Question and answer type fund transaction data analysis method and system

Abstract

The invention relates to the technical field of financial data analysis, in particular to a question and answer type fund transaction data analysis method and system, and the method comprises the steps: receiving a natural language query of a user, recognizing an intention through a natural language processing technology, and extracting key entity parameters (such as amount, time and a transaction opponent); maintaining a multi-round dialogue state, and converting intentions and parameters into structured query; calling a multi-dimensional analysis model according to the structured query parameters, performing multi-dimensional analysis on the fund transaction data, and generating a visual result; and based on the dialogue state and the intelligent decision strategy, the visualization result is returned to the user through a multi-round interaction mechanism. According to the technical scheme, the bottleneck of terminology analysis and complex dimension analysis in the financial field is overcome, normal form upgrading from passive query to active analysis is achieved, and an efficient and accurate end-to-end solution is provided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial data analysis, in particular to a question and answer type fund transaction data analysis method and system. BACKGROUND

[0002] In the financial field, especially in fund transaction data analysis, the traditional method highly depends on manual operation or limited query tools. Users must master specific commands or be familiar with complex operation interfaces to obtain the required analysis results, which not only has high operation threshold and complicated process, but also is difficult to flexibly cope with the increasing diversification and immediacy requirements of users for data insight. In recent years, the breakthrough of natural language processing (NLP) technology has promoted the landing application of intelligent question and answer systems in multiple fields, which provides the possibility of conveniently obtaining information through natural language interaction. However, in the specific scenario of fund transaction data analysis, its application still faces significant challenges. The data in this field is full of a large number of professional terms and complex business logic relationships, which requires the system to have a deep semantic understanding ability to accurately capture user intent and extract key analysis parameters (such as transaction counterparties, time intervals, amount ranges, transaction types, etc.). At the same time, fund analysis naturally involves the combination and drilling of multiple dimensions (such as time, amount, transaction type, transaction counterparties, channel, etc.), which urgently needs strong multi-dimensional analysis model support.

[0003] In the prior art, although some explorations have applied artificial intelligence to human-computer interaction and intent recognition. For example, Baidu Online Network Technology (Beijing) Co., Ltd. applied for “human-computer interaction method and system based on artificial intelligence” (CN105068661A) in 2015, which realized the transformation of interaction mode from tool command to personified natural language, supported natural language search and multi-round dialogue, but its design is general and has not been deeply optimized for the specific terminology system and analysis dimensions of fund transaction data analysis. The “intent distribution method, device, equipment, storage medium and program product” (CN120011507A) of China Merchants Bank and the “intent recognition method and device” (CN119312816A) of Lenovo have some innovations in the aspect of intent recognition technology, but they focus on the intent recognition link itself and lack a complete end-to-end technical solution covering from natural language understanding to professional parameter extraction, to complex multi-dimensional analysis calculation, and facing the fund transaction data analysis scenario. SUMMARY

[0004] In order to solve the above technical problems, the present application provides a question and answer type fund transaction data analysis method, and on the other hand, a question and answer type fund transaction data analysis system is also provided.

[0005] The technical problems solved by the present application can be realized by the following technical solutions:

[0006] A question and answer type fund transaction data analysis method, comprising:

[0007] Step S1, receiving a user input fund transaction data analysis related question, identifying user intent through natural language processing technology, and extracting entity parameters from query text;

[0008] Step S2, maintaining multi-round dialogue state, converting the user intent and the entity parameters into structured query parameters;

[0009] Step S3, calling a multi-dimensional analysis model according to the structured query parameters, performing multi-dimensional analysis on the fund transaction data, and generating a visualization result;

[0010] Step S4, based on dialogue state and intelligent decision-making strategy, returning the visualization result to the user through a multi-round interaction mechanism.

[0011] Preferably, the step S1 comprises:

[0012] Step S11, using a hybrid model of deep neural network and rule engine to classify the user intent;

[0013] Step S12, identifying the entity parameters based on BERT model and rule engine; wherein the entity parameters include amount, time, transaction type, account name and transaction counterparty.

[0014] Preferably, the step S2 comprises:

[0015] Step S21, recording historical query parameters and generated chart ID through dialogue state tracking technology; wherein the dialogue state tracking technology adopts a combined architecture of finite state machine and Transformer;

[0016] Step S22, converting the recognition results of the user intent and the entity parameters into structured query parameters in JSON format.

