Risk assessment method, device, equipment, storage medium and program product

Abstract

Embodiments of the present application provide a risk assessment method, device, equipment, storage medium and program product, relating to the field of financial technology. The method comprises: preprocessing and entity recognition on a policy text to be processed to determine a plurality of financial entities; determining a plurality of entity relationships between the plurality of financial entities based on the policy text; generating structured data according to the plurality of financial entities and the plurality of entity relationships; constructing a multi-dimensional risk index based on the structured data; and when a financial object to be evaluated is evaluated, evaluating the risk level of the financial object based on the multi-dimensional risk index. Through the method of the present application, the reliability of risk assessment on the financial object is improved.

Description

Technical Field

[0001] This application relates to the field of financial technology, and more particularly to a risk assessment method, apparatus, device, storage medium, and program product. Background Technology

[0002] In the financial sector, financial institutions face compliance and regulatory requirements for various types of financial instruments (such as funds, insurance, and trusts). Regulatory policies originate from diverse sources, encompass different regulatory bodies, and are characterized by dynamic updates, significant semantic differences, and overlapping constraints.

[0003] In related technologies, policy documents are typically interpreted manually. Based on the interpretation, risk indicators in the risk rule base are manually hard-coded to update the risk indicators. Then, a risk assessment of financial objects is conducted based on the updated risk rule base. However, in this method, manual interpretation of policy documents relies on individual understanding, which is prone to interpretation bias, omissions, and incomplete interpretations. This results in poor accuracy and comprehensiveness in the risk indicators constructed based on the manually interpreted results, ultimately making it impossible to conduct accurate and comprehensive risk assessments of financial objects based on these risk indicators.

[0004] As can be seen from the above, the reliability of risk assessment for financial objects is low. Summary of the Invention

[0005] This application provides a risk assessment method, apparatus, equipment, storage medium, and program product to address the problem of low reliability in risk assessment of financial objects.

[0006] Firstly, this application provides a risk assessment method, including:

[0007] The policy texts to be processed are preprocessed and entity identification is performed to identify multiple financial entities;

[0008] Based on the policy text, the relationships between the multiple financial entities are determined;

[0009] Based on the multiple financial entities and the relationships between them, structured data is generated;

[0010] A multidimensional risk indicator is constructed based on the structured data, and the multidimensional risk indicator is used to assess the risk level of the financial object to be assessed.

[0011] When assessing a financial object, the risk level of the financial object is evaluated based on the multidimensional risk indicators.

[0012] In one possible implementation, the policy text to be processed undergoes preprocessing and financial entity identification, identifying multiple financial entities, including:

[0013] The policy text is preprocessed to obtain multiple text sequences;

[0014] Domain labels are embedded into the multiple text sequences by a target model to obtain multiple fused text sequences. The fused text sequences incorporate domain information. The target model is a model based on a bidirectional encoder-representation converter.

[0015] Entity identification is performed on the multiple fused text sequences to determine the multiple financial entities.

[0016] In one possible implementation, the policy text is preprocessed to obtain multiple text sequences, including:

[0017] Based on a financial terminology dictionary, the policy text is coarse-grainedly segmented using Chinese lexical analysis tools to obtain a word sequence, which includes multiple terms from the policy text.

[0018] The long terms in the word sequence are further broken down using a word segmentation algorithm to obtain a sub-word sequence.

[0019] The word sequence is segmented using a sliding window to obtain the multiple text sequences.

[0020] In one possible implementation, determining multiple entity relationships among the plurality of financial entities based on the policy text includes:

[0021] Based on the relation types in the pre-established hierarchical domain ontology library, multiple entity pairs are identified among the multiple financial entities through the target model;

[0022] For any pair of entities, the entity relationships between the pairs are determined by combining the policy text and the pre-generated financial knowledge graph through the target model.

[0023] In one possible implementation, structured data is generated based on the plurality of financial entities and the relationships between the plurality of entities, including:

[0024] For any given financial entity, the entity is labeled using a bidirectional long short-term memory network-conditional random field model to obtain the entity's label.

[0025] For any entity relationship, the entity relationship is labeled using the bidirectional long short-term memory network-conditional random field model to obtain the label of the entity relationship;

[0026] The structured data is generated based on the multiple financial entities, the tags of the multiple financial entities, the multiple entity relationships, and the tags of the entity relationships.

[0027] In one possible implementation, determining the risk level of the financial object based on the multidimensional risk indicators includes:

[0028] The financial object to be evaluated is assessed based on the multidimensional risk indicators, and multiple initial risk values ​​corresponding to the multidimensional risk indicators are obtained.

[0029] The risk level of the financial object is determined based on the multiple initial risk values ​​corresponding to the multidimensional risk indicators.

[0030] In one possible implementation, the multidimensional risk indicators include compliance strictness, policy cross-constraints, user adaptability risk, and timeliness risk.

[0031] Based on the aforementioned multidimensional risk indicators, the financial object to be evaluated is assessed, resulting in multiple initial risk values, including:

[0032] Regarding compliance stringency, the initial risk value of the financial object under compliance stringency is determined based on the type and number of policies involved.

[0033] In response to policy cross-constraints, the initial risk value of the financial object under policy cross-constraints is determined based on the multiple related policies involved in the financial object.

[0034] To address user compatibility risk, the initial risk value of the financial object under user compatibility risk is determined based on the object level of the financial object and the user level of the user.

[0035] To address time-sensitive risks, the initial risk value of the financial entity is determined based on the release and effective dates of the policies related to that entity.

[0036] In one possible implementation, the risk level of the financial object is determined based on multiple initial risk values ​​corresponding to the multidimensional risk indicators, including:

[0037] Determine the risk weight corresponding to each risk indicator in the multidimensional risk index;

[0038] The comprehensive risk value of the financial object is determined by multiplying the initial risk value and the risk weight corresponding to each risk indicator.

[0039] The risk level is determined based on the comprehensive risk value.

[0040] In one possible implementation, if the financial object is a financial product, the method further includes:

[0041] For financial objects with a risk level of severe risk, the transaction of the financial object is blocked in real time, and an alarm message is sent to the target device;

[0042] For financial objects with a risk level of general risk, correction information is sent to the target device;

[0043] For financial objects with a risk level of potential risk, generate real-time voice prompts to remind users that the financial object has potential risks.

[0044] Secondly, this application provides a risk assessment apparatus, comprising: a processing module, a determining module, a generating module, and a constructing module, wherein,

[0045] The processing module is used to preprocess and identify entities from the policy text to be processed, and to determine multiple financial entities.

