Interactive financial data retrieval and analysis system based on big data

Abstract

The invention relates to the technical field of financial data analysis, in particular to an interactive financial data retrieval and analysis system based on big data. Comprising a data access module, a data unification module, a semantic understanding and intention recognition module, a retrieval execution module, a financial analysis module and an interactive visualization module. The system can access financial information from multi-source data and perform cleaning, standardization, main data alignment and management to generate data for query and analysis. Query intentions and entities are recognized through natural language analysis, structured semantic representation is constructed, and multi-dimensional aggregation, backtracking and predictive analysis are achieved. And the retrieval execution module and the financial analysis module cooperate to ensure that the result is reproducible and credible. And the interactive visualization module dynamically presents the analysis result. According to the method, accurate understanding and structured representation of natural language query of the user are realized, so that financial data retrieval can quickly respond to the intention of the user, the query accuracy and efficiency are improved, and the reliability and precision of financial decisions are improved.

Description

Technical Field

[0001] This invention relates to the field of financial data analysis technology, and more specifically to an interactive financial data retrieval and analysis system based on big data. Background Technology

[0002] With the rapid growth of enterprise data and the increasing demand for intelligent financial management, financial data analysis systems are transforming from traditional manual reports and offline statistical models to big data-driven, real-time interactive intelligent analysis models. Interactive financial data retrieval and analysis systems based on big data can achieve rapid acquisition, accurate analysis, and decision support of financial information by integrating multi-source financial data, unifying data governance, natural language query parsing, dynamic query execution, and visual analysis. As the types and sources of enterprise financial data become more diverse, traditional financial systems struggle to meet the demands for real-time, accurate, and multi-dimensional analysis.

[0003] Existing technologies suffer from the following problems: In current financial systems, data from different business systems and external data sources uses different formats and interfaces, leading to complex data access, difficulties in standardization, and challenges in direct information sharing between system layers, thus limiting the efficiency of cross-departmental and cross-system analysis. Traditional financial analysis relies on static reports and manual operations, lacking natural language interaction capabilities and structured semantic parsing mechanisms, making it impossible to achieve dynamic queries and automated analysis based on user intent. Financial data quality issues are prominent, with missing values, outliers, and inconsistent formats affecting the reliability of analysis results. Existing systems lack sufficient support for the traceability, reproducibility, and incremental updates of analysis results, making it difficult to conduct reliable audits and verify results for complex multidimensional analyses. Summary of the Invention

[0004] The purpose of this invention is to provide an interactive financial data retrieval and analysis system based on big data to solve the above-mentioned technical problems. This system enables accurate understanding and structured representation of users' natural language queries, allowing financial data retrieval to quickly respond to user intent, reduce manual analysis costs, improve query accuracy and efficiency, and enhance the reliability and precision of financial decision-making.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] An interactive financial data retrieval and analysis system based on big data, the system comprising:

[0007] The data access module is used to receive financial data from multiple data sources and encapsulate the financial data into financial events with timestamps, source metadata, and verification signatures.

[0008] The data unification module is used to clean, standardize, map fields, verify quality, align master data, manage metadata, and process consistency of financial events provided by the data access module, and generate governance data that can be used for subsequent querying and analysis.

[0009] The semantic understanding and intent recognition module is used to receive natural language queries input by users, identify entities, dimensions, metrics, analysis targets and query intents in the query, and generate structured semantic representations.

[0010] The retrieval execution module is used to construct a reconstructable query graph based on the structured semantic representation, schedule, calculate and incrementally execute the governance data, and output the retrieval execution results;

[0011] The financial analysis module is used to perform multidimensional data association, indicator calculation, aggregation analysis, backtracking analysis, and predictive operations on the governance data based on the query graph and the retrieval execution results, and output the financial analysis results. The retrieval execution module works in collaboration with the financial analysis module to ensure the interpretability, reproducibility, and incremental updates of the analysis results by recording the input and output metadata and execution evidence of the query nodes.

[0012] The interactive visualization module is used to receive the retrieval execution results and the financial analysis results, convert them into a visual display format, and support users to trigger further query and analysis requests based on the visual content.

[0013] Furthermore, the data access module includes:

[0014] The multi-source data acquisition unit is used to collect financial data from internal business systems, external interfaces, and documents.

[0015] A data format parsing unit is used to parse the financial data and convert it into structured data;

[0016] The data synchronization control unit is used to perform real-time, incremental, or full data synchronization.

[0017] The basic metadata recording unit is used to generate the source system identifier, timestamp, batch identifier, and verification information of the financial event.

[0018] The operation monitoring unit is used to record the operational status information of the data acquisition and analysis process.

[0019] Furthermore, the data governance module includes:

[0020] The master data management unit is used to uniformly manage the master data entities in the financial events provided by the data access module. The master data entities include accounts, organizations, customers, suppliers and product information for unified management.

[0021] The metadata management unit is used to manage structural metadata, source metadata, and version metadata, and provides metadata references to the master data management unit.

[0022] The data quality management unit is used to perform quality verification on the completeness, consistency, and accuracy of financial events based on the source metadata.

[0023] The data consistency management unit is used to unify the reporting standards and verify cross-process consistency of financial events from different systems.

[0024] The data lineage management unit is used to establish field-level and record-level data lineages based on the source metadata.

[0025] The data version management unit is used to generate data versions and data snapshots, and to provide version benchmarks to the data quality management unit.

[0026] The data access control unit is used to implement data access control and security management.

[0027] The governance rules unit is used to provide governance rules to the master data management unit, the data quality management unit, and the data consistency management unit in accordance with accounting standards, regulatory requirements, and business rules.

[0028] Furthermore, the semantic understanding and intent recognition module includes:

[0029] The natural language input receiving unit is used to receive natural language text to be parsed;

[0030] The text normalization processing unit is used to normalize the natural language text and generate a text sequence.

[0031] Encoding representation unit, used to perform vector encoding on the text sequence based on a semantic model and generate a context vector sequence. The context vector satisfies:

[0032]

[0033] in, This represents a text encoding model with parameter θ;

[0034] The entity recognition unit is used to perform sequence labeling based on the context vector sequence to generate an entity label sequence. And calculate the optimal label sequence:

[0035]

[0036] in, The label sequence scoring function is used to aggregate and calculate the optimal label sequence, outputting an entity set.

[0037] The intent recognition unit is used to generate an intent probability vector based on the context vector sequence and to identify the user's target intent based on the softmax calculation method. :

[0038]

[0039] in, Pooling operations are performed on the context vector sequence; and These represent the weight matrix and bias vector of the intent classifier, respectively; This represents the normalization function, which converts the output into a probability distribution.

[0040] The preliminary semantic representation generation unit is used to merge the outputs of the entity recognition unit and the intent recognition unit to form a preliminary structured semantic vector.

