Intelligent question answering method and device based on search enhancement generation technology, equipment and medium

Abstract

The application relates to the technical field of natural language processing and can be applied to the financial technology field, and discloses an intelligent question and answer method and device based on retrieval enhancement generation technology, equipment and a medium, which comprises the following steps: receiving a natural language query input by a user, performing heterogeneous retrieval on the natural language query by using a double-channel retrieval module, obtaining first index metadata and second index metadata, performing slot filling on candidate index metadata obtained by weighted fusion of the first index metadata and the second index metadata, generating a structured query statement according to slot information with a confidence degree, calling a preset index database to execute the structured query statement, and inputting target index structured data obtained by the query and context information of the natural language query into a retrieval enhancement generation model to output a natural language reply. The process considers semantic coverage and entity accuracy by means of double-channel heterogeneous retrieval, and can form a precise and efficient natural language reply for an insurance index professional question and answer scene.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and can be applied to the field of financial technology. In particular, it relates to an intelligent question-answering method, apparatus, device, and medium based on retrieval-enhanced generation technology. Background Technology

[0002] With the rapid development of natural language processing technology, intelligent question-answering systems based on retrieval-enhanced generation (RAG) technology have been widely used in the financial sector. These systems combine precisely retrieved domain data with generative models, providing financial professionals, investors, and ordinary users with convenient information access channels and significantly improving the efficiency and quality of financial services.

[0003] In specific application scenarios within the financial sector, insurance industry metrics are core data reflecting the operational status, risk level, and profitability of insurance businesses. The accuracy and efficiency of these metrics directly impact insurance companies' business decisions, regulatory compliance, and customer service quality. However, most existing intelligent question-and-answer systems rely on a single search model and suffer from rigidity and deficiencies in handling metric query dimensions. They struggle to balance the precision requirements of insurance terminology with the diverse nature of users' colloquial expressions, leading to a mismatch in search results and failing to effectively meet the professional question-and-answer needs related to insurance metrics. Summary of the Invention

[0004] This invention provides an intelligent question-answering method, device, computer equipment, and medium based on retrieval enhancement generation technology, in order to solve the problem that intelligent question-answering methods in related technologies rely on a single retrieval mode and have rigid dimensional processing, which cannot effectively adapt to the professional question-answering needs of insurance indicators.

[0005] Firstly, an intelligent question-answering method based on retrieval enhancement generation technology is provided, including: The system receives a natural language query input by the user, performs a heterogeneous search on the natural language query using a dual-channel retrieval module, and obtains a first indicator metadata that matches the semantics of the natural language query, and a second indicator metadata that matches the entity fields in the natural language query. Slot filling is performed using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to obtain slot information with confidence. The slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. A structured query statement is generated based on the slot information with confidence level, and the structured query statement is executed by calling the preset indicator database to obtain the target indicator structured data. The structured data of the target indicator and the contextual information of the natural language query are input into the retrieval enhancement generation model, and a natural language response is output.

[0006] Secondly, a smart question-answering device based on retrieval enhancement generation technology is provided, including: The retrieval module is used to receive natural language queries input by users, perform heterogeneous retrieval on the natural language queries using a dual-channel retrieval module, and obtain first indicator metadata that matches the semantics of the natural language queries, and second indicator metadata that matches the entity fields in the natural language queries. The filling module is used to perform slot filling using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata, to obtain slot information with confidence, wherein the slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. The execution module is used to generate a structured query statement based on the slot information with confidence level, call the preset indicator database to execute the structured query statement, and obtain the target indicator structured data; The generation module is used to input the structured data of the target indicator and the context information of the natural language query into the retrieval enhancement generation model and output a natural language response.

[0007] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the intelligent question-answering method based on retrieval enhancement generation technology described above.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the intelligent question-answering method based on retrieval enhancement generation technology described above.

