Intelligent financial question-answering system realized based on large language model

Abstract

The invention discloses an intelligent financial question and answer system based on a large language model, and relates to the technical field of financial question and answer, the system solves the problems of financial term ambiguity and context ambiguity through cooperative work of a plurality of modules, and the system comprises the following modules: a context recognition module, a query module and a query module, a natural language processing algorithm is combined with a context analysis technology in the financial field to automatically extract context clues in user questions; and a financial term knowledge base module. According to the method, context recognition and a financial term knowledge base are combined, and the accuracy of understanding the term ambiguity and the complex context is improved through the graph neural network. And the intelligent recommendation module dynamically provides personalized financial suggestions by using collaborative filtering and content recommendation algorithms in combination with user history and market data, and performs real-time adjustment. The risk assessment and prompt module assesses risks through a Bayesian network and a logistic regression model, and the adaptive learning module continuously optimizes risk prediction and helps users to avoid financial risks.

Description

Technical Field

[0001] The present invention relates to the technical field of financial question answering, and in particular to an intelligent financial question answering system based on a large language model. Background Art

[0002] The intelligent financial question-and-answer system based on a large language model refers to a financial question-and-answer platform built using advanced natural language processing technology (such as large language models such as GPT). The system can understand complex financial issues and provide accurate and fast answers based on existing financial knowledge and market information by intelligently analyzing large amounts of financial data, knowledge bases, and user questions. This system has self-learning capabilities and can continuously optimize the accuracy of answers, helping users solve practical problems in financial investment, financial management, risk assessment, etc., and improving users' financial decision-making capabilities and experience.

[0003] The prior art has the following deficiencies:

[0004] In intelligent financial question-answering systems based on large language models, the lack of polysemy and ambiguity in financial terms is a key issue. Although large language models have the ability to understand and generate natural language, when faced with some terms or phrases with multiple meanings in the financial field, especially when the context information is unclear, the model may find it difficult to accurately judge the specific context. This misjudgment may lead to the generation of wrong or misleading answers, which in turn has an adverse impact on users' investment decisions and financial operations, and even cause significant financial losses or legal risks. Therefore, solving this problem is crucial to the accuracy and security of the system.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0006] The purpose of the present invention is to provide an intelligent financial question-answering system based on a large language model. By combining a context recognition module with a financial terminology knowledge base module, a financial concept map is constructed using a graph neural network, thereby improving the system's understanding of the ambiguity and complex context of financial terms. The intelligent recommendation module combines user historical operations, market data and financial events, and uses collaborative filtering and content recommendation algorithms to dynamically recommend personalized and high-precision financial products and market analysis, and make real-time adjustments based on market changes. The risk assessment and prompt module uses a Bayesian network and a logistic regression model to assess the potential risks of financial recommendations, and combines historical data with an adaptive learning module to continuously optimize risk predictions, helping users avoid financial risks, so as to solve the problems in the above-mentioned background technology.

[0007] In order to achieve the above-mentioned purpose, the present invention provides the following technical solution: an intelligent financial question-answering system based on a large language model, wherein the system works together through multiple modules to solve the problems of polysemy and contextual ambiguity in financial terms, and the system includes the following modules:

[0008] The context recognition module uses natural language processing algorithms combined with context analysis technology in the financial field to automatically extract context clues from user questions;

[0009] The financial terminology knowledge base module combines a pre-trained large-scale financial knowledge base with self-expansion capabilities to store and update the multiple meanings of financial terms and their usage in different markets and countries. It automatically learns and infers the associations between different terms by introducing a financial concept graph based on a graph neural network.

[0010] The semantic disambiguation module uses advanced clustering algorithms combined with semantic embedding of large-scale language models to resolve the ambiguity of financial terms input by users;

[0011] The intelligent recommendation module uses collaborative filtering and content recommendation algorithms, combined with the user's historical operations, market data and financial events, to provide the most relevant financial advice. By combining deep learning networks with graph networks, it dynamically recommends relevant market analysis, financial products or operation strategies during the financial question-answering process, ensuring that the model gives advice after identifying the context.

[0012] Adaptive learning module, which uses online learning algorithms to continuously update and adjust the large language model during system use to adapt to new financial terms, market changes and user behaviors;

[0013] The risk assessment and prompt module conducts risk assessment on the financial recommendations generated by the system by introducing Bayesian networks and logistic regression models.

