Financial Product Knowledge Graph Construction and Intelligent Question Answering Customer Service System

Abstract

This invention relates to the field of intelligent question answering, specifically to a knowledge graph construction and intelligent question answering customer service system for financial products. The system includes: a text preprocessing module for splitting rule text, identifying core elements, and unifying cross-sentence entity attribute identifiers; an entity attribute extraction module for constructing entity hierarchies and extracting attribute constraints; an association condition parsing module for parsing condition result mappings, connecting multi-level associations, and splitting branch scenarios; an operation rule solidification module for converting operation formulas and constructing a constraint rule library; a graph fusion and verification module for integrating multi-dimensional knowledge and verifying logical consistency; and a graph update and optimization module for incrementally updating the graph and optimizing the extraction algorithm. This application significantly improves the quality of knowledge graph construction and the efficiency of intelligent question answering through the collaborative operation of multiple modules.

Description

Technical Field

[0001] This invention belongs to the field of intelligent question answering, specifically involving the construction of knowledge graphs for financial products and intelligent question answering customer service systems. Background Technology

[0002] The accurate response of intelligent financial Q&A customer service relies on the effective support of financial product knowledge graphs, among which rule-based knowledge is a core component that directly determines the Q&A system's ability to handle complex inquiries. Financial product rules are not linear text information, but rather a multi-level composite system including entities, relationships, attributes, and rules. There are clear hierarchical dependencies between entities, requiring intermediate carriers to form multi-level chain relationships. This is accompanied by multi-branch conditional mapping relationships and explicit mathematical operational logic. Some complex rules also involve dual constraints and multi-dimensional floating factors, and are often scattered across multiple sentences of text, exhibiting significant nested and logical characteristics.

[0003] Current mainstream knowledge extraction techniques are represented by Transformer-based sequence labeling models. Their core design focuses on entity recognition and simple relationship determination in linear text, relying on word-by-word labeling or pairwise entity relationship determination. These models have limitations in their context windows, making it difficult to handle rule information scattered across sentences, lacking the ability to chain multi-level nested associations, failing to effectively parse the binding mapping between conditions and results, and struggling to preserve numerical priorities and complex computational chains in the computational logic.

[0004] This technological characteristic creates a significant mismatch between the technology and the complexity of financial product rules, leading to fragmented, erroneous, or incomplete rule-based knowledge during the knowledge graph construction process. Isolated entity storage, confusing relational logic, and omitted conditional branches prevent intelligent question-answering systems from accurately interpreting customer inquiries involving multi-step reasoning. They struggle to perform compliant and accurate numerical calculations and provide logically coherent answers, resulting in customer complaints and compliance risks. This highlights the shortcomings of traditional extraction technologies in financial scenarios, necessitating targeted technological optimization to address this issue. Summary of the Invention

[0005] The purpose of this invention is to provide a financial product knowledge graph construction and intelligent question-and-answer customer service system to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Financial product knowledge graph construction and intelligent question-and-answer customer service system, including: The text preprocessing module is used to split the rule text, identify core elements, and unify cross-sentence entity attribute identification. The input includes the original rule text, and the output includes structured rules and core element identification results. The entity attribute extraction module is used to construct entity hierarchy and extract attribute constraints. The input includes structured rules and financial domain ontology, and the output includes entity hierarchy tree and "entity, attribute, constraint" three-dimensional mapping table. The association condition parsing module is used to parse condition result mapping, connect multi-level associations, and split branch scenarios. The input includes entity hierarchy tree and core element identification results, and the output includes conditions, result mapping diagram, multi-level association path, and branch logic structure diagram. The computation rule solidification module is used to convert computation formulas and build a constraint rule library. The input includes the computation description and constraint information in the rule text fragments, and the output includes standardized computation formulas and constraint rule library. The knowledge graph fusion and verification module is used to integrate multi-dimensional knowledge and verify logical consistency. The inputs include entity hierarchy tree, mapping table and operation formula, and the outputs include four-layer knowledge graph and logical verification results. The knowledge graph update and optimization module is used for incremental updates of the knowledge graph and optimization of the extraction algorithm. The inputs include the change rule text and question-and-answer feedback data, and the outputs include the updated knowledge graph and the optimized algorithm parameters.

[0007] Furthermore, the steps of splitting the rule text, identifying core elements, and unifying cross-sentence entity attribute identifiers are as follows: Based on the semantic logic of the financial product rule text, the original text is split into independent rule text fragments containing complete "condition, result" or "entity, relationship, constraint" logic; part-of-speech tagging combined with a financial domain dictionary is used to identify core elements such as entities, attributes, operators, conditional conjunctions, and constraint keywords; and a coreference resolution algorithm is used to process cross-sentence references, unifying the same entity and attribute identifiers scattered across multiple sentences and eliminating ambiguity.

[0008] Furthermore, the construction of entity hierarchy and extraction of attribute constraints specifically involves: receiving the output of the text preprocessing module, identifying the parent-child hierarchy among entities through dependency parsing based on the hierarchical framework of the financial domain ontology, and generating an entity hierarchy tree; and extracting multi-dimensional attributes and constraint information of each entity based on the core element identification results to establish a three-dimensional mapping table of "entity, attribute, and constraint".

[0009] Furthermore, the parsing of conditional result mapping, the connection of multi-level associations, and the splitting of branch scenarios specifically involve: loading the entity hierarchy tree, core element identification results, and structured rules; locating condition items and result items through semantic role annotation; constructing a condition-result mapping graph according to the priority of condition conjunctions; connecting multi-level association paths based on the entity hierarchy tree and core reference identifiers; and using a branch structure parsing algorithm to split multiple branches and nested conditions and generate a branch logic structure graph.

[0010] Furthermore, the conversion of the calculation formula and the construction of the constraint rule base specifically involve: receiving the output of the associated condition parsing; taking the calculation description and constraint information of the rule text fragment as the core processing object; identifying the calculation elements through regular expression matching and financial calculation dictionary; converting the natural language calculation logic into a standardized formula using Reverse Polish Notation; associating the formula parameters with information including entity attributes; constructing a constraint rule base using production rules; clarifying the rules for various constraint scenarios and dynamic floating factor information; marking anomalies found in the processing and storing them in the rule base; and indexing the rule base hierarchically by entity and attribute, forming a complete data link with the formula.

[0011] Furthermore, the integration of multi-dimensional knowledge and verification of logical consistency specifically involves: first, loading the output data of each module, then integrating it by building an index using entity identifiers and rule codes, and finally constructing a four-layer knowledge graph based on entity, attribute, relationship, and rule layers. Subsequently, multi-dimensional verification of entity associations and abnormal conditional branches is performed. If the abnormalities can be automatically corrected, the ontology rules are invoked for correction; otherwise, the abnormalities are marked for review. Finally, the compliance is compared with the expert rule base.

[0012] Furthermore, the incremental update of the graph and optimization of the extraction algorithm specifically involves: establishing a monitoring mechanism for changes in financial product rule texts, accurately locating the graph content corresponding to the changes, triggering the incremental extraction process to update relevant information, building a closed-loop optimization system based on question and answer feedback data, statistically analyzing core indicators to identify technical shortcomings and adjusting algorithm parameters, and periodically batch verifying the graph to supplement omissions and mark anomalies.