[0009] In the aforementioned intelligent question-answering method, device, computer equipment, and storage medium based on retrieval enhancement generation technology, the server can receive natural language queries input by the user, perform heterogeneous retrieval on the natural language queries using a dual-channel retrieval module to obtain first indicator metadata that matches the semantics of the natural language queries, and second indicator metadata that matches the entity fields in the natural language queries; use candidate indicator metadata obtained by weighted fusion of the first and second indicator metadata to perform slot filling, obtaining slot information with confidence scores, including indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions; generate structured query statements based on the slot information with confidence scores, call a preset indicator database to execute the structured query statements, and obtain target indicator structured data; input the target indicator structured data and the context information of the natural language queries into the retrieval enhancement generation model, output a natural language response, and send the natural language response to the client for display. In this invention, a dual-channel heterogeneous retrieval system balances semantic coverage and entity accuracy, improving the retrieval quality of indicator metadata. Furthermore, based on the dual-channel heterogeneous retrieval results, structured query statements are generated using slot information with confidence levels. This directly transforms non-standardized natural language requirements into standard query instructions adapted to a pre-defined insurance indicator database, directly calling the database to execute queries, significantly improving retrieval efficiency and ensuring the accuracy of the structured indicator data. Finally, the structured data and query context are input into the retrieval enhancement generation model. On the one hand, the professionalism of the structured data ensures that the response content aligns with the professional needs of insurance indicator queries; on the other hand, the contextual information ensures that the response matches the user's true query intent, achieving a precise balance between professional data and natural expression. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of an application environment for an intelligent question-answering method based on retrieval enhancement generation technology according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an intelligent question-answering method based on retrieval enhancement generation technology in one embodiment of the present invention; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S10; Figure 4 yes Figure 2A schematic diagram of a specific implementation method for step S20; Figure 5 This is a flowchart illustrating an intelligent question-answering method based on retrieval enhancement generation technology in another embodiment of the present invention; Figure 6 This is a flowchart illustrating an intelligent question-answering method based on retrieval enhancement generation technology in another embodiment of the present invention; Figure 7 yes Figure 5 A schematic diagram of a specific implementation method for step S50; Figure 8 yes Figure 5 A schematic diagram of a specific implementation method for step S60; Figure 9 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 10 This is a schematic diagram of a smart question-answering device based on retrieval enhancement generation technology in one embodiment of the present invention; Figure 11 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 12 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] The intelligent question-answering method based on retrieval enhancement generation technology provided in this invention can be applied to, for example... Figure 1In this application environment, the server receives a natural language query input by the user, performs a heterogeneous retrieval on the natural language query using a dual-channel retrieval module, and obtains first indicator metadata that matches the semantics of the natural language query, and second indicator metadata that matches the entity fields in the natural language query. The server then uses the candidate indicator metadata obtained by weighted fusion of the first and second indicator metadata to perform slot filling, obtaining slot information with confidence levels. This slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. Based on the slot information with confidence levels, a structured query statement is generated, and the structured query statement is executed by calling a preset indicator database to obtain the target indicator structured data. The target indicator structured data and the context information of the natural language query are input into the retrieval enhancement generation model, which outputs a natural language response and sends the natural language response to the client for display. In this invention, a dual-channel heterogeneous retrieval system balances semantic coverage and entity accuracy, improving the retrieval quality of indicator metadata. Furthermore, based on the dual-channel heterogeneous retrieval results, structured query statements are generated using slot information with confidence levels. This directly transforms non-standardized natural language requirements into standard query instructions adapted to a pre-defined insurance indicator database, directly calling the database to execute queries, significantly improving retrieval efficiency and ensuring the accuracy of the structured indicator data. Finally, the structured data and query context are input into the retrieval enhancement generation model. On the one hand, the professionalism of the structured data ensures that the response content matches the professional needs of insurance indicator queries; on the other hand, the contextual information ensures that the response matches the user's true query intent, achieving a precise balance between professional data and natural expression. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0014] Please see Figure 2 As shown, Figure 2 A flowchart illustrating an intelligent question-answering method based on retrieval enhancement generation technology provided in an embodiment of the present invention includes the following steps: S10: Receive a natural language query input by the user, perform heterogeneous retrieval on the natural language query using a dual-channel retrieval module, and obtain first indicator metadata that matches the semantics of the natural language query, and second indicator metadata that matches the entity fields in the natural language query.

[0015] The intelligent question-answering method based on retrieval enhancement generation technology provided by this invention can be applied to financial question-answering scenarios. Generally speaking, the financial field is complex and highly professional, and ordinary users find it difficult to accurately express their query needs. Through the collaboration of natural language understanding and precise retrieval, the professional needs behind colloquial and vague expressions can be accurately identified, transforming non-standardized queries into precise professional information retrieval and lowering the threshold for accessing financial information.

[0016] Natural language queries are queries that use everyday spoken language and non-standard written language input. They do not need to follow professional terminology standards or fixed query formats. The core is to fit the user's natural expression habits.

[0017] In this embodiment, the dual-channel retrieval module integrates two independent sub-channels. These two sub-channels can be started synchronously and operate in parallel, achieving heterogeneous retrieval collaboration. Specifically, asynchronous retrieval refers to the two retrieval channels employing differentiated retrieval logic and matching rules, rather than a single retrieval mode. By complementing each other with two different retrieval methods, both comprehensiveness and accuracy of the retrieval are ensured, avoiding the missed or false detection problems of a single retrieval method. The corresponding retrieval yields indicator metadata, which is basic descriptive data for financial indicators, including indicator name, indicator definition, related dimensions, statistical caliber, data source, and other information.

[0018] In practical applications, the aforementioned dual-channel retrieval module includes a semantic vector retrieval channel and an entity field retrieval channel. Specifically, for example... Figure 3 As shown, step S10, which involves using a dual-channel retrieval module to perform a heterogeneous retrieval on the natural language query to obtain first indicator metadata that matches the semantics of the natural language query, and second indicator metadata that matches the entity fields in the natural language query, includes the following steps: S11. Through the semantic vector retrieval channel, the natural language query is converted into a semantic vector, and vector similarity matching is performed in the preset indicator vector library to obtain indicator metadata with a similarity greater than a first threshold, which serves as the first indicator metadata that semantically matches the natural language query.

[0019] S12. Extract entity fields from the natural language query through the keyword field retrieval channel, match them in the preset index keyword index library, and obtain index metadata with a matching degree greater than the second threshold as the second index metadata that matches the entity fields in the natural language query.

