Financial knowledge intelligent question and answer method and device, storage medium and program product

Abstract

The invention discloses a financial knowledge intelligent question and answer method and device, a storage medium and a program product. The method comprises the steps that a user question text is obtained, semantic analysis is conducted on the user question text, the user question text is converted into a vector representation form, and a question text vector is obtained; determining a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library, and determining knowledge point information corresponding to the knowledge vector of which the similarity parameter is greater than a preset similarity threshold as target knowledge information; generating an initial answer result corresponding to the question text of the user according to the target knowledge information by adopting large language models of a plurality of different manufacturers; and fusing initial answer results generated by the large language models of different manufacturers for the user question text to obtain a target answer result. According to the method and the device, the technical problem of insufficient accuracy and specialty in professional questions and answers of a question and answer system in the financial field in related technologies is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method, device, storage medium and program product for intelligent question answering of financial knowledge. Background Art

[0002] At present, financial institutions will integrate large language models into existing business processes and service platforms, such as intelligent customer service systems, investment advisory platforms or internal knowledge bases, to improve overall operational efficiency and customer experience. However, the performance of large language models in related technologies in intelligent question-and-answer applications, especially in professional questions and answers in the financial field, is still insufficient in accuracy and contextual understanding. The model often does not accurately interpret financial terms and complex issues, and has a weak grasp of the context of multiple rounds of conversations. This causes users to often encounter vague answers and incoherent logic when using the intelligent question-and-answer system, affecting users' trust in the system. In addition, most of the existing large models are general-purpose, lacking specialized optimization and training for financial scenarios, and cannot fully meet users' deep needs for professional knowledge.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present application provide a method, device, storage medium and program product for intelligent question and answering of financial knowledge, so as to at least solve the technical problem of insufficient accuracy and professionalism of question and answer systems in the financial field in related technologies in professional question and answering.

[0005] According to one aspect of an embodiment of the present application, a method for intelligent question and answer of financial knowledge is provided, including: obtaining a user question text, and converting the user question text into a vector representation by semantically parsing the user question text to obtain a question text vector; determining a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library, and determining the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as target knowledge information, wherein the financial knowledge vector library contains multiple knowledge vectors, each of which corresponds to a piece of knowledge point information in the financial field; using large language models of multiple different manufacturers to generate an initial answer result corresponding to the user question text based on the target knowledge information, wherein each large language model generates an initial answer result; fusing the initial answer results generated by the large language models of different manufacturers for the user question text to obtain the target answer result.

[0006] Optionally, the user question text is converted into a vector representation by semantically parsing the user question text to obtain a question text vector, including: performing word segmentation on the user question text and determining the part of speech corresponding to each word obtained after the word segmentation; performing syntactic analysis on the user question text based on the part of speech corresponding to each word to obtain semantic structure information corresponding to the user question text, wherein the semantic structure information is used to characterize the contextual association relationship between each word in the user question text; using a language representation model, based on the semantic structure information, determining key sentence segments in the user question text, and mapping the key sentence segments to a preset vector space to obtain a question text vector.

[0007] Optionally, before using large language models from multiple different manufacturers to generate initial answer results corresponding to the user's question text based on target knowledge information, the method also includes: determining the question intent type corresponding to the user's question text, wherein the question intent type includes at least one of the following: policy interpretation, operational guidance, and market analysis; obtaining professional scores corresponding to the large language models from different manufacturers, wherein the professional scores are used to characterize the quality levels of historical answers given by various large language models to historical questions of different question intent types; and determining a preset number of large language models with the highest professional scores under the question intent type corresponding to the user's question text as the large language models used to answer the user's question text.

[0008] Optionally, using large language models from multiple different manufacturers to generate an initial answer result corresponding to the user's question text based on target knowledge information includes: determining an initial input prompt template corresponding to the question text vector, and adjusting the initial input prompt template based on the question intent type and the model characteristics of each large language model used to answer the user's question text to obtain a target input prompt template, wherein the target input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the user's question text; integrating key sentence segments in the user's question text into the target input prompt template, and adding the target knowledge information to the target field position of the target input prompt template to obtain the model input text, wherein the target field position is used to indicate the knowledge data based on which the large language model analyzes the user's question text; using the large language model to generate an initial answer result based on the model input text.

