Method and device for identifying financial risk control through self-evolution of large model

Abstract

The application discloses a large model self-evolution retrieval enhanced financial risk control identification method and related device, belongs to the field of electric digital data processing, and comprises the following steps: constructing a financial risk control corpus and dividing it into evidence blocks, constructing a self-induction evidence graph based on the evidence blocks; decomposing the current request into atomic retrieval sub-targets and retrieving historical experience memories; iteratively performing experience memory retrieval, corpus retrieval, evidence reading and graph neighbor search, and accumulating the current evidence state; generating a financial risk identification result when the stop condition is met. The application uses historical experience memories to locate the retrieval direction, uses a self-induction evidence graph to track multi-hop evidence, and uses iterative retrieval to ensure the dynamic selection of evidence reading and retrieval actions, realizes the closed-loop cooperation of evidence acquisition and risk reasoning, and improves the accuracy and interpretability of the identification result.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and in particular to a financial risk control identification method and related apparatus with large-model self-evolution retrieval enhancement. Background Technology

[0002] With the development of digital finance, mobile payment, online lending, and open banking, financial risk events are characterized by high frequency, cross-channel nature, long transmission chains, and complex sources of evidence. When identifying fraudulent transactions, account theft, false identities, cash-out schemes, money laundering, abnormal merchants, and credit default risks, financial institutions typically need to simultaneously refer to multi-source data, including transaction records, customer profiles, device fingerprints, network addresses, merchant information, blacklists, historical case studies, regulatory documents, business rules, and manual review records.

[0003] Existing financial risk control systems rely on fixed rules, supervisory models, or graph models for risk identification, while gradually introducing retrieval-enhanced generation techniques to generate interpretable risk control analysis conclusions using external knowledge bases. However, traditional retrieval-enhanced generation techniques typically assume that each request is independent: each request re-retrieves evidence from the corpus, ignoring key entities, bridging entities, evidence paths, search terms, and handling tags already exposed during previous risk control processes. In real-world risk control processes, the same customer, account, device, merchant, IP address, payee, or transaction pattern often appears repeatedly in multiple consecutive requests. Subsequent requests may reuse bridging clues from previous processes, but still require obtaining new transaction and rule evidence from the current corpus. Simply caching historical answers can easily lead to misapplying old conclusions to new events; completely ignoring historical experience can result in duplicate searches, low efficiency in evidence location, and breaks in the continuous risk chain.

[0004] Graph retrieval enhancement methods can improve corpus organization, and intelligent entity-based retrieval enhancement methods can perform multi-round tool calls. However, existing solutions typically lack an updatable, quality-controlled, and structured memory state that can retain retrieval experience across consecutive requests. Therefore, there is an urgent need for a financial risk control identification technology solution that can start from zero experience, evolve automatically with the risk control process, and forcibly acquire current evidence while reusing historical experience.

[0005] Furthermore, conventional retrieval enhancement technologies typically establish only a one-time retrieval link between the current request and the corpus, failing to identify whether the current request has bridging entities, shared evidence paths, or common risk rules with previous requests. While graph retrieval enhancement can extend evidence on static corpus graphs, its graph structure mainly describes the relationships between documents and does not record reusable retrieval experience formed during historical processing. Although agent-based retrieval enhancement can invoke multiple retrieval tools, without persistent and quality-controllable experience memory, it will still repeatedly attempt and fail in continuous risk control requests. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of the prior art and provide a financial risk control identification method and related device with large model self-evolution retrieval enhancement.

[0007] The objective of this invention is achieved through the following technical solution: a financial risk control identification method enhanced by large-model self-evolutionary retrieval, the method comprising the following steps: Construct a financial risk control corpus, including one or more of the following: transaction records, account information, device or network identifiers, risk rules, blacklists, historical handling cases, regulatory documents, and complaint or chargeback records; The financial risk control corpus is textualized and anonymized. The anonymized data is segmented according to a preset granularity, and an index is created for each segmentation unit to form multiple evidence blocks with unique identifiers. A self-induced evidence graph is constructed based on the multiple evidence blocks. The nodes of the self-induced evidence graph correspond to the evidence blocks, and the edges correspond to the semantic proximity relationship and / or lexical reference relationship between the evidence blocks. The current financial risk control analysis request is decomposed to obtain at least one atomic retrieval sub-target, and relevant historical risk control handling experience is retrieved from historical experience memory based on the current financial risk control analysis request; the atomic retrieval sub-target includes at least one of subject identifier query, relationship retrieval, subject attribute value acquisition, and risk rule-related evidence matching; Iteratively execute at least one of the following actions: perform an experience memory retrieval action on relevant historical risk control handling experience, perform keyword retrieval, semantic retrieval, or fuzzy retrieval actions on the financial risk control corpus, and perform evidence block reading and graph neighbor search actions on the self-induced evidence graph; in each iteration, select and execute one of the actions based on the current financial risk control analysis request, atomic retrieval sub-target, current evidence status, historical experience memory, and self-induced evidence graph, and accumulate and aggregate the observation results returned by each executed action to form a dynamic information set including the obtained evidence information, the source of the obtained evidence information, and the corresponding retrieval action record, as the current evidence status; When the iteration stopping condition is met, a financial risk control identification result is generated based on the current evidence status; wherein, the iteration stopping condition includes: at least one complete evidence fragment has been read and the generated result is not an unknown result, or the preset number of reading retries, the number of unknown result retries, or the maximum number of iteration steps has been reached; the financial risk control identification result includes one or more of the following: risk type, risk level, supporting evidence chain, confidence level, and disposal suggestion.

[0008] In one example, the construction of a self-induced evidence graph based on multiple evidence blocks includes: Treat each piece of evidence as a node to obtain the node set of the self-induced evidence graph; Each evidence block is represented by a vector representing its opening text, title, subject name, account identifier, device identifier, merchant name, transaction summary, or rule name. Retrieve the nearest neighbor evidence blocks of each evidence block in the vector space, and establish a semantic edge when the similarity between the current evidence block and the nearest neighbor evidence block is greater than a preset semantic threshold; the nearest neighbor evidence block is a preset number of other evidence blocks that are closest to the current evidence block in vector distance or have the highest vector similarity. Extract title-style query terms or financial entity terms from each evidence block, retrieve relevant evidence blocks based on sparse matching, and establish lexical reference edges when the matching score between the current evidence block and relevant evidence blocks is greater than a preset lexical threshold; the relevant evidence blocks are a preset number of other evidence blocks with the highest sparse matching score with the current evidence block. When two evidence nodes meet the semantic nearest neighbor condition and their semantic similarity exceeds a preset semantic threshold, a semantic edge is generated between the two evidence nodes, and a preset semantic edge weight is assigned to the semantic edge; when two evidence nodes meet the lexical reference condition and their lexical matching score exceeds a preset lexical threshold, a lexical reference edge is generated between the two evidence nodes, and a preset lexical reference edge weight is assigned to the lexical reference edge; the semantic edge weight and the lexical reference edge weight of the same node pair are summed to obtain the final edge weight of the node pair; A self-induced evidence graph is constructed based on the node set, semantic edges, lexical reference edges, and edge weights.