[0046] The determining module is used to determine multiple entity relationships between the multiple financial entities based on the policy text;

[0047] The generation module is used to generate structured data based on the plurality of financial entities and the relationships between the plurality of entities;

[0048] The construction module is used to construct multidimensional risk indicators based on the structured data, and the multidimensional risk indicators are used to assess the risk level of the financial object to be assessed.

[0049] The assessment module is used to assess the risk level of a financial object based on the multidimensional risk indicators when assessing the financial object to be assessed.

[0050] In one possible implementation, the processing module is specifically used for:

[0051] The policy text is preprocessed to obtain multiple text sequences;

[0052] Domain labels are embedded into the multiple text sequences by a target model to obtain multiple fused text sequences. The fused text sequences incorporate domain information. The target model is a model based on a bidirectional encoder-representation converter.

[0053] Entity identification is performed on the multiple fused text sequences to determine the multiple financial entities.

[0054] In one possible implementation, the processing module is specifically used for:

[0055] Based on a financial terminology dictionary, the policy text is coarse-grainedly segmented using Chinese lexical analysis tools to obtain a word sequence, which includes multiple terms from the policy text.

[0056] The long terms in the word sequence are further broken down using a word segmentation algorithm to obtain a sub-word sequence.

[0057] The word sequence is segmented using a sliding window to obtain the multiple text sequences.

[0058] In one possible implementation, the determining module is specifically used for:

[0059] Based on the relation types in the pre-established hierarchical domain ontology library, multiple entity pairs are identified among the multiple financial entities through the target model;

[0060] For any pair of entities, the entity relationships between the pairs are determined by combining the policy text and the pre-generated financial knowledge graph through the target model.

[0061] In one possible implementation, the generation module is specifically used for:

[0062] For any given financial entity, the entity is labeled using a bidirectional long short-term memory network-conditional random field model to obtain the entity's label.

[0063] For any entity relationship, the entity relationship is labeled using the bidirectional long short-term memory network-conditional random field model to obtain the label of the entity relationship;

[0064] The structured data is generated based on the multiple financial entities, the tags of the multiple financial entities, the multiple entity relationships, and the tags of the entity relationships.

[0065] In one possible implementation, the evaluation module is specifically used for:

[0066] The financial object to be evaluated is assessed based on the multidimensional risk indicators, and multiple initial risk values ​​corresponding to the multidimensional risk indicators are obtained.

[0067] The risk level of the financial object is determined based on the multiple initial risk values ​​corresponding to the multidimensional risk indicators.

[0068] In one possible implementation, the multidimensional risk indicators include compliance strictness, policy cross-constraints, user adaptability risk, and timeliness risk; the assessment module is specifically used for:

[0069] Regarding compliance stringency, the initial risk value of the financial object under compliance stringency is determined based on the type and number of policies involved.

[0070] In response to policy cross-constraints, the initial risk value of the financial object under policy cross-constraints is determined based on the multiple related policies involved in the financial object.

[0071] To address user compatibility risk, the initial risk value of the financial object under user compatibility risk is determined based on the object level of the financial object and the user level of the user.

[0072] To address time-sensitive risks, the initial risk value of the financial entity is determined based on the release and effective dates of the policies related to that entity.

[0073] In one possible implementation, the evaluation module is specifically used for:

[0074] Determine the risk weight corresponding to each risk indicator in the multidimensional risk index;

[0075] The comprehensive risk value of the financial object is determined by multiplying the initial risk value and the risk weight corresponding to each risk indicator.

[0076] The risk level is determined based on the comprehensive risk value.

[0077] In one possible implementation, if the financial object is a financial product, the device further includes an intervention module, the intervention module being used for:

[0078] For financial objects with a risk level of severe risk, the transaction of the financial object is blocked in real time, and an alarm message is sent to the target device;

[0079] For financial objects with a risk level of general risk, correction information is sent to the target device;

[0080] For financial objects with a risk level of potential risk, generate real-time voice prompts to remind users that the financial object has potential risks.

[0081] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0082] The memory stores the instructions that the computer executes;

[0083] The processor executes computer-executable instructions stored in memory to implement the method as described in any of the first aspects.

[0084] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method described in any of the first aspects.

[0085] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the first aspects.

[0086] This application provides a risk assessment method, apparatus, device, storage medium, and program product. In this embodiment, the policy text to be processed is preprocessed and entity identified to determine multiple financial entities; based on the policy text, multiple entity relationships between the multiple financial entities are determined; structured data is generated based on the multiple financial entities and their relationships; multidimensional risk indicators are constructed based on the structured data; and when assessing the financial object to be evaluated, the risk level of the financial object is assessed based on the multidimensional risk indicators. In this technical solution, electronic devices can interpret policy text, generate structured data, and construct multidimensional risk indicators based on the structured data. Compared to manual interpretation of policy text, this avoids interpretation bias due to individual comprehension issues, is less prone to omissions, and provides a more comprehensive interpretation. Therefore, it improves the accuracy and comprehensiveness of constructing multidimensional risk indicators, thereby enhancing the reliability of risk assessment of financial objects based on multidimensional risk indicators. Attached Figure Description

[0087] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0088] Figure 1 A schematic diagram illustrating an application scenario provided in an embodiment of this application;

[0089] Figure 2 A flowchart illustrating a risk assessment method provided in an embodiment of this application;

[0090] Figure 3 A flowchart illustrating another risk assessment method provided in this application embodiment;

[0091] Figure 4 This is a schematic diagram of the structure of a risk assessment device provided in an embodiment of this application;

[0092] Figure 5 This is a schematic diagram of another risk assessment device provided in an embodiment of this application;

[0093] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0094] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0095] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0096] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0097] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0098] It should be noted that the risk assessment methods, devices, equipment, storage media and products provided in this application can be used in the field of fintech, or in any field other than fintech. The application fields of the risk assessment methods, devices, equipment, storage media and products in this application are not limited.

[0099] Below, in conjunction with Figure 1 The application scenarios of this application will be explained.

[0100] Figure 1 This is a schematic diagram illustrating an application scenario provided by an embodiment of this application. Please refer to [link / reference]. Figure 1 Multidimensional risk indicators can be constructed based on policy texts. For example, multidimensional risk indicators can include compliance strictness, user adaptability risk, and timeliness risk.

[0101] Based on multidimensional risk indicators, risk assessments can be conducted on financial objects to determine their risk levels.

[0102] For example, if the financial object is a fund product, and the multi-dimensional risk indicators are as exemplified above, then based on these multi-dimensional risk indicators, the compliance rigor, user suitability risk, and timeliness risk of the fund product can be assessed, thereby comprehensively determining the risk level of the fund product as severe risk.