[0017] Preferably, the step S3 comprises:

[0018] Step S31, calling a multi-dimensional analysis model according to the user intent; wherein the multi-dimensional analysis model is based on a hybrid expert large model and a traditional algorithm, and fuses time, amount, transaction type, account and transaction counterparty entity dimensions;

[0019] Step S32, analyzing the fund transaction data through the multi-dimensional analysis model;

[0020] Step S33, automatically matching chart types according to the analysis results, and generating a visualization analysis result.

[0021] Preferably, the step S32 comprises:

[0022] Step S321, using a CNN-BiLSTM model, and combining discrete wavelet transform DWT and variational mode decomposition VMD for time series trend prediction;

[0023] Step S322, constructing a transaction graph network, and analyzing node relationships through a graph neural network to identify abnormal transaction patterns;

[0024] Step S323, using a random forest and a hybrid expert large model to classify transaction types and account behavior, and supporting multi-dimensional data fusion analysis.

[0025] Preferably, the step S321 comprises:

[0026] Step S3211, performing J-level discrete wavelet transform DWT decomposition on the original sequence x={x1,x2,...,x T},x t ∈R, and taking the highest-level approximation A j as the denoised sequence The specific formula is:

[0027]

[0028] Wherein,

[0029] DWT J represents a function of J-level discrete wavelet transform;

[0030] A j represents the low-frequency approximation coefficient vector obtained by j-level decomposition, with a dimension of T / 2 j ;

[0031] D j represents the high-frequency detail coefficient vector obtained by j-level decomposition, with a dimension of T / 2 j ;

[0032] Step S3212, inputting the denoised sequence into variational mode decomposition VMD to decompose into K intrinsic mode functions IMFs, and the specific formula is:

[0033]

[0034] Wherein,

[0035] VMD represents a variational mode decomposition function;

[0036] u k represents the kth intrinsic mode function obtained by decomposition, with a dimension of T';

[0037] T' represents the denoised sequence the length of T / 2 J ;

[0038] Step S3213, one-dimensional convolution operation is performed on each IMF sequence u k to extract local timing features f k , and the local timing features f k of all IMF sequences are spliced and then passed through a bidirectional LSTM to obtain timing hidden representations, and the specific formula is as follows:

[0039] f k = CNN (u k ) ∈ R T”×d ,

[0040] F = [f1; f2;...; f k ] ∈ R T”×(K·d)

[0041] wherein,

[0042] CNN represents a convolutional neural network operation;

[0043] u k represents the kth IMF sequence input to the CNN;

[0044] f k represents the feature matrix extracted by the CNN from the kth IMF sequence u k ;

[0045] d represents the number of convolution output channels T'<T;

[0046] F represents the comprehensive feature matrix formed by splicing the features f_k extracted from all K IMF sequences;

[0047] T' represents the length of the feature after splicing in the time dimension;

[0048] K·d represents the total length of the feature after splicing in the channel dimension;

[0049] Step S3214, attention is applied to the BiLSTM output sequence to obtain a weighted context vector c, and the specific formula is as follows:

[0050]

[0051] wherein,

[0052] Wh represents a learnable weight matrix;

[0053] b represents a learnable bias vector;

[0054] tanh() represents a hyperbolic tangent activation function;

[0055] v represents a learnable weight vector;

[0056] e t represents the unnormalized attention score calculated at time step t;

[0057] a t represents the normalized attention weight calculated at time step t;

[0058] c represents the calculated attention context vector;

[0059] Step S3215, the weighted context vector c is predicted for H steps in the future through several fully connected layers to obtain a predicted value vector, and the specific formula is represented as:

[0060]

[0061] wherein,

[0062] represents the prediction vector of the model for the future value;

[0063] represents the predicted value for the i-th time point after the original time point T, i = 1, 2,..., H;

[0064] H represents the number of future time steps to be predicted;

[0065] FC(c) represents a fully connected layer operation.

[0066] Preferably, the step S33 comprises: a trend analysis matching broken line chart, a risk distribution matching heat map, a fund flow direction matching Sanji chart or a graph neural network chart.

[0067] Preferably, the step S4 comprises:

[0068] Step S41, record the user intention, historical interaction record, generated chart ID and dialogue round in JSON format;

[0069] Step S42, dynamically select a reply form and optimize the reply content through a rule engine and deep reinforcement learning according to the user intention and the historical interaction record;

[0070] Step S43, trigger natural language follow-up questions through reference relationship identification and parameter missing detection;

[0071] Step S44, store high-frequency query results by using a cache mechanism, and directly return cached data for similar requests by combining multi-model phased processing technology.

[0072] The application also provides a question-answer type fund transaction data analysis system, which applies the question-answer type fund transaction data analysis method.