[0009] In one example, the historical experience memory is a graph-structured experience memory, where each experience memory node includes a historical request, a historical risk outcome, a set of evidence anchors, a set of evidence block identifiers, and a timestamp; the set of evidence anchors includes one or more of the following: customer identifier, account identifier, device identifier, IP address, merchant identifier, counterparty, transaction channel, risk rule name, fraud type, or handling label; When two experience memory nodes share at least one evidence anchor point, an experience edge is established between the two experience memory nodes.

[0010] In one example, retrieving relevant historical risk control handling experience from historical experience memory based on the current financial risk control analysis request includes: Dense retrieval of current and historical financial risk control analysis requests yields dense candidate experiences. Sparse search is performed on current financial risk control analysis requests and historical requests, evidence anchors or risk results to obtain sparse candidate experiences; The dense and sparse candidate experiences are sorted and fused based on the inverse sorting fusion method, and the direct experience node with the highest score is selected. When a direct experience node has an experience edge neighbor, select at least one neighbor experience node with the most shared evidence anchor points or the highest experience edge weight. Direct experience nodes and neighboring experience nodes are linearized into experience hints to guide the current evidence retrieval, but are not directly used as the current financial risk control identification result.

[0011] In one example, before generating the financial risk control identification result based on the current evidence status, it is determined whether the current evidence status includes at least one complete evidence block that has been read; if not and the number of read retries has not been reached, a correction instruction is generated to perform the evidence block reading action. If the candidate identification result generated based on the current evidence state is unknown or the confidence level is lower than the preset threshold, and the unknown retry count has not been reached, then a correction instruction is generated to perform precise retrieval, fuzzy retrieval, or evidence graph neighbor search.

[0012] In one example, after generating the financial risk control identification result based on the current evidence state, the process further includes: The financial risk control identification results are subjected to quality gating judgment. When the writing conditions are met, the current financial risk control analysis request, financial risk control identification results, evidence anchors, evidence block identifiers and timestamps are written into the historical experience memory to obtain the updated historical experience memory.

[0013] In one example, the quality gating decision includes: Determine whether the confidence level of the financial risk control identification result belongs to the preset high confidence interval, whether the identification result is unknown, whether it includes unresolved references or uncertainty markers, and whether the supporting evidence chain includes at least one evidence block in the current corpus.

[0014] In one example, the iteration stopping condition is that the confidence level of the candidate identification result generated based on the current evidence state reaches a preset threshold, or the maximum number of iterations has been reached; When the maximum number of iterations is reached, and the confidence level of the candidate recognition result generated based on the current evidence state has not yet reached the preset threshold, a forced candidate result is generated based on the collected current evidence state. When the forced candidate result is unknown or the confidence level is lower than the preset threshold, the current financial risk control analysis request is cleaned and fuzzy searched for remedies, and the financial risk control identification result and confidence level are regenerated based on the remedy search results.

[0015] It should be further noted that the technical features corresponding to the above examples can be combined or replaced to form new technical solutions.

[0016] The present invention also includes a computer program product comprising a computer program that, when executed by a processor, implements the steps of the large model self-evolutionary retrieval enhanced financial risk control identification method formed by any or a combination of the above examples.

[0017] The present invention also includes a terminal comprising a memory and a processor, wherein the memory stores computer instructions executable on the processor, and the processor executes the steps of the large model self-evolutionary retrieval enhanced financial risk control identification method formed by any or more of the above examples when executing the computer instructions.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1. A self-induced evidence graph is constructed based on multiple evidence blocks, encoding the potential relationships between evidence blocks into a graph structure. This enables the tracking of related transactions, related devices, or risk rules along the graph edges after discovering intermediate evidence, improving the ability to locate multi-hop risk links. This addresses the problem in existing graph retrieval enhancement methods where the graph structure only describes the relationships between documents and cannot record reusable retrieval experience formed during historical processing. By iteratively executing experience memory retrieval, corpus retrieval, evidence reading, and graph neighbor search, and accumulating the results into a dynamic evidence state, a structured risk identification result is generated when the stopping condition is met. In other words, this invention can determine the next retrieval action based on the dynamic changes of the currently acquired evidence until the evidence state meets quality requirements, avoiding transmission. This method overcomes the limitations of traditional retrieval enhancement methods, which terminate after a single retrieval and cannot adjust retrieval strategies based on intermediate findings. It also addresses the problem of repeated trial and error in consecutive requests due to the lack of persistent and quality-controlled experience memory in existing retrieval enhancement methods. This method achieves an evidence-driven adaptive reasoning process with self-evolutionary capabilities. Furthermore, by executing experience memory retrieval, it can reuse retrieval experience but not directly use it as the result of financial risk control identification, reducing the rate of repeated retrievals, improving the efficiency and accuracy of evidence location, enhancing the interpretability of results, and avoiding the problem of broken risk chains. By outputting risk type, risk level, evidence chain, and treatment recommendations, it enables financial institutions to obtain traceable, verifiable, and auditable risk control evidence while automatically identifying risks.

[0019] 2. By jointly constructing a self-induced evidence graph through semantic proximity and lexical reference relationships, the self-induced evidence graph can capture risk descriptions with similar semantics and accurately associate evidence blocks that share the same identifiers (such as account numbers and device fingerprints). This enhances the evidence graph's ability to represent multi-source heterogeneous data in financial risk control scenarios and provides more accurate association paths for subsequent graph neighbor searches.

[0020] 3. Through dense retrieval, sparse retrieval, and experience... Figure 1 The fusion approach of skipping and expanding to retrieve historical experience can capture semantically similar historical requests and accurately hit key identifiers and rule names in financial risk control; by limiting historical experience to search clues rather than answer caches, it can improve the accuracy and coverage of experience reuse.

