reasoning in the dynamic reasoning layer (graph reasoning unit):

[0097] The graph reasoning unit uses a graph attention network (GAT) for complex relationship modeling and reasoning, and its core formula is as follows:

[0098]

[0099]

[0100] Among them: Is the initial embedding vector of nodes And ; Is the learnable weight matrix; Is the weight vector of the attention mechanism; Denotes vector concatenation; Is the attention weight of node To its neighbor node ; Is node The set of neighbor nodes of is an activation function (such as ReLU); is the node The updated embedding vector.

[0101] Explanation and problem-solving:

[0102] Function of the formula: GAT dynamically calculates the contribution weights of each neighbor node to the current node through the attention mechanism and updates the node representation according to weighted aggregation This allows the framework to capture complex multi-hop relationships in the knowledge graph.

[0103] Problem solved: In the disposal of non-performing assets, the relationships between entities are complex (for example, the debt restructuring of enterprise A may involve legal risks, financial data, and market fluctuations). GAT discovers deep relationships through multi-hop reasoning (such as "Enterprise A - Debt restructuring - Legal risk - Involves - A certain regulation").

[0104] Three-dimensional risk assessment model of the dynamic reasoning layer:

[0105] The three-dimensional risk assessment model calculates the legal risk index , the market volatility index , and the asset liquidity index , and fuses them with weights into the total risk level. Its core formula is as follows:

[0106] 1. Legal risk index:

[0107]

[0108] Where: is the weight of the th type of legal violation; is the number of occurrences of this violation; is the amount involved.

[0109] 2. Market volatility index (predicted based on the ARIMA model):

[0110]

[0111] Where: is the actual transaction price; is the transaction price predicted by the ARIMA model; is the time window length; is the variance of the prediction error.

[0112] 3. Asset liquidity index:

[0113]

[0114] Where: TR is the asset turnover rate; DT is the days to liquidation; is the weight.

[0115] 4. Total risk level:

[0116]

[0117] Where: is the weighted coefficient, satisfying .

[0118] Explanation and problem solving:

[0119] Function of the formula: Legal risk index Comprehensively consider the frequency and amount of violations, and quantify legal risks.

[0120] Market volatility index Use the ARIMA model to predict asset price fluctuations and quantify market risks.

[0121] Asset liquidity index Comprehensively consider the asset turnover rate and days to liquidation, and evaluate the asset liquidation ability.

[0122] Total risk level Weightedly fuse the three-dimensional index to divide the risk level.

[0123] Problem solved: The disposal of non-performing assets requires a comprehensive assessment of risks, involving multi-dimensional factors such as law, market, and liquidity. The formula provides a quantitative risk assessment method to support scientific decision-making.

[0124] Multi-modal fusion model of the scenario adaptation layer:

[0125] The scenario recognition module adopts a multi-modal fusion model, which fuses text, historical records, time, and location information to calculate the probability of scenario template matching. Its core formula is as follows:

[0126]

[0127]

[0128] Where: , , , are the feature vectors of text, historical records, time, and location; is the vector concatenation operation; , , , are the learnable weights and biases; is the The matching probability of a scenario template.

[0129] Explanation and problem-solving:

[0130] Formula function: After concatenating the features of different modalities (text, historical records, etc.) in the multi-modal fusion model, the matching probability of each scenario template is calculated through a fully connected neural network and a function.

[0131] Problem solved: There are various scenarios for non-performing asset disposal (judicial auctions, debt restructurings, etc.), and it is necessary to automatically match scenario templates according to the user's context. The formula realizes the fusion of multi-modal information and scenario matching.

[0132] Specific applications:

[0133] Extract the text features of the user input and the feature of the historical query record and the time feature and the location feature .

[0134] Concatenate the feature vectors , and calculate the matching probability through a neural network .

[0135] Select the scenario template with the highest probability (such as the judicial auction template) for subsequent answer reconstruction.

[0136] Meta-learning (cross-scenario knowledge transfer) of the scenario adaptation layer:

[0137] The cross-scenario knowledge transfer module is based on the meta-learning algorithm (MAML), and its core formula is as follows:

[0138]

[0139]

[0140] Where: is the initial parameter of the model; is the task-specific parameter; is the loss function of the task; is the learning rate; is the task set.

[0141] Explanation and problem-solving:

[0142] Formula function: MAML learns the general knowledge representation across scenarios through two-stage optimization (inner-layer optimization and outer-layer optimization). The inner-layer optimization fine-tunes the parameters for each scenario task, and the outer-layer optimization updates the global parameters to enable the model to quickly adapt to new scenarios.