[0013] Furthermore, the change rule text includes adjustments to financial product rules throughout their entire lifecycle, and the Q&A feedback data originates from system interaction records.

[0014] Furthermore, the updated knowledge graph is based on a four-layer architecture, with only the changed content updated. Each layer is adjusted synchronously and updated with additional information, while the unchanged parts are retained. The optimized algorithm parameters are structured configurations, including adjustments to each core algorithm, with optimization justifications, and the format supports direct import.

[0015] This application also discloses an electronic device, including: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to execute the financial product knowledge graph construction and intelligent question-and-answer customer service system of the present invention.

[0016] Positive and Beneficial Effects: This application significantly improves the quality of knowledge graph construction and the efficiency of intelligent question answering through multi-module collaboration. The text preprocessing and entity attribute extraction module solves the problems of ambiguous entity reference and hierarchical dependency by splitting complete rules, resolving cross-sentence ambiguity through coreference, and constructing entity hierarchy trees, greatly reducing knowledge fragmentation. The association condition parsing module breaks through the limitations of traditional model context windows, accurately connects multi-level association paths, splits branch nested logic, and strengthens the binding mapping between conditions and results. The operation rule solidification module transforms natural language operation logic into standardized formulas, fully retains numerical priority and dynamic floating factor associations, and improves compliant numerical calculation capabilities. The graph fusion verification and update module ensures knowledge consistency through multi-dimensional logical verification, and achieves dynamic adaptation through incremental updates and feedback optimization, reducing compliance risks caused by rule changes. Overall, the completeness, logical relevance, and operational accuracy of knowledge extraction are continuously improved. The intelligent question answering system's ability to interpret and respond to complex multi-step reasoning inquiries is significantly enhanced, compliance risks and customer complaints show a downward trend, effectively addressing the shortcomings of traditional technologies in adapting to financial scenarios. Attached Figure Description

[0017] Figure 1 This is a block diagram of the financial product knowledge graph construction and intelligent question-and-answer customer service system of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a financial product knowledge graph construction and intelligent question-and-answer customer service system, such as... Figure 1 As shown, it includes: The text preprocessing module is used to split the rule text, identify core elements, and unify cross-sentence entity attribute identification. The input includes the original rule text, and the output includes structured rules and core element identification results. The entity attribute extraction module is used to construct entity hierarchy and extract attribute constraints. The input includes structured rules and financial domain ontology, and the output includes entity hierarchy tree and "entity, attribute, constraint" three-dimensional mapping table. The association condition parsing module is used to parse condition result mapping, connect multi-level associations, and split branch scenarios. The input includes entity hierarchy tree and core element identification results, and the output includes conditions, result mapping diagram, multi-level association path, and branch logic structure diagram. The computation rule solidification module is used to convert computation formulas and build a constraint rule library. The input includes the computation description and constraint information in the rule text fragments, and the output includes standardized computation formulas and constraint rule library. The knowledge graph fusion and verification module is used to integrate multi-dimensional knowledge and verify logical consistency. The inputs include entity hierarchy tree, mapping table and operation formula, and the outputs include four-layer knowledge graph and logical verification results. The knowledge graph update and optimization module is used for incremental updates of the knowledge graph and optimization of the extraction algorithm. The inputs include the change rule text and question-and-answer feedback data, and the outputs include the updated knowledge graph and the optimized algorithm parameters.

[0020] The specific implementation of splitting rule text, identifying core elements, and unifying cross-sentence entity attribute identification is as follows: First, based on the semantic logic of the financial product rule text, the original text is split into multiple independent rule text fragments. Each rule text fragment ensures that it fully includes the core logic of "condition and result" or "entity, relationship, and constraint," avoiding the fragmentation of rule information. Then, part-of-speech tagging technology combined with a financial domain-specific dictionary is used for matching to accurately identify various core elements in the text, including entities such as the name of the financial product and business indicators, attributes such as yield and term, operators such as × and ≥, conditional conjunctions such as "if," "and," and "then," and constraint keywords such as "maximum" and "floating." Finally, a coreference resolution algorithm is used to process cross-sentence referential relationships in the text, uniformly identifying the same entity or attribute scattered across multiple sentences, eliminating referential ambiguity, and providing a structured and unambiguous data foundation for subsequent knowledge extraction.

[0021] The input includes the original rule text, which in practice specifically includes various textual information related to rule-related knowledge throughout the entire lifecycle of a financial product. This includes product-side documents such as product prospectuses, business handling rules, and return calculation clauses, as well as regulatory constraints on financial products in regulatory policies and rule descriptions associated with high-frequency inquiries in customer service scenarios. The text can be in standard electronic document format or editable text after OCR recognition. Its core content must fully contain information related to the financial product, such as entities (e.g., product name, business indicators), attributes (e.g., yield, term), calculation logic (e.g., return calculation formula), conditional constraints (e.g., minimum purchase amount limit), and relationships. It must also include characteristic content such as cross-sentence dispersion, nested multi-branch conditions, and implicit multi-level relationships due to the requirements of rule expression, providing comprehensive raw data support for subsequent structured preprocessing and core element identification.

[0022] The output includes structured rules and core element identification results. Specifically, in implementation, this output is the core result of the structured preprocessing of the rule text. The structured rules are presented as independent text fragments, each with a unique identifier and core logic type label (e.g., "condition-result type," "entity-relationship-constraint type"). This ensures logical integrity within each text fragment and clear boundaries with other text fragments. The text fragment content retains the semantic information of the original rules while removing redundant modifying statements. The core element identification results are presented in a structured association format, identifying entities (e.g., financial product name, business indicators), attributes (e.g., yield, term), operators (e.g., ×, ≥), and conditions. Conjunctions (such as "if", "and", "then") and constraint keywords (such as "highest" and "floating") are assigned unique element type labels and identifiers. A clear association mapping is established between entities and attributes. The same entity or attribute referred to across sentences is identified by a coreference resolution algorithm, that is, the same unique identifier is used for association labeling. The final output is carried in the form of a structured data table, which clearly presents the scope, core logic and various core elements included in each rule text fragment, as well as their association relationships and identification information. This provides accurate and structured data input support for subsequent hierarchical extraction of entity attributes and multi-level association parsing, ensuring that the subsequent knowledge extraction process can directly and efficiently locate and process key information based on this output.

[0023] The specific implementation of constructing entity hierarchies and extracting attribute constraints involves: firstly, receiving the structured rules and core element identification results output by the text preprocessing module, and combining them with the hierarchical framework defined by the pre-defined financial domain ontology. Then, using dependency parsing technology, deeply mining the grammatical dependency relationships and semantic associations between entities, accurately identifying the parent-child hierarchical affiliation between entities. For example, clearly defining "wealth management products" as the parent entity, "cash management wealth management" as its next-level child entity, and "XX Cash Treasure Product" as belonging to "cash management wealth management". Based on this logic, a clear and well-defined entity hierarchy tree is generated, effectively solving the hierarchical dependency problem between entities.

