A deep learning-based financial question answering method, system, device, and medium.

Abstract

This invention relates to the field of intelligent question-answering technology, and discloses a deep learning-based financial question-answering method, system, device, and medium. The method includes: acquiring compliance rules for financial services and establishing a compliance rule base based on these rules; acquiring user query content and user personal information, retrieving them using a multi-source financial knowledge base, and generating a preliminary response text; constructing a compliance semantic firewall based on the compliance rule base, and using the compliance semantic firewall to review the preliminary response text to determine whether there is any non-compliant content; if non-compliant content is found in the preliminary response text, adjusting, supplementing, or reconstructing the non-compliant content according to a preset correction strategy to obtain the final response text. This invention avoids the compliance risks and legal liabilities that may arise when providing personalized services, significantly improves the accuracy, reliability, and compliance of financial service responses, and increases service efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology, specifically to a financial question answering method, system, device, and medium based on deep learning. Background Technology

[0002] In the financial services sector, intelligent question-and-answer systems are playing an increasingly important role in assisting customer service. However, with the increasing complexity of financial products, the rapid changes in market information, and increasingly stringent regulatory requirements, traditional question-and-answer systems face numerous challenges in handling customer inquiries. These challenges primarily manifest in the difficulty of accurately understanding the customer's underlying intentions, the inability to effectively integrate heterogeneous information from multiple sources, the difficulty in ensuring the real-time compliance of generated responses, and the difficulty in balancing information value with regulatory red lines when providing personalized guidance. Particularly when comparing and analyzing complex financial products, interpreting real-time market dynamics, or dealing with individual customer risk preferences, question-and-answer systems may provide fragmented, outdated, inaccurate, or even non-compliant information. This not only affects service efficiency and customer experience but may also pose potential compliance risks to financial institutions. Summary of the Invention

[0003] This invention provides a deep learning-based financial question answering method, system, device, and medium to address the problems of compliance risks and low service efficiency in existing financial question answering systems.

[0004] In a first aspect, the present invention provides a deep learning-based financial question-answering method, the method comprising: Obtain compliance rules for financial services and build a compliance rule base based on those rules; Obtain user query content and user personal information, use multi-source financial knowledge base for retrieval, and generate preliminary response text; A compliance semantic firewall is built based on a compliance rule base. The compliance semantic firewall is used to review the preliminary response text and determine whether there is any non-compliant content in the preliminary response text. If the initial response text contains non-compliant content, the non-compliant content will be adjusted, supplemented, or reconstructed according to the preset correction strategy to obtain the final response text.

[0005] The deep learning-based financial question-answering method provided by this invention builds a solid compliance foundation based on compliance rules for financial service scenarios. It comprehensively considers customer query content, customer personal information, and relevant information from multi-source financial knowledge bases to ensure the personalization and relevance of the initial response text. A compliance semantic firewall is introduced to conduct in-depth semantic analysis and review of the initial text, accurately determining the presence of non-compliant content. Non-compliant content is corrected, and a compliant final response text is output. This provides efficient and personalized financial services while strictly adhering to various compliance requirements, effectively avoiding the compliance risks and legal liabilities that may arise when traditional intelligent question answering methods provide personalized services. It significantly improves the accuracy, reliability, and compliance of financial service responses, avoiding potential compliance risks and providing financial institutions and customers with a safer and more professional intelligent question-answering experience.

[0006] In one alternative implementation, compliance rules include: financial regulatory provisions, product disclosure requirements, disclaimer templates, and a list of prohibited terms; Obtain compliance rules for financial services and build a compliance rule base based on these rules, including: In response to regulatory policy updates or expert instructions, update the compliance rule base and generate a new version number and corresponding timestamp; Priority values ​​are assigned to the updated compliance rules based on their nature, issuing organization, and severity.

[0007] The deep learning-based financial question-answering method provided by this invention clarifies the specific composition of compliance rules, providing a solid foundation for subsequent compliance semantic firewall review. It effectively reduces non-compliance risks caused by incomplete information, inappropriate expression, or insufficient risk warnings, making the scope and basis of review based on the compliance rule base clearer, improving the effectiveness, professionalism, comprehensiveness, and accuracy of the review, and ensuring that regulatory requirements are fully covered when handling various financial inquiries.

[0008] In one alternative implementation, the multi-source financial knowledge base includes: a static product database, a preset market data interface, and an unstructured document library; Obtain user query content and user personal information, conduct searches using multi-source financial knowledge bases, and generate preliminary response text, including: User profiles are generated based on users' personal information. These user profiles are used to characterize users' risk tolerance levels and investment portfolio information. Perform natural language understanding on user queries to identify user intent and query entities; Relevant information is retrieved using a multi-source financial knowledge base, and combined with user intent, query entities, and user profiles, a text generation model is used to generate an initial response text.

[0009] The deep learning-based financial question-answering method provided by this invention transforms unstructured customer queries into structured information that can be processed by the system, providing personalized background information for subsequent response generation. It utilizes a multi-source financial knowledge base to retrieve relevant information, obtaining comprehensive and real-time financial information that is highly relevant to the customer's query content and user profile. It then uses a text generation model to generate preliminary response text, presenting complex financial information to the customer in an easy-to-understand manner.

[0010] In one optional implementation, a compliance semantic firewall is used to review the initial response text to determine whether there is any non-compliant content in the initial response text, including: Perform semantic analysis on the initial response text to identify keywords and potentially biased statements; Based on keywords and potentially suggestive statements, the compliance rule base is traversed to determine whether there is any non-compliant content in the initial response text.

[0011] The financial question-answering method based on deep learning provided by this invention performs semantic analysis on the initial response text through a compliance semantic firewall, accurately identifying keywords such as financial product names and market indicators, as well as potentially biased statements. Combined with a compliance rule base, it completes a comprehensive comparison, quickly and accurately determining non-compliant content, significantly improving review efficiency and accuracy. It effectively identifies compliance risks such as missing risk warnings, prohibited words, and statements constituting investment advice, avoiding oversights and misjudgments in manual review, ensuring that financial question-answering responses are compliant, objective, and neutral, reducing compliance and legal risks for financial institutions, and adapting to dynamic requirements of financial regulation, making intelligent question-answering services safer.

