Financial intelligent question and answer method and system based on artificial intelligence

Abstract

The invention provides a financial intelligent question answering method and system based on artificial intelligence. The method comprises the steps that financial questions are collected for question text modeling, and keywords in the questions are analyzed; selecting a cross-platform data source related to the question by using a unified data access layer based on the structured financial question representation to perform retrieval query; constructing a following inference model according to the cross-platform data and the context information; when the current financial market condition changes, designing a timeliness dynamic adjustment mechanism to perform timeliness dynamic adjustment on the financial data; and collecting user feedback based on the preliminary intelligent decision-making result and the preliminary intelligent decision-making result to generate a final decision-making result. The problems of privacy disclosure, insufficient context understanding, limited data processing capability and the like in the prior art are effectively solved, and a new solution is provided for large-scale and cross-organization financial data analysis and intelligent question and answer.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and in particular relates to a financial intelligent question-answering method and system based on artificial intelligence. Background Art

[0002] With the advent of the big data era, massive amounts of text data are rapidly accumulating. How to effectively and automatically analyze and annotate these data has become a crucial technical issue. The text tag generation method is widely used in search engines, recommendation methods, natural language processing and other fields. It can generate brief tags suitable for the text through semantic analysis of text data, thereby improving the retrieval, classification and understanding capabilities of text data. Traditional text tag generation technology mainly relies on two methods: one is a tag generation method based on artificial rules, and the other is an automatic tag generation method based on machine learning and deep learning models.

[0003] The typical approach of the manual rule method is to generate labels for texts by having experts design a series of rules and keyword matching models based on domain knowledge. However, this method relies heavily on the experience of domain experts, has poor applicability and low scalability, and cannot be used flexibly in different fields. Especially today when the amount of data is increasing rapidly, relying on manual rules is both time-consuming and labor-intensive, and cannot meet the needs of high-frequency updates. In addition, rule methods are often only applicable to specific scenarios and cannot effectively handle complex semantics, polysemous words, or semantic changes in different contexts, resulting in the generated labels being difficult to accurately reflect the complex semantics of the text content.

[0004] The automatic label generation method based on machine learning and deep learning mainly relies on the training of large-scale data sets. By learning from massive annotated data, the model can extract the deep semantics of the text and generate corresponding labels. The emergence of deep learning technologies such as BERT, GPT and other language models has greatly improved the accuracy and efficiency of label generation. However, existing deep learning methods also face some limitations. First, the training process of such models usually relies on a large amount of annotated data, while in many emerging fields or fields with less data, the lack of annotated data makes it difficult to fully train the model. Secondly, most existing models are single-modal models, which mainly rely on text input for label generation, and it is difficult to combine other modal data (such as images, audio, etc.) to generate more comprehensive labels. Especially in scenarios with rich multimodal information such as social media and news, the label generation of the present invention that relies solely on text is difficult to capture all information. In addition, most of the current deep learning models are black box models, lacking transparency, and users cannot clearly understand the basis for the model to generate labels, especially in sensitive fields such as medical, finance, and law. The unexplainable label generation process may cause users to lose trust in the present invention. Therefore, although the existing technology can improve the level of automation of text label generation to a certain extent, it still has major shortcomings in terms of poor cross-domain adaptability, low transparency of label generation, and insufficient ability to integrate multimodal information. Summary of the invention

[0005] The purpose of the present invention is to design a financial intelligent question and answer method and system based on artificial intelligence. By combining differential privacy protection, distributed computing, hierarchical neural networks and timeliness adjustment mechanism, the present invention provides a financial question and answer system with high privacy protection and high intelligence level, which can effectively solve the problems of privacy leakage, insufficient context understanding and limited data processing capabilities in the prior art, and provides a new solution for large-scale, cross-organizational financial data analysis and intelligent question and answer.

[0006] In order to achieve the above object, a first aspect of the present invention provides a financial intelligent question-answering method based on artificial intelligence, the method comprising the following steps: S1. Collect financial questions to model the question text and analyze the keywords in the questions, and analyze the implicit financial goals in the questions according to the keywords based on the preset knowledge graph and rule-based reasoning method to obtain a structured financial question representation; S2. Based on the structured financial problem representation, a unified data access layer is used to select cross-platform data sources related to the problem for retrieval query. During the retrieval query process, noise is injected into the data through the differential privacy mechanism to prevent the leakage of sensitive information. Finally, the financial data returned by multiple platforms are aggregated through the aggregation operation. S3, build a contextual reasoning model based on cross-platform data and contextual information, supplement the missing and interpretable nature of the aggregated financial data through contextual information, enhance the aggregated financial data, and integrate multi-platform data sources to perform multi-level analysis and abnormal financial data detection. When abnormal financial data appears, use an adaptive anomaly detection algorithm to correct the abnormal financial data and input the final financial data into S4; S4. When the current financial market situation changes, a dynamic adjustment mechanism is designed to dynamically adjust the timeliness of the financial data generated by S3. According to the financial data after dynamic adjustment, an answer generation mechanism based on deep learning is introduced to generate intelligent decision-making results for financial problems, and the intelligent decision-making results are input into S5. S5. Collect user feedback and preliminary intelligent decision results based on preliminary intelligent decision results, design a user feedback mechanism based on user feedback, maximize the impact of user feedback on the financial decision-making system, and enable it to dynamically adapt to long-term changes in user preferences, while adaptively adjusting the model of the problem text modeling, and generate the final decision result after completing user feedback and adaptive adjustment.

[0007] Further, the structured financial problem representation includes financial goals, report types, time intervals and keywords; In S1, financial questions are collected to model the question text and analyze the keywords in the questions. The Transformer model fine-tuned based on the financial field is used to perform semantic modeling on the question text. The Transformer model fine-tuned based on the financial field uses the BERT model fine-tuned with financial field data to encode the question text. The BERT model can convert the vocabulary of each question into a semantic vector, thereby capturing the deep meaning of the question; The keyword analysis is as follows: First, we design the importance score of each word to analyze the output of the BERT model and obtain keywords; Then, these keywords and their contexts are weighted averaged based on the weight of context information to obtain semantic representation; The preset knowledge graph and rule-based reasoning method include: The knowledge graph includes the relationship between various financial indicators and financial goals; The rule is to infer possible financial goals based on the keywords in the question.