[0024] Subsequently, for each entity in the entity hierarchy tree, based on the core element identification results, the system sorts out and extracts corresponding attribute values ​​(such as "annualized rate of return 3.2%"), numerical ranges (such as "minimum purchase amount of 1 yuan to 1 million yuan"), time constraints (such as "holding period of no less than 7 days"), applicable groups (such as "only for individual investors"), and other multi-dimensional attribute and constraint information. This information is then linked and integrated to establish a three-dimensional mapping table of "entity, attribute, and constraint" with "unique entity identifier - attribute type - constraint conditions" as the core dimension, ensuring that the correspondence between entities and attributes and constraints is accurate and verifiable.

[0025] For multi-dimensional floating factors involved in the rules, such as descriptions like "fluctuating according to market interest rates" or "adjusted in conjunction with customer star ratings," it is necessary to further clarify the factor type (such as market factors or customer factors), the specific affected object (such as the yield attribute of a certain product), and the relationship with the attribute. At the same time, the factors' fluctuation trigger conditions and reference benchmarks should be recorded to provide hierarchical and attribute-complete entity data support for subsequent multi-level correlation analysis and solidification of calculation logic.

[0026] The input includes a financial domain ontology, which in practice is a structured knowledge system customized for scenarios of financial product rule knowledge extraction and graph construction. Its core content is built around the core elements of the entire life cycle of financial products, including multi-dimensional core entity categories such as financial product categories, business indicator categories, customer attribute categories, and regulatory constraint categories. It clearly defines the hierarchical relationship benchmark of various entities (such as the parent-child hierarchical specification of "financial product → wealth management product → cash management wealth management → specific product"), and pre-sets the standard attribute types corresponding to different entities (such as the attribute definitions of "yield, minimum purchase amount, holding period" associated with wealth management products) and attribute constraint rule dimensions (such as the unit specification of numerical attributes and the expression format of time attributes).

[0027] Meanwhile, the ontology includes fundamental definitions of logical relationships unique to the financial field, such as common relationship types between entities like "subordination, association, and constraint," as well as priority reference standards for conditional logic such as "AND / OR." It also integrates common financial industry terminology, core concepts clearly defined by regulatory policies, and product classification standards of mainstream financial institutions, ensuring the authority and compliance of the ontology content.

[0028] This ontology is constructed using an extensible structured format, including core elements such as classes, attributes, relationships, and axioms. It provides a clear hierarchical framework for entity hierarchical extraction, assists dependency parsing in accurately identifying parent-child relationships between entities, and provides a standard reference for attribute extraction, avoiding confusion of attribute types. Its content will be regularly iterated and optimized in accordance with updates to financial regulatory policies and the emergence of new financial products to ensure the accuracy and industry adaptability of the entity hierarchy tree construction, and to provide a unified knowledge benchmark for the establishment of a three-dimensional mapping table of "entities, attributes, and constraints".

[0029] The output includes an entity hierarchy tree and a three-dimensional mapping table of "entity, attribute, and constraint". In practice, the output is the core result of the entity attribute extraction module after processing based on structured rules, core element identification results, and financial domain ontology. It is a key data carrier that connects text preprocessing and association condition parsing.

[0030] The entity hierarchy tree is presented in both a visual hierarchical structure and a machine-readable tree data format. Each node in the tree corresponds to a financial entity. Each node encapsulates a unique entity identifier, standard name, category (such as "wealth management product" or "cash management wealth management"), and core feature description. The parent-child hierarchical relationship between nodes is clearly defined by directed edges. The edge attributes are labeled with relationship types such as "subordinate" and the rule text number on which the association is based. For example, "financial product" is the root node under which "wealth management product" is derived. "Wealth management product" is then attached to "cash management wealth management", and finally extends to specific product nodes such as "XX Cash Treasure Product", forming a logically coherent hierarchical system. Each node is also accompanied by a data source identifier (such as the paragraph number of the corresponding original rule text) and an extraction confidence score.

[0031] The "Entity, Attribute, Constraint" three-dimensional mapping table is constructed in the form of a structured data table. The unique identifier of the entity is used as the primary key to associate the corresponding nodes in the entity hierarchy tree. The attribute column strictly follows the standard attribute types preset by the financial domain ontology, clearly recording the specific attribute values ​​corresponding to the entity (such as "annualized return rate 3.2%" and "minimum purchase amount of 1 yuan"). At the same time, it distinguishes between fixed attributes and dynamic attributes and marks them. The constraint column integrates various constraint information corresponding to the attribute, including numerical range constraints (such as "holding period ≥ 7 days" and "minimum purchase amount ≤ 1 million yuan"), time constraints (such as "issuance period January 1, 2025 - December 31, 2025"), scope of application constraints (such as "only for individual investors"), and special restrictions (such as "partial redemption is not supported"). For attributes involving multi-dimensional floating factors, the constraint column is supplemented with factor type (such as market interest rate factor and customer star rating factor), the association rules between factors and attributes (such as "yield fluctuates in the same direction as market interest rate, with a fluctuation ratio of 0.1 times"), and the source of factor reference benchmark.

[0032] Each record in the mapping table includes a rule text identifier for attribute extraction and an ontology matching basis. A data verification identifier column is also included to indicate whether the record conforms to the attribute specifications of the financial ontology. Records with ambiguity or low confidence (such as attribute values ​​inconsistent with ontology standard units) are explicitly marked as pending review with a problem description. The entity hierarchy tree and the 3D mapping table achieve precise association through unique entity identifiers, ensuring that the hierarchical affiliation and attribute constraint information of entities form a complete data loop. This provides a clear foundation for the entity hierarchy relationship in the association condition parsing module to connect multi-level association paths, and provides precise data support for attribute values ​​and constraint boundaries for subsequent computational logic. Its data format supports seamless integration with subsequent modules and can be directly used as standardized input data for association condition parsing, graph fusion verification, and other processes.

[0033] The aforementioned parsing condition result mapping, multi-level association chaining, and branch splitting scenarios are implemented as follows: First, the entity hierarchy tree output by the entity attribute extraction module and the core element identification results generated by the text preprocessing module are loaded. Simultaneously, structured rules are retrieved as the basic data source. Semantic role labeling technology is used to perform deep semantic parsing on each rule text fragment to accurately locate the condition items (such as "minimum purchase amount ≥ 50,000 yuan" and "holding period ≥ 30 days") and result items (such as "annualized rate of return is 3.5%" and "supports early redemption"). Based on the priority of condition conjunctions preset in the financial field rule logic ("and" is higher than "or"), a condition-result mapping graph including logical weight information is constructed. The nodes in the graph are labeled with unique element identifiers, numerical attributes, and constraint types, and the edge attributes clearly define the logical association type and priority order.

[0034] To address the need for multi-level chain-like associations between entities, this approach uses an entity hierarchy tree as the topological foundation and combines the common identifiers in the core element identification results to identify intermediate carrier entities (such as "customer asset level" and "product risk rating") in the association chain of "product basic information → customer qualification requirements → rights and benefits enjoyment standards". Through the unique identifier of the entity, a cross-rule text fragment association mapping is established, which is then linked to form a complete multi-level association path. Each association edge in the path is accompanied by the rule text fragment number and semantic matching confidence of the association basis.