[0021] 4. By limiting historical experience to retrieval clues rather than answer caching, and requiring the reading of complete evidence blocks from the current corpus, the timeliness and reliability of evidence in continuous risk control scenarios are improved, avoiding the direct application of old risk conclusions to new transactions, new accounts, or new entities. At the same time, correction instructions are triggered for candidate results that are unknown or have low confidence, ensuring that iteration does not stop prematurely when there is no complete current evidence or the quality of the conclusion is insufficient. This ensures that the final output is sufficient and reliable current evidence, avoiding unreliable risk identification results due to insufficient evidence or low-quality conclusions.

[0022] 5. Perform quality gating judgment on financial risk control identification results, and only write the identification results with high confidence, non-unknown, non-ambiguous and supported by the current evidence chain into the graph structure experience memory. This ensures that the memory bank accumulates high-quality and reusable retrieval experience, and avoids low-confidence, speculative or insufficient evidence results from polluting subsequent risk control decisions. At the same time, by building and continuously updating the graph structure experience memory, reusable retrieval clues can be accumulated from zero experience as business operations progress. This solves the problem that traditional retrieval enhancement technology has to search from scratch for each request and cannot utilize historical evidence paths.

[0023] 6. In abnormal situations, generate forced candidate results based on the current evidence status, and clean and fuzzy search the request to remedy the situation when the forced candidate results are still unknown or have low confidence, and regenerate the identification results to avoid being unable to output any conclusions due to insufficient evidence. This ensures that usable risk control judgment results can be output even in extreme cases and improves the confidence level. Attached Figure Description

[0024] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The accompanying drawings are provided to provide a further understanding of the present application and constitute a part of the present application. The same reference numerals are used in these drawings to denote the same or similar parts. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

[0025] Figure 1 This is a flowchart of a method provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0028] This invention discloses a risk control identification method based on large model self-evolutionary retrieval enhancement, which can be applied to financial risk identification scenarios, network security identification scenarios, medical auxiliary diagnosis scenarios, etc. The embodiments of this invention specifically illustrate the technical concept of this invention in the context of financial risk identification scenarios.

[0029] In the context of continuous requests for financial risk identification, the preceding request is called the lead request, and the subsequent request is called the follow-up request. Follow-up requests can include reviews of new transactions from the same customer, horizontal investigations of accounts associated with the same device, extended analysis of transaction patterns of the same merchant, or further explanations of the reasons for triggering the same risk rule. The method of this invention does not require prior knowledge of whether the current request relies on historical experience; instead, it automatically determines whether to retrieve and use historical experience during the closed-loop control process.

[0030] In one embodiment, such as Figure 1 As shown, a financial risk control identification method enhanced by large-scale model self-evolutionary retrieval is presented. The method includes the following steps: S1: Construct a financial risk control corpus.

[0031] In step S1, a financial risk control corpus is constructed to provide the original data foundation for subsequent evidence retrieval. The financial risk control corpus C includes structured data and unstructured text data. The structured data may include transaction records, account registration information, login records, device fingerprints, IP addresses, payment and collection relationships, merchant information, and risk labels. The unstructured text data may include risk rule descriptions, historical audit opinions, customer complaints, refusal materials, regulatory documents, judicial cases, industry blacklists, and business system texts.

[0032] S2: The financial risk control corpus is processed by textualization, desensitization, segmentation and indexing to form multiple evidence blocks with unique identifiers. A self-induced evidence graph is constructed based on the multiple evidence blocks. The nodes of the self-induced evidence graph correspond to the evidence blocks, and the edges correspond to the semantic proximity relationship and / or lexical reference relationship between the evidence blocks.

[0033] In step S2, the structured or unstructured raw data in the financial risk control corpus is converted into unified natural language text data. Data containing personal information (such as customer name, ID number, account number) or sensitive financial information is anonymized or obfuscated. The anonymized data is then segmented according to business entities, time windows, document paragraphs or rule clauses, and a unique evidence block identifier (index) is assigned to each segmented unit to form multiple evidence blocks with unique identifiers, so that subsequent identification results can be traced back to the specific evidence source.

[0034] Furthermore, dense vector indexes and sparse term indexes are established separately. Dense vector indexes are used to capture risk descriptions, rule terms, and historical cases with similar semantics; sparse term indexes are used to accurately hit key financial risk control identifiers such as customer ID, account ID, device ID, IP address, merchant ID, rule name, blacklist name, and transaction serial number.

[0035] Dense vector indexes can use text embedding models to generate evidence block vectors, used to discover risk descriptions that are semantically similar but express differently; sparse term indexes can use inverted indexes or BM25 retrieval methods to accurately hit strong identifying fields in financial risk control. Both types of indexes work together to serve the closed-loop intelligent controller, enabling it to switch between semantic recall and precise recall.

[0036] Self-induced evidence graphs require no manual annotation of relationship types and no pre-trained graph retrieval model. Semantic edges in the self-induced evidence graph represent the proximity relationships of evidence blocks in the vector space, while lexical reference edges indicate shared title-style query terms, subject names, or rule names among evidence blocks. Therefore, after reading a local evidence block, one can extend along the evidence graph to upstream transactions, downstream transactions, related subjects, or relevant rule clauses.

[0037] S3: Receive financial risk control analysis requests and decompose them to obtain at least one atomic retrieval sub-target, and retrieve similar historical risk control handling experience from historical experience memory according to the current financial risk control analysis request.

[0038] In step S3, a closed-loop intelligent controller can be introduced to receive and decompose financial risk control analysis requests. The requests can be requests to determine whether a single transaction is fraudulent, requests to investigate whether an account has been stolen, requests for due diligence to determine whether a customer has money laundering risks, requests to identify whether a merchant is abnormal, or requests for early warning of whether a credit customer has default risks.

[0039] Each atomic retrieval sub-target corresponds to a minimum unit of information requirement that can be completed in a single retrieval (such as querying the identifier of an account, obtaining the attribute value of an entity, or matching a risk rule). Atomic retrieval sub-targets can be categorized from the perspective of the action performed into types such as entity identifier query, relationship retrieval, entity attribute value acquisition (such as querying account balance or transaction limit), and risk rule-related evidence matching.

[0040] Atomic retrieval sub-target decomposition can also be dynamically generated according to the question type. For bridging requests, the controller first identifies intermediate subjects or events, and then retrieves their attributes or related evidence; for comparison requests, the controller retrieves risk information for multiple subjects separately, and then generates a comparison conclusion; for binary classification requests, the controller retrieves evidence supporting whether the risk is established or not, and finally outputs yes / no or risk level.