[0014] Preferably, natural language processing algorithms are combined with context analysis technology in the financial field to automatically extract context clues in user questions, which is achieved by the following steps:

[0015] Perform preliminary text processing on financial questions input by users;

[0016] After completing the text preprocessing, semantic analysis technology is used to identify entities and terms in the financial field. Pre-trained financial word embedding models and named entity recognition technology are used to identify keywords and terms in the financial field from the input;

[0017] After identifying key entities, the context of the term is further determined through context association and dependency analysis. Dependency analysis helps the system understand the semantic structure of the input question by determining the sentence structure and identifying the grammatical relationship between words.

[0018] After completing the context analysis, the context clue generation module dynamically adjusts the semantic interpretation and extracts the final context clues based on the knowledge base in the financial field.

[0019] Preferably, a large pre-trained financial knowledge base and self-expansion capabilities are combined to build a financial concept graph through a graph neural network to store and reason about the multiple meanings and market differences of financial terms, which is achieved through the following four steps:

[0020] Build a comprehensive and large-scale financial knowledge base from multi-source data;

[0021] After the knowledge base is built, a financial concept graph is created using a graph neural network. The nodes in the graph represent financial terms, markets, or related financial events, and the edges represent the relationships between the terms. The complex relationships between different financial terms are automatically captured through the structured modeling of the graph neural network.

[0022] After the financial concept graph is constructed, term association learning and reasoning are performed through node feature propagation of the graph neural network. The graph neural network model aggregates the information of neighboring nodes layer by layer, so that each term node can learn contextual features from its associated terms.

[0023] A real-time update and self-expansion mechanism is designed to automatically detect new financial terms or changes in the meaning of existing terms through real-time synchronization with external data sources, and dynamically update them to the financial knowledge base and concept map. Through the self-learning characteristics of graph neural networks, the term association network can be automatically adjusted to keep it consistent with the latest market dynamics, ensuring continued accuracy in the understanding and reasoning of the terms.

[0024] Preferably, the ambiguity resolution of financial terms is handled by using advanced clustering algorithms combined with semantic embedding of large-scale language models, which is divided into the following three steps:

[0025] Use a pre-trained large-scale language model to perform semantic embedding on the financial terms input by users, convert the input text into a high-dimensional vector representation, and capture the contextual information and semantic features of the financial terms;

[0026] The generated semantic embedding is input into an advanced clustering algorithm, which clusters vectors with similar semantics by analyzing the distribution of semantic vectors.

[0027] After clustering is completed, the most appropriate semantic set is selected for resolution based on the specific context of the user's input. By analyzing the overall context of the user's question, the financial terminology that best fits the current context is selected from the clustering results.

[0028] Preferably, collaborative filtering and content recommendation algorithms are used to provide relevant financial advice in combination with the user's historical operations, market data and financial events, and dynamic recommendations are made by combining deep learning networks and graph networks. This is achieved through the following three steps:

[0029] Extract relevant features from users’ historical operation records, current market data, and financial events;

[0030] After feature extraction, user needs are analyzed through collaborative filtering algorithms and content recommendation algorithms;

[0031] In the recommendation process, dynamic recommendation optimization is achieved through the combination of deep learning network and graph neural network.

[0032] Preferably, an online learning algorithm is used to continuously update and adjust the large language model to adapt to new financial terms, market changes and user behaviors, which is specifically achieved through the following three steps:

[0033] First, real-time financial data and user interaction data are continuously collected through data pipelines;

[0034] After data preprocessing is completed, the large language model is dynamically updated through online learning algorithms without having to retrain the entire model. New financial terms, market changes, and user behavior patterns are used as incremental data to gradually adjust the model's parameter weights to adapt it to the latest market dynamics.

[0035] The performance of the model is evaluated in real time through user feedback and market reaction, and the user's interactive behavior is compared and analyzed with the model's recommendation results. If the suggestions generated by the system do not meet user needs, or there is a deviation between market data and predictions, the model parameters will be further tuned through the feedback mechanism.

[0036] Preferably, the risk assessment of the financial advice generated by the system is performed by introducing a Bayesian network and a logistic regression model, which is achieved through the following four steps:

[0037] Collect various data related to risk assessment, select appropriate features for Bayesian network node settings and logistic regression model inputs, and ensure that these input variables can reflect market changes and capture users' investment habits through data cleaning and feature engineering, laying a data foundation for subsequent risk assessment;

[0038] After the feature data is prepared, a risk assessment model is constructed through the Bayesian network. The Bayesian network derives potential risks by modeling key variables and user behaviors in the financial market as nodes and defining the probabilistic relationship between nodes.