[0009] Optionally, the method also includes: obtaining historical question and answer record information of the user corresponding to the user question text, wherein the historical question and answer record information includes at least one of the following: historical user question text, historical interaction feedback data, and historical satisfaction evaluation data; constructing a user feature portrait based on the historical question and answer record information and the user's account registration information, wherein the user feature portrait is used to characterize the user's behavioral tendencies and decision-making style; using sentiment analysis technology to determine the emotional state corresponding to the user question text, wherein the emotional state is used to characterize the user's emotional tendencies when inputting the user question text; using a large language model to analyze and process the model input text based on the user feature portrait and emotional state, and obtain an initial answer result that can meet the user's personalized needs.

[0010] Optionally, after generating an initial answer result corresponding to the user question text, the method further includes: determining a relevance parameter and an accuracy parameter corresponding to the initial answer result generated by each large language model, wherein the relevance parameter is used to characterize the degree of relevance between the initial answer result and the user question text, and the accuracy parameter is used to characterize the degree of consistency between the initial answer result and the factual information; determining a first score corresponding to the initial answer result based on the relevance parameter and the accuracy parameter; obtaining a second score corresponding to the initial answer result generated by each large language model of user feedback; and updating the corresponding professional score of each large language model under the question intent type corresponding to the user question text based on the first score and the second score.

[0011] Optionally, the initial answer results generated by large language models of different manufacturers for the user question text are integrated to obtain the target answer result, including: determining the common content between the initial answer results generated by each large language model, wherein the common content is a text segment whose similarity between the initial answer results exceeds a preset similarity threshold; determining the weighted weight corresponding to each initial answer result based on the first score and the second score corresponding to the initial answer result generated by each large language model; and integrating the content in each initial answer result into the common content based on the weighted weight corresponding to each initial answer result to obtain the target answer result.

[0012] According to another aspect of the embodiment of the present application, a financial knowledge intelligent question and answer device is also provided, including: a vector conversion module, used to obtain a user question text, and convert the user question text into a vector representation by semantically parsing the user question text to obtain a question text vector; a knowledge acquisition module, used to determine the similarity parameter between the question text vector and the knowledge vector in the financial knowledge vector library, and determine the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as the target knowledge information, wherein the financial knowledge vector library contains multiple knowledge vectors, each knowledge vector corresponds to a knowledge point information in the financial field; a model answer module, used to determine to use multiple large language models from different manufacturers, and generate an initial answer result corresponding to the user question text based on the target knowledge information, wherein each large language model generates an initial answer result; an answer fusion module, used to fuse the initial answer results generated by the large language models of different manufacturers for the user question text to obtain the target answer result.

[0013] According to another aspect of the embodiments of the present application, a non-volatile storage medium is provided, the non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the financial knowledge intelligent question-answering method by running the computer program.

[0014] According to another aspect of the embodiments of the present application, a computer program product is provided, including a computer program, which implements the steps of the financial knowledge intelligent question-answering method when the computer program is executed by a processor.

[0015] In an embodiment of the present application, a user question text is obtained, and the user question text is converted into a vector representation by semantically parsing the user question text to obtain a question text vector; a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library is determined, and the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold is determined as the target knowledge information, wherein the financial knowledge vector library contains multiple knowledge vectors, and each knowledge vector corresponds to a knowledge point information in the financial field; a plurality of large language models from different manufacturers are used to generate an initial answer result corresponding to the user question text based on the target knowledge information, wherein each large language model generates an initial answer result; a method of fusing the initial answer results generated by the large language models of different manufacturers for the user question text to obtain the target answer result, through the construction of a financial knowledge vector library and the use of a multi-model collaborative mechanism, the purpose of improving the coverage of professional knowledge and context understanding ability of the intelligent question-answering system in the financial field is achieved, thereby solving the technical problems of the accuracy and lack of professionalism of the financial field question-answering system in professional question-answering in the related technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0017] Figure 1 It is a hardware structure block diagram of a computer terminal (or electronic device) for implementing a method for intelligent question-answering of financial knowledge provided in an embodiment of the present application;

[0018] Figure 2 It is a schematic diagram of a method flow of intelligent question and answer of financial knowledge provided according to an embodiment of the present application;

[0019] Figure 3 It is a structural diagram of a financial knowledge intelligent question-answering device provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0021] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0022] Although the financial question-answering system in related technologies has made great progress by combining large language models, it still has some obvious shortcomings and deficiencies, as follows.