[0041] Atomic retrieval subtargets can also carry answer format constraints. For example, for a request to "whether account theft exists", the output format is constrained to "risk exists", "risk does not exist", or "insufficient evidence"; for a request to "what type of risk is it", the output format is constrained to one or more types from a preset risk type set; for a request to "how should it be handled", the output format is constrained to one or more types from a preset handling suggestion set.

[0042] The closed-loop intelligent controller decomposes the current financial risk control analysis request into atomic retrieval sub-targets, and the set of sub-targets. Represented as:

[0043] in, For the first Atomic retrieval sub-objectives. For example, for a request to "whether there is a risk of money laundering or money laundering after multiple small deposits into an account were quickly withdrawn recently", sub-objectives could include identifying the account holder, retrieving recent transaction patterns, retrieving payment counterparties, retrieving matching rule clauses, retrieving similar historical cases, and determining whether a risk evidence chain has been formed.

[0044] In step S3, similar historical risk control handling experiences are retrieved from historical experience memory based on the current financial risk control analysis request. Search clues can be obtained from historically successful cases. However, the retrieved historical experience is only used as a hint to guide the current evidence acquisition, rather than directly as the final risk identification result.

[0045] S4: Iteratively execute at least one of the following actions: perform experience memory retrieval on relevant historical risk control handling experience, perform keyword retrieval, semantic retrieval, or fuzzy retrieval on the financial risk control corpus, and perform evidence block reading and graph neighbor search on the self-induced evidence graph; In each iteration, select and execute one of the following actions based on the current financial risk control analysis request, atomic retrieval sub-target, current evidence status, historical experience memory, and self-induced evidence graph, and accumulate and aggregate the observation results returned by each executed action to form a dynamic information set including the obtained evidence information, the source of the obtained evidence information, and the corresponding retrieval action record, which serves as the current evidence status.

[0046] In step S4, a closed-loop intelligent controller can be introduced to iteratively perform retrieval and reading actions among three data sources: historical experience memory, financial risk control corpus, and self-induced evidence graph. In each iteration, the controller autonomously decides which action to perform next (such as retrieving historical experience, retrieving the corpus, reading the complete evidence block, or searching for neighboring evidence along the graph) based on the current request target, incomplete atomic retrieval sub-targets, the state of collected evidence, historical experience hints, and the structure of the evidence graph. The results returned by each action are accumulated and aggregated to form a dynamically updated current evidence state.

[0047] It should be noted that the method of this invention can start from a zero-experience state, i.e., the initial historical experience memory is empty. When processing the first financial risk control request, it mainly relies on the financial risk control corpus and the self-induced evidence graph to complete evidence acquisition, i.e., it iteratively performs the following actions: performing keyword retrieval, semantic retrieval, or fuzzy retrieval on the financial risk control corpus, and performing evidence block reading and graph neighbor search on the self-induced evidence graph. When the first request produces a high-confidence and sufficiently evidentiary financial risk control identification result, the evidence anchor point and evidence block identifier corresponding to the financial risk control identification result are written back into the historical experience memory, so that subsequent requests can utilize the bridging subject and evidence path exposed by the first request.

[0048] S5: When the iteration stopping condition is met, generate financial risk control identification results based on the current evidence status.

[0049] In step S5, the iteration stopping conditions include: at least one complete evidence fragment has been read and the generated result is not an unknown result, or the preset number of read retries, the number of retries for unknown results, or the maximum number of iteration steps has been reached. When the iteration stopping conditions are met, the final financial risk control identification result is generated based on the accumulated evidence status, including risk type, risk level, hit rule, key entity, key transaction, supporting evidence chain, confidence level, explanatory text, and disposal recommendations.

[0050] Risk types can include credit card fraud, account theft, identity theft, cash-out transactions, money laundering, money laundering, illegal money transfers, suspicious merchants, phishing attacks, and credit default risks. The evidence chain explicitly references the identifiers of evidence blocks, making the entire identification process traceable and auditable, facilitating manual review and subsequent analysis. Action recommendations can be generated based on the financial risk control identification results and business rules, including granting approval, enhanced verification, manual review, transaction interception, account limits, account freezing, supplementary due diligence, rule updates, or model review. For auditing and compliance scenarios, the identification results can also output evidence block identifiers and evidence paths, enabling auditors to trace back to original transaction records, rule clauses, or historical cases.

[0051] Preferably, the current evidence status is divided into four input categories: historical experience clues, current complete evidence blocks, current search summaries, and graph neighbor evidence. When generating explanatory text, facts from the current complete evidence block are cited first. If historical experience clues conflict with current evidence, the current evidence prevails, and the conflict is recorded in the audit log.

[0052] Preferably, for requests that hit multiple risk types, the risk types are sorted according to the completeness of the evidence chain, risk level, priority of the hit rule, and confidence level; for requests with insufficient evidence, an evidence insufficiency flag is output and the data types to be supplemented are suggested, such as device login logs, merchant qualification materials, counterparty information, or manual review notes.

[0053] In one embodiment, a self-induced evidence graph is constructed based on multiple evidence blocks, including: 1) Treat each piece of evidence as a node to obtain the node set of the self-induced evidence graph. .

[0054] 2) Represent the beginning text, title, subject name, account identifier, device identifier, merchant name, transaction summary or rule name of each evidence block using vectors.

[0055] 3) Retrieve the nearest neighbor evidence blocks for each evidence block in the vector space, and establish semantic edges when the cosine similarity between the current evidence block and its nearest neighbor evidence blocks is greater than a preset semantic threshold. The semantic edge weights are calculated as follows:

[0056] in, To extract evidence blocks To the evidence block Semantic edge weight values; Pre-set weight coefficients for semantic edges; For indicator functions; For evidence blocks Belongs to the evidence block The set of multiple nearest neighbor evidence blocks, where the nearest neighbor evidence blocks are a preset number of other evidence blocks that are closest to the current evidence block in vector distance or have the highest vector similarity. For evidence blocks vector representation With evidence blocks vector representation Cosine similarity between them; Located at the preset semantic threshold.

[0057] 4) Extract title-type query terms or financial entity terms from each evidence block. Title-type query terms include entity name, merchant name, rule name, account identifier, device identifier, address identifier, or event summary. Based on sparse matching, retrieve a preset number of related evidence blocks with the highest sparse matching score to the current evidence block, and establish lexical reference edges when the matching score between the current evidence block and related evidence blocks is greater than a preset lexical threshold.