[0073] A natural language understanding module is configured to receive a fund transaction data analysis related question input by a user, identify the user's intention through a natural language processing technology, and extract an entity parameter from the query text;

[0074] A dialogue management module is connected to the natural language understanding module and is configured to maintain a multi-round dialogue state and convert the user's intention and the entity parameter into a structured query parameter;

[0075] A data analysis module is connected to the dialogue management module and is configured to call a multi-dimensional analysis model according to the structured query parameter, analyze the fund transaction data, and generate a visual result;

[0076] An interaction optimization module is connected to the data analysis module and is configured to return the visual result to the user through a multi-round interaction mechanism based on a dialogue state and an intelligent decision-making strategy.

[0077] Preferably, the data analysis module is provided with an API interface, and the API interface is designed in a standardized RESTful style.

[0078] Beneficial effects: As the above technical scheme is adopted, the application fuses deep semantic understanding, multi-round dialogue management and a multi-dimensional analysis model for a fund transaction scenario, solves the pain points of traditional tools, such as complicated operation and insufficient flexibility, overcomes the bottleneck of general natural language processing technology (NLP) in semantic analysis and complex analysis support in a professional field, realizes the paradigm upgrade from passive query to active analysis and from single instruction to intelligent dialogue, and provides an efficient, accurate and expandable end-to-end solution for fund transaction data analysis. BRIEF DESCRIPTION OF DRAWINGS

[0079] Figure 1 is a flowchart of the method of the application;

[0080] Figure 2 is a flowchart of the method step S1 of the application;

[0081] Figure 3 is a flowchart of the method step S2 of the application;

[0082] Figure 4 is a flowchart of the method step S3 of the application;

[0083] Figure 5 is a flowchart of the method step S32 of the application;

[0084] Figure 6is a flow chart of method step S4 of the present application;

[0085] Figure 7 is a schematic diagram of the multi-round interaction mechanism and intelligent decision-making strategy of the present application;

[0086] Figure 8 is a system block diagram of the present application. DETAILED DESCRIPTION

[0087] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0088] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0089] The present application will be further described below in combination with the drawings and specific embodiments, but not as a limitation of the present application.

[0090] Referring to Figure 1 A question-and-answer type fund transaction data analysis method, comprising:

[0091] Step S1, receiving a user input fund transaction data analysis related question, identifying user intent through natural language processing technology, and extracting entity parameters from the query text;

[0092] Step S2, maintaining a multi-round dialogue state, converting the user intent and the entity parameters into structured query parameters;

[0093] Step S3, calling a multi-dimensional analysis model according to the structured query parameters, performing multi-dimensional analysis on the fund transaction data, and generating a visual result;

[0094] Step S4, based on the dialogue state and intelligent decision-making strategy, returning the visual result to the user through a multi-round interaction mechanism.

[0095] Specifically, in the embodiments of the present application, first, the query text input by the user is parsed by natural language processing technology (NLP) to identify and analyze the demand and key entity parameters (such as amount, time, transaction type, etc.); then, the query parameters are dynamically maintained and updated during the multi-round dialogue process, and are converted into structured query parameters; then, based on the structured parameters, a preset multi-dimensional analysis algorithm (such as trend analysis, risk detection, comparative query, etc.) is called to aggregate and calculate the fund transaction data and mine the pattern, and a visual chart (such as a heat map, a scatter plot, etc.) is automatically generated; finally, according to the dialogue context and decision rules, the analysis results are returned in an interactive question and answer form in stages, and result tracing and parameter adjustment are supported.

[0096] As a preferred embodiment of the present application, with reference to Figure 2 , the step S1 comprises:

[0097] Step S11, using a hybrid model of deep neural network and rule engine to classify the user intent;

[0098] Step S12, identifying the entity parameters based on the fine-tuned BERT model and rule engine; wherein the entity parameters include amount, time, transaction type, account name and transaction counterparty.

[0099] Wherein, the type of user intent includes trend analysis, risk detection or comparative query; and each type of user intent is further subdivided into more specific intents, such as "monthly spending trend", "high-risk transaction identification", "credit card and savings card consumption comparison". And the identification model based on the fine-tuned BERT model and rule engine is trained by a large amount of financial transaction text data, which can accurately understand professional terms and industry custom expressions.