[0039] The logistic regression model is used to quantitatively assess the risk of users' investment decisions. The logistic regression model uses the user's historical operation data, the characteristics of financial products and market changes as input variables, and generates a risk probability value by calculating the relationship between these characteristics and high-risk behaviors;

[0040] The results of Bayesian networks and logistic regression are combined to provide users with dynamic adjustments and risk warnings through the risk assessment module.

[0041] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0042] The present invention combines a context recognition module with a financial terminology knowledge base module, so that the system can more accurately extract and analyze the context information in the user input, and dynamically infer the ambiguity and meaning of financial terms in different markets and countries by combining the terminology base in the financial field. By introducing a graph neural network to construct a financial concept map, the system can not only obtain the definition of the current term from the knowledge base, but also understand the correlation between terms, thereby greatly improving the accuracy of context understanding when faced with complex or ambiguous contexts and avoiding misinterpretation of financial terms.

[0043] The present invention combines the user's historical operations, market data and financial events together through an intelligent recommendation module, and dynamically recommends the most relevant financial products and market analysis through collaborative filtering and content recommendation algorithms. The combination of deep learning networks and graph networks makes the recommendations not only relevant to user preferences, but also adjustable according to real-time changes in the market, ensuring that the recommendation results are not only personalized, but also provide operational strategies with high reference value based on the user's actual needs, thereby improving the accuracy and practicality of the recommendations.

[0044] The present invention integrates the risk assessment and prompt module with Bayesian network and logistic regression model, and conducts risk assessment on each recommendation, so that the system can assess the potential market fluctuations, policy changes and risks brought by user operation behaviors while generating financial suggestions, and provide risk prompts in time. By combining historical data and real-time feedback, the adaptive learning module ensures that the system can continuously adjust the risk prediction model, so that users can fully understand the potential risks before making decisions, thereby effectively avoiding investment mistakes or financial losses. This not only improves the security of the system, but also enhances the user's sense of trust. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0046] Figure 1 This is a module diagram of an intelligent financial question-answering system based on a large language model according to the present invention. DETAILED DESCRIPTION

[0047] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0048] The present invention provides Figure 1 The intelligent financial question-answering system based on a large language model is shown in FIG. The system works together through multiple modules to solve the problems of polysemy and contextual ambiguity in financial terms. The system includes the following modules:

[0049] The context recognition module uses natural language processing algorithms combined with context analysis technology in the financial field to automatically extract context clues from user questions;

[0050] Through advanced context-dependent models, such as the Transformer network based on the attention mechanism, historical conversations, market data, and user input are analyzed to ensure that the system can accurately identify the meaning of financial terms in context and reduce ambiguity. The context recognition module uses a semantic hierarchical structure and dynamic word vectors to distinguish the meaning of words in different contexts, thereby providing accurate initial input for the model.

[0051] The financial terminology knowledge base module combines a pre-trained large-scale financial knowledge base with self-expansion capabilities to store and update the multiple meanings of financial terms and their usage in different markets and countries. It automatically learns and infers the associations between different terms by introducing a financial concept graph based on a graph neural network (GNN);

[0052] Especially in the face of the rapidly changing financial market, we need to ensure the timeliness and breadth of the knowledge base and improve the model's ability to handle polysemous words;

[0053] The semantic disambiguation module uses advanced clustering algorithms (such as K-Means, DBSCAN, etc.) combined with the semantic embedding of large-scale language models (such as BERT, GPT) to resolve the ambiguity of financial terms input by users;

[0054] The semantic disambiguation module selects the most relevant meaning and excludes inapplicable interpretations by analyzing the context and related information in the knowledge base. By using a reinforcement learning algorithm, the semantic disambiguation module can also self-optimize and gradually improve the accuracy of disambiguation when processing complex contexts.

[0055] The intelligent recommendation module uses collaborative filtering and content recommendation algorithms, combines the user's historical operations, market data, and financial events, and provides the most relevant financial advice. Through the combination of a deep learning network and a graph network, it dynamically recommends relevant market analysis, financial products, or operation strategies during the financial Q&A process, ensuring that the advice given by the model is not only accurate but also practically instructive after identifying the context;

[0056] The adaptive learning module adopts an online learning algorithm to continuously update and adjust the large language model during the system's use to adapt to new financial terms, market changes, and user behaviors;

[0057] Through deep reinforcement learning and transfer learning, the model parameters can be adjusted according to user feedback to optimize the semantic understanding and disambiguation capabilities in real time, ensuring that the system can still maintain high accuracy when facing different users or emerging financial concepts.