[0143] Problems to be solved: It is difficult to reuse knowledge between different scenarios (such as judicial auctions and debt restructurings), and the framework needs to quickly adapt to new scenarios. Meta-learning extracts common features of scenarios to achieve knowledge transfer. Specific Embodiment 3:

[0145] Such as Figures 1-3 As shown below, the following are the specific application logic steps of each module and algorithm in the cross-scenario Q&A framework for non-performing assets based on the knowledge graph:

[0146] Knowledge Fusion Layer:

[0147] The multi-source data access module regularly or in real time obtains heterogeneous data, including legal provisions, judicial judgment documents, financial statements, asset transaction records, etc., from the legal regulations database, judicial case database, enterprise financial statement system, and asset trading platform through database connection protocols, API interfaces, etc., and temporarily stores the obtained data for processing. In the legal entity parsing engine, the legal concept extraction unit uses a model that combines a bidirectional encoder representation and conditional random fields to identify legal concepts in legal texts based on pre-training and fine-tuning on a large amount of legal text corpora; the relationship extraction unit uses a bidirectional long short-term memory network based on the attention mechanism combined with the legal domain ontology library to extract relationships between legal concepts; the attribute annotation unit completes the annotation of legal entity attributes through a combination of rule-based template matching and semi-supervised learning to form entity-relationship-attribute triples. The financial data standardization component converts the data of different enterprise financial statement systems into a unified semantic representation according to 128 predefined mapping rules covering financial subjects such as asset categories and liability categories, matches the source data format through regular expressions, and uses semantic mapping functions and data verification algorithms. The business process knowledge module performs process mining on historical disposal cases based on Petri net theory, extracts standardized disposal process templates such as asset inventory and value assessment, and clarifies the process steps, participating entities, execution standards, and time nodes of each template. The knowledge fusion engine uses graph embedding technology to perform semantic alignment on the legal, financial, and business triple knowledge subgraphs, eliminate semantic differences, and generate a fused knowledge graph to provide data support for subsequent reasoning.

[0148] Dynamic Reasoning Layer:

[0149] The query understanding module receives the user's natural language query, and through natural language processing technologies such as word segmentation, part-of-speech tagging, and semantic analysis, it converts it into a semantic query expression executable by the knowledge graph and transmits it to the multi-strategy inference engine. In the multi-strategy inference engine, the rule inference unit conducts reasoning based on the preset legal, financial, and business rule bases, adopting a reasoning strategy that combines forward and backward reasoning; the case inference unit uses a case retrieval algorithm based on cosine similarity to match similar cases from the historical handling case library; the graph inference unit uses a graph attention network model to model and reason about the complex relationships of the knowledge graph; the inference fusion unit integrates the results of the three inference units through a weighted voting mechanism. The three-dimensional risk assessment model calculates the legal risk index, market volatility index, and asset liquidity index based on the entity relationship network of the knowledge graph. The legal risk index is obtained by weighted summation according to the weights of different types of legal violations, the number of violations, and the amount. The market volatility index is obtained through predictive analysis of asset trading price data using the autoregressive integrated moving average model of time series analysis. The asset liquidity index is calculated by constructing an evaluation model containing indicators such as asset turnover rate and days to liquidate. Five risk levels are divided according to the weighted total score of the three indexes, and the weights are adjusted in combination with the degree of closeness and frequency of association between entities during the calculation. The answer generation module generates a preliminary answer containing at least 5 consecutive reasoning steps such as the parsing of the user's query intention and the retrieval of the knowledge graph subgraph and their knowledge sources according to the reasoning results and risk assessment information. The knowledge source points to the specific data source, and at the same time, the credibility of each reasoning step is evaluated and incorporated into the answer, and then the preliminary answer is transmitted to the scenario adaptation layer.

[0150] Scenario adaptation layer:

[0151] The scenario recognition module adopts a model based on multi-modal fusion, fuses context information such as the user's input text information, historical query records, current time, and geographical location, calculates the matching probabilities of scenario templates such as judicial auction, debt restructuring, and bankruptcy liquidation, and selects the template with the highest matching probability. The answer reconstruction engine conducts screening, classification, and formatting processing on the preliminary answer according to the formatting requirements of the selected scenario template, and converts it into a structured output. The multi-modal display module presents the structured output in three forms: text, chart, and knowledge graph visualization. The text form directly displays the content, the chart form is used to display numerical data, and the knowledge graph visualization uses different graphic layouts and color coding to distinguish entities and relationships. The cross-scenario knowledge transfer module constructs a meta-learning model based on the meta-learning algorithm, trains it on multi-scenario historical data, extracts common features and knowledge patterns, identifies reusable knowledge components, and invokes and applies them in new scenarios. The processed answer is transmitted to the decision verification layer as needed.