[0100] Specifically, in the embodiments of the present application, the user demand is first accurately classified into high-level categories such as trend analysis, risk detection or comparative query by a user intention classification framework, and then combined with a business rule library to subdivide specific analysis scenarios such as "monthly expenditure trend"; at the same time, a BERT (Bidirectional Encoder Representations from Transformers) model fine-tuned in the financial field is used to analyze the complex semantics of professional terms, and a rule engine is used to strongly constrain the processing of ambiguous expressions such as "large transaction" dynamically associated with user historical behavior thresholds, time implicit expressions such as "financial report season" automatically mapped to Q1 / Q4 and other financial scenario-specific phenomena, to ensure that in the highly specialized context of capital transactions, not only can industry jargon such as "cross-border capital allocation" be understood, but also the easily overlooked entity parameter boundaries can be accurately captured, effectively solving the problems of term ambiguity and logical missing in general NLP models in the financial vertical field.

[0101] As a preferred embodiment of the present application, with reference to Figure 3 , the step S2 comprises:

[0102] Step S21, recording historical query parameters and generated chart IDs through dialogue state tracking technology;

[0103] Step S22, converting the recognition results of the user intention and the entity parameters into structured query parameters in JSON format.

[0104] The dialogue state tracking (DST) technology adopts a combined architecture of finite state machine and Transformer, supports context association, can handle long-distance dependency, and ensures the consistency and coherence of the intention in multi-round dialogue.

[0105] Specifically, in the embodiments of the present application, the mandatory business rules of the financial analysis scenario (such as the need to associate the central bank blacklist library for anti-money laundering screening and the need to bind the risk level threshold for large transactions) are solidified by the finite state machine (FSM), and the self-attention mechanism of the Transformer is used to dynamically capture long-distance semantic dependencies in multi-round dialogue, for example, when the user changes from "Q1 securities transaction trend" to "abnormal transactions with counterparty A", the model automatically associates the time interval and transaction type parameters in the previous context; this hybrid architecture enables dialogue state tracking to be constrained within a professional analysis framework, while flexibly understanding complex references such as "separately list the cross-border data in the previous chart", and finally converts the dynamically maintained intention chain and parameter set into a machine-executable JSON structure, providing accurate and traceable instruction input for the multi-dimensional analysis engine in the back end.

[0106] As a preferred embodiment of the present application, with reference to Figure 4 , the step S3 comprises:

[0107] Step S31, calling a multi-dimensional analysis model according to the user intention.

[0108] Specifically, in the embodiment of the present application, the multi-dimensional analysis model adopts a hybrid expert large model (MoE, Mixture of Experts) and a traditional algorithm architecture, and through a dynamic routing mechanism, the user intention (such as trend prediction / risk detection) is distributed to sub-models such as financial time series trend prediction, risk prediction, classification and clustering, realizing the collaborative cutting and drilling of multiple dimensions such as amount, time, transaction type, account name, and transaction counterpart, and solving the rigid modeling limitations of traditional online analytical processing (OLAP, Online Analytical Processing) engines in complex financial scenarios.

[0109] Step S32, analyzing the fund transaction data through the multi-dimensional analysis model, with reference to Figure 5 , specifically comprising:

[0110] Step S321, using a CNN-BiLSTM model and combining discrete wavelet transform (DWT, Discrete Wavelet Transform) and variational mode decomposition (VMD, Variational Mode Decomposition) for time series trend prediction.

[0111] Step S322, constructing a transaction graph network and analyzing node relationships through a graph neural network to identify abnormal transaction patterns.

[0112] Specifically, in the embodiment of the present application, a transaction graph network is constructed, wherein the nodes represent accounts / partners, and the edges represent transaction relationships. Graph neural network (GNN, Graph Neural Network) is used to analyze node embedding and edge weight, identify hidden patterns such as "closed-loop money laundering cycle" (A→B→C→A fund loop) or "abnormal late-night large amount transfer", and break through the detection bottleneck of traditional rule engines for new risks.

[0113] Step S323, using a random forest and a hybrid expert large model to classify transaction types and account behaviors, and supporting multi-dimensional data fusion analysis.

[0114] Specifically, in this embodiment of the invention, structured transaction features (amount segmentation, frequency statistics, etc.) are processed by random forest for preliminary classification. The MoE model then dynamically integrates NLP experts (to parse transaction postscript semantics), graph experts (to analyze transaction network topology), and time series experts (to detect behavioral pattern mutations) to perform fine-grained risk assessment. Finally, a weighted fusion is used to output a multi-label classification result with confidence.

[0115] Specifically, step S321 includes:

[0116] Step S3211: Process the original sequence of the fund transaction data x = {x1, x2, ..., x...} T},x t Perform J-level discrete wavelet transform (DWT) decomposition on ∈R, and take the highest-level approximation A. j As the denoised sequence The specific formula is expressed as follows:

[0117]

[0118] in,

[0119] DWT J This represents a function that performs a J-order discrete wavelet transform.