[0058] The risk assessment and prompt module conducts risk assessment on the financial advice generated by the system by introducing Bayesian Networks and logistic regression models;

[0059] The risk assessment and prompt module combines the real-time data of the financial market and the user's investment preferences to provide risk prompts related to the generated advice, ensuring that users can fully understand potential financial risks during the decision-making process, thereby effectively avoiding legal risks and financial losses caused by misunderstandings.

[0060] Using natural language processing algorithms combined with context analysis techniques in the financial field to automatically extract context clues from the user's questions, which can usually be achieved through the following four steps:

[0061] 1. Preliminary text preprocessing

[0062] In this stage, the system first performs preliminary text processing on the financial questions input by the user, including operations such as word segmentation, stop word removal, and part-of-speech tagging. Through word segmentation, the system can break down the input financial questions into basic language units for subsequent analysis. Removing stop words (such as meaningless words like "of" and "is") can reduce interference and ensure that the system only processes key semantic information. Part-of-speech tagging helps the system distinguish between lexical types such as nouns and verbs and distinguish the functions of financial terms during further analysis. The preliminary text preprocessing provides a necessary basis for subsequent semantic understanding and context extraction.

[0063] 2. Semantic analysis and entity recognition

[0064] After completing the text preprocessing, the system uses semantic analysis technology to identify entities and terms in the financial field. Through the pre-trained financial word embedding model and named entity recognition (NER) technology, the system can identify financial keywords and terms such as "stocks", "options", and "interest rates" from the input. The key to this step is to judge the role and meaning of the terms through the contextual semantics of the vocabulary, especially when there are polysemous words. Combined with contextual clues, the system can make preliminary judgments on these terms. For example, the "crash" in "stock market crash" should be related to the market decline, not other meanings.

[0065] 3. Contextual association and dependency analysis

[0066] After identifying key entities, the system further determines the context of the terms through contextual association and dependency analysis. Dependency analysis helps the system understand the semantic structure of the input question by judging the sentence structure and identifying the grammatical relationship between words (such as subject, object, modifiers, etc.). For example, in the sentence "What impact will today's stock market crash have on my investment portfolio?", the system can identify through dependency relationships that "stock market crash" is a prerequisite for "investment portfolio impact". This step combines context and grammatical structure to further refine the semantic clues of the user's question.

[0067] 4. Contextual clue generation and dynamic adjustment

[0068] After completing the context analysis, the system dynamically adjusts the semantic interpretation and extracts the final context clues based on the knowledge base in the financial field through the context clue generation module. In this step, the system dynamically generates relevant financial scenarios, such as market fluctuations, economic policy changes, etc., based on the current input financial context. By leveraging the background knowledge in the financial field, the system can determine the specific meaning of the term in the current environment and generate context clues to guide the subsequent question-and-answer process. For example, when a user asks "What impact will the Fed's interest rate hike have on my mortgage?", the system can generate context clues as scenarios related to "rising interest rates" and "increased loan costs" based on the current financial background, ensuring that the question-and-answer system provides accurate answers.

[0069] Combining a large pre-trained financial knowledge base and self-expansion capabilities, a financial concept graph is constructed through a graph neural network (GNN) to store and reason about the multiple meanings and market differences of financial terms. This can be achieved mainly through the following four steps:

[0070] 1. Construction and preprocessing of financial knowledge base

[0071] First, the system needs to build a comprehensive and large-scale financial knowledge base from multi-source data. This knowledge base includes term definitions, historical data, legal terms, policy changes and other information from global financial markets. During the construction process, the system ensures the accuracy and consistency of the knowledge base content through steps such as data cleaning, normalization, and deduplication. Each financial term must not only store the basic meaning, but also record its usage in different markets and countries (such as the differences in tax terms in the United States and Europe). This process establishes the basic framework for the system to understand financial terms and lays the data foundation for subsequent associative reasoning.

[0072] 2. Graph Neural Network (GNN) Modeling Financial Concepts

[0073] After the knowledge base is built, the system uses graph neural networks (GNNs) to create a financial concept graph. The nodes in the graph represent financial terms, markets, or related financial events, and also represent the relationships between terms. Through the structured modeling of GNNs, the system can automatically capture the complex associations between different financial terms. For example, there is a market correlation between "interest rates" and "inflation", and there is a functional association between "stocks" and "options" in financial instruments. By propagating information layer by layer, GNN enables the system to infer the impact of terms in different markets and even predict future usage trends of terms.

[0074] 3. Financial terminology association learning and polysemy reasoning

[0075] After the financial concept graph is constructed, the system performs term association learning and reasoning through node feature propagation of the graph neural network. The GNN model aggregates the information of neighboring nodes layer by layer, allowing each term node to learn contextual features from its associated terms. For example, the meaning of the word "asset" in different markets can be accurately inferred through other terms associated with it (such as "stock assets" and "fixed assets"). In this way, the system can automatically identify the multiple meanings of terms in different markets and countries, and combine historical data to infer the potential meaning of terms in new contexts and solve the problem of term polysemy.