[0020] The specific semantic vector retrieval channel inputs a natural language query into a preset semantic encoding model. This model transforms the standardized query text into a high-dimensional semantic vector, which accurately represents the user's core query intent. Then, it retrieves a preset indicator metadata vector library and calculates the similarity between the semantic vector and the indicator metadata vectors in the library. This library contains semantic vectors of the same dimension after transformation from all indicator metadata. The library filters indicator metadata with a similarity greater than a first threshold to obtain the first indicator metadata that semantically matches the natural language query. The core characteristic of this first indicator metadata is its high degree of matching with the core semantics of the natural language query; it is not limited to literal wording but focuses on the user's core query intent, serving as a fallback to cover diverse query needs.

[0021] The specific keyword field retrieval channel can extract entity fields from natural language queries based on a domain-specific model. These entity fields include, but are not limited to, indicator-type entities and dimension-type entities. Indicator-type entities can be loss ratios, premium income, surrender rates, etc., while dimension-type entities can be insurance types, time periods, regions, etc. Then, a preset indicator metadata keyword retrieval database is retrieved, and the entity fields are matched with the indicator metadata in the database. Indicator metadata with a matching degree greater than a second threshold is filtered out to obtain second indicator metadata that matches the entity fields in the natural language query. The core feature of this second indicator metadata is its precise matching with the core entity fields in the natural language query, focusing on the specific objects and constraints of the user's query, used to lock in valid results and filter out invalid interference.

[0022] S20: Use the candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to perform slot filling, and obtain slot information with confidence.

[0023] In this embodiment, the slot filling process is the core link connecting candidate indicator metadata retrieval and subsequent precise query. Its purpose is to transform unstructured candidate indicator metadata into structured and verifiable slot information. By clarifying the correspondence between indicators and constraint dimensions and quantifying the filling reliability, it provides accurate input for subsequent dimension missing determination and structured query instruction generation.

[0024] Specifically, such as Figure 4 As shown, in step S20, which involves using the candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to perform slot filling and obtain slot information with confidence, the following steps are included: S21: Parse the candidate indicator metadata and extract the explicit constraint dimension information associated with the indicator information.

[0025] S22: Call the preset slot template, fill the corresponding indicator information into the indicator slot, and fill the corresponding explicit constraint dimension information into the dimension slot.

[0026] S23: Calculate the filling confidence of the indicator slot and the filling confidence of the dimension slot based on the indicator dimension co-occurrence probability model, and integrate the filled indicator slot, dimension slot and corresponding confidence to obtain slot information with confidence.

[0027] Specifically, in step S21, candidate indicator metadata, obtained by weighted fusion of the dual-channel retrieval results, is received. This candidate metadata has passed fusion screening and possesses high matching degree and high reliability. Further, based on preset parsing rules for financial indicator metadata, the candidate indicator metadata undergoes field splitting and information extraction. Indicator information refers to the unique identifier information representing the core target of the query in the candidate metadata, such as loss ratio and premium income. During extraction, it needs to be standardized using domain terminology to unify synonymous expressions and avoid ambiguity. Explicit constraint dimension information refers to the explicit indicator constraints and corresponding values ​​that exist in the candidate metadata and do not require derivation. In this embodiment, adapted to financial indicator query scenarios, explicit constraint dimensions can be insurance type (e.g., auto insurance, life insurance), time (e.g., Q3 2024, first half of the year), region (e.g., nationwide, Jiangsu Province), and channel (e.g., online channels, offline agents). During extraction, the dimension type and corresponding values ​​need to be obtained simultaneously. Finally, the extracted indicator information is bound to the explicit constraint dimension information to ensure that the dimension information is a specific constraint for the corresponding indicator, avoiding confusion between dimension information of different indicators.

[0028] Specifically, in step S22, a preset indicator query slot template can be retrieved. This slot template is a structured template adapted to financial scenarios, containing only two core slots: indicator slots and dimension slots. The two types of slots have a one-to-one binding relationship, meaning that one indicator slot corresponds to a set of dimension slots. Dimension slots support parallel filling of multiple dimensions, and each dimension slot is labeled with its dimension type, such as insurance type dimension slot, time dimension slot, etc., to ensure filling standardization. Further, based on the slot definitions in the slot template, the indicator information and explicit constraint dimension information are filled into slots separately: the indicator information is uniquely filled into the indicator slot in the slot template, and only one indicator is allowed to be filled into a single slot, ensuring the uniqueness of the query target; the explicit constraint dimension information is filled into the corresponding dimension slot in the slot template according to the dimension type matching principle. For example, "insurance type - auto insurance" is filled into the "insurance type dimension slot", and "time - Q3 of 2024" is filled into the "time dimension slot".

[0029] Specifically, in step S23, a pre-trained indicator dimension co-occurrence probability model can be retrieved. The training data for this model comes from massive historical indicator query data in the financial field, industry indicator statistical rule data, and indicator-dimension association annotation data. Its function is to quantify the reliability of slot filling results through statistical probability. For the process of calculating the confidence of indicator slot filling, the filling content of the indicator slot can be input into the model along with the original indicator field in the candidate indicator metadata. The model calculates the confidence based on the "indicator name matching degree". The higher the matching degree, the closer the confidence is to 1. For example, when the "COR indicator" completely matches the original indicator in the metadata, the confidence can reach 0.98. The process of calculating the confidence score for dimension slot filling involves inputting the filling content of each dimension slot and the indicator information of the indicator slot into the model. The model calculates the confidence score based on the historical co-occurrence frequency and industry statistical correlation between the dimension and the indicator. For example, "auto insurance" and "COR indicator" have a high co-occurrence frequency, and the corresponding dimension slot position confidence score can reach 0.95, while "region - Jiangsu Province" has a low correlation with "COR indicator," and the confidence score may be 0.6. The filled indicator slots and each dimension slot are then bound to the corresponding calculated filling confidence scores to obtain slot information with confidence scores.