[0120] A j This represents the low-frequency approximation coefficient vector obtained from the j-th level decomposition, which captures the main trends and low-frequency information of the original sequence. Its dimension is T / 2. j , where the length is halved as j increases;

[0121] D j This represents the high-frequency detail coefficient vector obtained from the j-th level decomposition. It captures the details, noise, and mutation information of the original sequence, and its dimension is also T / 2. j ;

[0122] Step S3212, denoise the sequence The input variational mode decomposition (VMD) is decomposed into K intrinsic mode functions (IMFs), as expressed by the following formula:

[0123]

[0124] in,

[0125] VMD stands for Variational Mode Decomposition Function;

[0126] u k This represents the k-th intrinsic mode function (IMF) obtained from the decomposition. It is a time series and represents... Oscillation modes at different scales, with a dimension of T' (the same length as the denoised sequence);

[0127] T' represents a noise reduction sequence The length is equal to T / 2 J ;

[0128] Step S3213, one-dimensional convolution operation is performed on each IMF sequence u k to extract local timing features f k , and all IMF local timing features f k are spliced and then passed through a bidirectional LSTM to obtain timing hidden representations, and the specific formula is as follows:

[0129] f k = CNN (u k ) ∈ R T”×d ,

[0130] F = [f1; f2;...; f k ] ∈ R T”×(K·d)

[0131] Wherein,

[0132] CNN represents a convolutional neural network operation;

[0133] u k represents the kth IMF sequence input to the CNN;

[0134] f k represents the feature matrix extracted by the CNN from the kth IMF sequence u k ;

[0135] d represents the number of convolution output channels T'<T;

[0136] F represents the comprehensive feature matrix formed by splicing all K IMF extracted features f_k;

[0137] T' represents the length of the feature after splicing in the time dimension (the same as the time length of each f k );

[0138] K·d represents the total length of the feature after splicing in the channel dimension (each f k contributes d channels);

[0139] Step S3214, attention is applied to the BiLSTM output sequence to obtain a weighted context vector c, and the specific formula is as follows:

[0140]

[0141] Wherein,

[0142] Wh represents a learnable weight matrix for converting hidden state h tprojected to the attention space;

[0143] b denotes a learnable bias vector;

[0144] tanh() denotes the hyperbolic tangent activation function;

[0145] v denotes a learnable weight vector used to compute the attention score;

[0146] e t denotes the unnormalized attention score (a scalar) computed at time step t;

[0147] a t denotes the normalized attention weight computed at time step t, which indicates how much the model should "focus" on time step t when generating the final representation;

[0148] c denotes the computed attention context vector, which is a weighted sum of all the hidden states ht, with weights determined by a t ; this vector contains summary information of the entire input sequence, and especially emphasizes the parts considered important by the attention mechanism;

[0149] In step S3215, the weighted context vector c is predicted through several fully connected layers to obtain a prediction value vector, and the specific formula is:

[0150]

[0151] wherein,

[0152] denotes the prediction vector of the model for the future value;

[0153] denotes the predicted value of the i-th time point after the original time point T, wherein i = 1, 2,..., H;

[0154] H denotes the number of future time steps to be predicted (prediction step, a preset integer);

[0155] FC(c) denotes a fully connected layer (Fully Connected layers) operation, usually containing one or more linear layers plus an activation function.

[0156] In step S33, the chart type is automatically matched according to the analysis result, and a visual analysis result is generated.

[0157] Specifically, in the embodiments of the present application, the trend analysis is matched with a broken line chart, the risk distribution is matched with a heat map, and the fund flow direction is matched with a Sankey chart or a GNN chart. The visualization technology adopts a D3.js or a GNNLens2 framework. The D3.js drives the Sankey chart to display the fund flow across levels, and the GNNLens2 framework is specifically used for deconstructing the black box decision of a graph neural network, such as highlighting the propagation path of a high-risk account, so that the complex analysis result is interpretable and drillable.

[0158] As a preferred embodiment of the present application, referring to Figure 6 , the step S4 comprises:

[0159] Step S41, record the user intention, historical interaction record, generated chart ID and dialogue round in a JSON format;

[0160] Step S42, dynamically select a reply form and optimize the reply content through a rule engine and deep reinforcement learning according to the user intention and the historical interaction record;

[0161] Step S43, trigger natural language follow-up questions through reference relationship identification and parameter missing detection;

[0162] Step S44, store high-frequency query results by using a cache mechanism, and directly return the cached data for similar requests by combining a multi-model staged processing technology.