[0023] The performance of existing large language models used in financial question-and-answer systems in related technologies in intelligent question-and-answer applications, especially in professional questions and answers in the financial field, is still insufficient in accuracy and context understanding. The models often do not accurately interpret financial terms and complex issues, and have a weak grasp of the context of multiple rounds of conversations. This leads to users often encountering vague answers and incoherent logic when using intelligent question-and-answer systems, which affects users' trust in the system. In addition, most existing large models are general-purpose, lacking specialized optimization and training for financial scenarios, and cannot fully meet users' in-depth needs for professional knowledge.

[0024] In addition, due to the limited depth of analysis and mining of user data by the financial question-and-answer systems in related technologies and the lack of sensitivity to users' real-time needs and dynamic behaviors, the accuracy and responsiveness of personalized services are still insufficient. As a result, the recommended content is often not accurate enough, making it difficult to effectively improve user satisfaction and stickiness.

[0025] In order to solve the above problems, relevant solutions are provided in the embodiments of the present application, which are described in detail below.

[0026] According to an embodiment of the present application, a method embodiment of intelligent question and answer of financial knowledge is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0027] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal (or electronic device) for implementing a financial knowledge intelligent question-answering method. Figure 1 As shown, the computer terminal 10 (or electronic device) may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0028] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10 (or electronic device). As involved in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0029] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the intelligent question-answering method for financial knowledge in the embodiment of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, the intelligent question-answering method for financial knowledge is realized. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0030] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0031] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or electronic device).

[0032] In the above operating environment, the embodiment of the present application provides a financial knowledge intelligent question answering method. Figure 2 is a schematic diagram of a method flow for intelligent question answering of financial knowledge provided in accordance with an embodiment of the present application, such as Figure 2 As shown, the method comprises the following steps:

[0033] Step S202, obtaining a user question text, and converting the user question text into a vector representation by semantically parsing the user question text to obtain a question text vector;

[0034] Step S204, determining a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library, and determining the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as the target knowledge information, wherein the financial knowledge vector library contains multiple knowledge vectors, each of which corresponds to a knowledge point information in the financial field;

[0035] Step S206, using large language models from multiple different vendors to generate an initial answer result corresponding to the user's question text based on the target knowledge information, wherein each large language model generates an initial answer result;

[0036] Step S208, integrating the initial answer results generated by the large language models of different manufacturers for the user's question text to obtain the target answer result.

[0037] Through the above steps, by building a financial knowledge vector library and utilizing a multi-model collaborative mechanism, the goal of improving the intelligent question-answering system's coverage of professional knowledge and contextual understanding in the financial field has been achieved, thereby solving the technical problem of insufficient accuracy and professionalism in professional question-answering in the financial field question-answering system in related technologies.

[0038] The following is a further introduction to the intelligent question-answering method for financial knowledge in steps S202 to S208 of the embodiment of the present application.

[0039] The intelligent application platform of the embodiment of the present application adopts a diversified large language model service architecture, which integrates large language models of various different manufacturers. Each model has its own advantages in natural language processing, data analysis, and intelligent recommendation, and can be accurately matched according to the specific needs of users. At the same time, a flexible model selection mechanism is provided, which can select the most appropriate model to handle different tasks, such as answers to the basic knowledge of digital RMB, policy interpretation, risk assessment, and customer service. This diverse model selection not only enhances the service capabilities of the platform, but also enables the digital currency bank intelligent application platform to better adapt to diverse user needs.

[0040] When faced with complex financial problems, efficient problem-solving capabilities can be achieved through the collaborative work of multiple models. The platform can call multiple models at the same time to analyze and answer the same problem from multiple angles, thereby providing a comprehensive solution. By integrating the advantages of different models, the platform not only improves the accuracy and professionalism of the answers, but also helps users find the best answers faster through comparison and optimization. For example, when dealing with policy interpretations and application scenario suggestions related to the digital RMB, the synergy of multiple models can significantly improve the depth and breadth of the answers.

[0041] The following is a detailed introduction to the model question-and-answer process of this mechanism.

[0042] First, after the user enters a question on the platform, the system will perform in-depth semantic analysis of the input content (i.e., the user's question text), as follows.