[0035] For scenarios with multiple branches, a branch structure parsing algorithm is used to split rule texts marked with phrases such as "if...then..." and "if...then...". This clarifies the independent triggering conditions of each branch (e.g., "holding period < 30 days", "30 days ≤ holding period < 90 days", "holding period ≥ 90 days") and their corresponding results. Through logical node splitting and edge reconstruction, a branch logic structure diagram is generated, clearly distinguishing between parallel and nested branches. For nested conditions like "if the customer is a VIP and the holding period ≥ 90 days, then the yield increases by 0.2%", the algorithm decomposes them into a superimposed constraint relationship between "customer identity condition" and "holding period condition" and labels the hierarchy. Simultaneously, rule text fragments with missing conditions, ambiguous results, or logical conflicts are automatically marked as anomalies and linked to their original text locations, along with preliminary conflict analysis (e.g., "condition 'market interest rate fluctuation' does not clearly correspond to the yield adjustment formula"). This is output synchronously with the parsing results, providing logically clear and structurally complete data support for subsequent rule solidification and graph verification, ensuring the accuracy of parsing multi-step reasoning-based consulting needs.

[0036] The input includes an entity hierarchy tree and core element identification results. Specifically, in implementation, this input is the basic data set for the core processing of the association condition parsing module. The two are accurately associated through unique entity identifiers, providing comprehensive support for condition result mapping and parsing, multi-level association chaining, and branch scenario decomposition. The entity hierarchy tree provides a clear foundation for entity topology relationships for the module. This tree is presented in structured data form, including the unique identifier, standard name, category, and core features of each financial entity. The nodes are clearly marked with parent-child hierarchical relationships such as "belonging" through directed edges. The edge attributes include the rule text fragment number and extraction confidence level of the association basis. For example, the complete hierarchical chain of "wealth management products → cash management wealth management products → XX cash treasure products" and the corresponding data source identifier can be directly used as the topological framework for multi-level association path chaining.

[0037] The core element identification results provide refined element information of the rule text, which is carried by a structured data table. This includes the split independent rule text fragments (with unique codes and logical type labels such as "condition-result type" and "entity-relationship-constraint type"), as well as exclusive type labels and unique identifiers for elements such as entities (e.g., product name, business indicators), attributes (e.g., rate of return, term), operators, conditional conjunctions ("if", "and", "then"), and constraint keywords ("maximum", "floating") within each rule text fragment. At the same time, the identification of the same entity and attribute across sentences is unified through the coreference resolution algorithm, clearly marking the element relationships scattered across multiple sentences.

[0038] With the combined effect of the two, the entity hierarchy tree provides hierarchical boundaries and subordinate logical references for element association, while the core element identification results provide precise targets for locating condition items, result items, and parsing logical relationships. For example, the module can rely on the logical connective words in the text fragments of the condition conjunctions in the core element identification results, combined with the entity subordinate relationships in the entity hierarchy tree, to accurately identify the intermediate carrier entities in multi-level associations, ensuring the logical integrity of the condition result mapping and the coherence of the multi-level association paths. Its data format and content strictly match the output specifications of the text preprocessing module and the entity attribute extraction module, ensuring the smoothness and consistency of data flow between modules.

[0039] The output includes conditions, result mapping diagram, multi-level association path, and branch logic structure diagram. In practice, this output is the core result of the association condition parsing module. The three are linked through entity unique identifier and rule text fragment encoding to provide logically clear structured association data for subsequent modules.

[0040] The condition and result mapping diagrams are presented in both a visual directed graph and a machine-readable data format. The condition nodes encapsulate the unique identifier of the element, numerical attributes (such as "minimum purchase amount of 50,000 yuan"), constraint type (such as "≥"), and the text fragment number of the rule to which they belong. The result nodes correspond to specific result information such as "annualized rate of return of 3.5%". The nodes are connected by directed edges labeled with logical types ("AND / OR"), priority weights, and association confidence, which clarifies the relationship that "AND" logic takes precedence over "OR" logic. The original text fragment identifiers on which the mapping is based are also attached.

[0041] The multi-level association path is presented in a chain-like data structure. The intermediate carrier entities are sequentially linked in the path (such as "product minimum purchase amount → customer asset level → yield level"). Each entity node includes hierarchical affiliation and core attributes. The association edges between entities are labeled with relationship types such as "association" and "subordination", the source of rule text fragments and semantic matching confidence, to ensure the integrity and traceability of the multi-level chain association.

[0042] The branch logic structure diagram adopts a hierarchical node layout, clearly distinguishing between parallel branches (such as "holding period < 30 days" and "30 days ≤ holding period < 90 days") and nested branches (such as "VIP customers and holding period ≥ 90 days"). Nested conditions are distinguished by hierarchical indentation or node color to differentiate constraint priorities. Each branch node is labeled with its triggering condition, corresponding result, and unique branch code. The logical mutual exclusion or parallel relationship between branches is also indicated. All three include data verification information. For abnormal rule text fragments found during the parsing process, such as missing conditions or ambiguous results, the abnormality type, original text location, and preliminary analysis conclusion are added in the form of associated annotations (such as "condition 'market interest rate fluctuation' is not associated with yield adjustment rules"). The output data format supports seamless integration with the operation rule solidification module and the graph fusion verification module, providing accurate associated logic support for operation logic extraction and knowledge graph construction.

[0043] The conversion calculation formula and the construction of the constraint rule base are implemented as follows: First, the conditions, result mapping diagram, multi-level association path and branch logic structure diagram output by the association condition parsing module are received. Simultaneously, the core element identification results and structured rules generated by the text preprocessing module are retrieved. The content involving operation description and constraint information in the rule text fragment is the core processing object. For the conversion of the calculation formula, regular expression matching combined with a financial domain operation dictionary is used to accurately identify the operators (such as ×, +, ≥), numerical variables (such as customer star rating coefficient, base interest rate), benchmark parameters (such as the central bank's benchmark interest rate for the same period), and operation priority prompts (such as "calculate first" and "calculate later") in the text. Then, the reverse Polish notation conversion algorithm is used to convert the operation logic described in natural language, such as "expected rate of return = base interest rate + floating interest rate × customer star rating coefficient", into the postfix expression form of "expected rate of return base interest rate floating interest rate customer star rating coefficient × + =", clarifying the priority order of numerical calculation. Finally, a standardized calculation formula that can be directly executed by the machine is generated. Each parameter in the formula is associated with the corresponding entity unique identifier and attribute type, and is accompanied by the rule text fragment number from which the operation logic comes and the conversion confidence score.

[0044] For the construction of the constraint rule base, for dual constraint scenarios such as "amount ≥ 50,000 and holding period ≥ 30 days", a production rule representation method is adopted to construct a rule structure of "IF constraint 1 (amount ≥ 50,000) AND constraint 2 (holding period ≥ 30 days) THEN result (enjoy an annualized rate of return of 3.5%)", integrating the threshold, comparison symbol, result information and related entity attributes in the constraint conditions into rule entries. For single constraint scenarios (such as "minimum purchase amount ≥ 1 yuan"), it is simplified to the basic structure of "IF constraint THEN result", and each rule entry is labeled with the constraint type (such as numerical constraint, time constraint), applicable entity scope and regulatory basis (such as the corresponding regulatory policy clause number).