[0041] Furthermore, the semantic understanding and intent recognition module also includes:

[0042] The entity linking and master data alignment unit is used to link the entity set with the master data entity, and to determine the optimal entity based on the following comprehensive scores of candidate entities:

[0043]

[0044] And calculate the link confidence:

[0045]

[0046] Where m represents a candidate entity mention in the entity set, and e represents a candidate entity in the master data entity. This represents the contextual semantic vector of entity mentions m. This represents the embedding representation of the candidate entity e. This represents a semantic similarity metric. Represents literal similarity based on string features. Let α, β, and γ represent the prior probability of entity e in the master data, and let α, β, and γ represent the weighting coefficients used to fuse multiple entity matching features.

[0047] The semantic slot filling and query logic generation unit is used to generate structured query logic based on intent recognition results and output logical expressions based on the generation probability model.

[0048]

[0049] in, This indicates the probability of predicting the current slot given the preceding slot and the input text.

[0050] Parse confidence assessment unit for use based on intent confidence Slot location reliability Confidence of entity links Calculate the overall analytical confidence level and generate the confidence level value:

[0051]

[0052] in, This represents the analytical confidence weighting parameter, satisfying... Used to measure the reliability of the output intent set I, as a basis for selection;

[0053] The parsing result output unit is used to select the structured query logic with the highest slot probability to generate an intent set based on the comprehensive parsing confidence and slot probability.

[0054] Furthermore, the retrieval execution module includes:

[0055] The request receiving and verification unit is used to receive the semantic vector and the intent set, verify user permissions, field access control, request format and semantic integrity, and when a missing slot is detected in the semantic vector or the intent set, perform semantic completion on the semantic vector and the intent set.

[0056] The graph compiler is used to receive the completed semantic vector and intent set and map the semantic vector and intent set into a query graph in the form of a logical directed acyclic graph. The graph compiler performs operator decomposition, type identification, data range labeling, and semantic source labeling on the semantic vector and intent set, and automatically injects privacy operators and lineage origin metadata according to the field sensitivity level to form a logical query graph with metadata and generate corresponding graph identifiers.

[0057] Cost and resource estimator for each node in the logical query graph Based on input / output operations Calculation consumption and network consumption Perform resource consumption estimation:

[0058]

[0059] The execution planner is used to query the graph and estimate the results based on the logic. Generate a physical execution plan, determine the operator fusion strategy, filter pushdown strategy, data partitioning strategy, and the binding relationship between nodes and the execution backend, and split the logical graph into multiple physically schedulable execution stages according to performance targets;

[0060] The execution scheduler is used to submit each execution stage in the physical execution plan to the corresponding execution engine for scheduling, manage parallelism, retry strategy, failure transfer, priority execution pool and resource allocation, and track the node status in real time during the execution process;

[0061] An execution unit is used to run the corresponding operator on a preset execution backend based on the physical execution plan. The execution unit performs node-level caching and partial result output processing on the operator execution process, and generates node-level data hashes based on the input semantic vector and node output.

[0062] The result fusion and refinement unit receives the node-level partial results and final results submitted by the execution unit, and merges the results based on the node execution status provided by the execution scheduling unit. During the merging process, the result fusion and refinement unit performs conflict resolution, replacement of approximate results with precise results, and assigns confidence level labels to the results based on the comprehensive resolution confidence level. The result fusion and refinement unit outputs the final retrieval execution result.

[0063] Furthermore, the financial analysis module includes:

[0064] The receiving layer is used to receive the retrieval execution results, parse the analysis dimensions, metrics, aggregation functions, filtering conditions and analysis types in the retrieval execution results, and generate a standardized execution description based on the master data entity association mapping;

[0065] The multidimensional model layer is used to build and maintain multidimensional cubes, dimension tables, metric tables, and scene caliber views, manage dimension levels and drill-down rules, and support multicaliber views and caliber transformation rules. The dimension tables are linked to the master data entities to associate cross-table semantics and source data.

[0066] The computation engine layer is used to perform multidimensional aggregation, slicing, drill-down, and summarization calculations. It adopts a calculation strategy of approximation followed by precision and incremental calculation, and uses vectorized columnar calculation and distributed parallel computing mode for analysis and processing. The operators of the analysis and processing include aggregation, window aggregation, dimension association, and approximate estimation.

[0067] The advanced analytics layer is used to perform causal simulation, time series forecasting, backtesting analysis, index attribution, anomaly detection, and scenario comparison. It manages the versions of training and inference models and outputs audit metadata. It also supports automatic selection of the optimal model based on the analytics request.

[0068] The results management and caching layer is used to manage financial analysis results, record the financial analysis results output by the computing engine layer, and make key analysis results available for subsequent backtracking and compliance checks.

[0069] The interface and presentation layer are used to provide standardized analysis results output, along with explanatory information, confidence level information, and audit information, and support front-end visualization and report export.

[0070] Furthermore, the calculation method of the calculation engine layer of the financial analysis module includes the following steps: receiving analysis requests from the semantic understanding module and the user interface, and parsing the dimensions in the analysis request. ,measure Aggregate functions It analyzes and generates standardized internal execution descriptions; it constructs multidimensional cubes and dimension tables based on the parsing results, and maps metric definitions to multidimensional models, supporting multi-caliber views and dimension hierarchy management.

[0071] Perform multidimensional analysis and aggregation:

[0072]

[0073] Where d represents the analysis dimension, m represents the metric to be aggregated, f represents the aggregation function, and slice(d) represents the slice data set under the specified dimension;

[0074]

[0075] Where w represents the time window. This represents the metric value at each point in time within the window. Indicates rolling summation;

[0076]

[0077] in This represents a slice of a lower level of dimension d. This indicates the total amount calculated by aggregation along the dimensional hierarchy;

[0078]

[0079] Where t represents the historical time range and h represents the prediction step size. This represents a time series forecasting model that outputs the predicted metric values ​​for the next h time steps.

[0080]

[0081] Where I represents the intervention operation. This represents the simulated metric under intervention conditions. Indicates the baseline metric. This indicates the difference in causal simulation calculations.

[0082] Furthermore, the interactive visualization module includes:

[0083] The front-end presentation layer is used to render the outputs of the retrieval execution module and the financial analysis module;

[0084] The interaction management layer is used to capture user operation events and generate interactive operation objects based on the events. Combining the data currently displayed by the front-end presentation layer with the master data entities and knowledge graph nodes provided by the data governance module, it generates secondary analysis suggestions to guide users to further drill down into analysis dimensions, metrics and causal paths.

[0085] The result parsing and rendering adaptation layer is used to receive the analysis results from the financial analysis module and the retrieval execution module, convert the results into a data structure that can be rendered on the front end, and support progressive updates, gradually replacing approximate results with accurate results.

[0086] The interaction triggering and query interface layer is used to transform the interaction operation object into a structured query representation that can be parsed by the semantic understanding and intent recognition module, call the semantic understanding and intent recognition module and the retrieval execution module to perform query execution, and push part and the final result to the front-end presentation layer through WebSocket;

[0087] The audit and interpretability layer is used to attach source information to each visualization result generated by the interactive visualization module and record all user interactions and analysis execution processes.