[0008] Furthermore, the S1 also includes setting a rule-based time recognition and standardization module to extract time information from the question and use regular expressions and simple rule sets to extract and convert it into a standard time interval.

[0009] Furthermore, the data is injected with noise through the differential privacy mechanism to prevent leakage of sensitive information, and finally the financial data returned by multiple platforms are aggregated through an aggregation operation; wherein the privacy budget of the differential privacy mechanism controls the noise intensity of the noise injection to ensure a balance between data privacy and data quality; Wherein, the differential privacy mechanism further includes: The privacy budget is dynamically adjusted according to the privacy budget threshold and the current privacy budget; Said S2 further comprises: Aggregate financial data is encrypted using the AES algorithm to generate encrypted financial data to ensure data security during storage and transmission.

[0010] Furthermore, the contextual reasoning model uses a dual attention mechanism to enhance the financial data according to the time dimension and context information, specifically including: First, the contextual reasoning model assigns a temporal attention weight to each data point, and the temporal attention weight is dynamically adjusted according to the timeliness of the data; Secondly, the contextual reasoning model gives attention weights to contextual information according to the industry background or macroeconomic environment of the data; Finally, based on the time attention weight and the attention weight of the context information, the financial data is enhanced by weighted sum calculation to obtain enhanced financial data; The comprehensive multi-platform data source integrates financial data by adopting a hierarchical weighted fusion mechanism, which is used to perform hierarchical processing on the financial data and gradually integrate the data characteristics of each level.

[0011] Furthermore, the adaptive anomaly detection algorithm dynamically adjusts the anomaly detection threshold according to the normalization of data distribution, including: During the detection process, the system calculates the degree of deviation of each data point from the mean and standard deviation of the current data set to determine whether it is an outlier: If the deviation exceeds the preset threshold, the data point is determined to be an outlier; For abnormal data, the system repairs it based on historical financial data and contextual information, and adopts a repair method based on multi-task learning, which can select the most appropriate repair strategy according to different types of abnormal data.

[0012] Furthermore, the dynamic adjustment mechanism of timeliness is constructed as follows: The time decay factor is designed to dynamically adjust the importance of the data in the model according to its historical timestamp; the industry impact factor is designed to adjust the impact of the current industry economic environment on the financial data; the external event adjustment factor is designed to introduce the real-time impact of external events on financial data. If a specific event affects the company's performance, the relevant financial data will be adjusted; The dynamic adjustment mechanism for timeliness is based on the original financial data, the financial data adjusted by the industry impact factor, and the external event adjustment factor, and performs a dynamic adjustment for timeliness on the financial data by using an exponentially weighted moving average based on a time decay factor to obtain the financial data that has been dynamically adjusted for timeliness; The answer generation mechanism is generated by combining a long short-term memory network with a self-attention mechanism.

[0013] Furthermore, the intelligent decision-making results include predicted net profits and industry risk assessments, providing decision makers with optimization suggestions and displaying time-adjusted data; the intelligent decision-making results are output in a structured JSON format.

[0014] Furthermore, the adjustment of the user feedback specifically includes: The preliminary financial decision generated after the dynamic adjustment of the collection timeliness is revised based on the current user's immediate feedback data and the user's long-term preferences to obtain the financial decision result adjusted by the user feedback, which is used to maximize the impact of the user's feedback on the financial decision system and enable it to dynamically adapt to the long-term changes in user preferences; The adaptive adjustment is based on an adaptive optimization mechanism, and the adjustment based on user feedback is optimized by gradient optimization, wherein the gradient of the gradient optimization is used to optimize the adjustment based on user feedback.

[0015] In a second aspect of the present invention, a financial intelligent question-answering system based on artificial intelligence is provided, the system comprising: The financial problem analysis module is used to model the problem text of financial problems and analyze the keywords in the problem. According to the preset knowledge graph and rule-based reasoning method, the implicit financial goals in the problem are analyzed according to the keywords to obtain a structured financial problem representation; The cross-platform data retrieval module is used to select cross-platform data sources related to the problem for retrieval query based on the structured financial problem representation using a unified data access layer. During the retrieval query process, noise is injected into the data through a differential privacy mechanism to prevent leakage of sensitive information. Finally, the financial data returned by multiple platforms is aggregated through an aggregation operation. The contextual reasoning integration module is used to build a contextual reasoning model based on cross-platform data and contextual information, supplement the missing and interpretable nature of aggregated financial data through contextual information, enhance the aggregated financial data, and integrate multi-platform data sources to perform multi-level analysis and abnormal financial data detection. When abnormal financial data appears, an adaptive anomaly detection algorithm is used to correct the abnormal financial data, and the final financial data is input into the timeliness dynamic adjustment and answer generation module; The module for dynamic adjustment of timeliness and answer generation is used to design a dynamic adjustment mechanism for timeliness to dynamically adjust the financial data generated by the contextual reasoning integration module when the current financial market situation changes. The module introduces a deep learning-based answer generation mechanism based on the financial data after dynamic adjustment of timeliness to generate intelligent decision-making results for financial problems, and inputs the intelligent decision-making results into the user feedback and answer optimization module. The user feedback and answer optimization module is used to collect user feedback and preliminary intelligent decision-making results based on the preliminary intelligent decision-making results, design a user feedback mechanism based on user feedback, maximize the impact of user feedback on the financial decision-making system, and enable it to dynamically adapt to long-term changes in user preferences. At the same time, the model for modeling the question text is adaptively adjusted, and the final decision result is generated after completing user feedback and adaptive adjustment.

[0016] The beneficial technical effects of the present invention are at least as follows: The present invention uses differential privacy technology to ensure that sensitive financial data will not be leaked when users query in a multi-party data source environment. Different from traditional privacy protection methods (such as encrypted transmission, access control, etc.), differential privacy can process query requests by adding noise to ensure the privacy of data while not affecting the validity of query results. This technology can solve the data leakage risk caused by the lack of privacy protection measures in traditional financial question-and-answer systems, especially in cross-organizational and cross-platform data exchange and analysis, ensuring the data security of users.