[0163] Specifically, in the embodiments of the present application, referring to Figure 7 , efficient multi-round interaction is realized through the cooperation of structured state recording and intelligent decision-making technology: first, a JSON state record containing user intention, historical interaction record, generated chart ID and dialogue round is constructed, and a self-attention mechanism is used to analyze the cross-round semantic association, such as automatically associating the chart ID of the previous query with "the above result"; then, based on a pre-defined rule engine and a deep reinforcement learning strategy (DDPG, Deep Deterministic Policy Gradient), the optimal reply form (such as a trend chart, a risk heat map or a table) is dynamically selected; when parameter missing or reference ambiguity is detected, the parameter missing or reference ambiguity scenario is identified through "function amount detection", semantic similarity calculation and slot filling are triggered, and professional follow-up questions in the financial field are generated, such as "Which account's transaction data do you need to view?" or "Please confirm the time range you need to analyze"; finally, query feature hash matching and a sliding time window mechanism are used to realize analysis result cache reuse, the cached response is directly returned for the request with a feature matching degree exceeding a threshold, and deep analysis is performed asynchronously to update the cache, so as to balance the response timeliness and analysis depth of the system.

[0164] Referring to Figure 8The application also provides a question-answer type fund transaction data analysis system 100, which applies the question-answer type fund transaction data analysis method as described above and comprises the following components:

[0165] A natural language understanding module 110 is configured to receive a fund transaction data analysis related question input by a user, identify the user's intention through a natural language processing technology, and extract entity parameters from the query text;

[0166] A dialogue management module 120 is connected to the natural language understanding module 110 and configured to maintain a multi-round dialogue state and convert the user's intention and the entity parameters into structured query parameters;

[0167] A data analysis module 130 is connected to the dialogue management module 120 and configured to call a multi-dimensional analysis model according to the structured query parameters, analyze the fund transaction data, and generate a visual result;

[0168] An interaction optimization module 140 is connected to the data analysis module 130 and configured to return the visual result to the user through a multi-round interaction mechanism based on a dialogue state and an intelligent decision-making strategy.

[0169] Specifically, in the embodiment of the application, the working process of the question-answer type fund transaction data analysis system is as follows:

[0170] First, the user inputs a fund transaction data analysis related question through natural language, such as "Please show the trend of large abnormal transactions of the Beijing branch last month".

[0171] Then, the natural language understanding module 110 analyzes the input through a natural language processing technology, accurately identifies the user's core query intention, such as "analyze abnormal transaction trends", and extracts key entity parameters, such as the time "last month", the region "Beijing branch", and the index "large abnormal transactions".

[0172] Subsequently, the dialogue management module 120 receives these intentions and parameters, converts them into structured query parameters that can be processed by the system, and dynamically maintains a dialogue state to support the context connection of multi-round interactions.

[0173] Then, the data analysis module 130 calls a pre-set multi-dimensional analysis model, such as a time series analysis, a regional comparison analysis, and an abnormal detection model, according to these structured parameters, performs real-time or batch calculation and analysis on the fund transaction data in the background, and generates corresponding analysis results.

[0174] Finally, the interaction optimization module 140 combines the current dialogue state and intelligent decision-making strategies, such as result complexity and user historical preferences, to efficiently and friendly return the generated visualization results to the user through a multi-round interaction mechanism (e.g., returning the core chart first, and then asking whether to drill down details or export the report), completing a complete question-and-answer analysis interaction.

[0175] As a preferred embodiment of the present application, in order to realize efficient transmission and system integration of analysis results, the data analysis module 130 is provided with an API interface.

[0176] Specifically, in the embodiment of the present application, the API interface is designed in a standardized RESTful style, follows the uniform resource location and operation specification, and supports request response through HTTP methods (such as GET / POST); the interface defines a structured data transmission format, and the transmitted analysis result data includes key parameters such as chart type (such as column chart, heat map, line chart, etc.), data points (covering timestamp, transaction amount, account ID, etc. Dimension field), color mapping rule (such as matching red-yellow-green gradient color value according to risk level), etc. Ensure that the visualization module can accurately render the analysis results based on parameters; at the same time, the interface has built-in parameter verification mechanism and error code return logic, which supports quick positioning and repair of abnormal data, and guarantees seamless connection with the visualization module.

[0177] In addition, the API interface uses SSL / TLS encryption transmission protocol and integrates Role-Based Access Control (RBAC, Role-Based Access Control) mechanism, which meets the requirements of financial systems for data security and compliance, and effectively realizes the compatible integration of analysis module and financial business system.