[0030] In practical applications, slot information with confidence levels relies on explicit constraint dimensions from the candidate indicator metadata. This means that only explicit dimensions explicitly mentioned in user queries and carried by the metadata can be extracted, resulting in a lack of implicit dimensions. To ensure that slot information meets the dimensional completeness requirements for financial indicator queries, further, such as... Figure 5 As shown, after step S20, the method further includes the following steps: S50: Using a pre-constructed dimensional correlation matrix, detect whether the slot information with confidence scores has missing dimensions.

[0031] S60: If the slot information has missing dimensions, the slot information with confidence is input into the Generative Adversarial Network (GAN) to dynamically complete the slot information and obtain complete slot information.

[0032] The dimensional association matrix, constructed based on a domain knowledge graph, is used to define dimensions with strong dependencies between metrics and dimensions. The specific process of constructing the dimensional association matrix is ​​as follows: Figure 6 As shown, correspondingly, before step S50, the method further includes the following steps: S70: Obtain standardized domain data, extract the indicator entities of the domain data as indicator nodes, extract the dimension entities of the domain data as dimension nodes, and extract the business relationship between indicators and dimensions as the relationship edges between nodes.

[0033] S80: Construct a domain knowledge graph based on the indicator nodes, the dimension nodes, and the associated edges.

[0034] S90: Construct an initial matrix with the index node as the matrix row dimension and the dimension node as the matrix column dimension, and assign the association weights corresponding to the association edges in the knowledge graph to the corresponding row and column intersection positions of the initial matrix to obtain the dimension association matrix.

[0035] The domain knowledge graph is used to represent the association logic between domain indicators and dimensions. The association logic includes whether there is a business relationship between the indicator and the dimension, that is, whether a certain indicator needs to be constrained by the corresponding dimension. It also includes the degree of closeness of the association between the indicator and the dimension, that is, the strength is represented by quantifying the association weight. The higher the weight, the closer the association. For example, the association weight between the COR indicator and the time dimension is 0.98, indicating that the association is extremely strong. The association weight between the COR indicator and the geographic dimension is 0.5, indicating that the association is relatively weak.

[0036] Correspondingly, such as Figure 7 As shown, step S50, which involves using a pre-constructed dimensional correlation matrix to detect whether the slot information with confidence scores has missing dimensions, includes the following steps: S51: Extract the target metrics and filled dimensions from the slot information with confidence levels.

[0037] S52: Match the target indicator with the row dimensions of the dimension association matrix, lock the target row corresponding to the target indicator, and extract the dimensions marked with strong dependencies in the target row as mandatory dimensions.

[0038] S53: Compare the filled dimensions with the required dimensions. If there are any unfilled required dimensions, it is determined that the slot information has missing dimensions.

[0039] In this embodiment, the target metric is the standardized domain metric filled in the metric slot. Each slot corresponds to a unique target metric, ensuring the uniqueness of the judgment object. The filled dimensions are all explicit constraint dimensions and their corresponding values ​​filled in the dimension slots. These dimensions originate from explicit information in the candidate metric metadata, either mentioned in user queries or inherent to the metadata itself. To avoid interference from invalid data and ensure that the dimensions participating in the comparison have actual query value, the extracted filled dimensions can be effectively filtered based on the confidence level of the slot information. Dimensions with a confidence level greater than a set confidence threshold are considered valid filled dimensions, while low-reliability filled dimensions are discarded.

[0040] Specifically, in step S52, the row dimensions of the dimension association matrix are pre-mapped with all domain indicators. Using the target indicator extracted above as the matching keyword, it is matched against the matrix row dimensions. Each target indicator corresponds to a row in the matrix, with no duplicate matches or failed matches, ensuring the uniqueness and determinism of the benchmark. After a successful match, the target row corresponding to the target indicator in the matrix is ​​locked. This target row stores the association strength information between the target indicator and all constraint dimensions. According to the matrix's preset rules, dimensions marked with strong dependencies in the target row are extracted, i.e., column dimensions with matrix element weights ≥ preset thresholds. These dimensions are essential for indicator queries as defined by financial industry business rules and indicator statistical specifications; their absence would prevent accurate data retrieval and statistics. Finally, the extracted strong dependency dimensions are integrated to form a list of required dimensions.

[0041] Understandably, if the valid filled dimensions completely cover the list of required dimensions, and no required dimension is missing, then the slot information is determined to have no missing dimensions. This indicates that the slot information has complete core dimension constraints and can be directly input into the subsequent structured query process. Conversely, if the valid filled dimensions do not completely cover the list of required dimensions, and at least one required dimension is missing, then the slot information is determined to have missing dimensions. The uncovered required dimensions are the implicit dimensions to be filled, and the required list is output simultaneously as the precise target for subsequent dynamic completion by the adversarial generative network.