[0017] The present invention designs a distributed data storage and computing architecture that supports joint query and processing of data from multiple distributed nodes (such as different enterprises, departments, financial platforms, etc.). Through this architecture, the system can effectively perform unified analysis on data from different data sources, avoiding the limitation that traditional systems can only process a single data source. While using differential privacy technology to ensure data privacy, the system can efficiently integrate and analyze data across organizations and platforms, comprehensively improving the intelligence and usability of the financial question-and-answer system.

[0018] The present invention improves the system's ability to understand the context of complex financial issues by introducing a hierarchical neural network architecture. Traditional financial question-and-answer systems are usually unable to effectively understand the contextual information in the questions, especially when it comes to multi-level, time-sensitive financial issues, and are unable to generate accurate answers. The hierarchical neural network of the present invention can parse financial issues at multiple levels, from basic financial items (such as income and expenditure) to more complex time-sensitive issues (such as quarterly comparisons of financial statements, annual trends, etc.), and gradually reason to generate more accurate and detailed answers. This innovation can effectively solve the understanding barriers in the prior art when dealing with complex and dynamic financial issues, and improve the intelligence level of the system.

[0019] In financial problems, timeliness is a key factor, especially when conducting data comparison and trend analysis, it is often necessary to dynamically adjust the data source and analysis method according to the time dimension in the question (such as "this year", "last year", "first quarter"). The present invention can automatically identify the time factor in the question by introducing a dynamic adjustment mechanism for timeliness, and adjust the analysis scope and processing method of the data according to the time node, thereby ensuring that the financial answer given can reflect the current real-time status and historical changes. This mechanism solves the defect that the existing financial question-and-answer system cannot cope with complex time dimension problems.

[0020] By combining differential privacy protection, distributed computing, hierarchical neural networks and timeliness adjustment mechanisms, the present invention provides a financial question-and-answer system with high privacy protection and high intelligence, which can effectively solve problems such as privacy leakage, insufficient context understanding and limited data processing capabilities in the prior art, and provides a new solution for large-scale, cross-organizational financial data analysis and intelligent question-and-answering. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0022] Figure 1 The present invention is a flowchart of a financial intelligent question-answering method based on artificial intelligence. DETAILED DESCRIPTION

[0023] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0024] In one or more embodiments, Figure 1 As shown, the present invention discloses a financial intelligent question-answering method based on artificial intelligence, and the method comprises the following steps: S1. Collect financial problems to model the problem text and analyze the keywords in the problems, and analyze the implicit financial goals in the problems according to the keywords based on a preset knowledge graph and a rule-based reasoning method to obtain a structured financial problem representation; wherein the structured financial problem representation includes financial goals, report types, time intervals and keywords.

[0025] Specifically, we first perform semantic analysis on this natural language question to identify the financial problem elements, time dimensions, related indicators, etc.

[0026] Among them, the semantic understanding of the question is to pay special attention to how to extract meaningful financial indicators, time dimensions and potential financial goals from natural language questions when dealing with financial problems. Therefore, the present invention uses the Transformer model fine-tuned based on the financial field to semantically model the question text.

[0027] Furthermore, the present invention uses the BERT model fine-tuned with financial data to encode the question text. The BERT model can convert the vocabulary of each question into a semantic vector, thereby capturing the deep meaning of the question. Assume that the present invention obtains a vector representation through the model , it can fully reflect the core meaning of the text.

[0028] ; in, BERT model fine-tuned for the financial domain, is a user question entered.

[0029] Furthermore, in financial problems, keyword extraction is crucial. The present invention first analyzes the output of the BERT model and identifies core financial information (such as "net profit" and "time") by calculating the "importance score" of each word. Then, these keywords are weighted averaged with their context to obtain a more accurate semantic representation. Weighted average formula: ; in, It is vocabulary The weight of yes Embedding vector in BERT model, is the total number of words in the question. The calculation method is as follows: ; in, It is vocabulary The "importance score" in the context is usually calculated based on factors such as frequency of occurrence and position in the context. This weighted average method can effectively enhance the weight of financial key elements (such as "net profit") in the question text and reduce ambiguity.

[0030] Furthermore, financial problems usually involve a specific time dimension, so the present invention designs a rule-based time identification and standardization module that can automatically extract time information from the problem and convert it into a standardized format.

[0031] The time standardization rule is for expressions such as "the first quarter of 2023", "Q12023" or "January 1, 2023 to March 31, 2023", etc. The present invention uses regular expressions and simple rule sets to extract and convert them into standard time intervals.

[0032] For example, a rule could be defined as follows: First quarter of 2023 -> [2023-01-01, 2023-03-31]; Q12023->[2023-01-01,2023-03-31]; January 1, 2023 -> [2023-01-01, 2023-01-01]; Through this rule set, the present invention can convert the time description in the problem into a clear time range, ensuring that the time dimension in subsequent processing is unified and standardized.

[0033] Furthermore, in financial question answering systems, many questions will indirectly contain specific financial goals (for example, "net profit" is usually related to the goal of "maximizing profit"). In order to infer these goals, the present invention uses knowledge graphs and rule-based reasoning methods to parse the implicit financial goals in the questions.

[0034] Furthermore, a knowledge graph in the financial field is constructed, which contains the relationship between various financial indicators and financial goals. For example, net profit, gross profit, operating income, etc. are all directly related to the goal of "profit maximization".

[0035] Then design a set of rules to infer possible financial goals based on the keywords in the question. For example, when "net profit" appears in the question, the system can automatically infer that the goal is "maximizing profit" and determine the type of report involved in the question (such as income statement).

[0036] ; In this way, the system can infer the user's financial needs from the question, thereby further accurately understanding the user's intention in asking the question.