[0152] Decision verification layer:

[0153] In the multi-dimensional evaluation module, the legal compliance evaluation unit uses a combination of a rule engine and a legal knowledge graph to match legal clauses with legal elements of the disposal plan to judge compliance; the financial feasibility evaluation unit conducts risk assessment on the financial indicators of the disposal plan based on the Monte Carlo simulation method; the process integrity evaluation unit models and analyzes the disposal process through a Petri net model to verify integrity. The executability verification module of the disposal plan constructs a simulation model containing elements such as resources, activities, and events based on discrete event simulation technology, simulates the execution process of the disposal process, and predicts resource consumption and time costs. The evidence chain tracing engine records the complete reasoning path from query to answer, including the data processing and reasoning processes at each layer. The feedback learning module collects user feedback data, updates the relationship weights and reasoning rules in the knowledge graph according to the feedback, optimizes the framework performance, and affects the subsequent knowledge fusion and reasoning processes.

[0154] Other modules:

[0155] In the knowledge evolution module, the knowledge update detection unit uses a timestamp-based incremental detection algorithm to monitor data source updates in real time; the knowledge incremental update engine uses a graph embedding-based knowledge fusion algorithm to integrate new knowledge into the existing knowledge graph; the knowledge conflict resolution module is based on a conflict resolution strategy library and uses priority sorting and negotiation strategies to resolve knowledge conflicts; the dynamic evolution mechanism of the knowledge graph adjusts the graph structure and parameters according to the disposal effect based on a reinforcement learning algorithm. In the cross-scenario migration module, the scenario feature extraction unit uses a convolutional neural network and a graph convolutional network to extract scenario text and knowledge graph structure features respectively; the meta-learning engine learns general knowledge representations based on model-agnostic meta-learning algorithms; the solution verification unit verifies the effectiveness of the cross-scenario migration solution through comparison between simulation and actual cases; the cross-modal knowledge representation module converts unstructured judicial documents and evaluation reports into semantic vectors that can be represented by the knowledge graph. The framework adopts a federated knowledge learning architecture. Local knowledge processing nodes preprocess local data and extract features. The secure aggregation center aggregates knowledge through a secure multi-party computing protocol. The privacy protection mechanism uses homomorphic encryption technology to ensure data security. The institutional credit evaluation module records the institutional knowledge contribution and data quality based on blockchain technology. The framework supports cross-jurisdiction reasoning. The legal conflict identification engine compares legal clauses and rules in the knowledge subgraphs of multiple jurisdictions to identify conflicts. The conflict resolution decision-making module migrates disposal experiences based on legal analogical reasoning technology to resolve conflicts. The framework function modules are implemented through standardized microservice interfaces. Based on RESTful API interfaces, they are deployed using Docker containers and horizontally scaled using Kubernetes. The Kafka message queue mechanism is used between layers, and different data topics are set to achieve data flow. The system supports 1000 concurrent user queries. The intelligent prompt module extracts and actively recommends legal clauses and disposal strategies from the knowledge system in the knowledge fusion layer based on the user's historical query behavior and the current scenario, using collaborative filtering algorithms and association rule mining techniques.

[0156] The system support module of the framework ensures the stable operation of the intelligent prompt module by providing an infrastructure with high availability and high scalability. The system adopts a standardized microservices architecture, realizes the interconnection of each functional module based on the RESTful API interface, deploys each microservice independently through Docker containers, and uses Kubernetes to achieve automated horizontal scaling and load balancing. The data flow relies on the Kafka message queue mechanism, sets up knowledge fusion data topics, inference request data topics, answer output data topics, and intelligent recommendation data topics to ensure real-time data processing and timely transmission of recommendation results. In addition, a distributed cache is equipped to accelerate the access to users' historical query behaviors, and Elasticsearch is integrated to support efficient association rule mining and recommendation algorithm calculations, jointly supporting the efficient query processing requirements of 1000 concurrent users, thus providing reliable computing and storage support for the intelligent prompt module. Specific Embodiment Four:

[0158] As Figures 1-3 shown, the following is the detailed hardware composition and hardware description of each module in Embodiment One:

[0159] The bad asset cross-scenario question answering framework based on the knowledge graph adopts a distributed cluster architecture, which consists of a server cluster, a storage device cluster, and network devices to form the overall hardware environment, and each module plays a different role in it. The multi-source data access module in the knowledge fusion layer realizes the efficient collection and temporary storage of multi-source heterogeneous data through a data acquisition server equipped with a high-performance multi-core CPU (such as the Intel Xeon Gold series), more than 128GB of memory, and multiple network interfaces above 10Gbps, and is paired with a solid-state storage array (SSD array) as a data cache device; modules such as the legal entity parsing engine rely on a cluster composed of multiple computing servers equipped with NVIDIA A100 GPUs, more than 256GB of memory, and NVMe SSDs, combined with the Ceph distributed storage system, to complete complex data processing and knowledge graph construction tasks.