[0042] Correspondingly, such as Figure 8 As shown, step S60 involves inputting the slot information with confidence levels into the Generative Adversarial Network (GAN) to dynamically complete the slot information and obtain the complete slot information. This includes the following steps: S61: Input the slot information with confidence and the list of dimensions to be completed into the generator of the adversarial generative network, and the generator generates candidate dimensions to be completed based on the domain business rules. S62: Combine the candidate completion dimension with the original slot information with confidence to form the slot information to be judged, and input it into the discriminator of the generative adversarial network. The discriminator performs business logic judgment on the slot information to be judged until the judgment meets the compliance conditions, and then outputs the complete slot information.

[0043] In this embodiment, slot dynamic completion relies on the collaborative work of the generator and discriminator of the Generative Adversarial Network. The purpose is to generate compliant dimensions that conform to domain business rules and adapt to the query requirements of indicators for the dimensions to be completed based on the dimension missing detection, thereby completing the integrity of the slot information and ensuring that the completed slot information can directly support accurate structured queries, thus solving the pain points of rigidity and poor adaptability of traditional rule-based completion.

[0044] Specifically, in step S61, the generator of the adversarial generative network has been pre-trained based on massive amounts of data in the financial field. Accordingly, after receiving slot information with confidence and a list of dimensions to be completed, it first extracts the types of dimensions to be completed to clarify the generation object and retrieves the corresponding domain business rules; then, it loads the rules to define compliance boundaries and removes non-compliant dimension types and invalid values ​​according to the set sorting rules; in the generation stage, it first generates basic compliant candidate dimensions according to the rules, and then combines the target indicator adaptation rules and the already filled dimensions for optimization and screening, retaining highly adaptable values ​​and generating 1-3 sets of differentiated candidate values; finally, it performs a quality screening, removes candidate values ​​with insufficient adaptability, assigns an initial confidence level, and outputs a set of candidate completed dimensions for subsequent discriminator verification.

[0045] Specifically, in step S62, the discriminator ensures that the output slot information conforms to the business logic of the financial field through multi-dimensional compliance verification and iterative optimization. Accordingly, upon receiving the slot information to be judged, several judgment logics are set: first, dimension value compliance, used to verify whether the dimension values ​​conform to industry statistical standards (e.g., time is a valid period); second, indicator-dimensional adaptability, used to confirm that the completed dimension is strongly correlated with the target indicator; and third, dimension combination consistency, used to check for logical conflicts between the completed dimension and the originally filled dimension. If the information to be judged passes all verifications, it meets the compliance conditions, and the complete slot information is directly output; if it fails, the discriminator outputs specific violation feedback and sends it back to the generator. The generator further adjusts its strategy to regenerate candidate dimensions, combines the new information to be judged, and inputs it back into the discriminator, repeating the above process until the judgment meets the compliance conditions, outputting the final complete slot information, providing high-quality input for subsequent structured queries.

[0046] S30: Generate a structured query statement based on the slot information with confidence level, call the preset indicator database to execute the structured query statement, and obtain the target indicator structured data.

[0047] In this embodiment, a structured query statement is generated based on slot information with confidence level and the target indicator data is retrieved. The core is to transform the structured slot information into a standardized query instruction that adapts to the preset indicator database, so as to achieve accurate and efficient retrieval of financial indicator data.

[0048] Specifically, such as Figure 9 As shown, step S30, which involves generating a structured query statement based on the slot information with confidence levels, executing the structured query statement by calling a preset indicator database, and obtaining the target indicator structured data, includes the following steps: S31: Analyze the slot information with confidence level, and extract the target indicators, effective dimension constraints and corresponding values.

[0049] S32: Based on the data storage specifications of the preset indicator database, the target indicator is used as the core of the query, and the effective dimension constraints and corresponding values ​​are used as the query filtering conditions to construct a structured query statement that is adapted to the preset indicator database.

[0050] S33: Call the preset indicator database to execute the structured query statement to obtain the target indicator structured data.

[0051] In step S31, the effective dimension constraints include the constraint values ​​corresponding to the filled explicit dimensions and the completed implicit dimensions, and the confidence level of all dimensions is not lower than a set threshold. Here, the filled explicit dimensions refer to the explicit constraint dimensions and their corresponding values ​​extracted from the original slot information with confidence, those mentioned in user queries, or those inherent in candidate metadata. For example, insurance type = car insurance, channel = offline. The completed implicit dimensions refer to the strongly dependent dimensions and their corresponding values ​​that are required for indicator queries but are missing in the original slots, completed by an adversarial generative network. For example, statistical period = 2024Q3.

[0052] In step S32, the construction of the structured query statement is based on the preset index database storage specification to ensure that the statement is compatible with the database execution logic. First, the database storage specification is loaded, the correspondence between the target index and the database table fields is extracted, and the matching rules between the effective dimension types / values ​​and the database filter fields / standardized values ​​are extracted. Then, the target index is mapped to the core query fields of the query statement, and the effective dimension constraints and values ​​are converted into standardized filter clauses. Finally, the fields and clauses are assembled according to the database syntax. The built-in validation module verifies the syntax integrity and field matching of the statement, corrects problems such as non-standard values ​​and syntax errors, and generates a structured query statement that can be executed directly.