[0037] Furthermore, after the above steps, the system will eventually output a structured financial problem representation, including: Financial objectives: maximize profits; Report type: income statement; Question keywords: Net profit; Time range: January 1, 2023 to March 31, 2023; This structured information provides high-quality input for subsequent financial data retrieval, analysis, and answer generation steps. The output structured representation is as follows: ; It can be understood that the design of this step closely combines the semantic parsing of financial problems with the time dimension and target reasoning. By using the BERT model for problem encoding, the weighted average method for semantic reinforcement, the rule-driven time parsing to ensure the correct time dimension of the problem, and the financial target reasoning based on the knowledge graph, the present invention establishes an effective financial problem understanding framework. This framework can not only identify the core financial elements of the problem, but also provide accurate time information and target reasoning, thereby providing accurate input for subsequent steps (such as financial data retrieval, answer generation, etc.).

[0038] S2. Based on the structured financial problem representation, a unified data access layer is used to select cross-platform data sources related to the problem for retrieval query. During the retrieval query process, noise is injected into the data through the differential privacy mechanism to prevent the leakage of sensitive information. Finally, the financial data returned by multiple platforms are aggregated through aggregation operations.

[0039] Specifically, the input of this step is the output of step 1, that is, the structured financial problem representation. This representation includes: Financial objectives (such as "maximize net profit" or "minimize costs").

[0040] The report type (such as Balance Sheet or Cash Flow Statement).

[0041] A time period (e.g., “first quarter of 2023”).

[0042] Keywords (such as "net profit", "operating income", etc.).

[0043] Furthermore, according to the structured question generated in step S1, the system selects data sources related to the question (such as internal ERP system, external financial API, etc.).

[0044] The unified data access layer (UDAL) generates data retrieval query expressions applicable to each platform. For example, the SQL statement for querying net profit is as follows: $$\text{SELECTnet\_profit}\\ \text{FROMfinancial\_data}\\ \text{WHEREreport\_type='IncomeStatement'}\\ \text{ANDdateBETWEEN'2023-01-01'AND'2023-03-31'}\qquad\text{(1)}$$ UDAL automatically adapts to different data sources and returns standardized structured data. If the data comes from multiple platforms, the system will format the returned data uniformly.

[0045] Furthermore, during the cross-platform data retrieval process, the system injects noise into the data through a differential privacy mechanism to prevent leakage of sensitive information. The noise injection formula is as follows: ; in, is the data point after adding noise, For the original financial data, The mean is 0 and the variance is Gaussian noise.

[0046] Privacy Budget Control the noise intensity to ensure a balance between data privacy and data quality. The noise intensity is adjusted by the following formula: ; in, is the sensitivity of the query function.

[0047] The privacy budget dynamic adjustment formula is: ; This formula is based on the privacy budget threshold and current privacy budgets Adjust your budget.

[0048] Furthermore, through aggregation operations, the financial data returned by multiple platforms are integrated. For example, the "net profit" returned by the internal system and the external API is weighted averaged: ; in, and is the weight of different data sources, and These are the net profit data from two data sources respectively.

[0049] Encrypt aggregated data to ensure data security during storage and transmission. The encryption formula is as follows: ; in, is the aggregated data, is the encryption key, The encrypted data.

[0050] Understandably, the output is: The final output is encrypted financial data retrieved across platforms, including: Aggregated financial data (such as net profit).

[0051] Encryption-protected data (ensuring data security).

[0052] For example, the final output data format is as follows:

[0053] Further, these data will be used for subsequent financial analysis and question and answer generation.

[0054] It can be understood that through cross-platform data retrieval and privacy protection mechanisms, this step efficiently integrates data from multiple platforms and performs privacy protection, aggregation and encryption on financial data. The system can automatically generate standardized query expressions based on financial questions and protect data security through differential privacy mechanisms. This provides reliable data support for subsequent financial intelligent question-answering and ensures data integrity and privacy.

[0055] S3. Build a contextual reasoning model based on cross-platform data and contextual information, supplement the missing and interpretable nature of the aggregated financial data through contextual information, enhance the aggregated financial data, conduct multi-level analysis based on multi-platform data sources, and detect abnormal financial data. When abnormal financial data appears, use an adaptive anomaly detection algorithm to correct the abnormal financial data.

[0056] Specifically, the input of this step is the output of step S2: cross-platform data retrieval and privacy protection, which includes encrypted and aggregated financial data and related contextual information. Specifically, it includes: Aggregated financial data, such as "Net Profit", "Revenue", etc.

[0057] Encrypted data: To ensure privacy protection, financial data is encrypted using AES.

[0058] Meta information of data query, including data source, encryption identifier, query time, query parameters, etc.

[0059] After privacy protection and cross-platform aggregation in step 2, these data lay the foundation for contextual reasoning and multi-level data integration.

[0060] Furthermore, the core of this step is to perform reasoning analysis based on cross-platform data and contextual information (such as industry trends and time series characteristics), extract potential patterns in financial data, and provide enhanced information for the next step of multi-level data integration. The goal of the contextual reasoning model is to supplement the missing and interpretable nature of financial data through contextual information, so that financial data is not just a single value, but content closely related to its background.

[0061] Furthermore, a dual attention mechanism is adopted to enhance financial data based on the time dimension and contextual information. First, the model assigns a time attention weight to each data point (for example, the revenue of a certain month), and this weight is dynamically adjusted according to the timeliness of the data. Secondly, the model gives the attention weight of the contextual information according to the industry background or macroeconomic environment of the data. The specific formula is as follows: ; in, For enhanced financial data. For the Financial data at a certain moment (e.g., revenue data for a certain month). is the attention weight of the time dimension, which controls the importance of financial data at each moment. For the Contextual information at the moment (e.g. industry averages, market trends, etc.). is the attention weight of contextual information, controlling the impact of industry background on financial data. The role of this inference model is to enhance the relationship between the timeliness of data and background information, especially the association reasoning of contextual information such as industry fluctuations and economic cycles.

[0062] Furthermore, after cross-platform data retrieval, in order to better integrate financial data from different data sources, the multi-level data integration model processes the financial data in layers and gradually integrates the data features of each layer. The goal of this model is to ensure that the most valuable information is extracted from each data source and reflect the integration results of different layers in the final output.