[0043] In some embodiments of the present application, the user question text is semantically parsed and converted into a vector representation to obtain a question text vector, which includes the following steps: performing word segmentation on the user question text and determining the part of speech corresponding to each word obtained after the word segmentation; performing syntactic analysis on the user question text based on the part of speech corresponding to each word to obtain semantic structure information corresponding to the user question text, wherein the semantic structure information is used to characterize the contextual association relationship between each word in the user question text; using a language representation model, based on the semantic structure information, determining the key sentence segments in the user question text, and mapping the key sentence segments to a preset vector space to obtain a question text vector.

[0044] Specifically, natural language processing technology is used to preprocess the user's question text, that is, to perform word segmentation, part-of-speech tagging and syntactic analysis to identify semantic structures. For example, tools such as SpaCy or NLTK can be used to effectively identify keywords and phrases, laying the foundation for subsequent vectorization processing; semantic vectorization is then performed, that is, the preprocessed text is converted into a high-dimensional semantic vector through a pre-trained model (such as BERT, Word2Vec). In this process, the model will use its learned contextual information (semantic structure information) to map each word or phrase into a vector space to ensure the preservation of semantics, thereby obtaining the question text.

[0045] Afterwards, the question text vector is matched with the knowledge vector in the financial knowledge vector library for similarity, and the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than the preset similarity threshold is determined as the target knowledge information, as follows.

[0046] In this embodiment, the above similarity parameters can be calculated using algorithms such as cosine similarity or Euclidean distance, and the generated question text vector can be quickly matched with the indicator vector in the system's built-in financial knowledge vector library. The financial knowledge vector library contains a variety of data sources, including but not limited to: financial policies, market dynamics, and user operation guides. By calculating the similarity between vectors, the most relevant knowledge items can be screened out to provide an information basis for subsequent model calls.

[0047] The following is an introduction to the construction and update process of the financial knowledge vector library.

[0048] The embodiment of the present application provides a powerful knowledge engine for the intelligent platform by constructing a financial knowledge vector library, realizing comprehensive knowledge integration, management and intelligent retrieval. Focusing on the core knowledge of financial professional terms such as digital RMB, including basic concepts, technical architecture, policies and regulations, usage procedures, user common problems and practical application scenarios, this information is systematically collected, organized and structured. By introducing embedding models such as Word2Vec and BERT, the platform can convert various types of text data into high-dimensional semantic vectors, thereby realizing fast matching and accurate response to user queries.

[0049] The financial knowledge vector library in the embodiment of the present application is not only a simple accumulation of information, but also a high-dimensional vector that can be understood by machines through deep semantic parsing technology, so that the platform can transcend traditional keyword retrieval and understand the true intention of user input. When the user asks a question, the platform first uses natural language processing (NLP) technology to perform deep semantic parsing on the input. This process first includes data preprocessing, such as noise removal, spelling correction and word segmentation. Subsequently, pre-trained models such as BERT and Word2Vec are used to extract features from user input. Specifically, the BERT model uses its Transformer architecture to capture contextual information through a self-attention mechanism and convert user input into a high-dimensional vector. This deep semantic parsing allows the platform to accurately locate the most relevant content from the knowledge base when facing complex and diverse questions from users, and provide professional and detailed answers. For example, when a user asks "How to open a digital RMB personal wallet", the platform can quickly parse the core of the problem, and provide step-by-step guidance to users in combination with policy requirements, operation steps and user scenarios.

[0050] At the same time, in order to achieve efficient knowledge retrieval, the embodiment of the present application also designs a high-dimensional vector similarity calculation algorithm, mainly based on cosine similarity and Euclidean Distance methods. These algorithms enable the system to quickly identify the knowledge point closest to the user query vector in the vector space, improving the response speed and accuracy.

[0051] In addition, the vector library has a dynamic and knowledge update mechanism to ensure the timeliness and comprehensiveness of the knowledge content. The update strategy includes regularly capturing the latest data from authoritative information sources, such as central bank policy dynamics, market news, system upgrade information, and user feedback, and processing them through automated workflows. After each update, the data will be classified, labeled, and vectorized, and the new knowledge will be seamlessly integrated into the vector library to ensure that users can always obtain the latest information related to the digital RMB.