[0088] Compared with the prior art, the present invention has the following technical effects:

[0089] By constructing a semantic understanding and intent recognition module, a structured semantic representation is generated through natural language input reception, text encoding, entity recognition, intent recognition, and entity linking. The optimal intent set is selected by combining the comprehensive score of candidate entities and the analytical confidence level. This achieves accurate understanding and structured representation of user natural language queries, enabling financial data retrieval to quickly respond to user intent, reducing manual analysis costs, and improving query accuracy and efficiency.

[0090] By analyzing dimensions, measures, aggregation functions, and analysis types, and combining the construction of multidimensional cubes and dimension tables, calculations can be performed along dimensions, scrolling windows, and hierarchical aggregation. This enables refined and multi-faceted statistical and summary analysis of financial data, improving the accuracy and comprehensiveness of the analysis. Attached Figure Description

[0091] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0092] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0093] like Figure 1 The system shown is an interactive financial data retrieval and analysis system based on big data. The system includes:

[0094] The data access module receives financial data from multiple data sources and encapsulates this data into financial events with timestamps, source metadata, and verification signatures. The module collects financial data from internal business systems (such as ERP and financial systems), external interfaces (such as bank interfaces and third-party data providers), and documents (such as CSV and Excel reports). It parses the collected raw data, converting unstructured or semi-structured data into a unified structured data format. The parsed data is then encapsulated into financial events, accompanied by timestamps, source metadata, and verification signatures, for subsequent tracking and verification of data integrity. This achieves unified collection and standardized encapsulation of multi-source financial data, improving data access efficiency and integrity, and ensuring data traceability and security through source metadata and verification signatures.

[0095] The data unification module is used to cleanse, standardize, map fields, perform quality checks, align master data, manage metadata, and handle consistency of financial events provided by the data access module, generating governance data that can be used for subsequent querying and analysis. This ensures high-quality, standardized, and consistent governance data, facilitating subsequent queries and multidimensional analysis, providing field-level and record-level lineage tracing, and improving data transparency and auditability. It also reduces analytical errors caused by data quality issues and improves the reliability of analytical results.

[0096] The semantic understanding and intent recognition module receives natural language queries from users, identifies entities, dimensions, metrics, analytical targets, and query intents within the query, and generates a structured semantic representation. It automatically identifies financial entities and analytical intents within the query, achieving the conversion from natural language to structured semantics. This enhances the intelligence and ease of use of user queries, reducing the burden of manual parsing and mapping. It supports semantic understanding of complex queries, providing reliable input for subsequent dynamic retrieval execution.

[0097] The retrieval and execution module is used to construct a reconfigurable query graph based on the structured semantic representation, schedule, calculate, and incrementally execute the governance data, and output the retrieval and execution results; it dynamically generates efficient execution plans to achieve rapid retrieval and calculation of large-scale financial data. Incremental execution and partial result output improve system response speed and support real-time interaction.

[0098] The financial analysis module is used to perform multidimensional data association, indicator calculation, aggregation analysis, backtracking analysis, and predictive operations on the governance data based on the query graph and the retrieval execution results, and output the financial analysis results. The retrieval execution module works collaboratively with the financial analysis module, ensuring the interpretability, reproducibility, and incremental updates of the analysis results by recording the input and output metadata and execution evidence of the query nodes. This enables multidimensional, scrolling window, hierarchical aggregation, and predictive analysis, improving the depth and accuracy of the financial analysis.

[0099] The interactive visualization module receives the retrieval results and the financial analysis results, converts them into a visual display, and allows users to trigger further queries and analysis requests based on the visualized content. It provides interactive analysis capabilities, enabling users to quickly conduct exploratory analyses.

[0100] Specifically, the data access module includes:

[0101] The multi-source data acquisition unit is used to collect financial data from internal business systems, external interfaces, and documents. It pulls or receives structured business data from internal business systems such as ERP, financial systems, budget systems, and settlement systems via API, message queues, or direct database connections. External interface components are configured to connect to bank interfaces, tax interfaces, and data sources from third-party data service providers, supporting multiple transmission protocols such as HTTPS, SFTP, and Webhook. The multi-source data acquisition unit uses file read / write components to read locally uploaded or system-archived CSV, Excel, TXT, and compressed files for batch file scanning and header recognition. A data acquisition task scheduler is configured with acquisition rules, triggering strategies, task cycles, and failure retry strategies. A data buffer queue is set up to cache the collected raw data in high-concurrency scenarios, ensuring no data loss and supporting asynchronous processing. It can stably and efficiently acquire financial data continuously from different sources, achieving unified access to multi-source heterogeneous data, avoiding data silos caused by system differences, and improving data coverage.

[0102] The data format parsing unit is used to parse the financial data and convert it into structured data. The data format parsing unit includes: a format recognizer, used to automatically identify the file type, encoding format, field separator, header structure, and content type of the data source; a parsing rule engine, which parses JSON, XML, CSV, and Excel format data field by field based on predefined templates or source rules; a field standardization component, used to perform field name cleaning, data type conversion, time format unification, and precision processing for amount fields; an anomaly detection component, used to detect parsing anomalies such as missing fields, type mismatches, illegal characters, and broken rows and columns; and a structured data generator, used to convert the parsed data into standard structured records and write them to the access buffer. This achieves automatic data format recognition and adaptive parsing, improving the intelligence of data access; and ensures a unified data structure after parsing, providing consistent input for subsequent governance, retrieval, and analysis.

[0103] The data synchronization control unit is used to perform real-time, incremental, or full data synchronization. Upon initial access or during large-scale data correction, the unit batch-accesses all data from a specified data source. It retrieves incremental data based on primary key change identifiers, timestamps, version numbers, or log sequence numbers. It receives incremental events in real-time via message queues, event streams, or push mechanisms. It configures synchronization cycles, priorities, timeout policies, and resource quotas. It handles duplicate data, delayed data, and cross-system timestamp inconsistencies. It records the progress, offset, and completion status of each synchronization, thereby achieving full, incremental, and real-time multi-mode synchronization.

[0104] The basic metadata recording unit is used to generate the source system identifier, timestamp, batch identifier, and verification information of the financial event; the basic metadata recording unit adds the collection source (such as system number, interface number, file path, etc.) to each record; records the collection time, parsing time, and the business occurrence time contained in the original data; combines information such as collection task number, file batch, and synchronization round into a batch identifier; calculates the checksum, hash value, or signature of the record for subsequent data integrity verification; stores metadata in key-value or document form to ensure subsequent traceability, and records the running status information of the data collection and parsing process.

[0105] Specifically, the data governance module includes:

[0106] The Master Data Management Unit (MDMU) is used to uniformly manage master data entities in financial events provided by the data access module. These master data entities include accounts, organizations, customers, suppliers, and product information. The master data management process includes identifying, mapping, and aligning master data entities in financial events. The system uses a master data identification component to parse the account, organization, customer, supplier, and product fields in financial events and performs master data mapping on entities from different source systems using a unified master data coding system. For entities that cannot be directly matched, the system enters a master data addition or candidate confirmation process to ensure the integrity and consistency of the master data. The aligned master data is then provided to subsequent governance steps by the master data referencing service. This process achieves unified management of master data across systems, eliminates discrepancies in entity definitions, and provides a consistent entity foundation for subsequent quality verification, consistency comparison, and multidimensional analysis, significantly improving data comparability and accuracy.