[0012] In one alternative implementation, semantic analysis is performed on the initial response text to identify keywords and potentially biased statements, including: Using text analysis tools, we identified financial product names, company names, market indicators, and dates in the initial response text as keywords. Using a pre-defined list of financial bias terms, statements containing bias terms in the initial response text are marked, and the marked statements are compared with the pre-defined bias terms to determine potential bias statements.

[0013] The financial question-answering method based on deep learning provided by this invention accurately extracts core keywords such as financial product names, company names, market indicators, and dates through text analysis tools. It combines a list of financial bias words to mark suspicious statements and locks down potentially biased statements through semantic similarity comparison. This achieves accurate analysis of the deep semantics of the response text, fully captures key information, and effectively identifies implicit investment advice, return promises, and other illegal tendencies. It avoids omissions and misjudgments caused by literal matching alone, and improves the depth and accuracy of compliance review.

[0014] In one optional implementation, based on keywords and potential tendencies, the compliance rule base is traversed to determine whether there is non-compliant content in the preliminary response text, including: Check the keywords in the initial response text to see if there are any risk warnings in the compliance rule base; Examine potentially biased statements to see if they contain prohibited words from the compliance rule base, and determine whether they constitute biased investment advice; If the preliminary response text does not contain any risk warnings or contains biased investment advice, then the preliminary response text is deemed to contain non-compliant content.

[0015] The financial question-answering method based on deep learning provided by this invention forms a dual compliance verification logic by checking risk warnings through keyword verification and prohibited words and investment advice through verification of potentially biased statements. The judgment criteria are clear and the judgment results are rigorous and reliable. It accurately identifies typical non-compliance situations such as missing mandatory risk warnings, use of illegal expressions, and inappropriate investment advice in the answers, which greatly reduces the probability of missed judgments and misjudgments. It ensures that the review fully covers regulatory requirements, triggers non-compliance judgments with clear rules, improves the automation and standardization of compliance review, and effectively prevents and controls financial question-answering compliance risks.

[0016] In one optional implementation, the non-compliant content is adjusted, supplemented, or reconstructed according to a preset correction strategy to obtain the final response text, including: Replace potentially biased statements with objective statements; Obtain the product type and user risk level, and insert the corresponding mandatory risk warning or disclaimer template into the preliminary response text based on the product type and user risk level; Adjust the sentence structure of prohibited words in potentially biased statements while retaining the core information.

[0017] The financial question-answering method based on deep learning provided by this invention accurately corrects non-compliant content by replacing biased statements, inserting corresponding risk warnings according to product and user risk levels, adjusting the structure of prohibited words and retaining core information. It can eliminate subjective investment advice and illegal expressions, ensuring objective and neutral responses, and can also supplement mandatory disclosure content in a targeted manner, fully meeting financial regulatory requirements. While ensuring compliance, it fully preserves information value, improves the standardization and professionalism of responses, effectively reduces institutional compliance risks, and adapts to the dual service needs of personalized and compliant financial scenarios.

[0018] In one alternative implementation, the method further includes: Record the user query content, preliminary response text, review results of the compliance semantic firewall, and final response text to form a compliance audit log; Analyze the compliance audit logs and optimize the parameters of the preset correction strategy based on the analysis results.

[0019] The deep learning-based financial question-answering method provided by this invention forms a complete compliance audit log by recording query content, preliminary answers, review results, and final answers, achieving full traceability and auditability of the question-answering process and meeting financial regulatory audit requirements. Based on the analysis and optimization of audit log parameters, it can continuously identify compliance risk patterns, improve review efficiency and judgment accuracy, and enable the system to adapt to new regulations and new scenario challenges, forming a closed-loop iterative mechanism of review-correction-recording-optimization, ensuring the long-term compliance stability of financial question-answering.

[0020] Secondly, this invention provides a deep learning-based financial question-answering system, the system comprising: The compliance rule base establishment module is used to acquire compliance rules for financial services and establish a compliance rule base based on these rules. The preliminary response generation module is used to obtain user query content and user personal information, and to generate preliminary response text by searching through a multi-source financial knowledge base. The compliance review module is used to build a compliance semantic firewall based on the compliance rule base, and to review the preliminary response text to determine whether there is any non-compliant content in the preliminary response text. The correction and response output module is used to adjust, supplement, or reconstruct non-compliant content in the initial response text according to a preset correction strategy to obtain the final response text.

[0021] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the method described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the first process of a deep learning-based financial question-answering method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a second process for a deep learning-based financial question-answering method according to an embodiment of the present invention; Figure 4 This is a structural block diagram of a deep learning-based financial question-answering device according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0026] As an optional application scenario of this invention, such as Figure 1 As shown, the intelligent question-answering system may include at least one terminal device and at least one server. Figure 1 The system is illustrated in the example, which includes a computer 101, a mobile terminal 102, and a server 103, and the terminal devices such as the computer 101 and the mobile terminal 102 are connected to the server 103 through a network 110.

[0027] Specifically, the terminal device can be a smartphone, tablet, laptop, PDA, desktop computer, game console, smart TV, smart wearable device, in-vehicle terminal, VR (Virtual Reality) device, AR (Augmented Reality) device, etc. Server 103 can be a standalone physical server, a server cluster, a distributed system, or a cloud server providing cloud services. Network 110 can be a wired or wireless network, examples of which include, but are not limited to, the Internet, corporate intranet, local area network, wide area network, mobile communication network, and combinations thereof.

[0028] In customer service scenarios within financial service institutions, existing intelligent question-answering systems struggle to design a framework that can identify and correct any content that might violate the latest financial regulations while integrating multi-source heterogeneous information and generating personalized responses. This framework would also need to strike a precise balance between providing valuable guidance and adhering to regulatory red lines for specific investment advice, thus avoiding inaccurate, inapplicable, or non-compliant information. These questions don't simply require information like historical gold prices or investment attributes; they implicitly reflect the customer's considerations regarding asset preservation and growth, and risk mitigation. Existing systems, primarily relying on textual semantic similarity for information matching, often fail to accurately capture the customer's deeper intent when faced with such open-ended queries with strong subjective biases. This failure to identify implicit intent leads to discrepancies between the system's answers and the customer's actual needs, reducing the effectiveness of the interaction.