[0053] In step S33, a compliant structured query statement can be sent to the execution engine of a preset indicator database. Here, the database is a standardized library in the financial field, which stores data in a three-dimensional structure of indicator-dimension-value, and supports precise conditional filtering. After receiving the statement, the execution engine parses the target indicator and effective dimension constraints, retrieves matching data in the database according to the effective dimension constraints, and filters low-quality data with confidence scores below a preset threshold to obtain the structured data of the target indicator.

[0054] S40: Input the structured data of the target indicator and the context information of the natural language query into the retrieval enhancement generation model, and output a natural language response.

[0055] In this embodiment, in order to ensure the validity and consistency of the model input, the structured data of the target indicator and the context information of the natural language query can be concatenated according to a preset format to form the integrated input data.

[0056] The integrated input data is then fed into the retrieval enhancement and generation model, which integrates two stages: retrieval enhancement and fusion generation. In the retrieval enhancement stage, the model calls upon its built-in financial domain knowledge base to retrieve matching enhancement information based on the input indicator information and contextual information, providing professional basis for response generation and avoiding data interpretation bias. In the fusion generation stage, the model deeply integrates the retrieved enhancement information with the input structured data and contextual context, organizing the content according to natural language expression logic. During generation, data accuracy is prioritized while adapting to the user's context to ensure that the response is easy to understand and fits the query requirements.

[0057] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0058] In one embodiment, an intelligent question-answering device based on retrieval enhancement generation technology is provided, which corresponds one-to-one with the intelligent question-answering method based on retrieval enhancement generation technology in the above embodiments.

[0059] Specifically, such as Figure 10 As shown, the intelligent question-answering device based on retrieval enhancement generation technology includes: a retrieval module 101, a filling module 102, an execution module 103, and a generation module 104. Detailed descriptions of each functional module are as follows: The retrieval module 101 is used to receive a natural language query input by the user, perform heterogeneous retrieval on the natural language query using a dual-channel retrieval module, and obtain first indicator metadata that matches the semantics of the natural language query, and second indicator metadata that matches the entity fields in the natural language query. The filling module 102 is used to perform slot filling using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata, to obtain slot information with confidence, wherein the slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. The execution module 103 is used to generate a structured query statement based on the slot information with confidence level, call the preset indicator database to execute the structured query statement, and obtain the target indicator structured data. The generation module 104 is used to input the structured data of the target indicator and the context information of the natural language query into the retrieval enhancement generation model and output a natural language response.

[0060] In one embodiment, the dual-channel retrieval module includes a semantic vector retrieval channel and a keyword field retrieval channel. The retrieval module 101 is specifically used for: The natural language query is converted into a semantic vector through the semantic vector retrieval channel. Vector similarity matching is performed in the preset indicator vector library to obtain indicator metadata with similarity greater than a first threshold, which serves as the first indicator metadata that semantically matches the natural language query. By using the keyword field retrieval channel, entity fields in the natural language query are extracted and matched in a preset index keyword index library to obtain index metadata with a matching degree greater than a second threshold, which serves as the second index metadata that matches the entity fields in the natural language query.

[0061] In one embodiment, the filling module 102 is specifically used for: Parse the candidate indicator metadata to extract the explicit constraint dimension information associated with the indicator information; Call the preset slot template, fill the corresponding indicator information into the indicator slot, and fill the corresponding explicit constraint dimension information into the dimension slot; Based on the co-occurrence probability model of the indicator dimensions, the filling confidence of the indicator slot and the filling confidence of the dimension slot are calculated respectively. The filled indicator slot, dimension slot and corresponding confidence are integrated to obtain the slot information with confidence.

[0062] In one embodiment, the device further includes: The detection module is used to perform slot filling on the candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to obtain slot information with confidence, and then use a pre-constructed dimension association matrix to detect whether there are missing dimensions in the slot information with confidence. The dimension association matrix is ​​constructed based on the domain knowledge graph and is used to define the dimensions that have strong dependencies between indicators and dimensions. The completion module is used to input the slot information with confidence into the Generative Adversarial Network (GAN) if the slot information has missing dimensions, so as to dynamically complete the slot information through the GAN to obtain complete slot information.

[0063] In one embodiment, the device further includes: Before using the pre-constructed dimension association matrix to detect whether there are missing dimensions in the slot information with confidence, standardized domain data is obtained, the indicator entities of the domain data are extracted as indicator nodes, the dimension entities of the domain data are extracted as dimension nodes, and the business association relationship between indicators and dimensions is extracted as the association edge between nodes. Based on the indicator nodes, the dimension nodes, and the associated edges, a domain knowledge graph is constructed, which is used to represent the association logic between domain indicators and dimensions. An initial matrix is ​​constructed using the indicator nodes as the matrix row dimension and the dimension nodes as the matrix column dimension. The association weights corresponding to the association edges in the knowledge graph are assigned to the corresponding row and column intersection positions of the initial matrix to obtain the dimension association matrix. Accordingly, the detection module is specifically used for: Extract the target metrics and filled dimensions from the slot information with confidence levels; Match the target metric with the row dimensions of the dimension association matrix, lock the target row corresponding to the target metric, and extract the dimensions marked with strong dependencies in the target row as the required dimensions. By comparing the filled dimensions with the required dimensions, if there are any unfilled required dimensions, it is determined that the slot information has missing dimensions.