[0063] Specifically, financial data may contain input data from multiple platforms, such as data from the "company financial system", industry data from the "external API", etc. Each type of data will be assigned different weights according to its importance and relevance. Through this hierarchical integration, the model can effectively avoid excessive information redundancy while ensuring data integrity. The hierarchical weighted fusion mechanism is used for data integration, and the formula is as follows: ; in, The final integrated financial data. is the input data of the i-th level (such as “company financial data” or “industry average data”). is the weighting coefficient of the i-th level data, which controls the importance of data at different levels. The advantage of data integration is that through hierarchical processing and weighted fusion, the source and function of data can be effectively distinguished, so that the advantages and disadvantages and differences of each data source can be reflected in the final result.

[0064] Furthermore, the next step after data integration is to detect and repair abnormal data to ensure the quality of the integrated data. Abnormal data detection is not only based on statistical features (such as mean and standard deviation), but also combines contextual information and business background to avoid simple numerical anomalies being ignored or misjudged. This step uses an adaptive anomaly detection algorithm, the core idea of ​​which is to dynamically adjust the threshold of anomaly detection based on the standardization of data distribution. Specifically, during the detection process, the system calculates the degree of deviation of each data point from the mean and standard deviation of the current data set. If the deviation exceeds the preset threshold, the data is judged to be abnormal. The detection formula for outliers is: ; in, For data points The standardized value of . and are the mean and standard deviation of the data set respectively. , then determine is abnormal data, where The threshold is set dynamically. For abnormal data, the system will repair it based on historical data, context information, etc. It uses a repair method based on multi-task learning and can select the most appropriate repair strategy based on different types of abnormal data.

[0065] It can be understood that after contextual reasoning, multi-level data integration and abnormal data repair, the system outputs integrated financial data, including the following: Enhanced financial data (e.g., “net profit” based on contextual reasoning models).

[0066] Weighted and hierarchically integrated data (such as the fusion results of "company data", "industry data", etc.).

[0067] Data after abnormal data is repaired (such as the repaired "income" data).

[0068] The output will be returned in structured JSON format for easy subsequent analysis and use. For example, the final JSON format is as follows:

[0069] It is understandable that the innovation of this step is to effectively combine financial data with industry background information through the contextual reasoning model to enhance the interpretability and timeliness of the data. At the same time, the use of a multi-level data integration model ensures the reasonable weighted fusion of various types of data to avoid information redundancy. Combined with the adaptive anomaly detection mechanism, the system can automatically repair possible abnormal data, thereby providing high-quality data support for the final intelligent question and answer generation. The final output data is in a structured, clear format that can be used for subsequent applications, which can provide efficient and accurate financial data support for the intelligent decision-making system in the entire patent.

[0070] S4. When the current financial market situation changes, in order to ensure that the analysis results can reflect the latest economic, industry and company conditions, a dynamic adjustment mechanism is designed to dynamically adjust the timeliness of the financial data generated by S3. According to the financial data after dynamic adjustment, an answer generation mechanism based on deep learning is introduced to generate intelligent decision-making results for financial problems, and the preliminary intelligent decision-making results are input into S5.

[0071] Specifically, the input of this step comes from the output of step S3: contextual reasoning and multi-level data integration. The present invention has obtained financial data through multi-level reasoning and cross-platform integration. , which includes company financials, industry trends, and macroeconomic background. Data enhanced by context , further reasoning and supplementing the context and details of the data. And, the data repaired after anomaly detection , providing a solid foundation for accurate financial analysis.

[0072] In this step, the data will be passed to the timeliness dynamic adjustment and answer generation module. The key task here is to ensure that all input data is not just static historical data, but can also be dynamically updated and inferred in real time based on timeliness factors. The specific data input is still: : Contains financial data after anomaly detection and timeliness correction; : Context-enhanced financial data; : Multi-dimensional financial data after cross-platform integration.

[0073] Furthermore, in the financial decision-making process, the time sensitivity of data is particularly important, especially for situations of rapid market changes and financial warnings. In order to ensure that the analysis results can reflect the latest economic, industry and company conditions, it is necessary to achieve dynamic adjustment of timeliness. The specific adjustment mechanism will consider the following important factors: Time decay factor : The time decay factor is used to dynamically adjust the importance of the data in the model based on its historical timestamp. The farther away the data is from the current moment, the smaller its decay factor. This ensures that historical data does not have an excessive impact on the model.

[0074] Impact of industry changes :Adjust according to the impact of the current industry economic environment on financial data. For example, if an industry has undergone major changes, its past financial data may no longer be fully applicable. The role of is to weight the data based on macro changes in the industry.

[0075] Event Driven Adjustment Factor :Introduce the real-time impact of external events (such as policy changes, industry risks) on data. If a specific event affects the company's performance (such as changes in national fiscal policies), adjust the relevant financial data.

[0076] Based on the above factors, the formula for the timeliness adjustment mechanism is: ; in, Financial data that has been dynamically adjusted based on timeliness. The timestamp is original financial data. is the time decay factor, which is adjusted according to the time interval of the data. It is the industry impact factor, which reflects the impact of industry changes on the data. It is an external event adjustment factor that reflects the immediate impact of major external events on the data. The time frame to be considered.

[0077] Understandably, this formula emphasizes how to dynamically adjust financial data based on time, industry, and external events. Through this adjustment, it is ensured that historical data will not affect the final forecast results due to a long time, thus ensuring the accuracy of real-time financial decisions.

[0078] Furthermore, after the dynamic adjustment of timeliness, the system needs to use the updated financial data, combined with contextual information, to generate intelligent decision support. At this point, the present invention introduces an answer generation mechanism based on deep learning, especially a model that combines the long short-term memory (LSTM) network with the self-attention mechanism. The generation model uses a self-attention mechanism based on multimodal data fusion, combining context-enhanced data, industry information, and time-adjusted financial data to generate answers. This approach can effectively capture the complex associations between different information sources. The generation model formula is as follows: ; in, To generate intelligent decision-making results, such as financial forecasts, risk assessments, optimization recommendations, etc. These are financial data that have been adjusted for timeliness. Enhancing the context of financial data for contextual reasoning data. It is the integrated multi-dimensional data. It is a time series feature that captures historical patterns and trends in financial data.

[0079] During the generation process, the timing characteristics It is extracted and used as the input of time series patterns to help the model understand the possible future performance of historical financial data. The self-attention mechanism can automatically weight the importance of different features, making the model more accurate in generating decision answers, especially when facing data across time and platforms.