[0052] Specifically, the update mechanism can include two modes: regular and real-time. In the regular update mode, the system sets a scheduled task (such as daily or weekly) to capture the latest financial data and policy dynamics from authoritative information sources. The captured data is cleaned and standardized, and then converted into a high-dimensional vector and updated to the semantic vector library; in the real-time update mode, the system can collect information and update vectors immediately when key policies are released or market dynamics occur in a time-driven manner to ensure the timeliness of the content obtained by users. The penalties of the real-time update mechanism are controlled by changes in external APIs or alarms from internal data monitoring systems.

[0053] In addition, in order to further improve the efficiency and accuracy of indication retrieval, as an optional implementation method, in the retrieval process, the semantic vector library is first efficiently indexed and constructed, using a multi-level index structure. Specifically, the platform uses approximate nearest neighbor search technology, and implements fast similarity search by using data structures such as local sensitive hashing and KD-Tree. These data structures can quickly locate the most likely related vector groups by mapping high-dimensional space to low-dimensional space, thereby significantly reducing calculation time. Secondly, the retrieval algorithm uses an incremental learning mechanism for dynamic model training. Incremental learning enables the model to improve adaptability and accuracy by adjusting the existing model weights when receiving new user interaction data without having to train from scratch. This process uses an online learning strategy combined with adaptive learning rate adjustment to ensure that the model can quickly adapt to changes and maintain high performance when new data enters. For example, when users frequently query specific questions, the system will record these interaction data and give these questions higher weights in model training, thereby improving the model's responsiveness to related questions. In order to further enhance the intelligence of the retrieval algorithm, the platform also introduces deep reinforcement learning technology to optimize the retrieval strategy by simulating user query behavior. Specifically, the platform learns the optimal retrieval strategy through a reinforcement learning agent, which is able to evaluate the satisfaction of the returned results after each retrieval and adjust its search strategy to improve the relevance and user satisfaction of future retrievals.

[0054] In addition, the platform implements a multi-level feedback mechanism in the search algorithm, and users can rate the search results. These feedbacks are not only used to update the model, but also to adjust the weights of different vectors in the vector library in real time, further improving the accuracy of subsequent searches. This multi-level feedback system ensures that the platform can continue to learn and evolve to cope with increasingly complex financial problems and diverse user needs.

[0055] After the semantic analysis of the question is completed and the target knowledge information is determined, the multi-model calling stage can be entered. At this time, the platform will select the appropriate question model based on the specific trigger mechanism. The specific steps are as follows.

[0056] In some embodiments of the present application, before using large language models of multiple different manufacturers to generate initial answer results corresponding to the user's question text based on target knowledge information, the method also includes the following steps: determining the question intent type corresponding to the user's question text, wherein the question intent type includes at least one of the following: policy interpretation, operational guidance, and market analysis; obtaining professional scores corresponding to the large language models of different manufacturers, wherein the professional scores are used to characterize the quality level of historical answers given by various large language models to historical questions of different question intent types; and determining a preset number of large language models with the highest professional scores under the question intent type corresponding to the user's question text as the large language models used to answer the user's question text.

[0057] Specifically, the question type (such as policy interpretation, operational guidance, market analysis, etc.) can be automatically identified based on the keywords and syntactic structure of the question. For example, if a user asks "How to open a digital RMB wallet?", the system will judge it as an operational guidance question; then, the optimal model can be selected through preset rules and machine learning algorithms, for example, by evaluating the performance of each model in historical data and considering its professionalism in a specific field (i.e., professionalism score). For example, if the question involves the latest financial policy, the system will give priority to models that perform well in policy interpretation.

[0058] In addition, in this embodiment, data transmission between different models can be achieved through RESTful API or message queue technology to ensure efficient and seamless information interaction, so that each model can independently process and return its analysis results after receiving the input.

[0059] After determining the large language models used to answer the user's question text, these large language models can be used to answer the questions. The specific steps are as follows.

[0060] In some embodiments of the present application, using large language models from multiple different manufacturers to generate an initial answer result corresponding to the user's question text based on target knowledge information includes the following steps: determining an initial input prompt template corresponding to the question text vector, and adjusting the initial input prompt template based on the question intent type and the model features of each large language model used to answer the user's question text to obtain a target input prompt template, wherein the target input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the user's question text; integrating key sentence segments in the user's question text into the target input prompt template, and adding the target knowledge information to the target field position of the target input prompt template to obtain a model input text, wherein the target field position is used to indicate the knowledge data based on which the large language model analyzes the user's question text; using the large language model to generate an initial answer result based on the model input text.