[0029] Furthermore, the financial industry is subject to stringent regulatory constraints, particularly regarding the provision of investment advice to clients. Any information issued by an institution, especially concerning investment decisions, must ensure its accuracy, suitability for clients, and strict adherence to disclosure requirements and compliance regulations. When intelligent question-and-answer systems need to synthesize responses from dynamic market data, complex product terms, and regulatory documents, they must not only guarantee the factual accuracy of the information but also ensure that the wording, structure, and included warnings of the responses fully comply with the latest regulatory requirements. Existing systems, in integrating information, may inadvertently generate a statement that is technically correct but could be misinterpreted as inappropriate advice in a regulatory context, or omit crucial risk warnings, thereby exposing institutions to potential compliance risks and legal liabilities.

[0030] This invention provides a deep learning-based financial question-answering method. By constructing a compliance rule base, combining user information with multi-source financial knowledge to generate preliminary answers, using a compliance semantic firewall to accurately determine non-compliant content, and correcting non-compliant content, the method aims to avoid potential compliance risks and improve service efficiency.

[0031] According to an embodiment of the present invention, a deep learning-based financial question answering method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0032] This embodiment provides a deep learning-based financial question-answering method, which can be used in the aforementioned computer system. Figure 2 This is a flowchart of a deep learning-based financial question-answering method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain compliance rules for financial services and establish a compliance rule base based on the compliance rules.

[0033] Specifically, deep learning refers to using multi-layered neural networks to learn from and extract features from data to achieve a higher level of abstraction and understanding. For example, in Natural Language Processing (NLP) tasks, deep learning models can better understand the semantics and context of text. A compliance rule base is a database that stores various financial regulatory requirements to guide intelligent question-answering systems in generating and reviewing compliance responses.

[0034] The first step is to receive compliance rules for financial service scenarios and store them in a compliance rule base. Compliance rules are regulations that financial institutions must follow when providing services. By establishing a comprehensive compliance rule base, a clear basis is provided for subsequent review work.

[0035] Compliance rules can be obtained in various ways. For example, compliance experts can manually input the latest regulatory policies, product disclosure requirements, disclaimer templates, and prohibited vocabulary lists through a manual input interface; or, data interfaces can be integrated to regularly scrape and parse the latest compliance documents from the websites or databases of official regulatory agencies and store them in a structured manner in a compliance rule base.

[0036] Step S202: Obtain the user's query content and personal information, perform a search using a multi-source financial knowledge base, and generate a preliminary response text.

[0037] Specifically, a multi-source financial knowledge base is a collection of knowledge that integrates various types of financial information, providing comprehensive and accurate financial knowledge support for intelligent question-and-answer systems.

[0038] Based on the customer's query, personal information, and relevant information retrieved from multi-source financial knowledge bases, an initial response text is generated. Leveraging the capabilities of deep learning models in natural language understanding and text generation, a personalized and informative initial response is provided to the customer. However, due to the complexity of the financial sector and the stringent regulations, the initially generated response text may contain potential compliance risks.

[0039] Deep learning-based text generation engines can be utilized, such as large-scale language models with the Transformer architecture. These models can be pre-trained on massive amounts of financial text data to enable them to understand customer intent, integrate multi-source information, and generate fluent and relevant responses. In practical applications, the system receives text input from customers, combines it with information such as personal risk preferences and investment history obtained from a customer profiling system, and uses a knowledge retrieval component to obtain relevant factual information from static product databases, real-time market data interfaces, and unstructured document libraries. Finally, a preliminary draft response is generated.

[0040] Step S203: Construct a compliance semantic firewall based on the compliance rule base, and use the compliance semantic firewall to review the preliminary response text to determine whether there is any non-compliant content in the preliminary response text.

[0041] Specifically, the compliance semantic firewall is an intelligent review mechanism that can perform semantic analysis on the initial response text, identify potential compliance risks, and adjust it according to preset correction strategies to ensure that the final output response text meets all relevant compliance requirements.

[0042] A pre-defined compliance semantic firewall performs semantic analysis on the initial response text to identify keywords and potential biases. For example, it identifies key information such as financial product names, company names, and market indicators mentioned in the initial response text and determines whether there are any subjective recommendations or misleading statements. Subsequently, the semantic analysis results are compared with the compliance rule base to determine whether the initial response text violates any compliance regulations. For example, it checks whether necessary risk warnings are omitted or whether prohibited words are used; these are just examples and are not exhaustive.

[0043] A compliance semantic firewall can contain multiple sub-modules, such as a semantic analyzer to identify keywords (e.g., financial product names, company names, market indicators, dates and times) and potential biases (e.g., subjective recommendations, exaggerated returns) in the text, and a rule comparison engine to compare the semantic analysis results with the various provisions in the compliance rule base in real time to determine the preliminary judgment result of whether the response text is compliant.

[0044] Step S204: If there is non-compliant content in the preliminary response text, adjust, supplement or reconstruct the non-compliant content according to the preset correction strategy to obtain the final response text.

[0045] Specifically, when the compliance semantic firewall determines that the preliminary response text contains non-compliant content, it adjusts, supplements, or reconstructs the non-compliant content according to a preset correction strategy to eliminate compliance risks and ensure that the final output response text not only meets the customer's information needs but also fully complies with regulatory requirements. For example, replacing biased statements with objective and neutral statements, inserting mandatory risk warnings, or adjusting the sentence structure to eliminate suggestive semantics are just examples and are not limited to this.

[0046] Utilizing a text editing and rewriting engine that integrates multiple correction algorithms, for example, for statements identified as potentially biased, a statement rewriting model is invoked to replace them with objective and neutral statements; for cases lacking risk warnings, preset mandatory risk warnings or disclaimer templates can be automatically inserted based on product type and customer risk level; for statements containing prohibited words, their structure can be adjusted to remove suggestive semantics while retaining core information. The final response text, after compliance correction, is then output to the user.

[0047] The deep learning-based financial question-answering method provided in this embodiment builds a solid compliance foundation based on compliance rules for financial service scenarios. It comprehensively considers customer query content, customer personal information, and relevant information from multiple sources of financial knowledge bases to ensure the personalization and relevance of the initial response text. A compliance semantic firewall is introduced to conduct in-depth semantic analysis and review of the initial text, accurately determining whether there is any non-compliant content. Non-compliant content is corrected, and a compliant final response text is output. This provides efficient and personalized financial services while strictly adhering to various compliance requirements, effectively avoiding the compliance risks and legal liabilities that may arise when traditional intelligent question answering methods provide personalized services. It significantly improves the accuracy, reliability, and compliance of financial service responses, avoiding potential compliance risks and providing financial institutions and customers with a safer and more professional intelligent question-answering experience.