[0064] In one embodiment, the completion module is specifically used for: The slot information with confidence and the list of dimensions to be completed are input into the generator of the generative adversarial network, and the generator generates candidate dimensions to be completed based on domain business rules. The candidate completion dimension is combined with the original slot information with confidence to form the slot information to be judged, which is then input into the discriminator of the generative adversarial network. The discriminator performs business logic judgment on the slot information to be judged until the judgment meets the compliance conditions, and then outputs the complete slot information.

[0065] In one embodiment, the execution module 103 is specifically used for: The slot information with confidence scores is analyzed to extract target indicators, effective dimension constraints and corresponding values. The effective dimension constraints include the constraint values ​​corresponding to the filled explicit dimensions and the completed implicit dimensions, and the confidence scores of all dimensions are not lower than the set threshold. Based on the data storage specifications of the preset indicator database, the target indicator is used as the core of the query, and the effective dimension constraints and corresponding values ​​are used as the query filtering conditions to construct a structured query statement that is adapted to the preset indicator database. The structured query statement is executed by calling the preset indicator database to obtain the structured data of the target indicator.

[0066] This invention provides an intelligent question-answering device based on retrieval enhancement generation technology. Through dual-channel heterogeneous retrieval, it balances semantic coverage and entity accuracy, improving the retrieval quality of indicator metadata. Furthermore, based on the dual-channel heterogeneous retrieval results, it generates structured query statements using slot information with confidence levels. This directly transforms non-standardized natural language requirements into standard query instructions adapted to a pre-defined insurance indicator database, directly calling the database to execute queries, significantly improving retrieval efficiency and ensuring the accuracy of the structured indicator data. Finally, the structured data and query context are input into the retrieval enhancement generation model. On the one hand, the professionalism of the structured data ensures that the response content aligns with the professional needs of insurance indicator queries; on the other hand, the contextual information ensures that the response matches the user's true query intent, achieving a precise balance between professional data and natural expression.

[0067] Specific limitations regarding the intelligent question-answering device based on retrieval enhancement generation technology can be found in the limitations of the intelligent question-answering method based on retrieval enhancement generation technology described above, and will not be repeated here. Each module in the aforementioned intelligent question-answering device based on retrieval enhancement generation technology can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0068] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 11 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side intelligent question-answering method based on retrieval enhancement generation technology.

[0069] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 12As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of an intelligent question-answering method based on retrieval enhancement generation technology.

[0070] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: The system receives a natural language query input by the user, performs a heterogeneous search on the natural language query using a dual-channel retrieval module, and obtains a first indicator metadata that matches the semantics of the natural language query, and a second indicator metadata that matches the entity fields in the natural language query. Slot filling is performed using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to obtain slot information with confidence. The slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. A structured query statement is generated based on the slot information with confidence level, and the structured query statement is executed by calling the preset indicator database to obtain the target indicator structured data. The structured data of the target indicator and the contextual information of the natural language query are input into the retrieval enhancement generation model, and a natural language response is output.

[0071] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: The system receives a natural language query input by the user, performs a heterogeneous search on the natural language query using a dual-channel retrieval module, and obtains a first indicator metadata that matches the semantics of the natural language query, and a second indicator metadata that matches the entity fields in the natural language query. Slot filling is performed using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to obtain slot information with confidence. The slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. A structured query statement is generated based on the slot information with confidence level, and the structured query statement is executed by calling the preset indicator database to obtain the target indicator structured data. The structured data of the target indicator and the contextual information of the natural language query are input into the retrieval enhancement generation model, and a natural language response is output.

[0072] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0073] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0074] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0075] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An intelligent question-answering method based on retrieval enhancement generation technology, characterized in that, include: The system receives a natural language query input by the user, performs a heterogeneous search on the natural language query using a dual-channel retrieval module, and obtains a first indicator metadata that matches the semantics of the natural language query, and a second indicator metadata that matches the entity fields in the natural language query. Slot filling is performed using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to obtain slot information with confidence. The slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. A structured query statement is generated based on the slot information with confidence level, and the structured query statement is executed by calling the preset indicator database to obtain the target indicator structured data. The structured data of the target indicator and the contextual information of the natural language query are input into the retrieval enhancement generation model, and a natural language response is output.

2. The intelligent question-answering method based on retrieval enhancement generation technology as described in claim 1, characterized in that, The dual-channel retrieval module includes a semantic vector retrieval channel and a keyword field retrieval channel. The dual-channel retrieval module performs heterogeneous retrieval on the natural language query to obtain first indicator metadata that semantically matches the natural language query, and second indicator metadata that matches entity fields in the natural language query, including: The natural language query is converted into a semantic vector through the semantic vector retrieval channel. Vector similarity matching is performed in the preset indicator vector library to obtain indicator metadata with similarity greater than a first threshold, which serves as the first indicator metadata that semantically matches the natural language query. By using the keyword field retrieval channel, entity fields in the natural language query are extracted and matched in a preset index keyword index library to obtain index metadata with a matching degree greater than a second threshold, which serves as the second index metadata that matches the entity fields in the natural language query.