[0107] The metadata management unit manages structural metadata, source metadata, and version metadata, and provides metadata references to the master data management unit. After master data alignment is completed, the system enters the metadata management process to uniformly register and manage the structural metadata, source metadata, and version metadata contained in financial events. The system reads information such as source system identifier, field structure, data format, and generation timestamp through the metadata collector, and the metadata registrar adds the field-level structure description to the metadata repository, while generating corresponding version metadata to record the time base for governance tasks. Subsequently, the metadata reference interface provides this metadata to processes such as quality management and lineage management. This process enables traceable management of data sources, structures, and versions, providing clear evidence for subsequent data verification, lineage construction, and version comparison, thereby improving the transparency and verifiability of the governance process.

[0108] The data quality management unit performs quality checks on the completeness, consistency, and accuracy of financial events based on the source metadata. The governance process continues into the data quality verification phase, where the system performs completeness, consistency, and accuracy checks on financial events based on metadata and master data information. Completeness checks examine whether any required fields or amount fields are missing from the financial events; consistency checks compare fields such as account, organization, and customer against master data; accuracy checks examine records according to the verification rules provided by the governance rules unit, including numerical values, formats, dates, and precision. The system marks records that fail the checks with quality indicators and automatically performs format repair and type conversion if repairable, while sending unrepairable abnormal records to the quality audit process. This process ensures that the governed data has high accuracy, high consistency, and high reliability, significantly reducing the risk of erroneous data entering the analysis process and enhancing the overall data reliability of the system.

[0109] The data consistency management unit is used to unify the terminology and perform cross-process consistency checks on financial events from different systems. After quality verification, the system performs standardization and cross-process consistency checks on financial events from different systems. Based on the terminology rules provided by the governance rules unit, the system standardizes the terminology of indicators from different source systems, such as amounts, taxes, and costs. Subsequently, using the master data alignment results, the system performs cross-process consistency checks on revenue-contracts, invoices-payments, and inventory-costs, identifying potential logical differences between different systems through correlation comparison and generating difference records for subsequent rectification. This process significantly improves the consistency of financial data across business processes, effectively avoids analytical biases caused by inconsistent terminology, and provides a rigorous and reliable data foundation for downstream retrieval and analysis modules.

[0110] The data lineage management unit is used to establish field-level and record-level data lineages based on the source metadata. Utilizing source metadata and operation records during the governance process, the data lineage management unit collects the field-level and record-level transformation relationships of financial events in the processing chain. A field dependency analyzer parses the source relationships of each field, a record-level traceability component records the mapping between different versions of financial events, and a lineage graph builder generates a visual lineage chain for retrieval by the execution module and the financial analysis module. This process makes the entire path of data traceable from source to governance to analysis, providing strong support for system auditing, error detection, and analytical interpretability, thereby improving system transparency and audit compliance capabilities.

[0111] The data version management unit generates data versions and data snapshots, and provides a version baseline to the data quality management unit. After each governance task, the system uses a version generator to create a new version number and records the time baseline of version generation. A snapshot generator periodically generates static snapshots of the governance data for restoring and replaying historical states. A version comparator supports comparing field-level differences between two versions, while the version baseline is provided to the quality management unit for time-series quality verification. This process ensures that the governance data is reproducible, comparable, and rollbackable, effectively improving the system's stability in data maintenance, auditing, and error recovery scenarios.

[0112] The Data Access Control (DAC) unit is used to implement data access control and security management. It establishes access control policies based on user roles, organizational scope, and data levels. The system uses a permission validator to perform permission verification each time governance data is read or modified, and writes access records to a security log to support auditing. This process enhances data security, prevents unauthorized access, ensures that sensitive financial data is effectively protected during governance and analysis, and provides audit trail capabilities, improving system security and compliance.

[0113] The governance rules unit provides governance rules to the master data management unit, the data quality management unit, and the data consistency management unit in accordance with accounting standards, regulatory requirements, and business rules. The governance rules unit is responsible for translating registered accounting standards, regulatory requirements, and internal control rules into executable formats using a rule parser, and then distributing the rules to the master data management, quality verification, and consistency management units via a rule distributor. The rule updater supports incremental updates and version management of rules, enabling the overall governance system to evolve synchronously with changes in regulations, systems, or business operations. This process ensures that the governance process complies with accounting and regulatory requirements, improving the standardization, consistency, and scalability of governance.

[0114] Specifically, the semantic understanding and intent recognition module includes:

[0115] The natural language input receiving unit receives the natural language text to be parsed. The system obtains the financial query text input by the user through this unit and sends it as data to be processed into the semantic parsing pipeline. Complete sentence structures are received through a standard input interface, eliminating the need for additional formatting by the user. This process ensures that the system accurately captures the user's original semantic expression, providing a complete text foundation for subsequent normalization and encoding processing.

[0116] The text normalization processing unit, upon receiving natural language text, performs normalization operations on the text, including word segmentation, case unification, removal of redundant symbols, standardization of numerical representation, and removal of unnecessary stop words. This is used to normalize the natural language text and generate a text sequence.

[0117] Encoding representation unit, used to perform vector encoding on the text sequence based on a semantic model and generate a context vector sequence. The context vector satisfies:

[0118]

[0119] in, This represents a text encoding model with parameter θ; the encoding model can adopt a bidirectional structure to capture contextual dependencies within the sequence. This process realizes the conversion of natural language text into high-dimensional semantic vectors, providing a sufficient input foundation for entity recognition and intent determination that fully expresses semantic and contextual relationships.

[0120] The entity recognition unit is used to perform sequence labeling based on the context vector sequence to generate an entity label sequence. And calculate the optimal label sequence:

[0121]

[0122] in, This is a label sequence scoring function; it performs boundary division and category labeling on semantic elements such as master data entities, financial indicators, time expressions, and organizational scope appearing in the text through the optimal label sequence, and performs entity-level aggregation on the labeling results to output an entity set.

[0123] The intent recognition unit is used to generate an intent probability vector based on the context vector sequence and to identify the user's target intent based on the softmax calculation method. :

[0124]

[0125] in, Pooling operations are performed on the context vector sequence; and These represent the weight matrix and bias vector of the intent classifier, respectively; This represents the normalization function, which transforms the output into a probability distribution. The probability vector output by the classifier represents the probability distribution of different intent categories, and the system determines the final intent label based on the maximum probability value. This process can be used to identify the target action of a user query, such as "query," "comparison," "statistic," and "prediction," ensuring that the system can accurately understand the type of operation the user wishes to perform.