[0048] This embodiment provides a deep learning-based financial question-answering method, which can be used in the aforementioned computer system. Figure 3 This is a flowchart of a deep learning-based financial question-answering method according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps: Step S301: Obtain compliance rules for financial services and establish a compliance rule base based on the compliance rules.

[0049] Specifically, compliance rules include: financial regulatory provisions, product disclosure requirements, disclaimer templates, and a list of prohibited terms.

[0050] Financial regulatory provisions refer to the various laws, regulations, rules, and normative documents issued and enforced by financial regulatory agencies. Their purpose is to regulate financial market behavior, protect investor rights, and maintain financial stability. For example, financial regulatory provisions may include specific clauses regarding product sales qualifications, information disclosure obligations, and risk warning requirements.

[0051] Product disclosure requirements refer to the specific regulations that financial institutions must clearly, accurately, and completely inform customers of key information such as product characteristics, risks, fees, and expected returns when providing financial products or services. The purpose is to ensure that customers make investment decisions based on a full understanding of the information. Specifically, product disclosure requirements may include mandatory disclosures in product prospectuses and format requirements for risk disclosure statements.

[0052] A disclaimer template is a standardized text format used in financial services to clarify the scope of a service provider's responsibilities and limit the assumption of specific risks. Its purpose is to reasonably delineate the rights, obligations, and risk responsibilities between the service provider and the client within the scope permitted by law. For example, when providing investment advice or market analysis, a disclaimer such as "This information is for reference only and does not constitute investment advice" is usually required.

[0053] A prohibited vocabulary list refers to a collection of words, phrases, or expressions that are explicitly prohibited or require caution in financial services scenarios. Its purpose is to prevent misleading advertising, excessive promises, or inappropriate inducements, ensuring the objectivity and neutrality of information dissemination. For example, a prohibited vocabulary list may include words with strong biases or misleading connotations, such as guaranteed returns, no risk, or sure-fire profits.

[0054] When the initial response text is input into a pre-defined compliance semantic firewall for review, the semantic analysis results can be accurately compared with these specific types of compliance rules. For example, by comparing it with financial regulatory provisions, it can be verified whether the initial response text meets the mandatory requirements for information disclosure; by comparing it with product disclosure requirements, it can be ensured that all necessary product information has been provided; by comparing it with disclaimer templates, it can be checked whether the necessary risk warnings have been correctly inserted; and by comparing it with a list of prohibited terms, it can effectively identify and avoid potentially misleading or non-compliant expressions.

[0055] Step S301 above includes: Step S3011: In response to regulatory policy updates or expert instructions, update the compliance rule base and generate a new version number and corresponding timestamp.

[0056] Specifically, in real-world financial service scenarios, compliance rules are not static; regulatory policies may be frequently updated, and multiple rules may apply simultaneously with varying priorities. Simply receiving and storing rules can lead to difficulties in maintaining and updating the rule base, or an inability to effectively adjudicate rule conflicts, thereby affecting the compliance and accuracy of intelligent question answering.

[0057] Compliance rules are stored in a predefined data format to form a compliance rule base. For example, each compliance rule can be broken down into multiple fields, including but not limited to: rule identifier (ID), rule text, applicable product type, applicable customer group, effective date, expiration date, risk level, and corresponding regulatory agency. This structured storage method facilitates subsequent automated retrieval, comparison, and processing, thereby improving the manageability and operability of the rule base.

[0058] In response to regulatory policy updates or expert directives, the compliance rule base is updated, generating new version numbers and timestamps. This can be understood as the compliance rule base being promptly modified, added to, or deleted when new regulatory policies are issued, existing policies are revised, or financial compliance experts propose new compliance requirements. Each update operation is accompanied by the generation of a new version number and timestamp, the purpose of which is to provide the rule base with historical version traceability capabilities, ensuring that the rule change process is auditable and that the rule state can be traced back to a specific point in time, thereby guaranteeing the transparency and accuracy of compliance reviews.

[0059] Step S3012: Assign priority values ​​to the updated compliance rules based on their nature, issuing organization, and severity.

[0060] Specifically, priority values ​​are assigned to multiple compliance rules that may be applicable simultaneously, with the rule with the higher priority value being executed first. This means that in the compliance rule base, a priority value is pre-set for multiple compliance rules that may simultaneously apply to a specific financial service scenario or customer inquiry. This priority value can be set based on the nature of the rule (e.g., the mandatory nature of laws and regulations is higher than industry self-regulation), the authority of the issuing institution, the severity of the rule, or expert experience. When reviewing the initial response text, if multiple compliance rules are identified as applicable, the rule with the higher priority value will be prioritized and executed. The purpose is to resolve rule conflicts and ensure that, in a complex and ever-changing compliance environment, judgments and corrections are always made based on the most important or strictest rule.

[0061] For example, suppose a financial institution's intelligent question-and-answer system needs to handle customer inquiries about the risks of fund products. First, the system receives and stores a series of structured compliance rules. For instance, one rule might stipulate that "all responses involving high-risk products must include a disclaimer stating 'Investment involves risk; proceed with caution'," and stores this as a record with specific fields (e.g., rule type: disclaimer, applicable product: high-risk fund, content: 'Investment involves risk; proceed with caution'). When a regulatory agency issues new regulations requiring additional risk warnings for specific types of funds, the system responds by adding the new risk warning rule to the compliance rule base, automatically generating a new version number and timestamp (e.g., updating from version 1.0 to version 1.1), and recording the update time. If two risk warning rules simultaneously apply to a particular fund product—one a general risk warning and the other a mandatory warning for a specific high-risk product—the system will prioritize the mandatory warning rule (e.g., mandatory warnings have higher priority than general warnings) based on a preset priority value, ensuring that the final response text contains the most stringent risk disclosure.

[0062] The deep learning-based financial question-answering method provided in this embodiment clarifies the specific composition of compliance rules, providing a solid foundation for subsequent compliance semantic firewall reviews. It effectively reduces non-compliance risks caused by incomplete information, inappropriate expression, or insufficient risk warnings, making the scope and basis of reviews based on the compliance rule base clearer. The dynamic updating of the compliance rule base enhances the dynamic management capabilities and execution efficiency of the rule base, improving the effectiveness, professionalism, comprehensiveness, and accuracy of the review, and ensuring that regulatory requirements are fully covered when handling various financial inquiries.