[0058] 5) When two evidence nodes satisfy the semantic nearest neighbor condition and their semantic similarity exceeds a preset semantic threshold, a semantic edge is generated between the two evidence nodes, and a preset semantic edge weight is assigned to the semantic edge; when two evidence nodes satisfy the lexical reference condition and their lexical matching score exceeds a preset lexical threshold, a lexical reference edge is generated between the two evidence nodes, and a preset lexical reference edge weight is assigned to the lexical reference edge; the semantic edge weight and the lexical reference edge weight of the same node pair are summed to obtain the final edge weight of the node pair. The expressions for calculating the lexical reference edge weight and the final edge weight are as follows:

[0059]

[0060] in, To extract evidence blocks To the evidence block Lexical reference edge weights; Preset weight coefficients for the basic lexical reference edges; To extract evidence blocks Title-style query terms extracted from After performing sparse search, the set of multiple evidence blocks with the highest matching scores is obtained; For title-based query terms With evidence blocks Content The calculated matching score; The preset lexical threshold; For evidence blocks To the evidence block The final edge weight.

[0061] 6) Construct a self-induced evidence graph based on the node set, semantic edges, lexical reference edges, and final edge weights. :

[0062] Among them, the node set Represents the set of evidence blocks and the set of edges. It includes semantic edges and lexical reference edges. The final edge weight is the attribute value on each edge. The edge set is used to represent the semantic proximity relationship and lexical reference relationship between evidence blocks.

[0063] In one embodiment, historical experience memory is a graph-structured experience memory, where each experience memory node... It is expressed as follows:

[0064] Among them, historical requests Historical risk results Anchor point set of evidence Evidence block identifier set and timestamp Together they form experiential memory nodes that can be reused but are not directly used as answers.

[0065] The evidence anchor set is used to preserve reusable retrieval clues exposed during historical handling processes, including customer identifiers, account identifiers, device identifiers, IP addresses, merchant identifiers, counterparties, risk rule names, fraud types, handling tags, transaction channels, and geographical regions. Among them, handling tags are used to mark the risk category or handling action to which the historical handling results belong.

[0066] Furthermore, if two experience memory nodes share at least one evidence anchor point, an experience edge is established between them, and the edge weight can be determined by the number of shared evidence anchor points.

[0067] In one embodiment, retrieving relevant historical risk control handling experience from historical experience memory based on the current financial risk control analysis request includes: 1) Perform dense retrieval on current financial risk control analysis requests and historical requests. Dense retrieval can be based on vector similarity to obtain dense candidate experiences.

[0068] 2) Sparse retrieval of current financial risk control analysis requests and historical requests, evidence anchors or risk results can be performed using inverted indexes or BM25 retrieval methods to obtain sparse candidate experiences.

[0069] 3) Dense and sparse candidate experiences are ranked and fused based on the inverse ranking fusion method, and the candidate experience score is calculated through inverse ranking fusion. :

[0070] in, Index the candidate experience nodes; This is a request for current financial risk control analysis. This is the set of candidate experience node indexes obtained through dense retrieval; This is the set of candidate experience node indexes obtained through sparse retrieval; This is a preset fusion constant; This is the sorting number of the candidate experience node in the dense search results; Candidate experience nodes The sorting number in the sparse search results; This is an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise.

[0071] Furthermore, the direct experience node with the highest experience score is selected and used as the experience clue for subsequent evidence retrieval.

[0072] 4) When a direct experience node has experience edge neighbors, select the neighbor experience node with the most shared evidence anchor points or the highest experience edge weight. That is, the one-hop extension of the graph structure experience memory only selects a small number of neighbor experience nodes that share the most evidence anchor points or the highest edge weight with the direct experience node, avoiding the injection of a large number of weakly related historical processing records into the context and causing noise. The selected neighbor experience nodes are used to suggest possible bridging subjects, search terms, or evidence paths, rather than replacing the current evidence.

[0073] 5) Linearize the direct experience nodes and neighbor experience nodes into experience hints:

[0074] in, A collection of experience tips; For direct experience nodes; This serves as a neighboring experience node. Experience hints include historical requests, historical risk outcomes, key evidence anchors, and relevant evidence block identifiers. It should be noted that these experience hints are not the final answer to the current request. The closed-loop intelligent controller only uses them as retrieval clues to more quickly locate relevant evidence in the current corpus and avoid erroneous transfer of historical answers to the current risk event.

[0075] In one embodiment, a similarity threshold is set for historical experience retrieval. When the similarity of dense candidate experiences is lower than a preset memory threshold and sparse candidate experiences are empty, an empty experience context is returned, causing the closed-loop intelligent controller to degenerate into a retrieval enhancement process based solely on the current corpus and self-induced evidence graph. This allows the system to continue operating normally when no historical experience is available.

[0076] In one embodiment, the closed-loop intelligent controller in the first... The round of iterations maintains the following state:

[0077] in, It is a state vector; This is the sequence number for the financial risk control analysis request; For the first One financial risk control analysis request, the first The current state of evidence accumulated during each iteration; This represents accumulated historical experience and memory. The current state of evidence is continuously updated as the tool is invoked.

[0078] When the controller performs an experience memory retrieval action, it returns relevant historical risk control handling experience; when it performs a keyword retrieval action, it performs a precise retrieval based on customer number, account number, device number, merchant number, rule name, or blacklist name; when it performs a semantic retrieval action, it retrieves relevant evidence blocks based on risk semantic descriptions; when it performs a fuzzy retrieval action, it expands the retrieval to synonymous descriptions, abnormal behavior patterns, or similar cases.

[0079] When the controller performs an evidence block reading action, it reads the complete content of the specified evidence block instead of just using the retrieval summary; when it performs an evidence graph neighbor search action, it returns the neighbor evidence block with the highest weight in the self-induced evidence graph based on the evidence blocks that have been read or retrieved, which is used to track intermediate entities, upstream and downstream transactions, related devices, related merchants or related risk rules.

[0080] After each action is executed, the result returned by the action is added to the current evidence state. The evidence state update formula is as follows:

[0081] in, For the first One financial risk control analysis request, the first The current state of evidence accumulated during each iteration; The results returned for the action may include the full content of the evidence block, a search summary, historical experience tips, etc.