[0045] For dynamic floating logic, such as "the yield fluctuates according to the market interest rate, and the yield increases by 0.05% for every 0.1% increase in the market interest rate", it is necessary to specify the floating factor type (market factor), the factor adjustment threshold (0.1%), the relationship of the calculation formula (yield = benchmark yield + (market interest rate increase / 0.1%) × 0.05%), and the factor data acquisition channel (such as the central bank's public data interface identifier), and supplement the corresponding rule entries as extended attributes.

[0046] During processing, if ambiguous calculation descriptions are found (e.g., the calculation method for "yield adjusted according to market conditions" is not clearly defined) or conflicting constraints are found (e.g., the same entity attribute has two constraints: "holding period ≥ 30 days" and "holding period ≥ 60 days"), the anomaly type must be automatically marked, the original rule text location and conflicting content must be associated, and an anomaly list must be generated and synchronously stored in the rule base. The final constraint rule base is hierarchically indexed according to entity category and attribute type. Standardized calculation formulas and rule entries are associated and bound through unique entity identifiers and attribute identifiers, forming a complete data link of "calculation logic - constraint conditions - entity attributes". This provides directly integrateable calculation and constraint data for the subsequent graph fusion verification module, ensuring that the intelligent question answering system can accurately call the corresponding formulas and constraint rules when processing numerical calculation-related inquiries, guaranteeing the compliance and accuracy of the calculation results.

[0047] The input includes the operational description and constraint information in the rule text fragments. Specifically, in implementation, this input is the core data source of the operational rule solidification module, extracted from the structured rules generated by the text preprocessing module, and linked with the core element identification results and the output data of the association condition parsing module. The operational description in the rule text fragments includes all natural language expressions involving numerical calculations in the financial product rules. This includes complete formulaic descriptions such as "expected rate of return = base interest rate + floating interest rate × customer star rating coefficient", as well as operational logic fragments scattered across multiple sentences. It also clearly defines the associated operators (×, +, ≥, etc.), numerical variables (base interest rate, customer star rating coefficient, etc.), benchmark parameters (such as the central bank's benchmark interest rate for the same period), and operational priority prompts (such as the priority expression in "deduct handling fees first and then calculate returns"). Each operational description is accompanied by a unique code for the corresponding rule text fragment and the associated element ID in the core element identification results.

[0048] The constraint information in the rule text fragments includes content related to various constraint scenarios, including dual constraints such as "amount ≥ 50,000 and holding period ≥ 30 days", single constraints such as "minimum purchase amount ≥ 1 yuan", and dynamic constraints such as "fluctuates according to market interest rates" and "adjusted based on customer qualifications". The constraint information clearly marks the constraint object (corresponding to the entity ID and attribute type in the entity hierarchy tree), constraint threshold (e.g., 50,000, 30 days), constraint logical relationship (AND / OR), and applicable constraint scenario (e.g., specific product, specific customer group). For dynamic constraints, the type identifier, impact attributes, and fluctuation reference benchmark of multi-dimensional floating factors are synchronously associated. All constraint information is associated with the corresponding record in the "entity, attribute, constraint" three-dimensional mapping table through the entity ID, ensuring accurate matching between constraint content and entity attributes, and providing comprehensive, accurate, and clearly associated data support for subsequent standardized calculation formula conversion and constraint rule base construction.

[0049] The output includes standardized calculation formulas and a constraint rule library. In practice, the output is the core result of the calculation rule solidification module after processing the calculation description and constraint information in the rule text fragment. It is a key data carrier that connects the correlation condition parsing and the graph fusion verification. The two are deeply associated through entity unique identifiers and attribute type identifiers, and together provide accurate data support for the construction of the knowledge graph rule layer.

[0050] The standardized calculation formulas are presented in a structured format that can be directly executed by machines. This includes the complete calculation logic generated by the Reverse Polish Notation (e.g., converting "Expected Rate of Return = Base Rate + Floating Rate × Customer Star Rating Coefficient" into a calculation expression with clear priorities), as well as the association information of each parameter in the formula. Each parameter is bound to a unique entity identifier (e.g., "Base Rate" is associated with the entity ID of "XX Cash Treasure Product"), attribute type labels (e.g., "Rate of Return" and "Coefficient"), and data source identifiers (corresponding to the paragraph number and rule text fragment code of the original rule text). At the same time, the confidence score and priority description of the formula conversion are attached (e.g., the logical label "Calculate the product first and then calculate the sum"). For dynamic formulas involving multi-dimensional floating factors, the factor association fields are additionally labeled to clarify the mapping relationship between factors such as "market interest rate" and "customer star rating" and the formula parameters.

[0051] The constraint rule base is constructed in a hierarchical structured entry format, with the core adopting a production rule notation. Each rule entry includes a complete link of "constraint condition - threshold - result - related information": For dual and multiple constraint scenarios, the logical relationship of each sub-condition is clearly broken down (such as the "AND" association between "amount ≥ 50,000" and "holding period ≥ 30 days") and its priority is marked. For single constraint scenarios, it is simplified to a basic "IF-THEN" structure. The entry clearly defines the constraint type (numerical constraint, time constraint, dynamic constraint, etc.), the scope of applicable entities (the node ID of the associated entity hierarchy tree), and the regulatory policy basis (such as the corresponding regulatory document clause number). For dynamic constraint entries, detailed information is added about the floating factor type (market type, customer type, etc.), factor adjustment threshold (such as "for every 0.1% increase in market interest rate"), the association rule between the factor and the calculation formula (such as "yield fluctuates in the same direction as the factor by 0.05%)), and the factor data acquisition channel identifier.

[0052] The rule base synchronously includes a list of anomalies discovered during computation and constraint processing. For ambiguous computation descriptions (such as "adjust according to market conditions" without a clear calculation method) and conflicting constraints (such as the same attribute having both "holding period ≥ 30 days" and "holding period ≥ 60 days"), the anomaly type, original rule location, and preliminary analysis conclusion are marked, and these are stored separately from valid rule entries. The entire output is primarily in a machine-readable format, supplemented by a manually verifiable structured explanatory document. This ensures that it can be directly used as standardized input for the knowledge graph fusion verification module, supporting the construction of the knowledge graph rule layer, and also provides clear traceability for subsequent cross-validation by domain experts, guaranteeing the accuracy and compliance of the intelligent question-answering system in numerical calculations and rule responses.

[0053] The integration of multi-dimensional knowledge and verification of logical consistency are implemented as follows: First, the multi-source data integration process is initiated, and the entity hierarchy tree and the three-dimensional mapping table of "entity, attribute, and constraint" output by the entity attribute extraction module are loaded simultaneously. The condition-result mapping diagram, multi-level association path, and branch logic structure diagram generated by the association condition parsing module are also loaded. The standardized calculation formula and constraint rule library output by the calculation rule solidification module are also loaded. A full-domain data association index is established through the unique entity identifier and rule text fragment encoding to ensure accurate matching of data in all dimensions.