[0126] The preliminary semantic representation generation unit is used to merge the outputs of the entity recognition unit and the intent recognition unit to form a preliminary structured semantic vector. The preliminary semantic representation generation unit merges the outputs of both to generate a structured semantic vector. The semantic vector contains entity type, entity value, logical relationships between entities, user intent category, and deep semantic information of the encoding model. This structured semantic result serves as input to the subsequent retrieval execution module for constructing a query plan.

[0127] Specifically, the semantic understanding and intent recognition module also includes:

[0128] The entity linking and master data alignment unit performs master data entity matching on the entity set after the entity recognition unit generates the entity set. This is used to link the entity set with the master data entities and calculate and determine the optimal entity based on the following comprehensive scores of candidate entities:

[0129]

[0130] And calculate the link confidence:

[0131]

[0132] Where m represents a candidate entity mention in the entity set, and e represents a candidate entity in the master data entity. This represents the contextual semantic vector of entity mentions m. This represents the embedding representation of the candidate entity e. This represents a semantic similarity metric. Represents literal similarity based on string features. α represents the prior probability of entity e in the master data, and β and γ represent the weighting coefficients used to fuse multiple entity matching features; this enables the final semantic query to be directly bound to standardized master data such as enterprise subjects, organizations, customers, and suppliers, achieving cross-system semantic consistency and ensuring the accuracy of subsequent retrieval execution.

[0133] The semantic slot filling and query logic generation unit, after completing entity linking, constructs a structured query logic form based on the intent recognition result. This structured query logic form is used to generate a logical expression based on the intent recognition result and outputs a logical expression based on the generation probability model.

[0134]

[0135] in, This indicates that given a preceding slot and input text, the system predicts the probability of the current slot. The system predicts slot fill values ​​sequentially according to slot dependencies, ultimately generating a logical expression that can be directly called by the query construction module. This expression explicitly represents the query objective, constraints, analysis dimensions, and computational requirements. This process achieves a structured mapping from natural language to executable logical queries, enabling the system to automatically generate query graphs.

[0136] Parse confidence assessment unit for use based on intent confidence Slot location reliability Confidence of entity links Calculate the overall analytical confidence level and generate the confidence level value:

[0137]

[0138] in, This represents the analytical confidence weighting parameter, satisfying... Used to measure the reliability of the output intent set I as a basis for selection; to identify uncertainties in semantic parsing and provide a reliable foundation for the system, improve the accuracy of subsequent retrieval execution, and allow the system to trigger a clarification mechanism under low confidence conditions.

[0139] The parsing result output unit is used to select the structured query logic with the highest slot probability to generate an intent set based on the comprehensive parsing confidence and slot probability. The generated intent set I contains structured logical expressions that satisfy the highest slot probability and highest parsing confidence, which can be directly used by the retrieval execution module to construct query nodes and execution sequences. This process achieves standardized output of natural language parsing results, ensuring that the system can stably transform user input into executable, traceable, and reproducible analytical queries.

[0140] Specifically, the retrieval execution module includes:

[0141] The request receiving and verification unit receives the semantic vector and the intent set, and verifies user permissions, field access control, request format, and semantic integrity. When a missing slot is detected in the semantic vector or the intent set, semantic completion is performed on the semantic vector and the intent set. The system first verifies the user's identity and role permissions, and checks whether the user has the access qualifications for the corresponding field according to the field access control policy provided by the unified data governance module. Subsequently, the semantic vector and intent set are checked for format and semantic integrity. If missing semantic slots, missing entities, or inconsistencies between analysis dimensions and metrics are detected, the request receiving and verification unit generates filler values ​​for the missing slots based on the semantic completion model, so that the completed semantic vector can meet the query construction requirements.

[0142] The graph compiler receives the completed semantic vectors and intent sets and maps them into a query graph in the form of a logically directed acyclic graph (RDA). The compiler performs operator decomposition, type identification, data range labeling, and semantic origination labeling on the semantic vectors and intent sets. It also automatically injects privacy operators and lineage origin metadata based on field sensitivity levels, forming a logical query graph with metadata and generating corresponding graph identifiers. The graph compiler receives the completed semantic vectors and intent sets and maps them into a query graph with a logically directed acyclic graph structure. This process first identifies the dimensions, metrics, analysis conditions, and calculation types involved in the query based on the semantic vectors and intent sets, and decomposes them into basic operators. The system then performs operator type identification, including filtering operators, aggregation operators, connection operators, projection operators, window operators, and restriction operators. For each operator node, the graph compiler generates field range labels, data source labels, and semantic origination labels, ensuring that each node contains traceable metadata. Based on field sensitivity levels, the system automatically injects privacy operators to perform desensitization, range compression, and identifier replacement. The final result is a logical query graph with kinship origin markers, privacy markers, and field range descriptions, and a unique graph identifier is generated.

[0143] The cost and resource estimator quantifies node-level resource consumption, providing a computable basis for optimizing the cost of the physical execution plan. This improves the performance controllability and time predictability of the entire query execution process, and is used to evaluate each node in the logical query graph. Based on input / output operations Calculation consumption and network consumption Perform resource consumption estimation:

[0144]

[0145] The execution planner is used to query the graph and estimate the results based on the logic. The system generates a physical execution plan, determines the operator fusion strategy, filter pushdown strategy, data partitioning strategy, and the binding relationship between nodes and the execution backend, and splits the logical graph into multiple physically schedulable execution stages according to performance targets. Based on the cost model, it identifies mergeable operators and performs operator fusion to improve execution efficiency. Subsequently, it identifies and marks the filter conditions that can be pushed down to the data source side. The system generates a data partitioning strategy based on data distribution, minimizing cross-node communication during the runtime of each operator node. The execution planner then determines the binding relationship of each node to the execution backend and divides the logical query graph into multiple execution stages, each corresponding to a physical operator tree that can be arranged in parallel.

[0146] The execution scheduler submits each execution stage in the physical execution plan to the corresponding execution engine for scheduling, manages parallelism, retry policies, failure transfer, priority execution pools, and resource allocation, and tracks node status in real time during execution. It determines the startup order of different execution stages according to stage dependencies and allocates execution resource pools based on resource occupancy. The scheduler tracks node running status in real time during execution, including queuing, execution, waiting, and completion states. If a node malfunctions, the system triggers a backup node to execute according to preset retry and failure transfer policies. Throughout the process, the scheduler manages operator parallelism, priority pools, and node lifecycles.

[0147] The execution unit is used to run the corresponding operator on a preset execution backend based on the physical execution plan. The execution unit performs node-level caching and partial result output processing on the operator execution process, and generates node-level data hashes based on the input semantic vector and node output. The execution unit executes the physical operator on the execution backend specified by the scheduler. During the execution process, the system maintains a node-level cache for each operator node in order to accelerate subsequent query requests corresponding to the same semantic vector.

[0148] The result fusion and refinement unit receives the node-level partial results and final results submitted by the execution unit, and merges the results based on the node execution status provided by the execution scheduling unit. During the merging process, the result fusion and refinement unit performs conflict resolution, replacement of approximate results with precise results, and assigns confidence level labels to the results based on the comprehensive resolution confidence level. The result fusion and refinement unit outputs the final retrieval execution result.