[0178] In particular, in the field of financial transaction data analysis, the high sensitivity of data and strict regulatory requirements make security and compliance the core considerations for system integration. This API interface integrates the SSL / TLS encryption transmission protocol, fundamentally ensuring the confidentiality and integrity of the analysis results during transmission. SSL / TLS establishes an end-to-end encryption tunnel, effectively resisting network eavesdropping, man-in-the-middle attacks and other threats, ensuring that analysis result data containing sensitive dimensions such as transaction amounts and account IDs is not leaked or tampered with when flowing from the data analysis module to visualization or business systems. At the same time, the integration of Role-Based Access Control (RBAC) mechanism provides a fine-grained management framework for API access permissions. RBAC defines clear user roles and assigns each role the minimum necessary permissions, strictly limiting the access range and operation permissions (such as query, export) of analysis result data for users of different identities. This fine-grained permission control not only effectively prevents unauthorized access and internal data abuse risks, but also directly meets the core compliance clauses of the financial industry, such as the data minimization principle, separation of duties (SoD), and strict access audit requirements. The combination of dual security mechanisms - transmission layer encryption ensures data security "on the way", and application layer RBAC ensures data security "at the end", together building a solid protection system, significantly reducing the risk of data leakage. This not only improves the overall security posture of the system, but more importantly, it enables the seamless, secure integration of the analysis module into the existing, highly secure financial business system ecosystem, meeting strict internal risk control strategies and external regulatory review requirements, thus truly realizing the secure, compliant, and efficient integration of analysis capabilities and business environments.

[0179] Here, those skilled in the art can understand that the specific functions and operations of each module in the above-mentioned one question and answer type fund transaction data analysis system 100 have been described in detail above with reference to the description of one question and answer type fund transaction data analysis method 100, and therefore repeated descriptions thereof will be omitted. Figure 1

[0180] ​In conclusion, the application provides a question and answer type fund transaction data analysis method and system, which accurately identifies financial professional terms and user intentions through a hybrid model of deep neural network and rule engine, dynamically maintains multi-round dialogue state by using finite state machine and Transformer architecture and converts it into structured query parameters, constructs an analysis engine based on a hybrid expert large model and traditional algorithm to realize multi-dimensional dynamic drilling of fund data, intelligently matches visual charts (such as line charts, heat maps, and Sankey diagrams) according to the analysis type, optimizes the interaction strategy by combining deep reinforcement learning, realizes a low-threshold, high-efficiency, and scenario-based intelligent analysis closed loop through parameter missing detection, cache mechanism, and multi-round follow-up questions, and effectively solves the problems of complex professional terms and coupled analysis dimensions in the financial field, which lead to cumbersome operation of traditional query tools and delayed response.

[0181] The above description is only the preferred embodiment of the application, and does not limit the implementation and protection scope of the application. It should be realized by those skilled in the art that any equivalent replacement and obvious changes made according to the description and drawings of the application should be included in the protection scope of the application.

Claims

1. A question-and-answer style fund transaction data analysis method, characterized in that, The method comprises the following steps: Step S1, receiving user input fund transaction data analysis related questions, identifying user intent through natural language processing technology, and extracting entity parameters from query text; Step S2, maintaining multi-round dialogue state, converting the user intent and the entity parameters into structured query parameters; Step S3, calling a multi-dimensional analysis model according to the structured query parameters, performing multi-dimensional analysis on the fund transaction data, and generating a visualization result; Step S4, based on the dialogue state and intelligent decision-making strategy, returning the visualization result to the user through a multi-round interaction mechanism.

2. The method of claim 1, wherein, The step S1 comprises: Step S11, using a hybrid model of deep neural network and rule engine to classify the user intent; Step S12, identifying the entity parameters based on the BERT model and the rule engine; wherein the entity parameters include amount, time, transaction type, account name and transaction counterparty.

3. The method of claim 1, wherein, The step S2 comprises: Step S21, recording historical query parameters and generated chart IDs through dialogue state tracking technology; wherein the dialogue state tracking technology adopts a combined architecture of finite state machine and Transformer; Step S22, converting the recognition results of the user intent and the entity parameters into structured query parameters in JSON format.

4. The method of claim 1, wherein, The step S3 comprises: Step S31, calling a multi-dimensional analysis model according to the user intent; wherein the multi-dimensional analysis model is based on a hybrid expert large model and a traditional algorithm, and fuses time, amount, transaction type, account and transaction counterparty entity dimensions; Step S32, analyzing the fund transaction data through the multi-dimensional analysis model; Step S33, automatically matching chart types according to the analysis results, and generating a visualization analysis result.