[0080] Furthermore, once the intelligent decision results are generated , these results will be output in a structured JSON format to ensure that they can be used by subsequent decision-making systems or manual review processes.

[0081] For example, the output JSON format may be as follows:

[0082] Among them, forecasted_net_profit: predicted net profit. This value is the future net profit value generated based on dynamically adjusted financial data and context information.

[0083] risk_level: Industry risk assessment, reflecting future risk predictions based on industry changes and historical data.

[0084] Recommendations: Optimization recommendations provided to decision makers, generated based on dynamic data and contextual reasoning results.

[0085] adjusted_data: Displays data that has been adjusted for timeliness to help understand how the model output was generated based on the adjusted data.

[0086] It is understandable that this step solves the challenge of reflecting dynamic changes in the financial decision-making process in real time through dynamic adjustment of timeliness and deep learning-based answer generation mechanism. Especially in complex financial data, timeliness and context processing capabilities directly affect the accuracy of prediction and decision-making. By introducing the time decay factor , industry impact factor , and external event adjustment factors ,The solution of the present invention not only ensures the real time of data ,but can also adapt to complex market environment and industry changes.

[0087] In addition, through the combination of deep learning generative models and self-attention mechanisms, the system can make full use of time series data and contextual information to automatically infer and generate high-precision financial decision answers. The final decision results can not only accurately reflect the current financial situation, but also provide strong support for future strategic decisions. Compared with traditional financial analysis based on historical data, this solution has strong real-time and adaptability, can dynamically respond to the ever-changing economic and industry environment, and provide decision makers with more scientific and timely financial decision support.

[0088] S5. Collect user feedback and preliminary intelligent decision-making results based on preliminary intelligent decision-making results, design a user feedback mechanism based on user feedback, maximize the impact of user feedback on the financial decision-making system, and enable it to dynamically adapt to long-term changes in user preferences, while adaptively adjusting the model of the problem text modeling, and generate the final decision result after completing user feedback and adaptive adjustment.

[0089] Specifically, in step S4: dynamic adjustment of timeliness and answer generation, the system has generated a decision result based on dynamic adjustment of timeliness. , including prediction values, risk assessments, and intelligent suggestions for future strategies. These results serve as input data for subsequent user feedback and intelligent adjustments. At the same time, the present invention also receives user feedback data , which reflects the user's acceptance of the results and whether there is a deviation in demand.

[0090] The input data includes: : Preliminary financial decision results generated through dynamic adjustment and answer generation based on timeliness.

[0091] : User feedback data at the current moment, indicating the user’s feedback on the accuracy of the prediction, the acceptance of the recommendation, etc.

[0092] : Historical feedback data helps the model adjust according to the user’s long-term behavior patterns.

[0093] The output of step S4 is closely related to the input of this step. It is based on the dynamic adjustment results of the previous stage. It directly affects the accuracy of subsequent decision-making adjustments. The previous step has already included consideration of real-time market changes. With the addition of user feedback, the system will further optimize decisions based on this information.

[0094] Furthermore, when designing the user feedback mechanism, the present invention mainly considers how to maximize the impact of user feedback on the financial decision-making system and enable it to dynamically adapt to long-term changes in user preferences. The weight of user feedback and the feedback correction factor play a crucial role in the entire optimization process.

[0095] User feedback weight :The present invention assigns different weights to each feedback according to the accuracy, historical behavior and authority of the user feedback This means that user feedback is not treated equally, but is weighted based on the stability and expertise of their long-term behavior.

[0096] Feedback correction factor :Combining user feedback with the system's initial prediction, the adjusted decision value is adjusted by the feedback correction factor The generation of feedback correction factors combines the user's immediate feedback and long-term preference model, allowing the system to dynamically adjust its response to user needs.

[0097] Long-term preference learning :The system also provides historical feedback data Used to train long-term preference learning models , thereby capturing the user's decision-making habits and tendencies over a longer period of time. The update formula is as follows: ; in, The financial decision results are adjusted based on user feedback. Initial financial decisions are generated after dynamic adjustments for timeliness. is the feedback weight, which is used to measure the impact of user feedback on the adjustment results. It will gradually increase as user feedback increases and accuracy improves. Provide instant feedback data for current users. It is a user's long-term preference model learned based on historical feedback, which is used to guide long-term decision adjustments. is historical feedback information, which is used to train the user's long-term preference model. Through this designed formula, the present invention ensures that the system can dynamically adjust the prediction results. , the decision output is modified according to the user's immediate feedback and long-term preference data, so that the model's prediction is closer to actual needs.

[0098] Furthermore, under the guidance of user feedback, the system will make adaptive adjustments to the original model. User feedback is not limited to adjusting the current results, but also promotes model learning and parameter optimization. By combining real-time feedback and historical data, the present invention adopts an adaptive optimization mechanism to ensure that the decision-making ability of the model continues to improve over time.

[0099] Adaptive adjustment factor : This factor reflects the strength of adjusting the model weights based on user feedback. In the early stage, The value of is small. With the accumulation of feedback and the gradual progress of model optimization, The value of will gradually increase, thus driving the system to adapt to new data and feedback more quickly.

[0100] Fine-tuning weight updates : The weight of the model are parameters optimized by back-propagation, these weights are adjusted by user feedback and adaptive factors With the help of , it is gradually optimized to adapt to the latest financial data and market changes.

[0101] The fine-tuning formula is as follows: ; in, For the current moment The weights of the model. is the adaptive adjustment factor, which is used to control the adjustment range. As user feedback increases, The value increases gradually. is the loss function with respect to weights The gradient of is used to optimize the model. is the adjusted decision result, The core of this step is to continuously optimize the internal parameters of the model through user feedback, so that the model can provide more accurate predictions and personalized decision support in the long run.

[0102] Furthermore, after completing user feedback and adaptive optimization, the final decision result is generated will be output. These decisions not only include real-time adjusted financial forecasts, but also incorporate user preference adjustments. Importantly, the new decision results will become the starting point for the next decision-making process and enter the next round of feedback loop.