[0103] In related technologies, policy documents are typically interpreted manually. Based on the interpretation, risk indicators in the risk rule base are manually hard-coded to update the risk indicators. Then, a risk assessment of financial entities is conducted based on the updated risk rule base. However, this method relies on individual understanding, which is prone to interpretation bias, omissions, and incomplete interpretations. This results in poor accuracy and comprehensiveness in constructing risk indicators based on manual interpretations, ultimately making it impossible to conduct accurate and comprehensive risk assessments of financial entities based on these indicators. Therefore, the reliability of risk assessments of financial entities is low.

[0104] In this embodiment, the electronic device can preprocess and identify entities from the policy text to be processed, determine multiple financial entities, and establish the relationships between these entities based on the policy text. The electronic device can generate structured data based on the multiple financial entities and their relationships, and then construct multidimensional risk indicators based on this structured data. When assessing a financial object, the risk level of the financial object can be evaluated based on these multidimensional risk indicators. In this technical solution, the electronic device can interpret the policy text, generate structured data, and construct multidimensional risk indicators based on this data. Compared to manual interpretation of policy text, this avoids interpretation biases due to individual comprehension issues, is less prone to omissions, and provides a more comprehensive interpretation. Therefore, it improves the accuracy and comprehensiveness of constructing multidimensional risk indicators, thereby enhancing the reliability of risk assessment of financial objects based on these indicators.

[0105] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0106] The execution subject in this application embodiment can be an electronic device or a risk assessment device installed in an electronic device. The risk assessment device can be implemented through software or a combination of software and hardware. The risk assessment device can be a processor in the electronic device. For ease of understanding, the following description uses an electronic device as the execution subject.

[0107] Figure 2This is a flowchart illustrating a risk assessment method provided in an embodiment of this application. Please refer to [link / reference]. Figure 2 The method may include:

[0108] S201. Preprocess and identify entities from the policy texts to be processed to determine multiple financial entities.

[0109] Optionally, there may be at least one policy text to be processed. The policy text to be processed is typically unstructured.

[0110] A financial entity refers to a specific, identifiable, and distinguishable entity or abstract concept within the financial sector. Examples of financial entities include: financial institutions, enterprises, financial products, risk assessments, and price-to-earnings ratios.

[0111] Optionally, the policy text to be processed can be preprocessed and financial entity identification can be performed to determine multiple financial entities in the following manner: the policy text is preprocessed to obtain multiple text sequences; domain labels are embedded into the multiple text sequences through a target model to obtain multiple fused text sequences; entity identification is performed in the multiple fused text sequences to determine multiple financial entities.

[0112] Optionally, preprocessing can be segmentation. Electronic devices can segment policy text to obtain multiple text sequences. For any given text sequence, the text sequence is a sequence of multiple words.

[0113] For example, if the policy text states that "private fund managers shall assess investors' risk identification and risk-bearing capabilities and confirm that investors meet the qualified investor standards," then the electronic device can segment the policy text to obtain eight text sequences as shown in Table 1:

[0114] Table 1

[0115]

[0116] For example, text sequence 1 is a sequence consisting of five words: "private equity, fund, manager, should, and to".

[0117] The target model can be based on a corpus and pre-trained on a Bidirectional Encoder Representations from Transformers (BERT). For example, the target model could be a financial BERT (BERT-Finance) model that incorporates span-based pre-training of BERT (Span BERT) technology.

[0118] Optionally, the corpus can include unstructured texts such as financial regulatory documents, self-regulatory rules of the Securities and Investment Fund Association, and product brochures. A domain-specific corpus of 200GB+ can be built, covering types such as legal provisions, business guidelines, and case precedents.

[0119] During model training, the BERT model can be trained using the Masked Language Model (MLM) task. This BERT model can then be trained on a corpus to obtain the BERT-Finance model. Finally, Span BERT technology can be incorporated to mask entire segments of continuous financial terms (such as "qualified investor identification standards"), improving the target model's ability to capture the semantics of specialized phrases. Furthermore, an entity type prediction task can be added to the target model to perform binary classification of the context corresponding to labeled entities (such as "cooling-off period duration"), strengthening entity boundary awareness.

[0120] The target model can be trained based on a corpus to output financial entities and the corresponding entity prediction probability for each financial entity. The entity prediction probability refers to the probability that the output entity prediction is correct. For entity samples whose entity prediction probability is less than a preset probability, samples with high uncertainty are selected and added to the labeling queue using a density-based sampling method, and the target model is updated periodically.

[0121] For any text sequence, a domain label can be embedded into the text sequence using a target model to obtain the corresponding fused text sequence. The fused text sequence incorporates domain information.

[0122] Specifically, one-hot encoding of regulatory agency types can be obtained, and this encoding can be concatenated into the text sequence at the input layer of the target model. For example, the one-hot encoding of the China Securities Regulatory Commission (CSRC) can be represented as (0, 1, 0).

[0123] For example, if the policy text in the above example is issued by the China Securities Regulatory Commission (CSRC), then for the eight text sequences in Table 1, the target model can be used to concatenate the CSRC's one-hot encoding (0, 1, 0) into each of the eight text sequences to obtain eight fused text sequences.

[0124] One-hot encoding provides a clear input representation for each regulatory agency type, which helps the target model better learn the correlation patterns between different regulatory agency types and other features, thereby improving the performance and generalization ability of the target model and enabling it to perform more accurate analysis and prediction when processing text data involving different regulatory agencies.

[0125] Entity identification can be performed on multiple fused text sequences using a target model to obtain multiple financial entities.

[0126] For example, if there are 8 fused text sequences, each derived from one of the 8 text sequences in Table 1, then the target model can identify 5 financial entities from these 8 fused text sequences, as shown in Table 2:

[0127] Table 2

[0128]

[0129] S202. Based on policy texts, determine the relationships between multiple financial entities.

[0130] Yes, multiple entity relationships between multiple financial entities can be determined based on policy texts in the following ways: Based on the relationship types in the hierarchical domain ontology, multiple entity pairs are identified among multiple financial entities using a target model; for any entity pair, the entity relationships between the entity pairs are determined using a target model, combining the policy text and a pre-generated knowledge graph.

[0131] Optionally, a hierarchical domain ontology can be pre-built using the OWL 2 Web Ontology Language based on a pre-established terminology mapping table. In the hierarchical domain ontology, the category system of financial entities (e.g., "regulatory institutions - financial institutions - investors", "products - rules - standards" and other levels) and potential association rules (e.g., "regulatory institutions → formulate → rules", "financial institutions → evaluate → investors") are pre-defined.

[0132] Optionally, the layered domain ontology library may include: a top-level ontology, a mid-level ontology, and a bottom-level ontology. The top-level ontology can define major categories of financial products (such as securities and insurance) and their common regulatory requirements. The mid-level ontology can define detailed business actions (such as risk assessment, dual recording, and after-sales follow-up) and related policies. The bottom-level ontology can map specific regulatory items to business scenarios.