[0063] Step S302: Obtain the user's query content and personal information, perform a search using a multi-source financial knowledge base, and generate a preliminary response text.

[0064] Specifically, the multi-source financial knowledge base includes: a static product database, a pre-defined market data interface, and an unstructured document library. Step S302 above includes: Step S3021: Generate a user profile based on the user's personal information. The user profile is used to characterize the user's risk tolerance level and investment portfolio information.

[0065] Specifically, user personal information can include a customer's age, occupation, income level, investment experience, risk preference, etc. By analyzing user personal information, the user's risk tolerance level can be assessed, for example, categorized as conservative, moderate, balanced, growth-oriented, or aggressive. Simultaneously, information about the user's current investment portfolio can be obtained, such as the types, proportions, and historical performance of assets held, including stocks, funds, and bonds. The generated user profile is a comprehensive description of customer characteristics, designed to provide personalized background information for generating subsequent responses.

[0066] Step S3022: Perform natural language understanding on the user's query content to identify the user's intent and the query entity.

[0067] Specifically, natural language understanding of customer queries refers to analyzing the text input by customers using natural language processing technology to accurately grasp their true intentions and the key entities they mention. For example, when a customer asks, "I want to know about the recent stock market trends, especially the performance of technology stocks," the natural language understanding module will be configured to recognize that the user's intention is to understand stock market trends, and the query entity is technology stocks. Its purpose is to transform unstructured customer queries into structured information that the system can process.

[0068] Step S3023: Retrieve relevant information using a multi-source financial knowledge base, and combine user intent, query entity, and user profile to generate a preliminary response text using a text generation model.

[0069] Specifically, the static product database can store basic information about various financial products, such as product name, code, yield, risk level, and issuing institution. The pre-defined market data interface can obtain real-time market data for stocks, funds, bonds, and other financial products, as well as macroeconomic data. The unstructured document library can contain textual materials such as financial product prospectuses, research reports, regulatory documents, and news articles. By searching these multi-source financial knowledge bases, comprehensive and real-time financial information highly relevant to customer queries and user profiles can be obtained.

[0070] Unstructured document libraries can include financial product brochures, research reports, and regulatory documents. An unstructured document library can be understood as a collection of unstructured text data; its content lacks a predefined, rigorous data model, but it is crucial for intelligent question answering in financial service scenarios. Financial product brochures are documents issued by financial institutions that detail the characteristics, risks, returns, fees, and purchase conditions of various financial products (such as funds, wealth management products, and insurance). These brochures typically contain a large amount of technical terminology and compliance requirements, serving as important references for customers when searching for product information. Research reports are usually written by financial analysts, economists, or research institutions, providing in-depth analysis and forecasts of specific industries, companies, market trends, or macroeconomic conditions. These reports provide customers with market insights and investment references, but their content may contain certain biases or predictions, requiring processing in conjunction with compliance review. Regulatory documents refer to relevant laws, regulations, rules, guidelines, notices, and announcements issued by financial regulatory agencies (such as the People's Bank of China, the State Financial Supervision and Administration Bureau, and the China Securities Regulatory Commission). These documents are the fundamental basis for the compliance of financial services, ensuring that the intelligent question answering system strictly adheres to all regulations when providing services, avoiding the generation of non-compliant content.

[0071] Finally, by combining user intent with query entities, user profiles, and relevant retrieved information, an initial response text is generated using a text generation model. This text generation model can be a deep learning-based language model, such as Transformer or GPT. Upon receiving structured user intent, query entities, personalized user profiles, and relevant information retrieved from a knowledge base, this model can generate a preliminary, coherent, and information-rich response text, presenting complex financial information to the customer in an easily understandable way.

[0072] The deep learning-based financial question-answering method provided in this embodiment transforms unstructured customer queries into structured information that can be processed by the system, providing personalized background information for subsequent response generation. It utilizes a multi-source financial knowledge base to retrieve relevant information, obtaining comprehensive and real-time financial information that is highly relevant to the customer's query content and user profile. It then uses a text generation model to generate preliminary response text, presenting complex financial information to the customer in an easy-to-understand manner.

[0073] Step S303: Construct a compliance semantic firewall based on the compliance rule base, and use the compliance semantic firewall to review the preliminary response text to determine whether there is any non-compliant content in the preliminary response text.

[0074] Specifically, step S303 includes: Step S3031: Perform semantic analysis on the preliminary response text to identify keywords and potentially biased statements.

[0075] In an optional implementation, step S3031 includes: Step a1: Use text analysis tools to identify financial product names, company names, market indicators, and dates in the initial response text as keywords.

[0076] Specifically, identifying financial product names, company names, market indicators, and dates as keywords in the initial response text using text analysis tools involves deep analysis of the initial response text using natural language processing technologies, such as Named Entity Recognition (NER) models. This text analysis tool can be configured to identify and extract entity information highly relevant to the financial sector from the text, such as specific fund product names, the names of companies issuing or managing the product, important market indices, and specific dates or time periods. This identified entity information is considered keywords and serves as crucial evidence for rule comparisons and risk assessments during subsequent compliance reviews.

[0077] Step a2: Using a pre-set list of financial bias terms, mark statements containing bias terms in the preliminary response text, compare the semantic similarity of the marked statements with the pre-set bias statements, and determine potential bias statements based on the comparison results.

[0078] Specifically, the list of biased terms can include a series of words that may imply guaranteed returns, risk aversion, or excessive optimism in financial marketing or consulting, such as "guaranteed profits," "high returns," "risk-free," and "absolute safety." When these biased terms appear in the initial response text, the corresponding statements are marked. Subsequently, these marked statements are compared semantically with a set of pre-set, objective, and neutral standard biased statements. For example, if the initial response text states that the product has stable and high returns, while the standard statement says that the product has a good historical performance but investment carries risks, the semantic similarity comparison can determine that the statements in the initial response text have a potential investment advice bias or an inappropriate promise bias, accurately identifying non-objective and non-neutral statements that may violate compliance requirements.

[0079] The deep learning-based financial question-answering method provided in this embodiment accurately extracts core keywords such as financial product names, company names, market indicators, and dates through text analysis tools. It also marks suspicious statements by combining a list of financial biased terms and locks down potentially biased statements through semantic similarity comparison. This achieves accurate analysis of the deep semantics of the response text, fully captures key information, and effectively identifies implicit investment advice, return promises, and other violations. It avoids omissions and misjudgments caused by literal matching alone, and improves the depth and accuracy of compliance review.