[0082] When the evidence status is sufficient to support the risk identification result, or when the maximum number of iterations is reached, the controller stops searching and, based on the evidence information already obtained in the current evidence status (including the complete content of the evidence block, search summary, historical experience hints, etc.), infers and generates candidate financial risk control identification results.

[0083] In one embodiment, to avoid generating conclusions solely based on short summaries or historical experience, the controller performs output verification before receiving the financial risk control identification results. If the current evidence state does not contain at least one complete evidence block read result and the read retries have not been reached, the controller adds a correction instruction requiring the execution of the evidence block read action; if the candidate result is unknown, lacks sufficient evidence, or has low confidence and the unknown retries have not been reached, the controller adds a correction instruction requiring the execution of precise retrieval, fuzzy retrieval, or evidence graph neighbor search.

[0084] The number of read retries and the number of unknown retries are set as configurable parameters. If there is no complete evidence block read result in the current evidence state, the evidence block read is triggered first; if the candidate result is unknown or has low confidence, precise retrieval, fuzzy retrieval, or graph neighbor search is triggered first. Through the above correction mechanism, the situation where conclusions are generated solely based on fragment summaries, historical experience, or model parameter knowledge can be reduced.

[0085] When the controller selects to stop the action, it can also perform lightweight validation on the output, including checking whether the risk type belongs to a preset enumeration set, whether the disposal suggestion belongs to a preset disposal set, whether the chain of evidence contains at least one current evidence block identifier, and whether the explanatory text references key facts in the evidence block. If the lightweight validation fails, the controller continues to perform the retrieval or reading action.

[0086] In one embodiment, after the step of generating a financial risk control identification result based on the current state of evidence, the method further includes: The financial risk control identification results are subjected to quality gating judgment. When the writing conditions are met, the current financial risk control analysis request, financial risk control identification results, evidence anchors, evidence block identifiers and timestamps are written into the historical experience memory to obtain the updated historical experience memory.

[0087] Optionally, quality gating judgments include: The confidence level of the financial risk control identification result is calculated or parsed to determine whether the confidence level of the financial risk control identification result belongs to a preset high confidence interval, whether the identification result is unknown, whether it includes unresolved referents (using ambiguous referential terms, such as "account," "transaction," "customer," etc.) or uncertainty markers, and whether the supporting evidence chain includes at least one evidence block from the current corpus. Correspondingly, the writing conditions for quality gating include: the confidence level of the financial risk control identification result belongs to a preset high confidence interval; the financial risk control identification result is not unknown; the current financial risk control analysis request does not contain unresolved referents; the financial risk control identification result does not contain uncertainty markers; the supporting evidence chain contains at least one evidence block from the current financial risk control corpus; and the length of the financial risk control identification result is within a preset range.

[0088] When the overall confidence level of the quality gating, the unknown result, the unresolved reference, the uncertainty marker, and the current evidence chain all meet the conditions, a pass signal is output. :

[0089] in, For financial risk control identification results Confidence level value; To preset a high confidence interval; The result is unknown. To determine whether a supporting chain of evidence contains at least one piece of evidence from the current corpus, a write-back of empirical memory is allowed when the signal is 1; otherwise, it is not, to avoid contaminating subsequent retrieval priors with low-confidence or insufficient evidence.

[0090] When the quality gate passes, evidence anchors are extracted from the current financial risk control analysis request, financial risk control identification results, and supporting evidence chain to form new experience nodes and update the historical experience memory:

[0091] in, For updated historical experience memory nodes; For the first A set of evidence anchors extracted from financial risk control analysis requests and financial risk identification results; For the first A set of evidence block identifiers for each financial risk identification result; For the first The timestamp of the completion of a financial risk control analysis request; To update the obtained set of historical experience memories; To process the first The collection of historical experience memories accumulated before each financial risk control analysis request.

[0092] Preferably, an optimized set of evidence anchors is obtained by filtering common words, weak pronouns, and low-discrimination words in the evidence anchors based on the stop word set; the set of evidence anchors is used to connect historical experience nodes that share customers, accounts, devices, IP addresses, merchants, counterparties, risk rules, or risk labels.

[0093] If a new experience node shares an evidence anchor with an existing experience node, then an experience edge is established and the edge weight is updated.

[0094] When a quality gate fails, the current identification result is not written into the experience memory; only the necessary audit logs are retained. This allows for the gradual accumulation of reliable experience during continuous risk control operations, while preventing low-confidence, ambiguous, unknown, or insufficiently evidenced results from becoming prior knowledge for future retrievals.

[0095] In one embodiment, when the maximum number of iterations is reached and the confidence level of the candidate identification result generated based on the current evidence state has not yet reached a preset threshold, a forced candidate result is generated based on the collected current evidence state. When the forced candidate result is unknown or the confidence level is lower than the preset threshold, the current financial risk control analysis request is cleaned and fuzzy searched for remedies, and the financial risk control identification result and confidence level are regenerated based on the remedy search results.

[0096] In one specific embodiment, historical experience shows that a certain device identifier has been associated with multiple fraudulent accounts, and the evidence path points to a specific IP address and an abnormal merchant. The current request is to determine whether multiple transactions of a new account pose a risk of account theft. First, historical experience is retrieved through the device identifier and abnormal merchant anchor points, but historical conclusions are not directly reused; then, the controller continues to search and read complete evidence blocks in the current transaction history, login logs, and merchant information, and finds that the new account is associated with the same device identifier, similar login IPs, and the same abnormal merchants. Finally, it outputs a high risk of account theft, supports the evidence chain, and suggests manual review / temporary limit.

[0097] In one specific embodiment, the first financial risk control analysis request is used to determine whether a merchant poses a cash-out risk. After reading transaction records, merchant categories, historical chargeback records, and risk rules, it outputs "high risk" and writes the merchant identifier, terminal number, receiving account, and hit rule into the experience memory. When a subsequent request reviews abnormal inflows into another account, it retrieves this historical experience through the receiving account and terminal number, and then continues to read the transaction records and login records of the current account. Ultimately, it discovers that the account has a closed-loop financial relationship with the aforementioned merchant, thus outputting a cash-out association risk.

[0098] In one specific implementation, the initial financial risk control analysis request is used to analyze whether a client has suspicious money laundering transactions. The evidence graph is expanded along the client, counterparty account, and geographical region to obtain multi-hop transaction evidence and write it back to high-confidence experience. Subsequent requests involving the same counterparty account but different clients quickly locate the counterparty account using experience anchors, and then read the current client's new transaction evidence, avoiding misjudgments caused by directly reusing the original client's conclusions.