[0103] Furthermore, based on the new decision results, users provide feedback again, which forms a continuous cycle. Each feedback will affect the adjustment and optimization of the system, making future decisions more accurate.

[0104] Furthermore, based on the user's long-term preferences , the system will provide more personalized suggestions, such as customized financial management plans or investment advice, thereby enhancing the user's decision-making experience.

[0105] In this cycle, user feedback The alternating effect with model optimization will continue to improve the accuracy and practicality of financial decisions.

[0106] Furthermore, by designing an adaptive optimization mechanism, the present invention combines the user's immediate feedback with long-term preferences to achieve dynamic adjustment of the system. In particular, by using the feedback correction factor and adaptive adjustment factor , the present invention can effectively adjust the prediction accuracy of the model and achieve long-term performance optimization. In addition, the present invention designs long-term preference learning , so that the system not only responds to current feedback, but also learns the user's long-term decision-making patterns, thereby providing customized decisions that are more in line with actual needs.

[0107] The core innovation of this step is the introduction of a complex user feedback mechanism, which combines instant feedback and historical data to achieve efficient and accurate dynamic adjustments through a feedback-driven adaptive optimization algorithm. This solution enables the financial decision-making system to not only adapt to market changes in real time, but also provide users with personalized and sustainable decision support through long-term learning.

[0108] Another embodiment of the present invention provides a financial intelligent question-answering system based on artificial intelligence, the system comprising: The financial problem analysis module is used to collect financial problems, model the problem text and analyze the keywords in the problem, and analyze the implicit financial goals in the problem according to the keywords based on the preset knowledge graph and rule-based reasoning method to obtain a structured financial problem representation; wherein the structured financial problem representation includes financial goals, report types, time intervals and keywords; The cross-platform data retrieval module is used to select cross-platform data sources related to the problem for retrieval query based on the structured financial problem representation using a unified data access layer. During the retrieval query process, noise is injected into the data through a differential privacy mechanism to prevent leakage of sensitive information. Finally, the financial data returned by multiple platforms is aggregated through an aggregation operation. The contextual reasoning integration module is used to build a contextual reasoning model based on cross-platform data and contextual information, supplement the missing and interpretable nature of aggregated financial data through contextual information, enhance the aggregated financial data, conduct multi-level analysis based on multi-platform data sources, and detect abnormal financial data. When abnormal financial data appears, an adaptive anomaly detection algorithm is used to correct the abnormal financial data; The module for dynamic adjustment of timeliness and generation of answers is used to design a dynamic adjustment mechanism for timeliness to dynamically adjust the financial data generated by the contextual reasoning integration module in order to ensure that the analysis results can reflect the latest economic, industry and company conditions when the current financial market situation changes. The module introduces a deep learning-based answer generation mechanism based on the financial data after dynamic adjustment of timeliness to generate intelligent decision results for financial problems, and inputs the preliminary intelligent decision results into the user feedback and answer optimization module; the preliminary intelligent decision results are output in a structured JSON format; The user feedback and answer optimization module is used to collect user feedback and preliminary intelligent decision-making results based on the preliminary intelligent decision-making results, design a user feedback mechanism based on user feedback, maximize the impact of user feedback on the financial decision-making system, and enable it to dynamically adapt to long-term changes in user preferences. At the same time, the model for modeling the question text is adaptively adjusted, and the final decision result is generated after completing user feedback and adaptive adjustment.

[0109] To summarize, the present invention provides a financial question-and-answer system with high privacy protection and high intelligence by combining differential privacy protection, distributed computing, hierarchical neural networks and timeliness adjustment mechanism. It can effectively solve the problems of privacy leakage, insufficient context understanding and limited data processing capabilities in the prior art, and provides a new solution for large-scale, cross-organizational financial data analysis and intelligent question-and-answer.

[0110] These are just some preferred embodiments of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of implementing the above embodiments and making equivalent changes according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A financial intelligent question-answering method based on artificial intelligence, characterized in that: The method comprises the following steps: S1. Collect financial questions to model the question text and analyze the keywords in the questions, and analyze the implicit financial goals in the questions according to the keywords based on the preset knowledge graph and rule-based reasoning method to obtain a structured financial question representation; S2. Based on the structured financial problem representation, a unified data access layer is used to select cross-platform data sources related to the problem for retrieval query. During the retrieval query process, noise is injected into the data through the differential privacy mechanism to prevent the leakage of sensitive information. Finally, the financial data returned by multiple platforms are aggregated through the aggregation operation. S3, build a contextual reasoning model based on cross-platform data and contextual information, supplement the missing and interpretable nature of the aggregated financial data through contextual information, enhance the aggregated financial data, and integrate multi-platform data sources to perform multi-level analysis and abnormal financial data detection. When abnormal financial data appears, use an adaptive anomaly detection algorithm to correct the abnormal financial data and input the final financial data into S4; S4. When the current financial market situation changes, a dynamic adjustment mechanism is designed to dynamically adjust the timeliness of the financial data generated by S3. According to the financial data after dynamic adjustment, an answer generation mechanism based on deep learning is introduced to generate intelligent decision-making results for financial problems, and the intelligent decision-making results are input into S5. S5. Collect user feedback and preliminary intelligent decision results based on preliminary intelligent decision results, design a user feedback mechanism based on user feedback, maximize the impact of user feedback on the financial decision-making system, and enable it to dynamically adapt to long-term changes in user preferences, while adaptively adjusting the model of the problem text modeling, and generate the final decision result after completing user feedback and adaptive adjustment.

2. The financial intelligent question-answering method based on artificial intelligence according to claim 1, characterized in that: The structured financial problem representation includes financial goals, report types, time periods and keywords; In S1, financial questions are collected to model the question text and analyze the keywords in the questions. The Transformer model fine-tuned based on the financial field is used to perform semantic modeling on the question text. The Transformer model fine-tuned based on the financial field uses the BERT model fine-tuned with financial field data to encode the question text. The BERT model can convert the vocabulary of each question into a semantic vector, thereby capturing the deep meaning of the question; The keyword analysis is as follows: First, we design the importance score of each word to analyze the output of the BERT model and obtain keywords; Then, these keywords and their contexts are weighted averaged based on the weight of context information to obtain semantic representation; The preset knowledge graph and rule-based reasoning method include: The knowledge graph includes the relationship between various financial indicators and financial goals; The rule is to infer possible financial goals based on the keywords in the question.