3. The intelligent question-answering method based on retrieval enhancement generation technology as described in claim 1, characterized in that, The process of filling slots using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to obtain slot information with confidence includes: Parse the candidate indicator metadata to extract the explicit constraint dimension information associated with the indicator information; Call the preset slot template, fill the corresponding indicator information into the indicator slot, and fill the corresponding explicit constraint dimension information into the dimension slot; Based on the co-occurrence probability model of the indicator dimensions, the filling confidence of the indicator slot and the filling confidence of the dimension slot are calculated respectively. The filled indicator slot, dimension slot and corresponding confidence are integrated to obtain the slot information with confidence.

4. The intelligent question-answering method based on retrieval enhancement generation technology as described in claim 1, characterized in that, After performing slot filling using the candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata to obtain slot information with confidence, the method further includes: Using a pre-constructed dimensional association matrix, the system detects whether there are missing dimensions in the slot information with confidence scores. The dimensional association matrix is ​​constructed based on a domain knowledge graph and is used to define dimensions that have strong dependencies on indicators. If the slot information has missing dimensions, the slot information with confidence is input into the Generative Adversarial Network (GAN) to dynamically complete the slot information and obtain complete slot information.

5. The intelligent question-answering method based on retrieval enhancement generation technology as described in claim 4, characterized in that, Before detecting whether the slot information with confidence scores has missing dimensions using a pre-constructed dimensional correlation matrix, the method further includes: Obtain standardized domain data, extract the indicator entities from the domain data as indicator nodes, extract the dimension entities from the domain data as dimension nodes, and extract the business relationship between indicators and dimensions as the relationship edges between nodes; Based on the indicator nodes, the dimension nodes, and the associated edges, a domain knowledge graph is constructed, which is used to represent the association logic between domain indicators and dimensions. An initial matrix is ​​constructed using the indicator nodes as the matrix row dimension and the dimension nodes as the matrix column dimension. The association weights corresponding to the association edges in the knowledge graph are assigned to the corresponding row and column intersection positions of the initial matrix to obtain the dimension association matrix. Accordingly, the step of using a pre-constructed dimensional correlation matrix to detect whether the slot information with confidence scores has missing dimensions includes: Extract the target metrics and filled dimensions from the slot information with confidence levels; Match the target metric with the row dimensions of the dimension association matrix, lock the target row corresponding to the target metric, and extract the dimensions marked with strong dependencies in the target row as the required dimensions. By comparing the filled dimensions with the required dimensions, if there are any unfilled required dimensions, it is determined that the slot information has missing dimensions.

6. The intelligent question-answering method based on retrieval enhancement generation technology as described in claim 4, characterized in that, The step of inputting the slot information with confidence scores into the Generative Adversarial Network (GAN) to dynamically complete the slot information through the GAN to obtain complete slot information includes: The slot information with confidence and the list of dimensions to be completed are input into the generator of the generative adversarial network, and the generator generates candidate dimensions to be completed based on domain business rules. The candidate completion dimension is combined with the original slot information with confidence to form the slot information to be judged, which is then input into the discriminator of the generative adversarial network. The discriminator performs business logic judgment on the slot information to be judged until the judgment meets the compliance conditions, and then outputs the complete slot information.

7. The intelligent question-answering method based on retrieval enhancement generation technology as described in any one of claims 1-6, characterized in that, The step involves generating a structured query statement based on the slot information with confidence levels, executing the structured query statement by calling a preset indicator database, and obtaining structured data of the target indicator, including: The slot information with confidence scores is analyzed to extract target indicators, effective dimension constraints and corresponding values. The effective dimension constraints include the constraint values ​​corresponding to the filled explicit dimensions and the completed implicit dimensions, and the confidence scores of all dimensions are not lower than the set threshold. Based on the data storage specifications of the preset indicator database, the target indicator is used as the core of the query, and the effective dimension constraints and corresponding values ​​are used as the query filtering conditions to construct a structured query statement that is adapted to the preset indicator database. The structured query statement is executed by calling the preset indicator database to obtain the structured data of the target indicator.

8. An intelligent question-answering device based on retrieval enhancement generation technology, characterized in that, include: The retrieval module is used to receive natural language queries input by users, perform heterogeneous retrieval on the natural language queries using a dual-channel retrieval module, and obtain first indicator metadata that matches the semantics of the natural language queries, and second indicator metadata that matches the entity fields in the natural language queries. The filling module is used to perform slot filling using candidate indicator metadata obtained by weighted fusion of the first indicator metadata and the second indicator metadata, to obtain slot information with confidence, wherein the slot information includes indicator slots for carrying indicators and dimension slots for carrying indicator constraint dimensions. The execution module is used to generate a structured query statement based on the slot information with confidence level, call the preset indicator database to execute the structured query statement, and obtain the target indicator structured data; The generation module is used to input the structured data of the target indicator and the context information of the natural language query into the retrieval enhancement generation model and output a natural language response.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent question-answering method based on retrieval enhancement generation technology as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent question-answering method based on retrieval enhancement generation technology as described in any one of claims 1 to 7.

Images