[0099] As can be seen from the above specific embodiments, the historical experience in this invention is not a static answer database, but a set of reusable search paths and evidence anchors. This design can reduce repeated searches in continuous financial risk control requests, while maintaining mandatory reading and quality verification of the evidence requested in the current request.

[0100] The method of this invention achieves the synergy between historical experience reuse and current evidence acquisition: historical experience helps to locate the search direction, self-induced evidence graph helps to track multi-hop evidence, closed-loop controller ensures the dynamic selection of evidence reading and retrieval actions, and quality gating ensures the reliability of memory updates.

[0101] In some embodiments, the method provided by this invention can be applied to cybersecurity risk identification scenarios. Accordingly, a cybersecurity risk identification method based on large-model self-evolutionary retrieval enhancement includes the following steps: Step S10: Construct a network security corpus; the network security corpus includes one or more of the following: network traffic records, access logs, device identifiers, network identifiers, IP addresses, MAC addresses, domain names, malicious IP address blacklists, malicious domain name blacklists, vulnerability intelligence, threat intelligence, and historical security incident handling cases.

[0102] Step S20: The network security corpus is processed into text and desensitized. The desensitized data is segmented according to a preset granularity, and a unique identifier is configured for each segmentation unit to obtain multiple evidence blocks. A self-induced evidence graph is constructed based on the multiple evidence blocks. The nodes in the self-induced evidence graph correspond to the evidence blocks, and the edges in the self-induced evidence graph include at least one of semantic edges generated based on semantic similarity and lexical reference edges generated based on lexical reference relations.

[0103] Step S30: Decompose the current network security analysis request to obtain at least one atomic retrieval sub-target, and retrieve relevant historical security handling experience from the historical experience memory based on the current network security analysis request; wherein, the atomic retrieval sub-target includes one or more of subject identifier query, association relationship retrieval, subject attribute value acquisition, threat feature evidence retrieval, and risk rule related evidence matching, and the subject attribute value includes one or more of IP address geographical location, host operating system version, port open status, and domain name resolution record.

[0104] Step S40: Iteratively execute at least one retrieval enhancement action, which includes: performing experience memory retrieval in the historical experience memory bank, performing keyword retrieval, semantic retrieval, or fuzzy retrieval in the network security corpus, reading the full text of the evidence block according to its unique identifier, and performing graph neighbor search on the self-induced evidence graph. In each iteration, based on the current network security analysis request, atomic retrieval sub-target, current evidence status, historical experience memory bank, and self-induced evidence graph, select and execute the corresponding retrieval enhancement action, and incorporate the results returned by the retrieval enhancement action into the current evidence status to form a dynamic information set including the acquired evidence information, the source of the evidence information, and the evidence block identifier.

[0105] Step S50: When the iteration stopping condition is met, a network security risk identification result is generated based on the current evidence status. The iteration stopping condition includes having read at least one full-text evidence block and generating a result that is not unknown, or having reached a preset number of read retries, unknown result retries, or the maximum number of iteration steps. The network security risk identification result includes one or more of the following: risk type, risk level, supporting evidence chain, confidence level, and handling recommendations. The risk type includes one or more of the following: distributed denial-of-service attack, data breach, unauthorized access, malicious scanning, account takeover, and abnormal external connections. The handling recommendations include one or more of the following: blocking IP addresses, isolating hosts, banning accounts, restricting port access, and triggering manual review.

[0106] Preferably, the method further includes a quality gating judgment step, namely: Step S60: When the network security risk identification result meets the quality gating conditions, the current network security analysis request, network security risk identification result, supporting evidence chain and corresponding handling experience are written into the historical experience memory for use in the risk identification of subsequent network security events.

[0107] In other embodiments, the method of the present invention can also be applied to industrial equipment anomaly detection, production quality anomaly identification, or other industrial data processing scenarios. Through retrieval enhancement, evidence state accumulation, self-induced evidence graph traversal, and historical experience memory update, the method can achieve interpretable and traceable identification of risk events.

[0108] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the risk identification method based on large model self-evolutionary retrieval enhancement described in any of the above method embodiments or combinations thereof. The risk identification method includes one or more of the following: financial risk identification method, cybersecurity risk identification method, industrial equipment anomaly identification method, and production quality anomaly identification method.

[0109] This invention also provides a terminal, which shares the same inventive concept as the risk identification method based on large model self-evolutionary retrieval enhancement described in any of the above method embodiments or combinations thereof. The terminal includes a memory and a processor. The memory stores computer instructions executable on the processor. When the processor executes the computer instructions, it implements the steps of the risk identification method based on large model self-evolutionary retrieval enhancement described above.

[0110] In some embodiments, the terminal is implemented as an electronic device, which includes at least one processing unit, at least one storage unit, a bus, an input / output interface, and a network adapter. The bus is used to connect the processing unit, the storage unit, the input / output interface, and the network adapter.

[0111] The storage unit is used to store program code, which can be executed by the processing unit to cause the processing unit to perform the steps described in the above method embodiments of this specification. The storage unit includes one or more of random access memory, cache memory, and read-only memory.

[0112] The electronic device is also capable of communicating with one or more external devices through the input / output interface, and / or with one or more other computing devices through the network adapter.

[0113] Those skilled in the art will understand that the above method embodiments can be implemented by computer program instructions in conjunction with corresponding hardware. The computer program instructions can be stored in a non-volatile computer-readable storage medium, and when executed by a processor, the processor performs all or part of the steps of any method embodiment of the present invention.

[0114] The above specific embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various modifications, substitutions, or improvements can be made without departing from the concept of the present invention, and all such modifications, substitutions, or improvements should fall within the scope of protection of the present invention.