[0160] The query understanding module in the dynamic inference layer realizes the rapid parsing of user queries through a natural language processing server equipped with an AMD EPYC series multi-core CPU and more than 64GB of memory, and is paired with Intel Optane DC Persistent Memory as a local cache; modules such as the multi-strategy inference engine rely on an inference computing cluster composed of servers equipped with NVIDIA H100 GPUs, are interconnected through an InfiniBand high-speed network, and are paired with an all-flash array to store data, ensuring the efficient operation of inference calculations.

[0161] Modules such as the scene recognition in the scene adaptation layer use an application server configured with an Intel Core i9 series multi-core CPU, more than 32GB of memory, and SSD storage, and are paired with an NVIDIA RTX 30 series GPU to complete the multi-modal display task; the cross-scene knowledge migration module uses a training server equipped with an NVIDIA A40 GPU and more than 128GB of memory, combined with the GlusterFS distributed file system, to achieve model training and data storage.

[0162] Modules such as the multi-dimensional evaluation in the decision verification layer rely on an evaluation computing server configured with an Intel Xeon Platinum series multi-core CPU and more than 256GB of memory, and are paired with an enterprise-level hard disk array (HDD array) to store data; modules such as the evidence chain traceability engine run a database management system through a data management server configured with a high-performance CPU and more than 64GB of memory, and are combined with a tape library or cloud storage for backup to ensure data management and storage security.

[0163] In other modules, the monitoring and update server of the knowledge evolution module is equipped with a multi-core CPU and more than 128GB of memory, and is combined with a distributed version control system (such as Git distributed storage) to achieve the update management of the knowledge graph; the cross-scene migration module shares hardware resources with related modules, and builds an independent test server cluster for solution verification; the local knowledge processing nodes of the federated knowledge learning architecture are composed of edge servers or local workstations, and the secure aggregation center uses a high-performance server cluster, and at the same time deploys hardware encryption devices (such as encryption network cards, HSM) to ensure data security; the framework deployment and performance-related modules use an x86 server cluster combined with virtualization technology, use high-performance switches, routers, and load balancing devices to build a network environment, adopt a hybrid storage architecture (combination of SSD and HDD), and use storage management software to achieve data storage and management, jointly supporting the efficient and stable operation of the framework. Specific Example Five:

[0165] Such as Figures 1-3 shown, the following provides a specific use case of the solution:

[0166] Case One: Comprehensive Evaluation of Judicial Auctions of Enterprise Non-Performing Assets

[0167] A certain bank holds the non-performing assets of an enterprise and plans to dispose of them through judicial auctions. Bank staff use this framework for relevant analysis. At the knowledge fusion layer, the multi-source data access module obtains the financial data such as the assets and liabilities and cash flow of the enterprise from the enterprise financial statement system, obtains cases of judicial auctions of non-performing assets of similar enterprises from the judicial case database, obtains transaction price data of similar assets in the current market from the asset trading platform, and obtains legal provisions related to judicial auctions from the laws and regulations database. After processing, a fused knowledge graph is generated. At the dynamic reasoning layer, the staff inputs "What are the risks and expected returns of the judicial auction of the non-performing assets of this enterprise?" The query understanding module converts it into a semantic query expression. The multi-strategy reasoning engine combines rule, case, and graph reasoning. The three-dimensional risk assessment model calculates the legal risk index (relatively high due to the existence of outstanding litigation cases of the enterprise), the market volatility index (medium due to the impact of the current economic situation), and the asset liquidity index (relatively high as the asset type is relatively easy to auction). The answer generation module gives a preliminary answer including the reasoning steps and knowledge sources, indicating that there is a risk of legal disputes in the auction and the expected return is uncertain due to market fluctuations. The scenario adaptation layer automatically matches the judicial auction scenario template. The answer reconstruction engine converts the preliminary answer into a structured output. The multi-modal display module presents the risk index in the form of a chart and explains the expected return situation in text. At the decision verification layer, the answer is evaluated. The legal compliance assessment unit checks whether the auction process complies with legal regulations, the financial feasibility assessment unit analyzes the reasonableness of the expected return, and the process integrity assessment unit verifies whether the steps of the auction process are complete. Finally, it is determined that the answer is reliable, providing a basis for the bank to formulate an auction strategy.