[0054] Based on a pre-defined four-layer knowledge graph architecture (entity layer, attribute layer, relationship layer, and rule layer), each layer is constructed sequentially: The entity layer is based on an entity hierarchy tree, importing the entity's unique identifier, standard name, category, and core features to form the core nodes of the graph; the attribute layer relies on a three-dimensional mapping table to associate the attribute values ​​and constraint information of the entity with the corresponding entity nodes, distinguishing between fixed and dynamic attributes and labeling floating factor association information; the relationship layer integrates the association logic of multi-level association paths and condition-result mapping graphs, encapsulating relationship types such as "subordinate," "association," and "condition trigger" and logical priorities as directed edges between nodes; the rule layer embeds standardized calculation formulas and constraint rule library entries, establishing an association mapping between rules and corresponding entities and attributes, forming a complete data link of "entity-attribute-relationship-rule".

[0055] After data integration, a logical consistency verification process is initiated, employing a multi-dimensional verification algorithm for layered processing: Entity association verification identifies issues such as "inconsistent hierarchical levels of the same entity in different rule text fragments" and "broken entity associations in intermediate carriers" by comparing entity hierarchy trees and multi-level association paths; Conditional branch verification combines branch logic structure diagrams and condition-result mapping diagrams to check for omissions or contradictions such as "a triggering condition in multiple branch conditions does not correspond to a result" and "incorrect priority of nested condition logic"; Operational logic verification verifies whether the order of numerical operations in the formula is consistent with the original rule description based on the priority annotation of the reverse Polish notation, and identifies issues such as "confusion of priority between multiplication / division and addition / subtraction operations" and "incorrect association between floating factors and operation parameters"; Constraint threshold verification discovers anomalies such as "threshold conflicts for the same attribute in different rules" and "mismatch between constraint conditions and attribute types" by cross-comparing the three-dimensional mapping table and the constraint rule base.

[0056] For identifiable anomalies (such as those with explicit ontology references in entity-level conflicts), pre-defined rules from the financial domain ontology are invoked for automatic correction. For anomalies that cannot be automatically determined (such as those with no explicit reference standards for constraint thresholds), the anomaly type is marked (e.g., "threshold conflict," "logical ambiguity"), the original rule text location and conflicting data fragments are associated, and a list to be reviewed is generated. Finally, a compliance verification step is initiated, cross-referencing the integrated knowledge graph with a domain expert rule base. This expert rule base includes the latest financial regulatory policy provisions, industry-wide rules, and product compliance standards. The focus is on verifying whether the rule-layer content complies with regulatory requirements (e.g., whether the yield calculation logic meets regulatory disclosure standards) and whether the association between entities and attributes conforms to industry classification standards. Non-compliant items are marked with compliance risk levels and associated with policy basis. The final output is a four-layer integrated knowledge graph (with data source identifiers and extraction confidence levels for each node / edge) and a complete logic verification result report. The report includes details of automatically corrected items, a list of anomalies to be reviewed, and compliance verification conclusions, providing basic data support for subsequent graph applications and incremental updates.

[0057] The input includes an entity hierarchy tree, a mapping table, and calculation formulas. In practice, this input is the core data support for the graph fusion and verification module to carry out multi-dimensional knowledge integration and logical verification. Each component is deeply associated through a unique entity identifier, which together ensures the integrity and accuracy of the knowledge graph construction.

[0058] The entity hierarchy tree is structured tree data, including the unique identifier, standard name, category, and core features of each financial entity. Directed edges clearly define hierarchical relationships such as "subordination" between nodes. Edge attributes are labeled with the rule text fragment number and extraction confidence level used for association, forming a complete hierarchical link of "financial products → wealth management products → cash management wealth management → specific products," providing a basic topological framework for constructing the entity layer of the knowledge graph. The mapping table specifically refers to a three-dimensional mapping table of "entity, attribute, and constraint." Using the entity's unique identifier as the primary key, and referencing preset attribute types in the financial domain ontology, it records the specific attribute values ​​(including fixed and dynamic attribute annotations), numerical range constraints, time constraints, and scope of application for each entity. For attributes involving floating factors, additional factor types, association rules, and reference benchmarks are provided. Each record includes the rule text fragment identifier used for attribute extraction and the ontology matching verification results.

[0059] The calculation formula is a standardized calculation formula, presented in a machine-executable structured format. It includes explicit calculation logic transformed by Reverse Polish Notation. Each parameter in the formula is bound to a unique entity identifier, attribute type label, and data source identifier. It also includes a transformation confidence score and calculation priority description. Dynamic formulas also indicate the mapping relationship between floating factors and parameters.

[0060] The three elements work together: the entity hierarchy tree provides the hierarchical benchmark for entity associations, the mapping table supplements the details of entity attributes and constraints, and the calculation formula carries the core calculation logic of the rule layer. All of them are associated with the rule text fragment encoding and the original text source identifier. This provides a comprehensive and accurate structured data foundation for the module to integrate the four layers of knowledge of "entity-attribute-relationship-rule", verify entity association conflicts, constraint threshold contradictions and calculation logic errors, and ensure that the integrated knowledge graph has logical coherence and traceable data from the same source.

[0061] The output includes a four-layer knowledge graph and logical verification results. In practice, the output is the core result of the graph fusion and verification module after integrating multi-source data and completing logical verification. It is the key carrier for connecting the knowledge graph construction and the intelligent question-and-answer customer service system. The two achieve deep linkage through data association index, and jointly provide underlying support for the system's accurate response.

[0062] The four-layer knowledge graph is structured around an "entity-attribute-relationship-rule" architecture, employing a dual format for efficient machine retrieval and visualization: The entity layer is based on an entity hierarchy tree, with each node encapsulating a unique entity identifier, standard name, category, and core features, along with paragraph numbers and extraction confidence scores for the corresponding original rule text, forming a logically coherent hierarchical node system such as "financial products → wealth management products → cash management products → XX Cash Treasure products"; The attribute layer relies on a three-dimensional mapping table of "entity, attribute, and constraint," associating the fixed / dynamic attribute values, numerical ranges, and time constraints of entities with the corresponding entity nodes. Dynamic attributes are additionally labeled with floating factor types, influencing rules, and reference benchmarks, and attribute entries include the rules upon which extraction is based. The text fragment identifier is matched with the ontology result; the relationship layer integrates the logic of multi-level association paths and condition-result mapping graphs, and uses directed edges to label relationship types such as "subordinate", "association", and "condition trigger", clarifying the logical priority and association confidence of "AND / OR" and other logics, and completely connecting multi-level association paths such as "product minimum purchase amount → customer asset level → rate of return level"; the rule layer embeds standardized calculation formulas and constraint rule library entries. The calculation formulas retain the priority logic after the transformation of the reverse Polish notation. Each parameter is bound to an entity and attribute identifier. The constraint rule entries are presented in the form of production rules, including the complete link of "constraint condition-threshold-result" and the labeling of regulatory basis. Each layer establishes a full-domain association through unique entity identifiers and rule text fragment codes to ensure data traceability.

[0063] The logical verification results are presented in a structured report, including three core parts: A detailed list of automatically corrected items clearly records issues corrected based on the financial domain ontology, such as "entity level conflict" and "operational priority error," with the basis for correction and original data indicated; a list of anomalies to be reviewed details issues that cannot be automatically determined, including types such as "constraint threshold conflict" and "logical ambiguity," along with the location of the anomaly in the original rule text, conflicting data fragments, and preliminary analysis conclusions; and compliance verification conclusions compared with the domain expert rule base (including regulatory policies and industry standards), indicating the compliance risk level of the rule layer content in the graph, and clearly specifying the policy clause number and rectification suggestions corresponding to the non-compliant items.