[0061] In the embodiment of the present application, Prompt Engineering technology can be used in the process of intelligent question-answering and answer generation, and the large language model can be effectively guided to generate high-quality responses that meet user needs through a carefully designed Prompt (prompt template) instruction set. Specifically, the Prompt instruction set can dynamically adjust the initial input prompt template according to the type of question input by the user (i.e., the type of question intent, such as policy interpretation, operation guidance, etc.). For example, for the question "How to open a digital RMB wallet" asked by the user, the platform may design a specific Prompt: "Please describe in detail how to open a digital RMB wallet based on the following information: {Relevant information}". The "relevant information" is the target field position of the knowledge data used to indicate the large language model to analyze the user's question text. This structure not only clearly indicates the content that the model needs to pay attention to, but also provides sufficient context to enable the model to generate more accurate and professional answers.

[0062] At the same time, in order to adapt to the characteristics of different large language models, a multi-model collaborative working mechanism can also be combined. Each large language model (such as ChatGLM, Tongyi Qianwen, Yayi and Baichuan) has its own unique algorithm and advantages. Therefore, in this embodiment, the Prompt instruction will be adjusted accordingly according to the characteristics of the model.

[0063] In addition, as an optional implementation, the embodiment of the present application can also adopt a multi-dimensional evaluation strategy, including user feedback, accuracy of answers, logic and professionalism, etc., to evaluate the impact of Prompt Engineering on the quality of answers, ensuring that the generated answers not only meet user needs, but also have a high degree of professionalism and practicality. This technical architecture not only ensures that the platform accurately understands user questions, but also generates logically rigorous and highly contextually relevant answers through collaborative processing of models.

[0064] On the other hand, in order to improve the accuracy and responsiveness of the model's personalized services in the intelligent question-answering process, the embodiment of the present application will also use deep learning and behavioral analysis technology to achieve in-depth mining of user historical queries, behavior patterns, and personality preferences when using a multi-model mechanism to answer user question texts, thereby providing more accurate and more personalized recommendations and real-time services that meet user needs, thereby improving user interactive experience and satisfaction. The specific steps are as follows.

[0065] In some embodiments of the present application, the method also includes the following steps: obtaining historical question and answer record information of the user corresponding to the user question text, wherein the historical question and answer record information includes at least one of the following: historical user question text, historical interaction feedback data, and historical satisfaction evaluation data; constructing a user feature portrait based on the historical question and answer record information and the user's account registration information, wherein the user feature portrait is used to characterize the user's behavioral tendencies and decision-making style; using sentiment analysis technology to determine the emotional state corresponding to the user question text, wherein the emotional state is used to characterize the user's emotional tendencies when inputting the user question text; using a large language model, based on the user feature portrait and emotional state, to analyze and process the model input text to obtain an initial answer result that can meet the user's personalized needs.

[0066] Specifically, historical question-and-answer record information (including user query content, response time, satisfaction evaluation, etc.) such as user historical queries, interactive feedback, and preference data is collected, and data mining technology is used to analyze and identify user demand patterns and characteristics, thereby building a user feature profile. Based on the constructed user feature profile, the model can apply recommendation algorithms such as collaborative filtering and content recommendation. Collaborative filtering recommends content that other users like to the current user by analyzing the behavior of similar users, while content recommendation pushes information related to the user's interests based on the user's previous query content.

[0067] In addition, as an optional implementation, machine learning algorithms can also be combined to provide users with intelligent decision-making support. For example, when a user asks how to optimize the use of digital RMB, the system will consider the user's historical transaction data and preferences and automatically generate personalized suggestions, such as recommending the use of a specific type of wallet or providing relevant policy information. On the other hand, the content recommendation algorithm analyzes the keywords and contextual information of the user's query, combined with the content features in the knowledge base, to provide suggestions that are highly matched with the user's current needs.

[0068] After obtaining the initial answer results output by each large language model, it is necessary to conduct a quality assessment on the output of each model. When conducting the assessment, multiple dimensions such as relevance, accuracy, and user satisfaction can be considered. The specific steps are as follows.