[0076] 4. Real-time update and self-expansion mechanism

[0077] In order to ensure that the system can keep up with the dynamic changes in the financial market, the system is designed with real-time update and self-expansion mechanisms. Through real-time synchronization with external data sources (such as market news, government policy updates, financial reports, etc.), the system can automatically detect new financial terms or changes in the meaning of existing terms, and dynamically update them to the financial knowledge base and concept map. Through the self-learning characteristics of GNN, the system can automatically adjust the term association network to keep it consistent with the latest market dynamics, thereby ensuring continued accuracy in the understanding and reasoning of terms. This self-expansion capability ensures that the system can still provide effective analysis and explanation when faced with emerging markets or sudden financial changes.

[0078] Using advanced clustering algorithms (such as K-Means, DBSCAN) combined with semantic embedding of large-scale language models (such as BERT, GPT) to deal with the polysemy resolution of financial terms can be divided into the following three main steps:

[0079] 1. Semantic Embedding Generation

[0080] First, the system uses pre-trained large-scale language models (such as BERT and GPT) to perform semantic embedding processing on the financial terms input by the user. Models such as BERT can convert the input text into high-dimensional vector representations that capture the contextual information and semantic features of financial terms. For example, the input term "option" will generate different semantic embeddings through the model combined with the context to reflect the specific meanings in different contexts such as "stock options" or "futures options". In this way, the system can convert the multiple meanings of each term into semantic vectors, providing basic data for subsequent clustering processing.

[0081] 2. Application of clustering algorithm

[0082] Next, the system inputs the generated semantic embeddings into advanced clustering algorithms, such as K-Means or DBSCAN. Clustering algorithms cluster vectors with similar semantics by analyzing the distribution of these semantic vectors. K-Means divides semantic vectors into preset classes, while DBSCAN is suitable for handling semantic distributions of different densities and can automatically detect abnormal meanings with sparse semantics. For example, the term "bond" will be clustered in the two common classes of "corporate bonds" and "government bonds", while the rarer "convertible bonds" may be identified separately by DBSCAN. Through clustering, the system is able to classify different financial terms with similar semantics together and effectively distinguish the ambiguity of terms.

[0083] 3. Semantic resolution and final selection

[0084] After clustering is completed, the system selects the most appropriate semantic set for resolution based on the specific context of the user's input. By analyzing the overall context of the user's question, such as combining grammatical structure, historical conversations, or other financial terms, the system selects the financial terms that best fit the current context from the clustering results. For example, when a user asks "how to trade options", the system recognizes through context that the user is discussing "stock options" rather than other meanings. At this stage, the system feeds back the specific meaning after semantic resolution to the user to ensure the accuracy and relevance of the answer. This process significantly improves the accuracy of the financial question-answering system in handling polysemous words in complex contexts.

[0085] Using collaborative filtering and content recommendation algorithms, combining users’ historical operations, market data, and financial events to provide relevant financial advice, and combining deep learning networks with graph networks for dynamic recommendations, can be achieved through the following three main steps:

[0086] 1. Feature extraction of user data and market information

[0087] First, the system extracts relevant features from the user's historical operation records, current market data, and financial events. The user's historical operation data includes their investment behavior, browsing history, preferences, etc., market data covers dynamic changes in stocks, funds, interest rates, etc., and financial events involve major policy changes or economic fluctuations. Through the feature extraction technology in deep learning, the system can convert this multidimensional information into structured data, laying the foundation for subsequent personalized recommendations. For example, by analyzing the user's past investment preferences, the system can extract their risk tolerance in the market, thereby screening out suitable financial products or strategies.

[0088] 2. Combined application of collaborative filtering and content recommendation algorithm

[0089] After feature extraction, the system analyzes the user's needs through collaborative filtering algorithms and content recommendation algorithms. The collaborative filtering algorithm infers the financial products or operation strategies that the current user may be interested in based on the historical behavior of similar users. For example, if a group of users chooses a stock investment strategy, the system will give priority to recommending it to users with similar behaviors. At the same time, the content recommendation algorithm recommends financial products or market analysis related to the user's investment portfolio or preferences by analyzing the latest changes in market data and financial events. Through the combination of these two algorithms, the system can provide users with multi-angle, personalized and current market financial advice.