[0064] Each node and edge of the knowledge graph is accompanied by a data source identifier (such as the code of the rule text fragment and the original text paragraph number) and an extraction confidence level. The logical verification results are synchronously associated with the corresponding graph node / entry identifier. The output format can be directly called by intelligent question answering systems, providing them with accurate knowledge query and numerical calculation support. It also facilitates technical personnel and domain experts to carry out review and optimization. At the same time, it provides a clear iterative direction for the graph update and optimization module, ensuring the accuracy, compliance and practicality of the knowledge graph.

[0065] The incremental update graph and optimized extraction algorithm are implemented as follows: First, a dynamic monitoring mechanism for changes in financial product rule texts is established. Through text comparison algorithms and keyword triggering technology, newly added, modified, or abolished content in the original rule text is captured in real time. Combined with the unique identifier of the rule text fragment and the entity association index, the knowledge graph nodes, relationships, and rule entries corresponding to the changed content are accurately located, avoiding resource waste and efficiency loss caused by full updates.

[0066] In response to changes, an incremental extraction process is triggered, from structured preprocessing of rule text to knowledge graph fusion and logical verification. Semantic decomposition, core element identification, and coreference resolution are performed only on the rule text fragments affected by the changes. Simultaneously, the hierarchical affiliation of corresponding nodes in the entity hierarchy tree, attribute constraint information in the 3D mapping table, and the associated content of the condition-result mapping graph, calculation formulas, and constraint rule base are updated to ensure accurate synchronization of change information to the knowledge graph. Subsequently, a closed-loop optimization system is constructed based on the interactive feedback data from the intelligent question-answering system. Data tracking technology is used to collect rule-based queries from users, system response results, and user satisfaction evaluations. Key indicators such as response accuracy, numerical calculation accuracy, and logical coherence score are statistically analyzed to pinpoint the technical shortcomings behind low-accuracy scenarios. For example, if element identification deviations are caused by the lack of coverage of new financial terms, the domain dictionary is expanded and new term attributes are labeled; if hierarchical judgment errors are caused by insufficient adaptation of dependency parsing parameters, the syntactic weights and threshold settings of the algorithm are adjusted.

[0067] Simultaneously, a batch verification process for the knowledge graph is initiated regularly. An association integrity scanning algorithm is used to traverse multi-level association paths, conditional branch logic, and operational rule chains between entities, supplementing any missing associations, unsplit branch scenarios, and unfixed operational details. For ambiguous rules or conflicting information discovered during verification, anomaly types are marked and an optimization suggestion list is generated. Finally, combined with the review opinions of domain experts on the anomaly list, the model parameters and logical judgment rules of the extraction algorithm are further iterated to achieve a closed-loop process of "change monitoring - incremental update - feedback analysis - algorithm optimization - batch verification." This ensures that the knowledge graph can dynamically adapt to changes in financial product rules and continuously improve the response accuracy and compliance of the intelligent question-answering system.

[0068] The inputs include the rule text and question-and-answer feedback data. In practice, these inputs are the core data support for the graph update and optimization module to achieve incremental updates and algorithm iterations. They correspond to the two core requirements of "dynamic rule adaptation" and "system closed-loop optimization" respectively, and both are precisely related to the knowledge graph and the data of the preceding modules. The change rule text includes dynamic adjustments to financial product rules throughout their entire lifecycle. This includes product-side rule changes such as revisions to product prospectuses and updates to business processing details, as well as changes in constraints and regulations resulting from regulatory policy adjustments. The text can be in the form of a complete new rule document or a partially modified text fragment. Each change text includes a clear change type identifier (new / modified / repealed), a unique code associated with the original rule text fragment, and a corresponding knowledge graph node ID. Content-wise, it fully encompasses the entities involved in the change (such as new product names, adjusted business indicators), attributes (such as updated yield rates, modified minimum purchase amounts), calculation logic (such as adjusted yield calculation formulas), and constraints (such as newly added holding period restrictions). It retains the inherent characteristics of rules, such as cross-sentence reference, nested multi-branch conditions, and multi-level implicit associations, ensuring consistency with the input specifications of the rule text structured preprocessing stage, and can directly trigger the incremental extraction process.

[0069] The question-and-answer feedback data originates from the full interaction records of the intelligent question-and-answer system, and includes three core types of information: First, rule-related query details, including the specific content of user inquiries (especially complex rule queries involving multi-step reasoning, such as "How is the rate of return calculated if I have held XX Cash Treasure product for 30 days and my assets are over 50,000 yuan"), the corresponding knowledge graph call node identifier, and the encoding of associated rule text fragments; Second, response result data, including the system-generated reply content, the called calculation formula ID, constraint rule entries, and numerical calculation process; Third, feedback evaluation information, including both user-initiated feedback (such as complaints, follow-up questions, and satisfaction ratings) and system behavior feedback (such as response timeouts, secondary queries caused by deviations between calculation results and user expectations), along with quantitative indicators such as response accuracy, calculation result correctness, and logical coherence scores. All feedback data is linked to the rule text fragment encoding, changed rule text, and knowledge graph nodes, providing precise targets for locating algorithm shortcomings (such as missing domain dictionaries and deviations in syntactic analysis parameters) and missing graph content, ensuring the targeted nature of incremental updates and algorithm optimization.

[0070] The output includes the updated knowledge graph and the optimized algorithm parameters. In practice, the output is the core result of the graph update and optimization module after incremental updates and algorithm iterations. The two are deeply associated through rule text fragment encoding and entity unique identifiers, which together support the long-term accurate operation of the intelligent question answering system.

[0071] The updated knowledge graph is based on a four-layer architecture of "entity-attribute-relationship-rule". Only the content involved in the changed rule text is precisely updated: In the entity layer, new entity nodes are added to supplement complete identifiers, categories and features, the hierarchical affiliation and core information of entity nodes are modified, and deprecated entity nodes are marked as invalid while retaining historical data; In the attribute layer, the attribute values ​​and constraints of the corresponding entities are adjusted synchronously, and the floating factor association rules and reference benchmarks of dynamic attributes are updated; In the relationship layer, the association edge information of multi-level association paths is corrected, and the logical priority and association confidence of the condition-result mapping graph are updated; In the rule layer, standardized calculation formulas are replaced or added, and the entries of the constraint rule library and the labeling of regulatory basis are updated. All nodes and edges of the graph are attached with the change type, timestamp and data source identifier of this update. The parts not involved in the change retain the original structure and confidence score, ensuring the continuity and accuracy of the knowledge graph.