5. The method of claim 4, wherein, The step S32 comprises: Step S321, using a CNN-BiLSTM model, and combining discrete wavelet transform DWT and variational modal decomposition VMD to perform time series trend prediction; Step S322, constructing a transaction graph network, and analyzing node relationships through a graph neural network to identify abnormal transaction patterns; Step S323, using a random forest and a hybrid expert large model to classify transaction types and account behaviors, and supporting multi-dimensional data fusion analysis.

6. The method of claim 5, wherein, The step S321 comprises: Step S3211, do J-level discrete wavelet transform DWT decomposition on the original sequence x = {x1, x2,..., x T},x t ∈R, and take the highest-level approximation A j as the sequence after noise reduction The specific formula is: Wherein, DWT J represents a function that performs a J-level discrete wavelet transform; A j represents the low-frequency approximation coefficient vector obtained by the jth-level decomposition, with dimension T / 2 j ; D j denotes the high frequency detail coefficient vector obtained by the jth level decomposition, with dimension T / 2 j ; Step S3212, sequence after noise reduction Input the variational modal decomposition VMD, and decompose it into K intrinsic modal functions IMFs, and the specific formula is: satisfy Wherein, VMD represents a variational modal decomposition function; u k denotes the kth eigenmode function obtained by decomposition, with dimension T'; T' denotes a noise reduction sequence of length equal to T / 2 J ; Step S3213, for each IMF sequence u k perform one-dimensional convolution operation to extract local timing features f k , of all IMF sequences k After splicing, pass through bidirectional LSTM to obtain timing hidden representation, and the specific formula is represented as: f k = CNN(u k )∈R T”×d , F = [fl; f2;... ; fn] e R k ] e R T”×(K·d) Wherein, CNN represents a convolutional neural network operation; u k represents the k-th IMF sequence input to the CNN; f k represents the CNN for the k-th IMF sequence u k extracted feature matrix; d represents a convolution output channel number T'<T; F represents a comprehensive feature matrix formed by splicing all K IMF extracted features f_k; T' represents the length of the spliced features in the time dimension; K·d represents the total length of the spliced features in the channel dimension; Step S3214, the BiLSTM output sequence Attention is applied to obtain a weighted context vector c, and the specific formula is represented as: e t = v T tanh(W t h t + b), t = 1,..., T' Wherein, Wh represents a learnable weight matrix; b represents a learnable bias vector; tanh() represents a hyperbolic tangent activation function; v represents a learnable weight vector; e t denotes the unnormalized attention scores computed at time step t; a t denotes the normalized attention weights computed at time step t; c represents a calculated attention context vector; Step S3215, passing the weighted context vector c through several fully connected layers to predict the future H steps, and obtaining a predicted value vector, which is specifically represented by the formula: Wherein, represents a prediction vector of the model for future values; represents the predicted value at the i-th time point after the original time point T, i = 1, 2,..., H; H represents the number of future time steps to be predicted; FC(c) denotes a fully connected layer operation.

7. The method of claim 4, wherein, The step S33 comprises: trend analysis matching a broken line chart, risk distribution matching a heat map, fund flow direction matching a Sankey diagram or a graph neural network diagram.

8. The method of claim 1, wherein, The step S4 comprises: Step S41, record the user's intention, historical interaction record, generated chart ID and dialogue round in JSON format; Step S42, dynamically select the reply form and optimize the reply content through the rule engine and deep reinforcement learning according to the user's intention and the historical interaction record; Step S43, trigger natural language follow-up questions through reference relationship identification and parameter missing detection; Step S44, store high-frequency query results by using a cache mechanism, and directly return cached data for similar requests by combining multi-model phased processing technology.

9. A question and answer based data analysis system for capital transaction data, characterized by, The application of the fund transaction data analysis method of the question and answer type according to any one of claims 1-8 comprises: A natural language understanding module is configured to receive a user input fund transaction data analysis related question, identify a user intention through a natural language processing technology, and extract entity parameters from a query text; A dialogue management module is connected to the natural language understanding module and configured to maintain a multi-round dialogue state, and convert the user intention and the entity parameters into structured query parameters; A data analysis module is connected to the dialogue management module and configured to call a multi-dimensional analysis model according to the structured query parameters, analyze the fund transaction data, and generate a visual result; An interaction optimization module is connected to the data analysis module and configured to return the visual result to the user through a multi-round interaction mechanism based on a dialogue state and an intelligent decision-making strategy.

10. The question and answer funds transaction data analysis system of claim 9, wherein, The data analysis module is provided with an API interface, and the API interface is designed in a standardized RESTful style.

Images