[0168] Case 2: Formulation of a non-performing asset debt restructuring plan

[0169] An asset management company takes over a batch of non-performing assets and considers debt restructuring. In the knowledge fusion stage, the system integrates the financial distress data of the enterprises involved in the non-performing assets, the industry development trend, the successful and failed cases of previous debt restructurings, and the relevant laws and policies. When the asset manager asks "How to design a feasible debt restructuring plan for the non-performing assets of a certain enterprise", the dynamic reasoning layer responds quickly. The multi-strategy reasoning engine synthesizes knowledge from all parties, refers to the debt restructuring cases of similar enterprises, combines the current economic situation and legal regulations, and reasons out various possible restructuring methods; the three-dimensional risk assessment model assesses the risks of each method and weighs the pros and cons. The scenario adaptation layer matches the debt restructuring scenario template, structurally arranges the reasoning results, and presents the key elements, expected effects, and potential risks of different restructuring plans in clear texts and charts. The decision verification layer strictly verifies the plan from aspects such as legal compliance, financial feasibility, and process integrity, simulates the problems that may occur during the restructuring process, and ensures that the plan is practical and feasible. Finally, based on the plan provided by the system and in combination with the actual situation, the asset manager negotiates with the debtor enterprise to determine and implement the debt restructuring plan, successfully revitalizing the non-performing assets.

[0170] Case 3: Consultation on Cross-regional Non-performing Asset Disposal Strategies

[0171] A large financial group has non-performing assets in multiple regions. Due to differences in local laws, policies, and market environments, the disposal is difficult. The group's employees use this framework and input "What issues should be noted and what strategies should be adopted for disposing of similar non-performing assets in different regions". The knowledge fusion layer converges relevant data from each region and constructs a multi-jurisdiction knowledge subgraph. The legal conflict identification engine in the dynamic reasoning layer compares the multi-jurisdiction knowledge subgraphs, finds out the differences and potential conflicts in legal clauses. The conflict resolution decision-making module, based on legal analogical reasoning technology, migrates the successful disposal experiences of other regions. The multi-strategy reasoning engine combines the actual situations of each region and gives disposal strategy suggestions for different regions. For example, in Region A, the law provides strong protection for creditors, and it is recommended to adopt judicial litigation; in Region B, the market is active, and it is recommended to give priority to asset transfer. The scenario adaptation layer presents the strategy suggestions in a personalized manner according to the characteristics of different regions. The decision verification layer verifies the feasibility and compliance of the strategies. Based on these suggestions, the group formulates targeted cross-regional non-performing asset disposal strategies, improving the disposal efficiency and success rate. Specific Embodiment 6:

[0173] As Figures 1-3 shown, the following provides complete experimental data:

[0174] Experimental Objectives: Verify the performance of the framework in cross-scenario Q&A for non-performing assets, including knowledge graph construction efficiency, reasoning accuracy, risk assessment effect, scenario matching accuracy, decision verification results, etc.

[0175] Experimental scenarios: Three typical non-performing asset disposal scenarios are selected: judicial auction, debt restructuring, and bankruptcy liquidation.

[0176] Datasets: Legal regulations database: containing 100,000 legal provisions. Judicial case database: containing 50,000 historical disposal cases. Enterprise financial statement system: containing 10,000 financial statements. Asset trading platform: containing 20,000 transaction records.

[0177] Evaluation metrics: Knowledge graph construction time (seconds), entity-relationship extraction accuracy rate (%), reasoning accuracy rate (%), reasoning time (seconds), risk assessment error (%), scenario matching accuracy rate (%),

[0178] Decision verification passing rate (%), feedback learning optimization effect (improvement percentage of reasoning accuracy rate before and after weight update).

[0179] Experimental conditions:

[0180] Hardware: 16-core CPU, 64GB memory, GPU support.

[0181] Number of concurrent users: 1000 (meeting the requirement of supporting 1000 concurrent users mentioned in the technical solution).