[0090] 3. Dynamic recommendation optimization of deep learning map network

[0091] In the recommendation process, the system optimizes dynamic recommendations by combining deep learning networks with graph neural networks (GNNs). Graph networks model users, financial products, market events, and other related information as a network of nodes and their surroundings. Through the propagation and learning of relationships between nodes, the system can capture potential recommendation relationships on a larger scale. For example, GNNs can analyze the correlation between different stocks and funds in the market, and recommend relevant market analysis or emerging investment opportunities in combination with the user's portfolio. At the same time, the deep learning model optimizes the recommendation results through real-time feedback, ensuring that the system can continuously adjust the recommendation strategy based on the user's real-time needs, market dynamics, and feedback, thereby providing the most relevant and instructive financial advice.

[0092] Use online learning algorithms to continuously update and adjust the large language model to adapt to new financial terms, market changes, and user behaviors. This can be achieved through the following three steps:

[0093] 1. Real-time data collection and preprocessing

[0094] The system first continuously collects real-time financial data and user interaction data through data pipelines, including newly emerging financial terms, market dynamics, user behavior patterns, etc. In order to ensure the availability of data, the system cleans, normalizes and denoises the collected data to remove invalid or repeated information. Through natural language processing algorithms, the system conducts preliminary analysis of newly emerging financial terms, compares these terms with the existing financial knowledge base, and determines whether they are new terms or updates to existing terms. For example, when new financial products or policy terms appear in the market, the system can identify and process them in a short time, laying the foundation for subsequent online learning.

[0095] 2. Model incremental update and online learning

[0096] After data preprocessing is completed, the system dynamically updates the large language model through online learning algorithms (such as incremental gradient descent or incremental learning models) without having to retrain the entire model. The system uses new financial terms, market changes, and user behavior patterns as incremental data, and gradually adjusts the parameter weights of the model to adapt it to the latest market dynamics. For example, when a certain type of financial term appears frequently in the market, the system will gradually enhance its understanding of the term through incremental learning and associate its meaning with related terms. At the same time, the online learning algorithm will continue to optimize the model's semantic embedding and prediction capabilities to ensure that the model can maintain high accuracy in the face of different contexts and changes.

[0097] 3. Real-time feedback and model self-tuning

[0098] The system also evaluates the performance of the model in real time through user feedback and market reactions, and compares and analyzes the user's interactive behavior (such as clicks, browsing, and consultation history) with the model's recommendation results. If the recommendations generated by the system do not meet user needs, or if there is a deviation between market data and forecasts, the system will further tune the model parameters through a feedback mechanism. For example, when the system recognizes that certain financial terms or strategies do not accurately match user expectations, it will update the model's internal weights through reinforcement learning or adaptive adjustment techniques to make it more accurate when dealing with similar problems in the future. This feedback and self-tuning process ensures that the model always maintains the ability to quickly adapt to the latest financial information in a changing market environment.

[0099] By introducing Bayesian Networks and logistic regression models to conduct risk assessment on financial recommendations generated by the system, it can be achieved mainly through the following four steps:

[0100] 1. Data collection and feature selection

[0101] First, the system collects various data related to risk assessment, including user investment behavior, financial product characteristics, historical market fluctuations, economic indicators, etc. In order to accurately conduct risk assessment, the system must select appropriate features, such as the volatility of investment products, yields, market risk factors (such as inflation rate, interest rate, etc.), user risk preferences, etc. These features will be used for node settings in the Bayesian network and inputs to the logistic regression model. Through data cleaning and feature engineering, the system ensures that these input variables can not only reflect market changes, but also capture users' investment habits, laying a data foundation for subsequent risk assessment.

[0102] 2. Bayesian Network Modeling

[0103] After the feature data is prepared, the system constructs a risk assessment model through Bayesian Networks. Bayesian networks derive potential risks by modeling key variables in the financial market (such as market volatility, investment product performance, etc.) and user behavior as nodes and defining the probabilistic relationship between nodes. For example, the system can set up a node to represent "market volatility" and another node to represent "investment product returns". The relationship between market volatility and return volatility is inferred through the Bayesian network, and the future risk level is predicted based on historical data. Bayesian networks can effectively handle uncertainty through their probabilistic structure and dynamically update risk assessment results based on changing inputs (such as new market data or user behavior).

[0104] 3. Application of Logistic Regression Model

[0105] At the same time, the system uses a logistic regression model to conduct a quantitative risk assessment of the user's investment decision. The logistic regression model uses the user's historical operation data, the characteristics of financial products, and market changes as input variables, and generates a risk probability value by calculating the relationship between these characteristics and high-risk behaviors. For example, logistic regression can analyze the historical gains and losses of an investment strategy (such as high leverage operations), and combined with the current market conditions, derive the possibility of the strategy incurring losses under existing conditions. The logistic regression model performs well in handling classification problems and can generate a specific risk probability value to help users make more informed investment decisions.