[0080] Step S3032: Based on keywords and potentially biased statements, traverse the compliance rule base to determine whether there is any non-compliant content in the preliminary response text.

[0081] In some optional implementations, step S3032 above includes: Step b1: Check the keywords in the preliminary response text to see if there are any risk warnings in the compliance rule base.

[0082] Specifically, based on the identified financial product names, company names, market indicators, and dates in the preliminary response text, verifying whether the preliminary response text mentions the risk warnings stipulated in the compliance rule base refers to comparing the key information such as financial product names, company names, market indicators, and dates identified by text analysis tools during the semantic analysis phase with the pre-set risk warning requirements for these specific financial entities or market situations in the compliance rule base. This ensures that the preliminary response text fully and accurately discloses the corresponding risks when mentioning specific financial products or market information, avoiding compliance risks caused by incomplete information disclosure. For example, when the preliminary response text mentions a stock or fund, it checks whether there are specific risk warning requirements for that stock or fund in the compliance rule base and verifies whether the preliminary response text already includes these risk warnings.

[0083] Step b2: Check for potentially biased statements, whether there are prohibited words in the compliance rule base, and determine whether they constitute biased investment advice.

[0084] Specifically, statements with potential biases identified during the semantic analysis phase are analyzed in depth. These statements are compared with a pre-defined list of prohibited words in the compliance rule base to identify any explicitly forbidden words or phrases. Simultaneously, a comprehensive assessment of the statement's context, tone, and expression is conducted to evaluate whether it conveys a certain investment inclination or advice to the client, whether this advice is explicitly stated or implicitly implied. This prevents the intelligent question-answering system from unintentionally or intentionally providing investment advice in its responses, thereby circumventing relevant legal and regulatory responsibilities. For example, statements containing phrases such as "strong recommendation," "guaranteed returns," or "sure-fire profits," or using suggestive language to guide clients in making specific investment decisions, may be deemed to constitute investment advice.

[0085] Step b3: If the preliminary response text does not contain any risk warnings or contains biased investment advice, then the preliminary response text is deemed to contain non-compliant content.

[0086] Specifically, if the preliminary response text contains direct or indirect investment advice but fails to mention any risk warnings, it is deemed non-compliant. Only when the preliminary response text simultaneously meets both conditions—containing direct or indirect investment advice and failing to mention any risk warnings as stipulated in the compliance rule base—will it be ultimately deemed non-compliant. This logical judgment mechanism aims to ensure the rigor and accuracy of compliance reviews, avoid misjudgments, and emphasize the importance of risk warnings when providing investment-related information.

[0087] By identifying key information such as financial product names, company names, market indicators, and dates in the initial response text and precisely comparing it with the pre-set risk warning requirements in the compliance rule base, the system ensures that necessary risk disclosures are fully covered when specific financial entities or market situations are involved. Secondly, for statements with potential bias identified by semantic analysis, the system not only checks for prohibited words but also determines whether they constitute direct or indirect investment advice, thus avoiding inappropriate guidance at the deeper semantic level of language expression. Finally, by setting the presence of investment advice without mentioning risk warnings as the final condition for non-compliance, a rigorous logical chain is formed. This ensures that compliance review is no longer a single-dimensional word matching process but comprehensively considers the completeness of information disclosure and the compliance of suggestive expressions, thereby significantly improving the accuracy and reliability of the review.

[0088] The deep learning-based financial question-answering method provided in this embodiment forms a dual compliance verification logic by checking risk warnings through keywords and checking prohibited words and investment advice through potentially biased statements. The judgment criteria are clear and the judgment results are rigorous and reliable. It accurately identifies typical non-compliance situations such as missing mandatory risk warnings, using illegal expressions, and constituting inappropriate investment advice in the answers, which greatly reduces the probability of missed judgments and misjudgments. It ensures that the review fully covers regulatory requirements, triggers non-compliance judgments with clear rules, improves the automation and standardization of compliance review, and effectively prevents and controls financial question-answering compliance risks.

[0089] The deep learning-based financial question-answering method provided in this embodiment performs semantic analysis on the initial response text through a compliance semantic firewall. It accurately identifies keywords such as financial product names and market indicators, as well as potentially biased statements. Combined with a compliance rule base, it completes a comprehensive comparison, quickly and accurately determining non-compliant content. This significantly improves review efficiency and accuracy, effectively identifying compliance risks such as missing risk warnings, prohibited words, and statements constituting investment advice. It avoids oversights and misjudgments in manual review, ensuring that financial question-answering responses are compliant, objective, and neutral, reducing compliance and legal risks for financial institutions. At the same time, it adapts to dynamic requirements of financial regulation, making intelligent question-answering services more secure.

[0090] Step S304: If there is non-compliant content in the preliminary response text, adjust, supplement or reconstruct the non-compliant content according to the preset correction strategy to obtain the final response text.

[0091] In practice, non-compliant content comes in various forms, including potential biases, missing risk warnings, or the inclusion of prohibited words. Simply making general adjustments may not completely eliminate the risks of non-compliance, or excessive modifications during the correction process may result in the loss of core information, thus affecting the accuracy of responses and user experience. It is also impossible to handle different types of non-compliance issues in a more refined and systematic way.

[0092] Specifically, step S304 includes: Step S3041: Replace the potentially biased statement with an objective statement.

[0093] Specifically, when the pre-defined compliance semantic firewall identifies potentially biased statements in the initial response text—for example, statements that might imply certainty of investment returns or overemphasize product advantages while ignoring risks—these statements need to be replaced with objective and neutral statements. This replacement aims to eliminate subjective and misleading expressions, ensuring the fairness of information transmission. Objective and neutral statements are typically based on facts, data, or generally accepted market principles, avoiding the use of highly emotional or predictive language.

[0094] Step S3042: Obtain the product type and user risk level, and insert the corresponding mandatory risk warning or disclaimer template into the preliminary response text according to the product type and user risk level.