[0182] The table is as follows:

[0183] Table 1: Performance of the knowledge fusion layer and the dynamic reasoning layer:

[0184] Scenario Knowledge graph construction time (seconds) Entity-relationship extraction accuracy rate (%) Inference accuracy rate (%) Risk assessment error (%) Inference time (seconds) Judicial auction 120 92.5 88.7 5.4 2.5 Debt restructuring 105 90.8 87.3 5.5 2.8 Bankruptcy liquidation 135 93.2 89.1 5.2 2.3

[0185] Table 2: Performance of the scenario adaptation layer and the decision verification layer:

[0186] Scenario Scenario matching accuracy rate (%) Cross-scenario knowledge transfer success rate (%) Verification pass rate (%) Feedback learning optimization effect (improvement in inference accuracy rate) Executability verification resource error (%) Judicial auction 94.5 85.2 90.6 3.5 6.2 Debt restructuring 93.8 83.9 89.9 3.2 6.5 Bankruptcy liquidation 95.1 86.7 91.3 3.8 5.9

[0187] In the above tables:

[0188] The knowledge graph construction time is between 105 - 135 seconds, and the entity-relationship extraction accuracy rate is stable above 90%, indicating that the framework can efficiently process heterogeneous data.

[0189] The reasoning accuracy rate is between 87% - 89%, the risk assessment error is controlled within 5.2% - 5.5%, and the reasoning time is between 2.3 - 2.8 seconds, meeting the real-time requirements.

[0190] The scenario matching accuracy rate exceeds 93%, and the cross-scenario knowledge transfer success rate is between 83% - 86%, indicating that the framework has strong adaptability to multiple scenarios.

[0191] The verification passing rate is between 89% - 91%, the feedback learning optimization effect is about 3%, and the resource prediction error is between 5.9% - 6.5%, indicating that the decision support is reliable and can be optimized.

[0192] Figure 3 This is the evaluation simulation diagram of the method. Horizontal axis: represents three scenarios, namely judicial auction, debt restructuring, and bankruptcy liquidation. Vertical axis: represents the risk value (unit: risk index, range 0 - 180). Legend: includes legal risk index (blue), market volatility index (orange), asset liquidity index (yellow), and total risk (purple).

[0193] It can be concluded from the figure that:

[0194] The legal risk indices of debt restructuring and bankruptcy liquidation are relatively high (about 160 - 170), significantly higher than that of judicial auction (about 40), indicating that legal risk is the main influencing factor.

[0195] The market volatility and asset liquidity indices are relatively low (about 5 - 15), contributing less to the total risk.

[0196] The total risk (purple) in judicial auction (about 70) is lower than that in debt restructuring and bankruptcy liquidation (about 80), reflecting that the former has lower risk.

[0197] It should be noted that in this article, relational terms such as first and second are only used 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 term "comprising", "including" or any other variant thereof is intended to cover non - exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "comprising a reference structure" does not exclude the existence of additional identical elements in the process, method, article or device comprising the element.

[0198] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An impaired asset cross-scenario question-answering framework based on a knowledge graph, including the framework, characterized in that: The framework includes a knowledge fusion layer, a dynamic reasoning layer, a scenario adaptation layer, a decision verification layer, a knowledge evolution module, a cross-scenario migration module, and a federated knowledge learning architecture. Each layer works together through standardized service interfaces. The specific composition and data flow are as follows: The knowledge fusion layer is used to integrate heterogeneous data and construct a unified knowledge graph, including a multi-source data access module, a legal entity parsing engine, a financial data standardization component, a business process knowledge module, and a knowledge fusion engine. Among them: The multi-source data access module is configured to obtain heterogeneous data from a legal regulations database, a judicial case database, an enterprise financial statement system, and an asset trading platform; The legal entity parsing engine, based on ontology mapping technology, converts unstructured legal texts into entity-relationship-attribute triples that can be represented by the knowledge graph; The financial data standardization component converts financial data in different formats into a unified semantic representation through predefined financial indicator mapping rules; The business process knowledge module analyzes historical disposal cases based on process mining technology and extracts a standardized non-performing asset disposal process template; The knowledge fusion engine uses graph embedding technology to achieve semantic alignment of legal, financial, and business knowledge subgraphs and generates a fused knowledge graph; The dynamic reasoning layer is used to process user queries and generate preliminary answers, including a query understanding module, a multi-strategy reasoning engine, a three-dimensional risk assessment model, and an answer generation module. Among them: The query understanding module converts the user's natural language query into a semantic query expression executable by the knowledge graph; The multi-strategy reasoning engine performs cross-entity multi-hop reasoning, supporting rule reasoning, case reasoning, and graph reasoning; The three-dimensional risk assessment model calculates a legal risk index, a market volatility index, and an asset liquidity index based on the entity relationship network in the knowledge graph; The answer generation module generates a preliminary answer containing an evidence chain.