3. The financial intelligent question-answering method based on artificial intelligence according to claim 1, characterized in that: Said S1 also includes setting a rule-based time recognition and standardization module to extract time information from the question and use regular expressions and simple rule sets to extract and convert it into a standard time interval.

4. The financial intelligent question-answering method based on artificial intelligence according to claim 1, characterized in that: The differential privacy mechanism is used to inject noise into the data to prevent leakage of sensitive information, and finally the financial data returned by multiple platforms are aggregated through an aggregation operation; wherein the privacy budget of the differential privacy mechanism controls the noise intensity of the noise injection to ensure a balance between data privacy and data quality; Wherein, the differential privacy mechanism further includes: The privacy budget is dynamically adjusted according to the privacy budget threshold and the current privacy budget; Said S2 further comprises: Aggregate financial data is encrypted using the AES algorithm to generate encrypted financial data to ensure data security during storage and transmission.

5. The financial intelligent question-answering method based on artificial intelligence according to claim 1, characterized in that: The contextual reasoning model uses a dual attention mechanism to enhance financial data based on the time dimension and context information, including: First, the contextual reasoning model assigns a temporal attention weight to each data point, and the temporal attention weight is dynamically adjusted according to the timeliness of the data; Secondly, the contextual reasoning model gives attention weights to contextual information according to the industry background or macroeconomic environment of the data; Finally, based on the time attention weight and the attention weight of the context information, the financial data is enhanced by weighted sum calculation to obtain enhanced financial data; The comprehensive multi-platform data source integrates financial data by adopting a hierarchical weighted fusion mechanism, which is used to perform hierarchical processing on the financial data and gradually integrate the data characteristics of each level.

6. The financial intelligent question-answering method based on artificial intelligence according to claim 5, characterized in that: The adaptive anomaly detection algorithm dynamically adjusts the anomaly detection threshold according to the normalization of data distribution, including: During the detection process, the system calculates the degree of deviation of each data point from the mean and standard deviation of the current data set to determine whether it is an outlier: If the deviation exceeds the preset threshold, the current data point is determined to be an outlier; For abnormal data, the system repairs it based on historical financial data and contextual information, and adopts a repair method based on multi-task learning, which can select the most appropriate repair strategy according to different types of abnormal data.

7. The financial intelligent question-answering method based on artificial intelligence according to claim 1, characterized in that: The dynamic adjustment mechanism of timeliness is constructed as follows: The time decay factor is designed to dynamically adjust the importance of the data in the model according to its historical timestamp; the industry impact factor is designed to adjust the impact of the current industry economic environment on the financial data; the external event adjustment factor is designed to introduce the real-time impact of external events on financial data. If a specific event affects the company's performance, the relevant financial data will be adjusted; The dynamic adjustment mechanism for timeliness is based on the original financial data, the financial data adjusted by the industry impact factor, and the external event adjustment factor, and performs a dynamic adjustment for timeliness on the financial data by using an exponentially weighted moving average based on a time decay factor to obtain the financial data that has been dynamically adjusted for timeliness; The answer generation mechanism is generated by combining a long short-term memory network with a self-attention mechanism.

8. The financial intelligent question-answering method based on artificial intelligence according to claim 7, characterized in that: The intelligent decision-making results include predicted net profits and industry risk assessments, providing decision makers with optimization suggestions and displaying data adjusted for timeliness; the intelligent decision-making results are output in a structured JSON format.

9. The financial intelligent question-answering method based on artificial intelligence according to claim 1, characterized in that: The adjustments based on user feedback are as follows: The preliminary financial decision generated after the dynamic adjustment of the collection timeliness is revised based on the current user's immediate feedback data and the user's long-term preferences to obtain the financial decision result adjusted by the user feedback, which is used to maximize the impact of the user's feedback on the financial decision system and enable it to dynamically adapt to the long-term changes in user preferences; The adaptive adjustment is based on an adaptive optimization mechanism, and the adjustment based on user feedback is optimized by gradient optimization, wherein the gradient of the gradient optimization is used to optimize the adjustment based on user feedback.

10. A financial intelligent question-answering system based on artificial intelligence, characterized in that: The system comprises: The financial problem analysis module is used to model the problem text of financial problems and analyze the keywords in the problem. According to the preset knowledge graph and rule-based reasoning method, the implicit financial goals in the problem are analyzed according to the keywords to obtain a structured financial problem representation; The cross-platform data retrieval module is used to select cross-platform data sources related to the problem for retrieval query based on the structured financial problem representation using a unified data access layer. During the retrieval query process, noise is injected into the data through a differential privacy mechanism to prevent leakage of sensitive information. Finally, the financial data returned by multiple platforms is aggregated through an aggregation operation. The contextual reasoning integration module is used to build a contextual reasoning model based on cross-platform data and contextual information, supplement the missing and interpretable nature of aggregated financial data through contextual information, enhance the aggregated financial data, and integrate multi-platform data sources to perform multi-level analysis and abnormal financial data detection. When abnormal financial data appears, an adaptive anomaly detection algorithm is used to correct the abnormal financial data, and the final financial data is input into the timeliness dynamic adjustment and answer generation module; The module for dynamic adjustment of timeliness and answer generation is used to design a dynamic adjustment mechanism for timeliness to dynamically adjust the financial data generated by the contextual reasoning integration module when the current financial market situation changes. The module introduces a deep learning-based answer generation mechanism based on the financial data after dynamic adjustment of timeliness to generate intelligent decision-making results for financial problems, and inputs the intelligent decision-making results into the user feedback and answer optimization module. The user feedback and answer optimization module is used to collect user feedback and preliminary intelligent decision-making results based on the preliminary intelligent decision-making results, design a user feedback mechanism based on user feedback, maximize the impact of user feedback on the financial decision-making system, and enable it to dynamically adapt to long-term changes in user preferences. At the same time, the model for modeling the question text is adaptively adjusted, and the final decision result is generated after completing user feedback and adaptive adjustment.

Images