[0133] In the technical solution of this application, the definitional differences among different regulatory entities can be resolved based on a terminology mapping table. Identifying entity pairs among multiple financial entities using a hierarchical domain ontology avoids meaningless entity combinations (e.g., excluding logically unrelated entity pairs such as "qualified investor criteria" and "private fund manager address"), directly focusing on entity pairs that conform to the semantic logic of the financial domain, thus improving the efficiency of subsequent relationship determination.

[0134] Optionally, a scalable financial knowledge graph can be pre-generated using the TransEdge algorithm based on a hierarchical domain ontology and terminology mapping table.

[0135] Specifically, the TransEdge algorithm can be applied to dynamically generate cross-ontology edges in the financial knowledge graph. The weight of the edges can be determined by both term similarity and policy effectiveness. First, data input is obtained from independently constructed ontologies by each regulatory agency, along with the source ontologies and policy sources for each term. Then, domain-enhanced Word2Vec is used to train term vectors for vector space mapping, mapping the terms into the vector space. This is done to enable subsequent calculation of semantic similarity between terms.

[0136] Alternatively, the semantic similarity (cosine similarity) between two terms can be calculated using the following formula (1):

[0137] Formula (1)

[0138] in, For the term The vector representation of , For the term The vector representation of semantic similarity. The closer the semantic similarity is to 1, the more similar the semantics.

[0139] Optionally, when calculating the policy effectiveness weights, the weights for laws, administrative regulations, departmental rules, and self-regulatory rules can be preset (the weights can be defined manually). Simultaneously, a time-effect decay factor can be introduced. This is to reduce the weight of historical policies.

[0140] For example, if 2023 is used as the base year, then the aging decay factor... It can be expressed by the following formula (2)

[0141] Formula (2)

[0142] in, e is the natural constant.

[0143] Formula (2) is used to characterize the greater the weight decay of a policy as the time of its release is further away from the base year.

[0144] Based on the above formulas (1) and (2), the weight of the associated edge can be expressed by the following formula (3):

[0145] Formula (3)

[0146] in, , which is a weighting coefficient used to balance the impact of semantic similarity and policy effectiveness on the final edge weight; and Is it related to policy? , Related weights, It is a comprehensive consideration of policies and Factors after time decay. Through formula (3), semantic similarity and policy effectiveness can be combined to obtain the final weight of cross-ontology related edges.

[0147] It should be noted that the financial knowledge graph is constantly being accumulated and expanded. Whenever new policy documents or financial information are released, the financial knowledge graph will be expanded based on these new documents or information.

[0148] When determining entity relationships, the pre-generated financial knowledge graph serves as a "verification and supplement to the relationship": the financial knowledge graph can provide historical relationship references or general relationship references (for example, common relationships between "financial institutions" and "investors" include "evaluation", "service", "fundraising", etc.), to help verify whether the relationships extracted from policy texts conform to common sense in the field, or to supplement relationships that are not explicitly stated in policy texts but are logically implied (which need to be carefully judged in conjunction with the context).

[0149] It should be noted that the target model has the ability to extract entity relationships. During the training of the target model, joint training for relationship extraction can be performed. Specifically, in the BERT output vector, relative position embeddings can be performed on the head and tail entity positions in the original text (e.g., head entity start / end position + sine encoding). Secondly, a Transformer Decoder structure can be used, with the head entity vector as the query and the tail entity context as the key / value, to calculate cross-entity semantic relationships. A multi-task loss function can be established in the model as follows:

[0150]

[0151] W1 and W2 are the corresponding weights, which can be preset by humans.

[0152] This weighted combination method allows the model to simultaneously optimize the accuracy of entity boundary recognition and entity relationship classification during training, achieving synergistic improvement in both tasks.

[0153] For example, if multiple financial entities are as shown in Table 2, then based on the relation types in the hierarchical domain ontology library, the target model can be used to determine 5 entity pairs among these 5 financial entities, as shown in Table 3:

[0154] Table 3

[0155]

[0156] For the entity pair "private fund manager - investor", the relationship between the entities can be determined as "assessment" by combining policy texts and pre-generated financial knowledge graphs through the target model. This means that the private fund manager needs to assess the investor. Similarly, the entity relationships between other entity pairs can be determined as shown in Table 3.

[0157] S203. Generate structured data based on multiple financial entities and their relationships.

[0158] Optionally, structured data can be generated based on multiple financial entities and multiple entity relationships in the following manner: For any financial entity, the financial entity is labeled using a Bidirectional Long Short-Term Memory - Conditional Random Field (BiLSTM-CRF) model to obtain the entity's label; for any entity relationship, the entity relationship is labeled using the same model to obtain the relationship's label; and structured data is generated based on multiple financial entities, their labels, the relationships, and the labels of the relationships.

[0159] BiLSTM can capture semantic dependencies through context (e.g., distinguishing the different roles of "bank" in "Bank of China" and "bank deposits"); CRF can improve annotation accuracy by leveraging constraints between labels (e.g., "financial institution" is more likely to be followed by "business name" rather than "date").

[0160] Optionally, if the financial entity is a policy provision, the policy provision can be attribute-labeled using a BiLSTM-CRF model to obtain the corresponding label. Optionally, the attributes can include mandatory, prohibited, and recommended.

[0161] If the financial entity is a financial product, attributes such as the product's scope of application and product type can be labeled using BiLSTM-CRF. For example, the product types of financial products can include: funds, insurance, and wealth management products.

[0162] Alternatively, the structured data can be JSON data, existing as key-value pairs.

[0163] For example, if multiple financial entities are as shown in Table 2, then each financial entity can be labeled using BiLSTM-CRF, and the corresponding labels for each financial entity are shown in Table 4:

[0164] Table 4

[0165]

[0166] If multiple relationships between multiple financial entities are as shown in Table 3, then BiLSTM-CRF can be used to label each relationship, resulting in the labels corresponding to each relationship, as shown in Table 5.

[0167] Table 5

[0168]

[0169] Then, based on the multiple financial entities and their labels shown in Table 4, the multiple entity relationships and their labels shown in Table 5, structured data can be generated to enable the interpretation and processing of policy texts.

[0170] S204. Construct multidimensional risk indicators based on structured data.

[0171] Multidimensional risk indicators can be used to assess the risk level of a financial object to be evaluated.

[0172] Optionally, the financial object to be evaluated can be a financial product or a financial-related video, etc.

[0173] For example, if the structured data is generated based on the content shown in Tables 4 and 5 above, then multi-dimensional risk indicators can be constructed, including compliance strictness, policy cross-constraints, user adaptation risk, and timeliness risk.