[0069] In some embodiments of the present application, after generating an initial answer result corresponding to the user question text, the method also includes the following steps: determining a relevance parameter and an accuracy parameter corresponding to the initial answer result generated by each large language model, wherein the relevance parameter is used to characterize the degree of relevance between the initial answer result and the user question text, and the accuracy parameter is used to characterize the degree of consistency between the initial answer result and the factual information; determining a first score corresponding to the initial answer result based on the relevance parameter and the accuracy parameter; obtaining a second score corresponding to the initial answer result generated by each large language model of user feedback; and updating the corresponding professional score of each large language model under the question intent type corresponding to the user question text based on the first score and the second score.

[0070] Specifically, machine learning models (such as random forests or neural networks) can be used to analyze the quality (relevance and accuracy) of different answers and generate a first score for each model output. At the same time, the second score can be determined by collecting user rating feedback. Finally, the professionalism score corresponding to each large language model under the question intent type corresponding to the user's question text is determined and updated.

[0071] Finally, the answer results of different models are dynamically integrated. The specific steps are as follows.

[0072] In some embodiments of the present application, fusing the initial answer results generated by large language models of different manufacturers for user question text to obtain a target answer result includes the following steps: determining the common content between the initial answer results generated by each large language model, wherein the common content is a text segment whose similarity between the initial answer results exceeds a preset similarity threshold; determining the weighted weight corresponding to each initial answer result based on the first score and the second score corresponding to the initial answer result generated by each large language model; and fusing the content in each initial answer result into the common content based on the weighted weight corresponding to each initial answer result to obtain the target answer result.

[0073] In this embodiment, the outputs of each model can be fused by using weighted average, voting mechanism or other integrated learning strategies according to the professional rating and historical feedback information of the model. The dynamic fusion strategy ensures that the weights of each model can be flexibly adjusted under different contexts, thereby generating a more accurate final answer, that is, the target answer result.

[0074] Specifically, a dynamic fusion strategy is adopted to integrate and optimize the outputs of each model. The answer of each model will be sent to an integration layer, which uses an algorithm to score and sort the credibility, relevance and logical consistency of the output for multiple rounds, and gradually screen out the best answer. In this step, the platform applies technologies such as Bayesian reasoning, integrated learning and adaptive attention mechanism to comprehensively evaluate the model output to ensure that the final generated answer is not only accurate but also highly professional and practical. For example, when a user asks "Which cities are promoting the digital RMB pilot?", the platform will not only provide a list of pilot cities, but also combine the latest policies, pilot progress and specific operational suggestions that users are concerned about to form a comprehensive and accurate answer.

[0075] In addition, user feedback will become an important basis for further optimization of the system. By continuously collecting user feedback on answers, the platform can automatically adjust the model selection algorithm and answer integration strategy to continuously improve the intelligence level of the service and user satisfaction. Finally, the system returns the generated answers to the user and displays them through a friendly interface. In this process, the platform ensures that the answers are not only professional but also easy to understand, supplemented by necessary charts or step-by-step instructions to enhance the user experience.

[0076] This multi-model collaborative mechanism enables the Digital Currency Bank platform to handle user questions from multiple angles. For example, in response to users' inquiries about the "digital RMB pilot process", the platform can combine the analytical capabilities of different models to provide multi-dimensional answers on policy background, practical applications and future trends. By combining the strengths of each model, the platform not only improves the accuracy and depth of the answers, but also adjusts the focus of the answers according to user needs, providing more personalized and comprehensive solutions.

[0077] In addition, the platform's multi-model collaborative working method can dynamically respond to users' complex needs. Whether it is policy interpretation, technical operation guidance, or market dynamic analysis, the Digital Coin Bank platform can quickly integrate the output of multiple models to present users with a variety of choices. Users can compare multiple answers and choose the answer that best suits their needs. This multi-angle, personalized service significantly improves user experience and problem-solving efficiency. Through multi-model collaboration and semantic retrieval, the Digital Coin Bank Intelligent Platform not only achieves accurate matching of user needs, but also promotes the comprehensive upgrade of intelligent financial services.

[0078] This application scheme can provide users with a personalized and efficient service experience. Through the intelligent question-and-answer system and the collaborative work of multiple models, the platform can accurately understand and quickly respond to users' various needs for financial concepts such as digital RMB, and can provide professional and detailed answers from basic operations to complex financial consultations. The highly personalized recommendation system ensures that users receive tailored solutions, significantly improving service satisfaction, and at the same time enhancing user trust and loyalty through strict data protection measures.