[0106] 4. Dynamic adjustment and risk warning

[0107] Finally, the system combines the results of Bayesian networks and logistic regression to provide users with dynamic adjustments and risk warnings through the risk assessment module. For example, when the system finds that the risk probability of a certain investment suggestion is high, it will issue a warning to the user through the risk warning function to remind the user of the potential risks of the investment strategy. The system can also dynamically adjust the parameters of the risk assessment model based on user feedback and real-time changes in the market, so that it can provide more accurate risk predictions when facing different market conditions. By combining the probabilistic reasoning of the Bayesian network and the precise calculation of logistic regression, the system can generate a risk assessment report with high reference value to help users effectively avoid potential financial risks.

[0108] The present invention combines a context recognition module with a financial terminology knowledge base module, so that the system can more accurately extract and analyze the context information in the user input, and dynamically infer the ambiguity and meaning of financial terms in different markets and countries by combining the terminology base in the financial field. By introducing a graph neural network to construct a financial concept map, the system can not only obtain the definition of the current term from the knowledge base, but also understand the correlation between terms, thereby greatly improving the accuracy of context understanding when faced with complex or ambiguous contexts and avoiding misinterpretation of financial terms.

[0109] The present invention combines the user's historical operations, market data and financial events together through an intelligent recommendation module, and dynamically recommends the most relevant financial products and market analysis through collaborative filtering and content recommendation algorithms. The combination of deep learning networks and graph networks makes the recommendations not only relevant to user preferences, but also adjustable according to real-time changes in the market, ensuring that the recommendation results are not only personalized, but also provide operational strategies with high reference value based on the user's actual needs, thereby improving the accuracy and practicality of the recommendations.

[0110] The present invention integrates the risk assessment and prompt module with Bayesian network and logistic regression model, and conducts risk assessment on each recommendation, so that the system can assess the potential market fluctuations, policy changes and risks brought by user operation behaviors while generating financial suggestions, and provide risk prompts in time. By combining historical data and real-time feedback, the adaptive learning module ensures that the system can continuously adjust the risk prediction model, so that users can fully understand the potential risks before making decisions, thereby effectively avoiding investment mistakes or financial losses. This not only improves the security of the system, but also enhances the user's sense of trust.

[0111] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0112] It should be noted that, in this article, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or still includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including a..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0113] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0114] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each application, but such implementation should not be considered to be beyond the scope of this application.

[0115] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0116] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0117] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0118] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0119] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks or optical disks.

[0120] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An intelligent financial question-answering system based on a large language model, characterized in that: The system works together through multiple modules to solve the problems of polysemy and contextual ambiguity in financial terms. The system includes the following modules: The context recognition module uses natural language processing algorithms combined with context analysis technology in the financial field to automatically extract context clues from user questions; The financial terminology knowledge base module combines a pre-trained large-scale financial knowledge base with self-expansion capabilities to store and update the multiple meanings of financial terms and their usage in different markets and countries. It automatically learns and infers the associations between different terms by introducing a financial concept graph based on a graph neural network. The semantic disambiguation module uses advanced clustering algorithms combined with semantic embedding of large-scale language models to resolve the ambiguity of financial terms input by users; The intelligent recommendation module uses collaborative filtering and content recommendation algorithms, combined with the user's historical operations, market data and financial events, to provide the most relevant financial advice. By combining deep learning networks with graph networks, it dynamically recommends relevant market analysis, financial products or operation strategies during the financial question-answering process, ensuring that the model gives advice after identifying the context. Adaptive learning module, which uses online learning algorithms to continuously update and adjust the large language model during system use to adapt to new financial terms, market changes and user behaviors; The risk assessment and prompt module conducts risk assessment on the financial recommendations generated by the system by introducing Bayesian networks and logistic regression models.

2. An intelligent financial question-answering system based on a large language model according to claim 1, characterized in that: Using natural language processing algorithms combined with contextual analysis technology in the financial field, we can automatically extract context clues from user questions. This is achieved through the following steps: Perform preliminary text processing on financial questions input by users; After completing the text preprocessing, semantic analysis technology is used to identify entities and terms in the financial field. Pre-trained financial word embedding models and named entity recognition technology are used to identify keywords and terms in the financial field from the input; After identifying key entities, the context of the term is further determined through context association and dependency analysis. Dependency analysis helps the system understand the semantic structure of the input question by determining the sentence structure and identifying the grammatical relationship between words. After completing the context analysis, the context clue generation module dynamically adjusts the semantic interpretation and extracts the final context clues based on the knowledge base in the financial field.