Claims

1. A financial risk control identification method enhanced by large-scale model self-evolutionary retrieval, characterized in that, Includes the following steps: Construct a financial risk control corpus, including one or more of the following: transaction records, account information, device or network identifiers, risk rules, blacklists, historical handling cases, regulatory documents, and complaint or chargeback records; The financial risk control corpus is textualized and anonymized. The anonymized data is segmented according to a preset granularity, and an index is created for each segmentation unit to form multiple evidence blocks with unique identifiers. A self-induced evidence graph is constructed based on the multiple evidence blocks. The nodes of the self-induced evidence graph correspond to the evidence blocks, and the edges correspond to the semantic proximity relationship and / or lexical reference relationship between the evidence blocks. The current financial risk control analysis request is decomposed to obtain at least one atomic retrieval sub-target, and relevant historical risk control handling experience is retrieved from historical experience memory based on the current financial risk control analysis request; the atomic retrieval sub-target includes at least one of subject identifier query, relationship retrieval, subject attribute value acquisition, and risk rule-related evidence matching; Iteratively execute at least one of the following actions: perform an experience memory retrieval action on relevant historical risk control handling experience, perform keyword retrieval, semantic retrieval, or fuzzy retrieval actions on the financial risk control corpus, and perform evidence block reading and graph neighbor search actions on the self-induced evidence graph; in each iteration, select and execute one of the actions based on the current financial risk control analysis request, atomic retrieval sub-target, current evidence status, historical experience memory, and self-induced evidence graph, and accumulate and aggregate the observation results returned by each executed action to form a dynamic information set including the obtained evidence information, the source of the obtained evidence information, and the corresponding retrieval action record, as the current evidence status; When the iteration stopping condition is met, a financial risk control identification result is generated based on the current evidence status; wherein, the iteration stopping condition includes: at least one complete evidence fragment has been read and the generated result is not an unknown result, or the preset number of reading retries, the number of unknown result retries, or the maximum number of iteration steps has been reached; the financial risk control identification result includes one or more of the following: risk type, risk level, supporting evidence chain, confidence level, and handling suggestions; The historical experience memory is a graph structured experience memory. Each experience memory node includes a historical request, a historical risk result, a set of evidence anchors, a set of evidence block identifiers, and a timestamp. The set of evidence anchors includes one or more of the following: customer identifier, account identifier, device identifier, IP address, merchant identifier, counterparty, transaction channel, risk rule name, fraud type, or handling label. When two experience memory nodes share at least one evidence anchor point, an experience edge is established between the two experience memory nodes. Based on the current financial risk control analysis request, relevant historical risk control handling experience is retrieved from historical experience memory, including: Dense retrieval of current and historical financial risk control analysis requests yields dense candidate experiences. Sparse search is performed on current financial risk control analysis requests and historical requests, evidence anchors or risk results to obtain sparse candidate experiences; The dense and sparse candidate experiences are sorted and fused based on the inverse sorting fusion method, and the direct experience node with the highest score is selected. When a direct experience node has an experience edge neighbor, select at least one neighbor experience node with the most shared evidence anchor points or the highest experience edge weight. Direct experience nodes and neighboring experience nodes are linearized into experience hints to guide the current evidence retrieval, but are not directly used as the current financial risk control identification result.

2. The financial risk control identification method enhanced by large-model self-evolutionary retrieval according to claim 1, characterized in that, The construction of a self-induced evidence graph based on multiple evidence blocks includes: Treat each piece of evidence as a node to obtain the node set of the self-induced evidence graph; Each evidence block is represented by a vector representing its opening text, title, subject name, account identifier, device identifier, merchant name, transaction summary, or rule name. Retrieve the nearest neighbor evidence blocks of each evidence block in the vector space, and establish a semantic edge when the similarity between the current evidence block and the nearest neighbor evidence block is greater than a preset semantic threshold; the nearest neighbor evidence block is a preset number of other evidence blocks that are closest to the current evidence block in vector distance or have the highest vector similarity. Extract title-style query terms or financial entity terms from each evidence block, retrieve relevant evidence blocks based on sparse matching, and establish lexical reference edges when the matching score between the current evidence block and relevant evidence blocks is greater than a preset lexical threshold; the relevant evidence blocks are a preset number of other evidence blocks with the highest sparse matching score with the current evidence block. When two evidence nodes meet the semantic nearest neighbor condition and their semantic similarity exceeds a preset semantic threshold, a semantic edge is generated between the two evidence nodes, and a preset semantic edge weight is assigned to the semantic edge; when two evidence nodes meet the lexical reference condition and their lexical matching score exceeds a preset lexical threshold, a lexical reference edge is generated between the two evidence nodes, and a preset lexical reference edge weight is assigned to the lexical reference edge; the semantic edge weight and the lexical reference edge weight of the same node pair are summed to obtain the final edge weight of the node pair; A self-induced evidence graph is constructed based on the node set, semantic edges, lexical reference edges, and edge weights.

3. The financial risk control identification method enhanced by large-scale model self-evolutionary retrieval according to claim 1, characterized in that, Before generating the financial risk control identification result based on the current evidence status, it is determined whether the current evidence status includes at least one complete evidence block that has been read; if not and the number of read retries has not been reached, a correction instruction is generated to perform the evidence block reading action. If the candidate identification result generated based on the current evidence state is unknown or the confidence level is lower than the preset threshold, and the unknown retry count has not been reached, then a correction instruction is generated to perform precise retrieval, fuzzy retrieval, or evidence graph neighbor search.

4. The financial risk control identification method enhanced by large-model self-evolutionary retrieval according to claim 1, characterized in that, After generating the financial risk control identification result based on the current evidence status, the process also includes: The financial risk control identification results are subjected to quality gating judgment. When the writing conditions are met, the current financial risk control analysis request, financial risk control identification results, evidence anchors, evidence block identifiers and timestamps are written into the historical experience memory to obtain the updated historical experience memory.

5. The financial risk control identification method enhanced by large-model self-evolutionary retrieval according to claim 4, characterized in that, The quality gating judgment includes: Determine whether the confidence level of the financial risk control identification result belongs to the preset high confidence interval, whether the identification result is unknown, whether it includes unresolved references or uncertainty markers, and whether the supporting evidence chain includes at least one evidence block in the current corpus.

6. The financial risk control identification method enhanced by large-scale model self-evolutionary retrieval according to claim 1, characterized in that, When the maximum number of iterations is reached, and the confidence level of the candidate recognition result generated based on the current evidence state has not yet reached the preset threshold, a forced candidate result is generated based on the collected current evidence state. When the forced candidate result is unknown or the confidence level is lower than the preset threshold, the current financial risk control analysis request is cleaned and fuzzy searched for remedies, and the financial risk control identification result and confidence level are regenerated based on the remedy search results.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the financial risk control identification method enhanced by large model self-evolution retrieval as described in any one of claims 1-6.

8. A terminal comprising a memory and a processor, wherein the memory stores computer instructions executable on the processor, characterized in that, When the processor executes the computer instructions, it performs the steps of the financial risk control identification method enhanced by large model self-evolution retrieval as described in any one of claims 1-6.

Images