[0095] Specifically, to ensure the comprehensive compliance of financial service responses, the system automatically inserts corresponding mandatory risk warnings or disclaimer templates at appropriate locations in the initial response text, based on the type of product involved and the customer's risk tolerance level. For example, for high-risk financial products, it may be necessary to insert warnings about market volatility and the risk of principal loss; for specific customer groups, it may be necessary to insert investment warnings that match their risk tolerance. The insertion positions of these templates are usually preset to ensure their prominence and readability in the text.

[0096] Step S3043: Adjust the sentence structure of prohibited words in the potential tendency statement while retaining the core information.

[0097] Specifically, when the initial response text contains prohibited words, the remediation strategy is not to simply delete these words, but to adjust the sentence structure to eliminate their suggestive meaning while preserving the core information of the statement as much as possible. For example, if a word is considered to have the nature of investment advice in a specific context, the statement will be restructured to express it as a statement of fact or information provision, rather than a direct recommendation or guarantee, thus complying with compliance requirements without affecting the client's access to necessary information.

[0098] For example, suppose a customer inquires about the value of a high-yield investment product. The initial response might contain a statement like, "This product has an excellent track record with an expected return of X%, making it an ideal choice for asset appreciation." First, a pre-defined compliance semantic firewall will recognize that the statement "an expected return of X% makes it an ideal choice for asset appreciation" carries a potential bias, as it may imply certainty of returns and investment advice. According to the correction strategy, this statement will be replaced with an objective and neutral statement, such as, "This product has a good track record, and its expected return may reach X% depending on market conditions. However, investment involves risk; please evaluate carefully." Second, if the investment product is classified as medium-to-high risk and the customer's risk tolerance is conservative, a mandatory risk warning template will be inserted at the end of the initial response or in relevant paragraphs, such as, "This product is a non-principal-protected floating-rate product. Past performance does not guarantee future results. Investors should fully understand the product's risk characteristics and invest prudently according to their own risk tolerance." Finally, if the initial response inadvertently uses prohibited terms such as "guaranteed profit," the system will not delete it directly but will adjust the sentence structure. For example, "Investing in this product is a sure thing" could be changed to "This product aims to reduce risk through diversified asset allocation, but market volatility risk still exists." In this way, non-compliant suggestive semantics are eliminated while retaining the core message that the product aims to reduce risk.

[0099] With the above revisions, the final response text will be compliant, objective, and complete in information.

[0100] The deep learning-based financial question-answering method provided in this embodiment accurately corrects non-compliant content by replacing biased statements, inserting corresponding risk warnings according to product and user risk levels, adjusting the structure of prohibited words and retaining core information. It can eliminate subjective investment advice and illegal expressions, ensuring objective and neutral responses, and can also supplement mandatory disclosure content in a targeted manner, fully meeting financial regulatory requirements. While ensuring compliance, it fully preserves information value, improves the standardization and professionalism of responses, effectively reduces institutional compliance risks, and adapts to the dual service needs of personalized and compliant financial scenarios.

[0101] In practical applications, simply outputting the final response text may not provide a sufficient mechanism for continuous monitoring, auditing, and optimization of the compliance performance of the entire Q&A process. Without recording and analyzing historical interaction data and review results, the system will struggle to identify potential compliance risk patterns and will be unable to effectively iterate and improve its correction strategies. This could lead to stagnation in the efficiency and accuracy of compliance reviews, or even inadequacy in the face of new compliance challenges.

[0102] Therefore, follow-up processing is needed after the final response text is output. By recording key data and analyzing it, the compliance review and correction strategies can be continuously optimized.

[0103] In some alternative implementations, after the step of outputting the final response text, the method further includes: Step S305: Record the customer query content, preliminary response text, review results of the preset compliance semantic firewall, and final response text to form a compliance audit log.

[0104] Specifically, the system records the customer query content, the initial response text, the review results of the preset compliance semantic firewall, and the final response text, providing a comprehensive data foundation for subsequent system performance evaluation and compliance audit. The customer query content is the user's original input, the initial response text is the system's original output before compliance review, the review results of the preset compliance semantic firewall record in detail the non-compliant content identified during the review process and the basis for the judgment, and the final response text is the final information output to the user after correction. This information is integrated into a compliance audit log, which serves as a complete record of the system's operational history and can be used to trace the compliance processing of each query and response.

[0105] Step S306: Based on the analysis of the compliance audit logs, optimize the parameters of the preset correction strategy.

[0106] Specifically, based on the analysis of the aforementioned compliance audit logs, the parameters of the remediation strategy can be optimized. The compliance audit logs contain a large amount of historical data. By performing statistical analysis, pattern recognition, or machine learning on this data, the performance of the remediation strategy in different scenarios can be identified. For example, it can be identified which types of initial response text are frequently marked as non-compliant, which remediation strategies perform poorly on specific types of non-compliant content, or which remediation efforts fail to completely eliminate non-compliance risks. Through these analytical results, the internal parameters of the remediation strategy can be adjusted in a targeted manner, such as adjusting keyword weights, semantic similarity thresholds, risk warning trigger conditions, or the priority of reconstruction rules, to improve the accuracy and efficiency of the remediation strategy.

[0107] In one specific implementation, suppose a customer inquires about high-yield financial products. The intelligent question-answering system generates an initial response text, which may contain suggestive phrases such as "guaranteed high returns" and "risk-free." During review, a pre-defined compliance semantic firewall identifies these phrases as conflicting with prohibited words in the compliance rule base and determines that the initial response text contains non-compliant content. Subsequently, a correction strategy is triggered, replacing "guaranteed high returns" with "expected returns are high, but investment risks exist," and inserting a disclaimer, ultimately outputting a compliant response text. The customer's query, the initial response text, the firewall's review result, and the final response text are all recorded in the compliance audit log. Over time, the audit log accumulates numerous similar cases. Through analysis of these logs, the system discovers that although the correction strategy can replace obvious prohibited words, in some cases, the initial response text may still contain some subtle and misleading expressions. Although these expressions are not directly marked as prohibited words, their semantic bias may still lead to non-compliance. For example, if the audit logs show that in Q&As involving high-yield products, the corrected text is still frequently misunderstood by users as low-risk, this indicates that the current correction strategy may be insufficiently configured in terms of the strength or placement parameters of the "risk warning." Based on this analysis, the system can automatically or semi-automatically adjust the parameters of the correction strategy, for example, by increasing the strength of risk warnings for statements related to high-yield products, or by mandating the insertion of a more prominent risk warning template in all responses involving high-risk products.