[0149] Specifically, the financial analysis module includes:

[0150] The receiving layer is used to receive the retrieval execution results, parse the analysis dimensions, metrics, aggregation functions, filtering conditions and analysis types in the retrieval execution results, and generate a standardized execution description based on the master data entity association mapping;

[0151] The multidimensional model layer is used to build and maintain multidimensional cubes, dimension tables, metric tables, and scene caliber views, manage dimension levels and drill-down rules, and support multicaliber views and caliber transformation rules. The dimension tables are linked to the master data entities to associate cross-table semantics and source data.

[0152] The computation engine layer is used to perform multidimensional aggregation, slicing, drill-down, and summarization calculations. It adopts a calculation strategy of approximation followed by precision and incremental calculation, and uses vectorized columnar calculation and distributed parallel computing mode for analysis and processing. The operators of the analysis and processing include aggregation, window aggregation, dimension association, and approximate estimation.

[0153] The advanced analytics layer is used to perform causal simulation, time series forecasting, backtesting analysis, index attribution, anomaly detection, and scenario comparison. It manages the versions of training and inference models and outputs audit metadata. It also supports automatic selection of the optimal model based on the analytics request.

[0154] The results management and caching layer is used to manage financial analysis results, record the financial analysis results output by the computing engine layer, and make key analysis results available for subsequent backtracking and compliance checks.

[0155] The interface and presentation layer are used to provide standardized analysis results output, along with explanatory information, confidence level information, and audit information, and support front-end visualization and report export.

[0156] The financial analysis module operates as follows: First, the receiving layer receives the structured query results generated by the retrieval and execution module, and parses the analysis dimensions, measures, aggregation methods, filtering restrictions, and analysis types in the results. After parsing, the system maps the input content according to the master data structure, such as standard accounts, organizations, customers, suppliers, and products in the master data entity, generating a standardized execution description in a unified format to ensure that the subsequent analysis model maintains consistent semantics and scope under cross-source data conditions. Next, the multidimensional model layer constructs or calls the corresponding multidimensional data model based on the standardized description. This layer maintains multidimensional cubes, dimension tables, measure tables, and scenario scope views, and uniformly manages dimension levels, drill-down paths, and transformation scopes. Each dimension table is linked to the master data entity, thereby automatically associating cross-table semantic relationships and corresponding data source information during analysis, and supporting conversion and verification between different management scopes.

[0157] Subsequently, the computation engine layer performs multidimensional aggregation, slicing, drill-down, and summarization calculations based on a multidimensional model, employing vectorized columnar computation and distributed parallel processing methods according to task scale and real-time requirements. During the analysis process, the engine uses a strategy of approximation followed by precision to quickly predict large-scale data, and then combines incremental computation to reduce redundant processing and improve overall computational efficiency. The analytical operators that this layer can execute include basic aggregation operators, window aggregation operators, dimensional correlation operators, and approximate estimation operators to support different types of financial analysis needs. After the basic analysis is completed, the advanced analysis layer further performs in-depth inference on the output results, including causal simulation, time series forecasting, indicator attribution analysis, backtesting analysis, anomaly detection, and scenario comparison. This layer manages the versions of the training and inference models used and generates audit metadata after each model inference, making model selection, model source, and analysis chain completely traceable. The system can also automatically select the optimal model based on user requests to balance accuracy and response performance.

[0158] The management of analysis results is handled by the results management and caching layer. This layer stores key analysis results output by the computing engine layer and the advanced analysis layer, managing the version, generation time, traceability, and audit information of the results. It provides a reproducible data foundation for subsequent backtracking, recalculation, or compliance checks. Simultaneously, this layer has a caching mechanism to cache frequently used or interim analysis results for rapid response to similar analysis requests. Finally, the interface and presentation layer standardizes the output of the analysis results, including necessary explanatory information, confidence scores, and audit information, ensuring users have a full understanding of the source, reliability, and calculation logic of the analysis results. This layer also provides front-end visualization support, including multidimensional tables, graphical displays, and key indicator cards, and supports exporting reports according to regulatory requirements or enterprise needs, enabling the analysis results to be directly used for management decisions, audit reports, or external disclosures.

[0159] Specifically, the calculation method of the calculation engine layer of the financial analysis module includes the following steps: receiving analysis requests from the semantic understanding module and the user interface, and parsing the dimensions in the analysis request. ,measure Aggregate functions It analyzes and generates standardized internal execution descriptions; it constructs multidimensional cubes and dimension tables based on the parsing results, and maps metric definitions to multidimensional models, supporting multi-caliber views and dimension hierarchy management.

[0160] Perform multidimensional analysis and aggregation:

[0161] Aggregate calculations on metric data m under a specified dimension d, where f represents the aggregation function (such as summation, average, maximum, minimum, etc.). slice(d) represents the slice set of data selected under dimension d, such as selecting data from a specific department, month, or business unit:

[0162]

[0163] Where d represents the analysis dimension, m represents the metric to be aggregated, f represents the aggregation function, and slice(d) represents the slice data set under the specified dimension; it can flexibly summarize and calculate financial data in multiple dimensions, realize statistical analysis of key indicators and report generation, and provide quantitative basis for decision-making.

[0164] Accumulate and sum the metric m over a time or series window w:

[0165]

[0166] Where w represents the time window. This represents the metric value at each point in time within the window. It represents rolling summation; it supports rolling statistics and time series analysis, such as rolling summation and moving average, which facilitates the discovery of trends and fluctuation patterns and provides basic data for financial forecasting and monitoring.

[0167] Aggregate calculations along the hierarchical structure of dimension d:

[0168]

[0169] in This represents a slice of a lower level of dimension d. It represents the total amount calculated by aggregation along the dimension hierarchy; it can summarize the data of the lower level to the upper level to form hierarchical statistical results, realize hierarchical summary analysis of financial data, and support multi-caliber reports and management decisions.

[0170] Measurement data based on historical time range t Construct a time series forecasting model to predict metrics for the next h time steps:

[0171]

[0172] Where t represents the historical time range and h represents the prediction step size. This represents a time series forecasting model that outputs predicted values ​​for the next h time steps; it enables trend forecasting of financial data, such as estimating future changes in revenue, costs, and cash flow, to assist budgeting and business decisions and improve forecast accuracy.

[0173] Causal simulation analysis is performed on metric m to simulate the changes in metric caused by intervention operation I (such as adjusting the budget or modifying the cost policy). Compared with baseline value The differences.

[0174]

[0175] Where I represents the intervention operation. This represents the simulated metric under intervention conditions. Indicates the baseline metric. It indicates the differences in causal simulation calculations; quantifies the causal impact of intervention measures on financial indicators, supports scenario analysis, policy testing and strategy optimization, and improves the scientific nature of decision-making and risk management capabilities.

[0176] Specifically, the interactive visualization module includes:

[0177] The front-end presentation layer is used to render the outputs of the retrieval execution module and the financial analysis module;

[0178] The interaction management layer is used to capture user operation events and generate interactive operation objects based on the events. Combining the data currently displayed by the front-end presentation layer with the master data entities and knowledge graph nodes provided by the data governance module, it generates secondary analysis suggestions to guide users to further drill down into analysis dimensions, metrics and causal paths.