[0174] Compliance rigor can be used to measure the degree to which a financial entity complies with regulatory rules such as laws and regulations.

[0175] Cross-policy constraints can be used to assess the risk level of a financial entity being subject to multiple overlapping policy restrictions simultaneously.

[0176] User fit risk can be used to measure the degree of risk caused by insufficient matching between financial objects and target users.

[0177] Timeliness risk can be used to measure the degree of risk to financial entities caused by the timeliness of policies.

[0178] S205. When assessing a financial object, the risk level of the financial object shall be assessed based on multidimensional risk indicators.

[0179] For example, if the multidimensional risk indicators include compliance strictness, policy cross-constraints, user adaptability risk, and timeliness risk, then the financial object to be evaluated can be assessed based on these four risk indicators to determine the risk level of the financial object.

[0180] In this embodiment, the electronic device can preprocess and identify entities from the policy text to be processed, determine multiple financial entities, and, based on the policy text, determine multiple entity relationships between these entities. Then, it can generate structured data based on the multiple financial entities and their relationships. The electronic device can construct multidimensional risk indicators based on this structured data and assess the risk level of the financial object being evaluated based on these multidimensional risk indicators. In this technical solution, the electronic device can interpret the policy text, generate structured data, and construct multidimensional risk indicators based on this data. Compared to manual interpretation of policy text, this avoids interpretation biases due to individual comprehension issues, is less prone to omissions, and provides a more comprehensive interpretation. Therefore, it improves the accuracy and comprehensiveness of constructing multidimensional risk indicators, thereby enhancing the reliability of risk assessment of financial objects based on these indicators.

[0181] Below, in Figure 2 Based on the illustrated embodiments, combined with Figure 3 The above risk assessment methods will be explained in detail.

[0182] Figure 3 This is a flowchart illustrating another risk assessment method provided in an embodiment of this application. Please refer to... Figure 3 The methods may include:

[0183] S301. Based on a financial terminology dictionary, the policy text is coarsely segmented using Chinese lexical analysis tools to obtain a word sequence.

[0184] A financial terminology dictionary can include multiple financial terms. It can provide a basis for word segmentation processing.

[0185] Optionally, the policy text can be coarsely segmented using a Chinese lexical analysis (LAC) tool to obtain a word sequence. The word sequence includes multiple terms from the policy text. A word sequence is a holistic sequence composed of multiple words corresponding to the policy text.

[0186] For example, if the policy text is as shown in the example above, then based on a financial terminology dictionary, after coarse-grained segmentation of the policy text using the LAC tool, the resulting word sequence is: "private fund manager, should, to, investor, of, risk identification ability, and, risk-bearing ability, conduct, assess, and, confirm, investor, comply with, qualified investor, standard". This word sequence can include 16 terms, among which "private fund manager" and "investor" are policy terms.

[0187] S302. Using a word segmentation algorithm, long terms in the word sequence are broken down into sub-word sequences.

[0188] Alternatively, the word segmentation algorithm can be the WordPiece algorithm.

[0189] Optionally, based on a financial terminology dictionary, the word sequence can be further refined and split using the WordPiece algorithm to obtain a sub-word sequence. The sub-word sequence is a complete sequence composed of multiple sub-words corresponding to the policy text.

[0190] For example, if the word sequence is as shown in the example above, where "private fund manager," "risk identification capability," "risk-bearing capability," and "qualified investor" are all long terms, then the WordPiece algorithm can be used to refine and break down these four long terms in the word sequence, resulting in the sub-word sequence "private, fund, manager, should, to, investor, of, risk, identification, capability, and, risk, bear, capability, conduct, assessment, and, confirm, investor, comply with, qualified, investor, standard". This sub-word sequence includes 23 words.

[0191] By using word segmentation algorithms to refine and split long terms in a word sequence, the out-of-vocabulary (OOV) problem can be avoided.

[0192] S303. By using a sliding window, the sub-word sequence is segmented to obtain multiple text sequences.

[0193] Optionally, the window size and step size of the sliding window can be preset. For example, the window size can be 5 sub-words, and the step size can be 4 sub-words.

[0194] For example, if the sub-word sequence is as shown in the example above, and the sliding window size is 5 sub-words and the step size is 4 sub-words, then the sub-word sequence can be segmented through the sliding window to obtain multiple text sequences, as shown in Table 1.

[0195] S304. Based on policy texts, determine the entity relationships between multiple financial entities.

[0196] S305. Generate structured data based on multiple financial entities and the relationships between them.

[0197] S306. Construct multidimensional risk indicators based on structured data.

[0198] It should be noted that the specific execution process of steps S304 to S306 can be found in steps S202 to S204, and will not be repeated here.

[0199] S307. Evaluate the financial object to be evaluated based on multidimensional risk indicators to obtain multiple initial risk values ​​corresponding to the multidimensional risk indicators.

[0200] For example, multidimensional risk indicators may include compliance strictness, policy cross-constraints, user adaptability risk, and timeliness risk.

[0201] Regarding compliance stringency, the initial risk value of the financial entity under compliance stringency is determined based on the type and number of policies involved.

[0202] Different types of policies have different weights. For example, a mandatory policy can have a weight of 1, a prohibitive policy can have a weight of 0.8, and a policy of recommendation can have a weight of 0.3.

[0203] Optionally, the initial risk value of a financial object under strict compliance can be expressed as follows: (4)

[0204] Formula (4)

[0205] Where x represents the number of mandatory policies involving the financial entity, y represents the number of prohibitive policies involving the financial entity, and z represents the number of advisory policies involving the financial entity.

[0206] For example, if a financial entity involves one mandatory policy, two prohibitive policies, and one advisory policy, then the initial risk value of that financial entity under strict compliance requirements can be determined. .

[0207] In response to policy cross-constraints, the initial risk value of a financial entity under policy cross-constraints is determined based on the multiple related policies involved in that entity.

[0208] Optionally, if there is no policy conflict among the multiple related policies involved in the financial object, the initial risk value of the financial object under policy cross-constraint can be set to 0; if there is a policy conflict among the multiple related policies involved in the financial object, the initial risk value of the financial object under policy cross-constraint can be set to 5.

[0209] Optionally, when using a policy cross-reference model to detect whether there are policy conflicts among related policies involving financial objects, a local subgraph can be constructed based on the financial knowledge graph first, and then the policy conflict can be detected based on the local subgraph.

[0210] Optionally, the policy crossover model is constructed based on a directed graph with attributes, whereby the directed graph G can be formally defined as: V is a set of vertices, including three types of nodes: policy documents, constraining entities, and business actions.