2. The cross-scenario question answering framework for non-performing assets based on the knowledge graph according to claim 1, wherein: The scenario adaptation layer is used to adjust the answer output according to the user context, including a scenario recognition module, a cross-scenario knowledge migration module, an answer reconstruction engine, and a multi-modal display module. Among them: The scenario recognition module automatically matches the corresponding disposal scenario template based on the user context information; The cross-scenario knowledge migration module identifies reusable knowledge components between different scenarios; The answer reconstruction engine converts the preliminary answer into a structured output according to the requirements of the scenario template; The multi-modal display module supports three output forms: text, charts, and knowledge graph visualization; The decision verification layer is used to verify the accuracy and feasibility of the answer, including a multi-dimensional evaluation module, a disposal plan executability verification module, an evidence chain traceability engine, and a feedback learning module. Among them: The multi-dimensional evaluation module verifies the legal compliance, financial feasibility, and process integrity of the answer based on predefined evaluation indicators; The disposal plan executability verification module predicts the resource consumption and time cost of the disposal process through simulation technology; The evidence chain traceability engine records the complete reasoning path from the query to the answer; The feedback learning module updates the relationship weights and reasoning rules of the knowledge graph according to the user feedback; The knowledge evolution module is used to dynamically update and optimize the knowledge graph, including a knowledge update detection unit, a knowledge incremental update engine, a knowledge conflict resolution module, and a knowledge graph dynamic evolution mechanism, where: the knowledge update detection unit monitors the update of the data source in real time; the knowledge incremental update engine integrates the newly added knowledge into the existing knowledge graph; the knowledge conflict resolution module resolves the conflicts in the knowledge graph; The cross-scenario migration module is used to support the migration of knowledge between different scenarios, including a scenario feature extraction unit, a meta-learning engine, a solution verification unit, and a cross-modal knowledge representation module, where: the scenario feature extraction unit extracts features from scenario-related information; the meta-learning engine learns the general knowledge representation between different scenarios; the solution verification unit verifies the effectiveness of the cross-scenario migration solution; The federated knowledge learning architecture is used for distributed knowledge processing, including local knowledge processing nodes, a secure aggregation center, a privacy protection mechanism, and an institutional credit assessment module, where: the local knowledge processing nodes perform preprocessing and feature extraction on local data; the secure aggregation center aggregates the knowledge uploaded by each node; the privacy protection mechanism ensures privacy security during data transmission and aggregation; Each functional module of the framework is implemented through a standardized microservice interface, supporting independent deployment and horizontal expansion; the data flow between layers adopts a message queue mechanism, and different topics are set for the message queue to distinguish different types of data, including knowledge fusion data topics, inference request data topics, and answer output data topics; the system as a whole supports high-concurrency query processing; The legal entity parsing engine in the knowledge fusion layer includes a legal concept extraction unit, a relationship extraction unit, and an attribute annotation unit, where: the legal concept extraction unit uses a pre-trained deep learning model, combined with fine-tuning of legal domain corpora, to identify legal concepts in legal texts; the relationship extraction unit extracts the relationships between legal concepts based on a neural network combined with a legal ontology library; the attribute annotation unit completes the attribute annotation of legal entities through a combination of rule templates and semi-supervised learning; The financial data standardization component contains a variety of financial indicator mapping rules, covering asset, liability, equity, revenue, and expense accounts, and supports data format parsing, conversion, and verification; The disposal process templates extracted by the business process knowledgeization module include asset inventory, value assessment, judicial auction, debt restructuring, and bankruptcy liquidation, and each template contains process steps, participating parties, execution standards, and time nodes.

3. The cross-scenario Q&A framework for non-performing assets based on the knowledge graph according to claim 1, wherein: The multi-strategy inference engine in the dynamic inference layer includes a rule inference unit, a case inference unit, a graph inference unit, and an inference fusion unit, where: the rule inference unit is based on legal, financial, and business rule libraries and uses forward and backward reasoning; the case inference unit matches historical disposal cases through a similarity algorithm; the graph inference unit uses a graph neural network to model the relationships in the knowledge graph; the inference fusion unit integrates the inference results through a weighting mechanism; The three-dimensional risk assessment model calculates the legal risk index, market volatility index, and asset liquidity index through weighted calculation, and adjusts the weights by combining the entity relationship tightness and frequency; the evidence chain includes at least five reasoning steps, including query parsing, sub-graph retrieval, rule matching, case comparison, and result integration.