[0179] The result parsing and rendering adaptation layer is used to receive the analysis results from the financial analysis module and the retrieval execution module, convert the results into a data structure that can be rendered on the front end, and support progressive updates, gradually replacing approximate results with accurate results.

[0180] The interaction triggering and query interface layer is used to transform the interaction operation object into a structured query representation that can be parsed by the semantic understanding and intent recognition module, call the semantic understanding and intent recognition module and the retrieval execution module to perform query execution, and push part and the final result to the front-end presentation layer through WebSocket;

[0181] The audit and interpretability layer is used to attach source information to each visualization result generated by the interactive visualization module and record all user interactions and analysis execution processes.

[0182] The interactive visualization module renders the analysis content output by the retrieval and execution module and the financial analysis module through the front-end presentation layer. Based on a pre-defined visualization component library, the front-end presentation layer presents indicator results, trend charts, dimensional distribution charts, analysis tables, causal link diagrams, and prediction results in graphical form. This layer supports switching between multiple visualization views, enabling users to obtain a stable and consistent presentation experience in different analysis scenarios.

[0183] Building upon the content displayed on the front end, the interaction management layer is responsible for capturing user actions such as clicks, filtering, drill-down, dragging, switching perspectives, switching dimensions, time selection, and highlighting or filtering graphical components. Based on these captured events, the interaction management layer generates standardized interactive objects and, combined with the data currently displayed on the front end, reads master data entities and knowledge graph nodes provided by the data governance module to generate actionable secondary analysis suggestions for the user. These suggestions include analytical dimensions for further drill-down, expandable metrics, potential causal paths, and related data sources, helping users conduct deeper analysis along the business logic.

[0184] Subsequently, the results parsing and rendering adaptation layer receives the analysis results from the financial analysis module and the retrieval execution module, converting the internally used structured results into a data structure that the front end can directly render. This includes flat table data, hierarchical dimensional structures, graph node sets and edge sets, indicator card data, and trend sequences. Simultaneously, this layer supports progressive content updates: while the execution plan is still in the calculation process, the front end can prioritize displaying approximate calculations or partial results; as the execution unit and the results fusion layer produce precise results, the front end automatically updates to the final results, thus maintaining a continuous interactive experience for the user and the system's real-time responsiveness.

[0185] The interaction triggering and query interface layer transforms the interactive operation objects generated by the interaction management layer into a structured query representation that can be processed by the semantic understanding and intent recognition module. This layer calls the semantic understanding and intent recognition module to parse the interaction request, and then calls the retrieval execution module to complete the execution scheduling and multi-source data query after data governance. During query execution, this layer pushes the interim and final results to the front-end presentation layer in real time via WebSocket, allowing users to receive immediate feedback even in complex analysis scenarios.

[0186] Finally, the audit and interpretability layer adds source descriptions to each visualization result generated by the interactive visualization module, recording the data version, execution node, master data mapping relationship, analysis model version used, and governance rules involved in generating the result. Simultaneously, this layer records every user interaction and its corresponding execution chain throughout the analysis process, enabling the system to reconstruct the complete analysis process when compliance checks, internal audits, or result reproduction are required, ensuring interpretability, transparency, and traceability.

[0187] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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 or all of the technical features; and these 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.

Claims

1. An interactive financial data retrieval and analysis system based on big data, characterized in that, The system includes: The data access module is used to receive financial data from multiple data sources and encapsulate the financial data into financial events with timestamps, source metadata, and verification signatures. The data unification module is used to clean, standardize, map fields, verify quality, align master data, manage metadata, and process consistency of financial events provided by the data access module, and generate governance data that can be used for subsequent querying and analysis. The semantic understanding and intent recognition module is used to receive natural language queries input by users, identify entities, dimensions, metrics, analysis targets and query intents in the query, and generate structured semantic representations. The retrieval execution module is used to construct a reconstructable query graph based on the structured semantic representation, schedule, calculate and incrementally execute the governance data, and output the retrieval execution results; The financial analysis module is used to perform multidimensional data association, indicator calculation, aggregation analysis, backtracking analysis, and predictive operations on the governance data based on the query graph and the retrieval execution results, and output the financial analysis results. The retrieval execution module works in collaboration with the financial analysis module to ensure the interpretability, reproducibility, and incremental updates of the analysis results by recording the input and output metadata and execution evidence of the query nodes. The interactive visualization module is used to receive the retrieval execution results and the financial analysis results, convert them into a visual display format, and support users to trigger further query and analysis requests based on the visual content.

2. The system according to claim 1, characterized in that, The data access module includes: The multi-source data acquisition unit is used to collect financial data from internal business systems, external interfaces, and documents. A data format parsing unit is used to parse the financial data and convert it into structured data; The data synchronization control unit is used to perform real-time, incremental, or full data synchronization. The basic metadata recording unit is used to generate the source system identifier, timestamp, batch identifier, and verification information of the financial event. The operation monitoring unit is used to record the operational status information of the data acquisition and analysis process.

3. The system according to claim 2, characterized in that, The data governance module includes: The master data management unit is used to uniformly manage the master data entities in the financial events provided by the data access module. The master data entities include accounts, organizations, customers, suppliers and product information for unified management. The metadata management unit is used to manage structural metadata, source metadata, and version metadata, and provides metadata references to the master data management unit. The data quality management unit is used to perform quality verification on the completeness, consistency, and accuracy of financial events based on the source metadata. The data consistency management unit is used to unify the reporting standards and verify cross-process consistency of financial events from different systems. The data lineage management unit is used to establish field-level and record-level data lineages based on the source metadata. The data version management unit is used to generate data versions and data snapshots, and to provide version benchmarks to the data quality management unit. The data access control unit is used to implement data access control and security management. The governance rules unit is used to provide governance rules to the master data management unit, the data quality management unit, and the data consistency management unit in accordance with accounting standards, regulatory requirements, and business rules.

4. The system according to claim 1, characterized in that, The semantic understanding and intent recognition module includes: The natural language input receiving unit is used to receive natural language text to be parsed; The text normalization processing unit is used to normalize the natural language text and generate a text sequence. ; Encoding representation unit, used to perform vector encoding on the text sequence based on a semantic model and generate a context vector sequence. The context vector satisfies: ； in, This represents a text encoding model with parameter θ; The entity recognition unit is used to perform sequence labeling based on the context vector sequence to generate an entity label sequence. And calculate the optimal label sequence: ； in, The label sequence scoring function is used to aggregate and calculate the optimal label sequence, outputting an entity set. The intent recognition unit is used to generate an intent probability vector based on the context vector sequence and to identify the user's target intent based on the softmax calculation method. : ； in, Pooling operations are performed on the context vector sequence; and These represent the weight matrix and bias vector of the intent classifier, respectively; This represents the normalization function, which converts the output into a probability distribution. The preliminary semantic representation generation unit is used to merge the outputs of the entity recognition unit and the intent recognition unit to form a preliminary structured semantic vector.