[0072] The optimized algorithm parameters are presented in a structured configuration file, including adjustments to each core algorithm in the extraction process: The domain dictionary is supplemented with new financial terms, product names, and regulatory terms, clearly labeling the attribute types and associated entity categories of these new terms; the dependency parsing algorithm optimizes the syntactic weight coefficients and hierarchical judgment thresholds to improve the accuracy of entity parent-child attribution identification; the semantic role labeling algorithm adjusts the recognition weights of conditional and result items and optimizes the judgment parameters for logical conjunction priority; the coreference resolution algorithm updates the similarity threshold of the referential matching model to enhance the accuracy of unified cross-sentence entity identification. All parameters include optimization justifications, clearly defining low-accuracy scenarios in the corresponding question-and-answer feedback data (such as element identification deviations and hierarchical judgment errors), comparisons of effects before and after parameter adjustments, and expected improvement targets. The parameter format supports direct import into various algorithm modules, enabling rapid iteration of extraction capabilities. The two work together to provide real-time, compliant knowledge support for the intelligent question-answering system, while the optimized algorithm parameters ensure the efficiency and quality of subsequent knowledge extraction, forming a closed-loop mechanism of "data feedback - algorithm optimization - graph upgrade."

[0073] This application also provides an embodiment of an electronic device. The electronic device is manifested in the form of a general-purpose computing device. The components of the electronic device may include, but are not limited to: one or more processors or processing units, memory, and buses connecting different components (including memory and processing units).

[0074] A bus refers to one or more of several bus architectures, including memory buses or memory controllers, peripheral buses, graphics acceleration ports, processors, or local buses using any of the various bus architectures. Examples of these architectures include, but are not limited to, Industry Standard Architecture (ISA) buses, Micro Channel Architecture (MCA) buses, Enhanced ISA buses, Video Electronics Standards Association (VESA) local buses, and Peripheral Component Interconnect (PCI) buses.

[0075] Electronic devices typically include a variety of computer-readable media. These media can be any available media that can be accessed by the electronic device, including volatile and non-volatile media, and removable and non-removable media.

[0076] Memory may include computer-readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic devices may further include other removable / non-removable, volatile / non-volatile computer device storage media. By way of example only, storage may be used to read and write non-removable, non-volatile magnetic media.

[0077] The electronic device can also communicate with one or more external devices (e.g., keyboard, pointing device, camera, etc.), may include a display, and may communicate with one or more devices that enable a user to interact with the electronic device, and / or with any device that enables the electronic device to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via an input / output (I / O) interface. Furthermore, the electronic device can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN)) and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. The processor executes various functional applications and data processing by running programs stored in memory, such as implementing the financial product knowledge graph construction and intelligent question-and-answer customer service system provided in the above embodiments of the present invention.

[0078] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A financial product knowledge graph construction and intelligent question-and-answer customer service system, characterized in that, include: The text preprocessing module is used to split the rule text, identify core elements, and unify cross-sentence entity attribute identification. The input includes the original rule text, and the output includes structured rules and core element identification results. The entity attribute extraction module is used to construct entity hierarchy and extract attribute constraints. The input includes structured rules and financial domain ontology, and the output includes entity hierarchy tree and a three-dimensional mapping table of "entity, attribute, constraint". The association condition parsing module is used to parse condition result mapping, connect multi-level associations, and split branch scenarios. The input includes entity hierarchy tree and core element identification results, and the output includes conditions, result mapping diagram, multi-level association path, and branch logic structure diagram. The computation rule solidification module is used to convert computation formulas and build a constraint rule library. The input includes the computation description and constraint information in the rule text fragments, and the output includes standardized computation formulas and constraint rule library. The knowledge graph fusion and verification module is used to integrate multi-dimensional knowledge and verify logical consistency. The inputs include entity hierarchy tree, mapping table and operation formula, and the outputs include four-layer knowledge graph and logical verification results. The knowledge graph update and optimization module is used for incremental updates of the knowledge graph and optimization of the extraction algorithm. The inputs include the change rule text and question-and-answer feedback data, and the outputs include the updated knowledge graph and the optimized algorithm parameters.

2. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The specific steps for splitting the rule text, identifying core elements, and unifying cross-sentence entity attribute identifiers are as follows: Based on the semantic logic of the financial product rule text, the original text is split into independent rule text fragments containing complete "condition, result" or "entity, relationship, constraint" logic; part-of-speech tagging combined with a financial domain dictionary is used to identify core elements such as entities, attributes, operators, conditional conjunctions, and constraint keywords; and a coreference resolution algorithm is used to process cross-sentence references, unifying the identifiers of the same entity and attribute scattered across multiple sentences to eliminate ambiguity.

3. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The construction of entity hierarchy and extraction of attribute constraints are specifically as follows: receiving the output of the text preprocessing module, identifying the parent-child hierarchy between entities through dependency parsing based on the hierarchical framework of the financial domain ontology, and generating an entity hierarchy tree; based on the core element identification results, extracting multi-dimensional attributes and constraint information of each entity, and establishing a three-dimensional mapping table of "entity, attribute, constraint".

4. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The parsing of conditional result mapping, multi-level association chaining, and branch splitting scenarios are specifically as follows: loading entity hierarchy tree, core element identification results, and structured rules; locating condition items and result items through semantic role annotation; constructing a condition-result mapping graph according to the priority of condition conjunctions; relying on entity hierarchy tree and core reference identifiers to chain multi-level association paths; using a branch structure parsing algorithm to split multiple branches and nested conditions and generate a branch logic structure graph.

5. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The conversion of the operation formula and the construction of the constraint rule base are specifically as follows: receiving the output of the associated condition parsing, taking the operation description and constraint information of the rule text fragment as the core processing object, identifying the operation elements through regular expression matching and financial operation dictionary, converting the natural language operation logic into a standardized formula using reverse Polish notation, associating information including entity attributes with the formula parameters, constructing a constraint rule base using production rules, clarifying the rules for various constraint scenarios and dynamic floating factor information, marking the anomalies found in the processing and storing them in the rule base, and indexing the rule base hierarchically by entity and attribute, forming a complete data link with the formula.

6. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The integration of multidimensional knowledge and verification of logical consistency are specifically implemented as follows: First, the output data of each module is loaded, and indexes are built and integrated through entity identifiers and rule codes. Then, a four-layer knowledge graph is constructed according to entity, attribute, relationship, and rule layers. Subsequently, entity association and condition branch anomalies are verified in multiple dimensions. If they can be automatically corrected, the ontology rules are called to correct them; otherwise, they are marked for review. Finally, compliance is compared with the expert rule base.

7. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The incremental update of the graph and optimization of the extraction algorithm are specifically as follows: establish a monitoring mechanism for changes in financial product rule texts, accurately locate the graph content corresponding to the changes, trigger the incremental extraction process to update relevant information, build a closed-loop optimization system based on question and answer feedback data, statistically analyze core indicators to locate technical shortcomings and adjust algorithm parameters, and regularly perform batch verification of the graph to supplement omissions and mark anomalies.

8. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The change rule text includes adjustments to financial product rules throughout their entire lifecycle, and the Q&A feedback data originates from system interaction records.

9. The financial product knowledge graph construction and intelligent question-and-answer customer service system according to claim 1, characterized in that, The updated knowledge graph is based on a four-layer architecture, with only the changed content updated. Each layer is adjusted synchronously and updated with additional information, while the unchanged parts are retained. The optimized algorithm parameters are structured configurations, including adjustments to each core algorithm, with optimization justifications, and the format supports direct import.

Images