[0211] A set of directed edges representing entity relations, where edges contain explicit conflict markers, defined by the function Φ(e). If edge e is a conflict relation, then... Otherwise, it is 0.

[0212] It is a property function. Store the metadata of vertex v, such as the original text of the terms and the effective date; Store the relevant attributes of edge e.

[0213] When performing policy conflict detection, a local subgraph GP based on policy P can be constructed. The local subgraph GP can be composed of a vertex set. and edge set Composition, that is .

[0214] First, we can define a set. ,gather It can include all vertices (policy clauses) associated with policy P, and can be represented as follows:

[0215]

[0216] Edge set Including The edges between the vertices in the middle can be represented as follows:

[0217]

[0218] in, and Let V represent any two vertices in the vertex set V that constitute the graph. Where, in the edge...

[0219] middle, It is the starting vertex of this directed edge. It is the endpoint vertex, used to indicate the vertex from which the endpoint originates. To the top The existence of an edge.

[0220] Optionally, a bidirectional breadth-first search (BFS) algorithm is used to detect policy conflicts. Two queues are initialized, each starting from a potentially conflicting policy node. One queue searches forward along the edge direction, while the other searches backward along the opposite edge direction. The levels of the two queues are expanded alternately, recording visited nodes and paths to avoid loops. After each expansion step, it is checked whether the searches in the two directions meet (i.e., whether there is a common node). If they meet, it is checked whether there is a conflicting edge in the path connecting the two directions; if so, a policy conflict is confirmed.

[0221] For example, if a financial object involves 5 policies, and a directed graph GP with policy A and policy B as vertices is constructed based on the financial knowledge graph using the policy crossover model, if a conflict between policy A and policy B is detected, then there is a policy conflict risk between the two policies, and the initial risk value S2 of the financial object under the policy crossover constraint can be determined to be 5.

[0222] To address user-adaptability risk, the initial risk value of the financial object is determined based on the object level of the financial object and the user level of the user.

[0223] Optionally, basic matching rules between user levels and financial object levels can be established in advance to determine a reasonable matching range, which can serve as the basis for judging mismatches. For example, user level C1 matches object level R1, and user level C2 matches object levels R1 and R2.

[0224] For example, if the user level is C1, the financial object level is R3, and the deviation value is 2, then the initial risk value S3 of the financial object under the user suitability risk can be determined to be 2.

[0225] Optionally, a dynamic deviation matrix can be used to establish a matching benchmark between user levels and object levels. The dynamic deviation matrix can have user levels as rows and financial object levels as columns, with matrix cells labeled "fit / misfit" and the misfit weight. The matrix rules are dynamically updated based on changes in customer ratings or product risk levels. A score is recorded for each misfit point.

[0226] To address time-sensitive risks, the initial risk value of the financial entity is determined based on the release and effective dates of the policies related to that entity.

[0227] Optionally, if the policy concerning the financial object has been issued but not yet in effect, and the financial object does not comply with the policy, the initial risk value can be set to 1; if the policy concerning the financial object has been in effect, and the financial object does not comply with the policy, the initial risk value can be set to 5; if the policy concerning the financial object has been in effect, and the financial object complies with the policy, the initial risk value can be set to 0.

[0228] If a financial entity is involved in multiple policies, its risk value under each policy can be determined. Then, a weighted sum is calculated based on the relationships between the policies, the weight of each policy, and its corresponding risk value to obtain the initial risk value of the financial entity under time-sensitive risk. For any given policy, a time-sensitive decay factor based on the policy's release date can be incorporated into the policy's weight.

[0229] For example, if the current time is September 10, 2025, and if policy C related to the financial object has already taken effect and the financial object complies with the relevant policy, then the initial risk value of the financial object under policy C can be determined to be 0; if policy D related to the financial object has been issued but has not yet taken effect, and the financial object does not comply with policy D, then the initial risk value of the financial object under policy D can be determined to be 1; if policy C and policy D are superimposed policies, and if the weight of policy C is 0.2 and the weight of policy D is 1.2, then the initial risk value S4 of the financial object under timeliness risk can be determined to be 0.2*0+1.2*1=1.2.

[0230] S308. Determine the risk level of a financial object based on multiple initial risk values ​​corresponding to multidimensional risk indicators.

[0231] Optionally, the risk weights corresponding to each risk dimension can be determined, and the comprehensive risk value of the financial object can be determined based on the product of the initial risk value and the risk weight corresponding to each risk dimension. Then, the risk level of the financial object can be determined based on the comprehensive risk value of the financial object.

[0232] Optionally, the overall risk value of a financial object can be the maximum value among the products of the initial risk value and the risk weight corresponding to each risk dimension, or it can be the sum of the products of the initial risk value and the risk weight corresponding to each risk dimension.

[0233] For example, the initial risk values ​​and risk weights corresponding to compliance strictness, policy cross-constraints, user adaptation risks, and timeliness risks are shown in Table 6:

[0234] Table 6

[0235]

[0236] Then the overall risk value of the financial object can be determined. .

[0237] If the risk level includes 3 levels, namely severe risk, general risk, and potential risk, then optionally, the following settings can be made: for severe risk, the overall risk value is greater than 8; for general risk, the overall risk value is in the range of [2, 8]; and for potential risk, the overall risk value is less than 2.

[0238] For example, if the overall risk value of the aforementioned financial object is 2, then the risk level of the financial object can be determined as general risk.

[0239] S309. Based on the risk level of the financial object, intervene and handle the financial object.

[0240] Optionally, if the financial object is a financial product, for financial objects with a risk level of severe risk, the transaction of the financial object is blocked in real time and an alarm message is sent to the target device; for financial objects with a risk level of moderate risk, a correction message is sent to the target device; for financial objects with a risk level of potential risk, a real-time voice prompt is generated to remind the user that the financial object has potential risks.

[0241] Alternatively, the target device can be a computer belonging to the risk control department.

[0242] Alarm information may include the financial entity's identifier, risk level, and risk details.

[0243] Corrections can include supplementing relevant information or improving related details. For example, if the financial object is a financial product, the corrected information could be "Please supplement the risk disclosure statement."

[0244] If the financial object is a financial product, and a user is purchasing a financial product, a real-time voice prompt can be generated for financial objects with a potential risk level to remind the user of the potential risk. For example, the real-time voice prompt could be, "The currently recommended product does not match the user's level; it is recommended to switch to an R2-level financial product."