[0108] By recording query content, preliminary responses, review results, and final responses to form a complete compliance audit log, the entire Q&A process is made traceable and auditable, meeting the audit requirements of financial regulators. Based on the analysis and optimization of audit log parameters, the system can continuously identify compliance risk patterns, improve review efficiency and judgment accuracy, and adapt to new regulations and new scenario challenges. This forms a closed-loop iterative mechanism of review-correction-recording-optimization, ensuring the long-term compliance stability of financial Q&A.

[0109] This embodiment also provides a deep learning-based financial question-answering system, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the systems described in the following embodiments are preferably implemented in software, hardware implementations, or a combination of software and hardware, are also possible and contemplated.

[0110] This embodiment provides a deep learning-based financial question-answering system, such as... Figure 4 As shown, it includes: The compliance rule base establishment module 401 is used to obtain compliance rules for financial services and establish a compliance rule base based on these rules.

[0111] The preliminary response generation module 402 is used to obtain the user's query content and personal information, and to generate a preliminary response text by searching through a multi-source financial knowledge base.

[0112] The compliance review module 403 is used to build a compliance semantic firewall based on the compliance rule base, and to review the preliminary response text to determine whether there is any non-compliant content in the preliminary response text.

[0113] The correction and response output module 404 is used to adjust, supplement, or reconstruct non-compliant content in the preliminary response text according to a preset correction strategy to obtain the final response text.

[0114] The deep learning-based financial question-answering system provided in this embodiment of the invention can execute the deep learning-based financial question-answering method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution. Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0115] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0116] The following is a detailed reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 501, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 502 or a program loaded from memory 508 into random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device. The processor 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0117] Typically, the following devices can be connected to I / O interface 505: input devices 506 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 508 including, for example, magnetic tapes, hard disks, etc.; and communication devices 509. Communication device 509 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0118] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 509, or installed from a memory 508, or installed from a ROM 502. When the computer program is executed by the processor 501, it performs the functions defined in the deep learning-based financial question-answering method of the embodiments of the present invention.

[0119] Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0120] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the deep learning-based financial question-answering method shown in the above embodiments is implemented.

[0121] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A deep learning-based financial question-answering method, characterized in that, The method includes: Obtain compliance rules for financial services and establish a compliance rule base based on those rules; Obtain user query content and user personal information, use multi-source financial knowledge base for retrieval, and generate preliminary response text; A compliance semantic firewall is constructed based on the aforementioned compliance rule base. The compliance semantic firewall is then used to review the preliminary response text and determine whether there is any non-compliant content in the preliminary response text. If the preliminary response text contains non-compliant content, the non-compliant content will be adjusted, supplemented, or reconstructed according to a preset correction strategy to obtain the final response text.

2. The method according to claim 1, characterized in that, The compliance rules include: financial regulatory provisions, product disclosure requirements, disclaimer templates, and a list of prohibited terms; The acquisition of compliance rules for financial services and the establishment of a compliance rule base based on the compliance rules include: In response to regulatory policy updates or expert instructions, update the compliance rule base and generate a new version number and corresponding timestamp; Priority values ​​are assigned to the updated compliance rules based on their nature, issuing organization, and severity.

3. The method according to claim 1, characterized in that, The multi-source financial knowledge base includes: a static product database, a preset market data interface, and an unstructured document library; The process of obtaining user query content and user personal information, retrieving data using a multi-source financial knowledge base, and generating a preliminary response text includes: A user profile is generated based on the user's personal information, and the user profile is used to characterize the user's risk tolerance level and investment portfolio information; Natural language understanding is performed on the user's query content to identify the user's intent and query entities; Relevant information is retrieved using the multi-source financial knowledge base, and combined with user intent, query entities, and user profiles, a preliminary response text is generated using a text generation model.

4. The method according to claim 1, characterized in that, The step of reviewing the preliminary response text using the compliance semantic firewall to determine whether there is any non-compliant content in the preliminary response text includes: Semantic analysis was performed on the preliminary response text to identify keywords and potentially biased statements; Based on the keywords and potential biased statements, the compliance rule base is traversed to determine whether there is any non-compliant content in the preliminary response text.

5. The method according to claim 4, characterized in that, Semantic analysis was performed on the preliminary response text to identify keywords and potentially biased statements, including: Using text analysis tools, financial product names, company names, market indicators, and dates in the preliminary response text were identified as keywords. Using a pre-defined list of financial bias terms, statements containing bias terms in the preliminary response text are marked, and the marked statements are compared with the pre-defined bias terms in terms of semantic similarity. Based on the comparison results, potential bias terms are determined.

6. The method according to claim 4, characterized in that, Based on the keywords and potential tendencies, the compliance rule base is traversed to determine whether there is any non-compliant content in the preliminary response text, including: Check the keywords in the preliminary response text to see if there are any risk warnings in the compliance rule base; Examine the potentially biased statements to see if they contain prohibited words in the compliance rule base, and determine whether they constitute biased investment advice; If the preliminary response text does not contain any risk warnings or contains biased investment advice, then the preliminary response text is deemed to contain non-compliant content.

7. The method according to claim 6, characterized in that, The process of adjusting, supplementing, or reconstructing non-compliant content according to a preset correction strategy to obtain the final response text includes: Replace the potentially biased statements with objective statements; Obtain the product type and user risk level, and insert the corresponding mandatory risk warning or disclaimer template into the preliminary response text based on the product type and user risk level; Adjust the sentence structure of prohibited words in the potentially biased statements while retaining the core information.

8. The method according to claim 1, characterized in that, The method further includes: Record the user query content, preliminary response text, review results of the compliance semantic firewall, and final response text to form a compliance audit log; The compliance audit logs are analyzed, and the parameters of the preset correction strategy are optimized based on the analysis results.

9. A deep learning-based financial question-answering system, characterized in that, The system includes: The compliance rule base establishment module is used to acquire compliance rules for financial services and establish a compliance rule base based on the compliance rules. The preliminary response generation module is used to obtain user query content and user personal information, and to generate preliminary response text by searching through a multi-source financial knowledge base. The compliance review module is used to build a compliance semantic firewall based on the compliance rule base, and to use the compliance semantic firewall to review the preliminary response text to determine whether there is any non-compliant content in the preliminary response text. The correction and response output module is used to adjust, supplement, or reconstruct non-compliant content in the preliminary response text according to a preset correction strategy to obtain the final response text.

10. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 8.

Images