4. The cross-scenario Q&A framework for non-performing assets based on the knowledge graph according to claim 2, characterized in that: The scenario recognition module of the scenario adaptation layer adopts a multi-modal fusion model to calculate the matching probability of each scenario template based on the text information, historical query records, current time, and geographical location input by the user. The scenario templates of the scenario adaptation layer include a judicial auction scenario template, a debt restructuring scenario template, and a bankruptcy liquidation scenario template. Each template contains multiple structured sub-templates, and each sub-template defines at least 10 standardized fields. Among them, the sub-templates of the judicial auction scenario template include an auction asset information sub-template, an auction process stage sub-template, a participating entity information sub-template, an auction result record sub-template, and a risk reminder sub-template. The auction asset information sub-template contains fields such as asset name, asset category, appraised value, starting price, and ownership certificate number. When visualizing the knowledge graph, the multi-modal display module uses different graph layouts and color coding to distinguish entity and relationship types.

5. The cross-scenario Q&A framework for non-performing assets based on the knowledge graph according to claim 2, wherein: The multi-dimensional evaluation module of the decision verification layer includes a legal compliance evaluation unit, a financial feasibility evaluation unit, and a process integrity evaluation unit. Among them: the legal compliance evaluation unit matches legal clauses with the disposal plan through a rule engine and a legal knowledge graph to judge compliance; the financial feasibility evaluation unit evaluates the financial risks of the disposal plan based on a simulation method; the process integrity evaluation unit verifies the integrity of the disposal process through process modeling technology. The disposal plan executability verification module constructs a simulation model containing resource, activity, and event elements based on discrete event simulation technology to predict resource consumption and time costs.

6. The cross-scenario question answering framework for non-performing assets based on the knowledge graph according to claim 2, wherein: The knowledge update detection unit of the knowledge evolution module adopts an incremental detection algorithm to monitor the updates of data sources such as the laws and regulations database and the enterprise financial statement system in real time. The knowledge incremental update engine uses graph embedding technology to integrate new knowledge into the existing knowledge graph. The knowledge conflict resolution module resolves conflicts based on a conflict resolution strategy library through priority sorting and negotiation strategies. The knowledge graph dynamic evolution mechanism is based on a reinforcement learning algorithm to adjust the knowledge graph structure and parameters according to the disposal effect.

7. The cross-scenario question-answering framework for non-performing assets based on the knowledge graph according to claim 2, wherein: The scenario feature extraction unit of the cross-scenario migration module uses a neural network to extract features from scenario text information and knowledge graph structure. The meta-learning engine learns the general knowledge representation between different scenarios based on the model-agnostic meta-learning algorithm. The solution verification unit verifies the effectiveness of the cross-scenario migration solution by comparing simulation with actual cases. The cross-modal knowledge representation module converts unstructured judicial documents and evaluation reports into semantic vectors that can be represented by the knowledge graph.

8. The cross-scenario Q&A framework for non-performing assets based on the knowledge graph according to claim 2, characterized in that: The local knowledge processing node of the federated knowledge learning architecture preprocesses and extracts features from local non-performing asset data. The secure aggregation center aggregates the knowledge uploaded by each node through a secure multi-party computing protocol. The privacy protection mechanism adopts encryption technology to ensure privacy security during data transmission and aggregation; The institutional credit assessment module is based on blockchain technology and records the knowledge contribution and data quality of each participating institution.

9. The cross-scenario Q&A framework for non-performing assets based on the knowledge graph according to claim 1, characterized in that: The framework supports cross-jurisdictional reasoning, including a multi-jurisdiction knowledge subgraph, a legal conflict identification engine, a conflict resolution decision-making module, and a legal knowledge migration module. Among them: The legal conflict identification engine compares legal provisions in different jurisdiction knowledge subgraphs to identify potential conflicts; The conflict resolution decision-making module, based on legal analogical reasoning technology, migrates the disposal experience of one jurisdiction to another jurisdiction; The legal knowledge migration module is based on legal analogical reasoning technology and supports cross-jurisdiction knowledge migration.

10. The cross-scenario Q&A framework for non-performing assets based on the knowledge graph according to claim 1, characterized in that: The framework is configured with a system support module. When the framework uses the system support module, each functional module of the framework is interconnected through a standardized microservice interface, supporting independent deployment and horizontal expansion; The data flow between layers adopts the Kafka message queue mechanism. The message queue sets up knowledge fusion data topics, inference request data topics, and answer output data topics. The framework is equipped with an intelligent prompt module; The intelligent prompt module, based on the user's historical query behavior and the current disposal scenario, adopts collaborative filtering algorithms and association rule mining techniques to recommend legal provisions and disposal strategies.

Images