[0245] In this embodiment, the electronic device can perform coarse-grained segmentation of policy text using a Chinese lexical analysis tool based on a financial terminology dictionary to obtain a word sequence; then, using a word segmentation algorithm, it can further refine and split long terms within the word sequence to obtain a sub-word sequence; subsequently, it can segment the sub-word sequence using a sliding window to obtain multiple text sequences. Based on the policy text, the electronic device can determine the entity relationships between multiple financial entities and generate structured data based on these entities and their relationships. Based on the structured data, the electronic device can construct multi-dimensional risk indicators and evaluate the financial object to be assessed, obtaining multiple initial risk values ​​corresponding to the multi-dimensional risk indicators. Then, based on these initial risk values, it can determine the risk level of the financial object. Finally, based on the multi-dimensional risk indicators constructed from the structured data, the electronic device can intervene in the financial object to be assessed according to its risk level. In the technical solution of this application, electronic devices can interpret policy texts, generate structured data, and construct multi-dimensional risk indicators based on the structured data. Compared with manual interpretation of policy texts, this method avoids interpretation deviations due to individual comprehension issues, is less prone to omissions, and provides a more comprehensive interpretation. Therefore, it improves the accuracy and comprehensiveness of constructing multi-dimensional risk indicators, thereby enhancing the reliability of risk assessment of financial objects based on multi-dimensional risk indicators.

[0246] The technical solution of this application also has the following effects:

[0247] (1) Electronic devices can interpret policy texts, generate structured data, and construct multidimensional risk indicators based on structured data. Compared with human interpretation of policies, this improves the efficiency of processing policy texts. Electronic devices can automatically construct multidimensional risk indicators based on structured data. Compared with human coding and updating the rule base, this improves the efficiency of constructing multidimensional risk indicators. Therefore, by combining these two aspects, the efficiency of constructing multidimensional risk indicators is comprehensively improved, thereby improving the efficiency of risk assessment of financial objects.

[0248] (2) It solves the problem of integrating multi-source regulatory policies, namely the problem of regulatory fragmentation faced by banks, securities and other institutions when dealing with various financial products (funds, wealth management, insurance, etc.). It uses semantic fusion technology to uniformly parse policies from multiple departments, eliminate differences in terminology and conflicts in clauses, and realize the integrated mapping of compliance requirements across regulatory bodies.

[0249] (3) It has improved the dynamic compliance adaptability, that is, it overcomes the shortcomings of traditional rule engines that rely on static configuration, and uses NLP incremental learning and knowledge graph dynamic updates to adapt to policy revisions in real time, reducing the policy update response time from several weeks to hours of automation.

[0250] (4) Achieve penetrating control across sales scenarios, that is, support compliance cross-verification of mixed product combinations (such as "fund + insurance") sold on consignment, and use semantic coupling to deduce policy superposition constraints (such as suitability and sales suitability dual requirements) in complex transaction scenarios to avoid the risk of compliance of a single product but non-compliance of the combination.

[0251] (5) Reduced the risks of human intervention and operation, namely, the automated closed loop from policy interpretation and behavior monitoring to real-time intervention reduces human interpretation bias and increases the accuracy of compliance inspection from 70%-80% in traditional schemes to over 95%.

[0252] (6) It meets the requirements of penetrating supervision, that is, through fine-grained semantic association (for example, mapping "high-risk customers" to the amount and experience threshold of specific products), the regulatory rules are accurately implemented at the sales terminal.

[0253] In summary, the technical solution of this application transforms the dispersed, dynamic, and polymorphic policy semantics into executable compliance logic, thus solving the pain points of high compliance costs and incomplete coverage for financial institutions when selling diversified products.

[0254] Figure 4 This is a schematic diagram of a risk assessment device provided in an embodiment of this application. Please refer to... Figure 4 The risk assessment device 40 includes: a processing module 41, a determining module 42, a generating module 43, a constructing module 44, and an assessment module 45, wherein...

[0255] Processing module 41 is used to preprocess and identify entities from the policy text to be processed, and to determine multiple financial entities;

[0256] Module 42 is used to determine the relationships between multiple financial entities based on policy texts;

[0257] The generation module 43 is used to generate structured data based on multiple financial entities and the relationships between them.

[0258] Module 44 is used to construct multidimensional risk indicators based on structured data. These multidimensional risk indicators are used to assess the risk level of the financial object to be assessed.

[0259] The assessment module 45 is used to assess the risk level of the financial object based on the multidimensional risk indicators when assessing the financial object to be assessed.

[0260] The risk assessment device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0261] In one possible implementation, the processing module 41 is specifically used for:

[0262] The policy text is preprocessed to obtain multiple text sequences;

[0263] By embedding domain labels into multiple text sequences using a target model, multiple fused text sequences are obtained. These fused text sequences incorporate domain information. The target model can be a model based on a bidirectional encoder-representation converter.

[0264] Entity identification is performed on multiple fused text sequences to identify multiple financial entities.

[0265] In one possible implementation, the processing module 41 is specifically used for:

[0266] Based on a financial terminology dictionary, the policy text is coarsely segmented using Chinese lexical analysis tools to obtain a word sequence, which includes multiple terms from the policy text.

[0267] By using a word segmentation algorithm, long terms in a word sequence are broken down into sub-word sequences.

[0268] By using a sliding window, the word sequence is segmented to obtain multiple text sequences.

[0269] In one possible implementation, the determining module 42 is specifically used for:

[0270] Based on the relation types in a pre-established hierarchical domain ontology library, multiple entity pairs are identified among multiple financial entities through the target model;

[0271] For any pair of entities, the entity relationships between the pairs are determined by combining policy texts and pre-generated financial knowledge graphs through a target model.

[0272] In one possible implementation, the generation module 43 is specifically used for:

[0273] For any given financial entity, a label is generated by labeling the financial entity using a bidirectional long short-term memory network-conditional random field model.

[0274] For any entity relation, the entity relation is labeled using a bidirectional long short-term memory network-conditional random field model to obtain the entity relation label;

[0275] Structured data is generated based on multiple financial entities, tags of multiple financial entities, multiple entity relationships, and tags of entity relationships.

[0276] In one possible implementation, the evaluation module 45 is specifically used for:

[0277] The financial object to be evaluated is assessed based on the multidimensional risk indicators, and multiple initial risk values ​​corresponding to the multidimensional risk indicators are obtained.

[0278] The risk level of the financial object is determined based on the multiple initial risk values ​​corresponding to the multidimensional risk indicators.

[0279] In one possible implementation, the multidimensional risk indicators include compliance strictness, policy cross-constraints, user adaptability risk, and timeliness risk; the assessment module 45 is specifically used for:

[0280] Regarding compliance stringency, the initial risk value of the financial object under compliance stringency is determined based on the type and number of policies involved.

[0281] In response to policy cross-constraints, the initial risk value of the financial object under policy cross-constraints is determined based on the multiple related policies involved in the financial object.