3. An intelligent financial question-answering system based on a large language model according to claim 1, characterized in that: Combining a large pre-trained financial knowledge base and self-scaling capabilities, we build a financial concept graph through a graph neural network to store and reason about the multiple meanings and market differences of financial terms. This is achieved through the following four steps: Build a comprehensive and large-scale financial knowledge base from multi-source data; After the knowledge base is built, a financial concept graph is created using a graph neural network. The nodes in the graph represent financial terms, markets, or related financial events, and the edges represent the relationships between the terms. The complex relationships between different financial terms are automatically captured through the structured modeling of the graph neural network. After the financial concept graph is constructed, term association learning and reasoning are performed through node feature propagation of the graph neural network. The graph neural network model aggregates the information of neighboring nodes layer by layer, so that each term node can learn contextual features from its associated terms. A real-time update and self-expansion mechanism is designed to automatically detect new financial terms or changes in the meaning of existing terms through real-time synchronization with external data sources, and dynamically update them to the financial knowledge base and concept map. Through the self-learning characteristics of graph neural networks, the term association network can be automatically adjusted to keep it consistent with the latest market dynamics, ensuring continued accuracy in the understanding and reasoning of the terms.

4. An intelligent financial question-answering system based on a large language model according to claim 1, characterized in that: Using advanced clustering algorithms combined with semantic embedding of large-scale language models to resolve the ambiguity of financial terms is divided into the following three steps: Use a pre-trained large-scale language model to perform semantic embedding on the financial terms input by users, convert the input text into a high-dimensional vector representation, and capture the contextual information and semantic features of the financial terms; The generated semantic embedding is input into an advanced clustering algorithm, which clusters vectors with similar semantics by analyzing the distribution of semantic vectors. After clustering is completed, the most appropriate semantic set is selected for resolution based on the specific context of the user's input. By analyzing the overall context of the user's question, the financial terminology that best fits the current context is selected from the clustering results.

5. An intelligent financial question-answering system based on a large language model according to claim 1, characterized in that: Using collaborative filtering and content recommendation algorithms, we combine the user's historical operations, market data, and financial events to provide relevant financial advice, and combine deep learning networks with graph networks for dynamic recommendations. This is achieved through the following three steps: Extract relevant features from users’ historical operation records, current market data, and financial events; After feature extraction, user needs are analyzed through collaborative filtering algorithms and content recommendation algorithms; In the recommendation process, dynamic recommendation optimization is achieved through the combination of deep learning network and graph neural network.

6. An intelligent financial question-answering system based on a large language model according to claim 1, characterized in that: Use online learning algorithms to continuously update and adjust the large language model to adapt to new financial terms, market changes, and user behaviors. This is achieved through the following three steps: First, real-time financial data and user interaction data are continuously collected through data pipelines; After data preprocessing is completed, the large language model is dynamically updated through online learning algorithms without having to retrain the entire model. New financial terms, market changes, and user behavior patterns are used as incremental data to gradually adjust the model's parameter weights to adapt it to the latest market dynamics. The performance of the model is evaluated in real time through user feedback and market reaction, and the user's interactive behavior is compared and analyzed with the model's recommendation results. If the suggestions generated by the system do not meet user needs, or there is a deviation between market data and predictions, the model parameters will be further tuned through the feedback mechanism.

7. An intelligent financial question-answering system based on a large language model according to claim 1, characterized in that: The risk assessment of the financial recommendations generated by the system is carried out by introducing Bayesian networks and logistic regression models, which is achieved through the following four steps: Collect various data related to risk assessment, select appropriate features for Bayesian network node settings and logistic regression model inputs, and ensure that these input variables can reflect market changes and capture users' investment habits through data cleaning and feature engineering, laying a data foundation for subsequent risk assessment; After the feature data is prepared, a risk assessment model is constructed through the Bayesian network. The Bayesian network derives potential risks by modeling key variables and user behaviors in the financial market as nodes and defining the probabilistic relationship between nodes. The logistic regression model is used to quantitatively assess the risk of users' investment decisions. The logistic regression model uses the user's historical operation data, the characteristics of financial products and market changes as input variables, and generates a risk probability value by calculating the relationship between these characteristics and high-risk behaviors; The results of Bayesian networks and logistic regression are combined to provide users with dynamic adjustments and risk warnings through the risk assessment module.

Images