5. The system according to claim 4, characterized in that, The semantic understanding and intent recognition module also includes: The entity linking and master data alignment unit is used to link the entity set with the master data entity, and to determine the optimal entity based on the following comprehensive scores of candidate entities: ； And calculate the link confidence: ； Where m represents a candidate entity mention in the entity set, and e represents a candidate entity in the master data entity. This represents the contextual semantic vector of entity mentions m. This represents the embedding representation of the candidate entity e. This represents a semantic similarity metric. Represents literal similarity based on string features. Let α, β, and γ represent the prior probability of entity e in the main data, and let α, β, and γ represent the weighting coefficients used to fuse multiple entity matching features. The semantic slot filling and query logic generation unit is used to generate structured query logic based on intent recognition results and output logical expressions based on the generation probability model. ； in, This indicates the probability of predicting the current slot given the preceding slot and the input text. Parse confidence assessment unit for use based on intent confidence Slot location reliability Confidence of entity links Calculate the overall analytical confidence level and generate the confidence level value: ； in, This represents the analytical confidence weighting parameter, satisfying... Used to measure the reliability of the output intent set I, as a basis for selection; The parsing result output unit is used to select the structured query logic with the highest slot probability to generate an intent set based on the comprehensive parsing confidence and slot probability. .

6. The system according to claim 5, characterized in that, The retrieval execution module includes: The request receiving and verification unit is used to receive the semantic vector and the intent set, verify user permissions, field access control, request format and semantic integrity, and when a missing slot is detected in the semantic vector or the intent set, perform semantic completion on the semantic vector and the intent set. The graph compiler is used to receive the completed semantic vector and intent set and map the semantic vector and intent set into a query graph in the form of a logical directed acyclic graph. The graph compiler performs operator decomposition, type identification, data range labeling, and semantic source labeling on the semantic vector and intent set, and automatically injects privacy operators and lineage origin metadata according to the field sensitivity level to form a logical query graph with metadata and generate corresponding graph identifiers. Cost and resource estimator for each node in the logical query graph Based on input / output operations Calculation consumption and network consumption Perform resource consumption estimation: ； The execution planner is used to query the graph and estimate the results based on the logic. Generate a physical execution plan, determine the operator fusion strategy, filter pushdown strategy, data partitioning strategy, and the binding relationship between nodes and the execution backend, and split the logical graph into multiple physically schedulable execution stages according to performance targets; The execution scheduler is used to submit each execution stage in the physical execution plan to the corresponding execution engine for scheduling, manage parallelism, retry strategy, failure transfer, priority execution pool and resource allocation, and track the node status in real time during the execution process; An execution unit is used to run the corresponding operator on a preset execution backend based on the physical execution plan. The execution unit performs node-level caching and partial result output processing on the operator execution process, and generates node-level data hashes based on the input semantic vector and node output. The result fusion and refinement unit receives the node-level partial results and final results submitted by the execution unit, and merges the results based on the node execution status provided by the execution scheduling unit. During the merging process, the result fusion and refinement unit performs conflict resolution, replacement of approximate results with precise results, and assigns confidence level labels to the results based on the comprehensive resolution confidence level. The result fusion and refinement unit outputs the final retrieval execution result.

7. The system according to claim 6, characterized in that, The financial analysis module includes: The receiving layer is used to receive the retrieval execution results, parse the analysis dimensions, metrics, aggregation functions, filtering conditions and analysis types in the retrieval execution results, and generate a standardized execution description based on the master data entity association mapping; The multidimensional model layer is used to build and maintain multidimensional cubes, dimension tables, metric tables, and scene caliber views, manage dimension levels and drill-down rules, and support multicaliber views and caliber transformation rules. The dimension tables are linked to the master data entities to associate cross-table semantics and source data. The computation engine layer is used to perform multidimensional aggregation, slicing, drill-down, and summarization calculations. It adopts a calculation strategy of approximation followed by precision and incremental calculation, and uses vectorized columnar calculation and distributed parallel computing mode for analysis and processing. The operators of the analysis and processing include aggregation, window aggregation, dimension association, and approximate estimation. The advanced analytics layer is used to perform causal simulation, time series forecasting, backtesting analysis, index attribution, anomaly detection, and scenario comparison. It manages the versions of training and inference models and outputs audit metadata. It also supports automatic selection of the optimal model based on the analytics request. The results management and caching layer is used to manage financial analysis results, record the financial analysis results output by the computing engine layer, and make key analysis results available for subsequent backtracking and compliance checks. The interface and presentation layer are used to provide standardized analysis results output, along with explanatory information, confidence level information, and audit information, and support front-end visualization and report export.

8. The system according to claim 7, characterized in that, The calculation method of the calculation engine layer of the financial analysis module includes the following steps: receiving analysis requests from the semantic understanding module and the user interface, and parsing the dimensions in the analysis request. ,measure Aggregate functions And analyze the type, and generate standardized internal execution descriptions; Based on the parsing results, a multidimensional cube and dimension table are constructed, and the metric definition is mapped to the multidimensional model, supporting multi-caliber views and dimension hierarchy management; Perform multidimensional analysis and aggregation: ； Where d represents the analysis dimension, m represents the metric to be aggregated, f represents the aggregation function, and slice(d) represents the slice data set under the specified dimension; ； Where w represents the time window. This represents the metric value at each point in time within the window. Indicates rolling summation; ； in This represents a slice of a lower level of dimension d. This indicates the total amount calculated by aggregation along the dimensional hierarchy; ； Where t represents the historical time range and h represents the prediction step size. This represents a time series forecasting model that outputs the predicted metric values ​​for the next h time steps. ； Where I represents the intervention operation. This represents the simulated metric under intervention conditions. Indicates the baseline metric. This indicates the difference in causal simulation calculations.

9. The system according to claim 8, characterized in that, The interactive visualization module includes: The front-end presentation layer is used to render the outputs of the retrieval execution module and the financial analysis module; The interaction management layer is used to capture user operation events and generate interactive operation objects based on the events. Combining the data currently displayed by the front-end presentation layer with the master data entities and knowledge graph nodes provided by the data governance module, it generates secondary analysis suggestions to guide users to further drill down into analysis dimensions, metrics and causal paths. The result parsing and rendering adaptation layer is used to receive the analysis results from the financial analysis module and the retrieval execution module, convert the results into a data structure that can be rendered on the front end, and support progressive updates, gradually replacing approximate results with accurate results. The interaction triggering and query interface layer is used to transform the interaction operation object into a structured query representation that can be parsed by the semantic understanding and intent recognition module, call the semantic understanding and intent recognition module and the retrieval execution module to perform query execution, and push part and the final result to the front-end presentation layer through WebSocket; The audit and interpretability layer is used to attach source information to each visualization result generated by the interactive visualization module and record all user interactions and analysis execution processes.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in memory to run the system according to any one of claims 1 to 9.

Images