[0282] To address user compatibility risk, the initial risk value of the financial object under user compatibility risk is determined based on the object level of the financial object and the user level of the user.

[0283] To address time-sensitive risks, the initial risk value of the financial entity is determined based on the release and effective dates of the policies related to that entity.

[0284] In one possible implementation, the evaluation module 45 is specifically used for:

[0285] Determine the risk weight corresponding to each risk indicator in the multidimensional risk index;

[0286] The comprehensive risk value of the financial object is determined by multiplying the initial risk value and the risk weight corresponding to each risk indicator.

[0287] The risk level is determined based on the comprehensive risk value.

[0288] The risk assessment device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0289] Figure 5 This is a schematic diagram of another risk assessment device provided in an embodiment of this application. Please refer to... Figure 5 ,exist Figure 4 Based on the illustrated embodiment, the risk assessment device 40 further includes an intervention module 46, which is used for:

[0290] For financial instruments with a risk level of severe risk, the transaction of the financial instrument is blocked in real time, and an alarm message is sent to the target device;

[0291] For financial objects with a risk level of general risk, send correction information to the target device;

[0292] For financial instruments classified as "potential risk," real-time voice prompts are generated to remind users of the potential risks associated with these instruments.

[0293] The risk assessment device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0294] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Please refer to... Figure 6 The electronic device 60 may include a processor 61 and a memory 62. Exemplarily, the processor 61 and the memory 62 are interconnected via a bus 63.

[0295] Memory 62 is used to store program instructions;

[0296] The processor 61 is used to execute the program instructions stored in the memory, so that the electronic device 60 performs the method shown in the above method embodiment.

[0297] The electronic device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be repeated here.

[0298] This application provides a computer-readable storage medium storing computer-executable instructions, which are used to implement the above-described method when executed by a processor.

[0299] This application embodiment may also provide a computer program product, including a computer program that, when executed by a processor, can implement the above-described method.

[0300] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0301] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0302] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0303] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0304] When integrated units / modules are implemented in hardware, the hardware can be digital circuits, analog circuits, etc. The physical implementation of the hardware structure includes, but is not limited to, transistors, memristors, etc. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC, etc. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc.

[0305] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0306] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0307] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0308] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A risk assessment method, characterized in that, include: The policy texts to be processed are preprocessed and entity identification is performed to identify multiple financial entities; Based on the policy text, the relationships between the multiple financial entities are determined; Based on the multiple financial entities and the relationships between them, structured data is generated; Multidimensional risk indicators are constructed based on the structured data; When assessing a financial object, the risk level of the financial object is evaluated based on the multidimensional risk indicators.

2. The method according to claim 1, characterized in that, The policy texts to be processed undergo preprocessing and financial entity identification, identifying multiple financial entities, including: The policy text is preprocessed to obtain multiple text sequences; Domain labels are embedded into the multiple text sequences by a target model to obtain multiple fused text sequences. The fused text sequences incorporate domain information. The target model is a model based on a bidirectional encoder-representation converter. Entity identification is performed on the multiple fused text sequences to determine the multiple financial entities.

3. The method according to claim 2, characterized in that, The policy text is preprocessed to obtain multiple text sequences, including: Based on a financial terminology dictionary, the policy text is coarse-grainedly segmented using Chinese lexical analysis tools to obtain a word sequence, which includes multiple terms from the policy text. The long terms in the word sequence are further broken down using a word segmentation algorithm to obtain a sub-word sequence. The word sequence is segmented using a sliding window to obtain the multiple text sequences.

4. The method according to any one of claims 1-3, characterized in that, Based on the policy text, the relationships between the multiple financial entities are determined, including: Based on the relation types in the pre-established hierarchical domain ontology library, multiple entity pairs are identified among the multiple financial entities through the target model; For any pair of entities, the entity relationships between the pairs are determined by combining the policy text and the pre-generated financial knowledge graph through the target model.

5. The method according to any one of claims 1-3, characterized in that, Based on the multiple financial entities and their relationships, structured data is generated, including: For any given financial entity, the entity is labeled using a bidirectional long short-term memory network-conditional random field model to obtain the entity's label. For any entity relationship, the entity relationship is labeled using the bidirectional long short-term memory network-conditional random field model to obtain the label of the entity relationship; The structured data is generated based on the multiple financial entities, the tags of the multiple financial entities, the multiple entity relationships, and the tags of the entity relationships.

6. The method according to any one of claims 1-3, characterized in that, The risk level of the financial object is determined based on the aforementioned multidimensional risk indicators, including: The financial object to be evaluated is assessed based on the multidimensional risk indicators, and multiple initial risk values ​​corresponding to the multidimensional risk indicators are obtained. The risk level of the financial object is determined based on the multiple initial risk values ​​corresponding to the multidimensional risk indicators.

7. The method according to claim 6, characterized in that, The multidimensional risk indicators include compliance strictness, policy cross-constraints, user adaptability risk, and timeliness risk. Based on the aforementioned multidimensional risk indicators, the financial object to be evaluated is assessed, resulting in multiple initial risk values, including: Regarding compliance stringency, the initial risk value of the financial object under compliance stringency is determined based on the type and number of policies involved. In response to policy cross-constraints, the initial risk value of the financial object under policy cross-constraints is determined based on the multiple related policies involved in the financial object. To address user compatibility risk, the initial risk value of the financial object under user compatibility risk is determined based on the object level of the financial object and the user level of the user. To address time-sensitive risks, the initial risk value of the financial entity is determined based on the release and effective dates of the policies related to that entity.

8. The method according to claim 6, characterized in that, If the financial object is a financial product, the method further includes: For financial objects with a risk level of severe risk, the transaction of the financial object is blocked in real time, and an alarm message is sent to the target device; For financial objects with a risk level of general risk, correction information is sent to the target device; For financial objects with a risk level of potential risk, generate real-time voice prompts to remind users that the financial object has potential risks.

9. A risk assessment device, characterized in that, include: The module includes a processing module, a determining module, a generating module, and a building module. The processing module is used to preprocess and identify entities from the policy text to be processed, and to determine multiple financial entities. The determining module is used to determine multiple entity relationships between the multiple financial entities based on the policy text; The generation module is used to generate structured data based on the plurality of financial entities and the relationships between the plurality of entities; The construction module is used to construct multidimensional risk indicators based on the structured data, and the multidimensional risk indicators are used to assess the risk level of the financial object to be assessed. The assessment module is used to assess the risk level of a financial object based on the multidimensional risk indicators when assessing the financial object to be assessed.

10. An electronic device, characterized in that, include: The processor, and the memory that is in communication with the processor; The memory stores the instructions that the computer executes; The processor executes computer execution instructions stored in memory to implement the method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 8.

12. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 8.

Images