[0079] According to an embodiment of the present application, an embodiment of a financial knowledge intelligent question and answer device is also provided. Figure 3 Schematic diagram of a structure of a financial knowledge intelligent question-answering device provided according to an embodiment of the present application. Figure 3 As shown, the device comprises:

[0080] A vector conversion module 30 is used to obtain a user question text, and convert the user question text into a vector representation by semantically parsing the user question text to obtain a question text vector;

[0081] The knowledge acquisition module 32 is used to determine the similarity parameter between the question text vector and the knowledge vector in the financial knowledge vector library, and determine the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as the target knowledge information, wherein the financial knowledge vector library contains multiple knowledge vectors, each of which corresponds to a knowledge point information in the financial field;

[0082] The model answering module 34 is used to determine to use a plurality of large language models of different manufacturers, and to generate an initial answer result corresponding to the user's question text according to the target knowledge information, wherein each large language model generates an initial answer result;

[0083] The answer fusion module 36 is used to fuse the initial answer results generated by the large language models of different manufacturers for the user's question text to obtain the target answer result.

[0084] Optionally, the user question text is converted into a vector representation by semantically parsing the user question text to obtain a question text vector, including: performing word segmentation on the user question text and determining the part of speech corresponding to each word obtained after the word segmentation; performing syntactic analysis on the user question text based on the part of speech corresponding to each word to obtain semantic structure information corresponding to the user question text, wherein the semantic structure information is used to characterize the contextual association relationship between each word in the user question text; using a language representation model, based on the semantic structure information, determining key sentence segments in the user question text, and mapping the key sentence segments to a preset vector space to obtain a question text vector.

[0085] Optionally, before using large language models from multiple different manufacturers to generate initial answer results corresponding to the user's question text based on target knowledge information, the method also includes: determining the question intent type corresponding to the user's question text, wherein the question intent type includes at least one of the following: policy interpretation, operational guidance, and market analysis; obtaining professional scores corresponding to the large language models from different manufacturers, wherein the professional scores are used to characterize the quality levels of historical answers given by various large language models to historical questions of different question intent types; and determining a preset number of large language models with the highest professional scores under the question intent type corresponding to the user's question text as the large language models used to answer the user's question text.

[0086] Optionally, using large language models from multiple different manufacturers to generate an initial answer result corresponding to the user's question text based on target knowledge information includes: determining an initial input prompt template corresponding to the question text vector, and adjusting the initial input prompt template based on the question intent type and the model characteristics of each large language model used to answer the user's question text to obtain a target input prompt template, wherein the target input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the user's question text; integrating key sentence segments in the user's question text into the target input prompt template, and adding the target knowledge information to the target field position of the target input prompt template to obtain the model input text, wherein the target field position is used to indicate the knowledge data based on which the large language model analyzes the user's question text; using the large language model to generate an initial answer result based on the model input text.

[0087] Optionally, the method also includes: obtaining historical question and answer record information of the user corresponding to the user question text, wherein the historical question and answer record information includes at least one of the following: historical user question text, historical interaction feedback data, and historical satisfaction evaluation data; constructing a user feature portrait based on the historical question and answer record information and the user's account registration information, wherein the user feature portrait is used to characterize the user's behavioral tendencies and decision-making style; using sentiment analysis technology to determine the emotional state corresponding to the user question text, wherein the emotional state is used to characterize the user's emotional tendencies when inputting the user question text; using a large language model to analyze and process the model input text based on the user feature portrait and emotional state, and obtain an initial answer result that can meet the user's personalized needs.

[0088] Optionally, after generating an initial answer result corresponding to the user question text, the method further includes: determining a relevance parameter and an accuracy parameter corresponding to the initial answer result generated by each large language model, wherein the relevance parameter is used to characterize the degree of relevance between the initial answer result and the user question text, and the accuracy parameter is used to characterize the degree of consistency between the initial answer result and the factual information; determining a first score corresponding to the initial answer result based on the relevance parameter and the accuracy parameter; obtaining a second score corresponding to the initial answer result generated by each large language model of user feedback; and updating the corresponding professional score of each large language model under the question intent type corresponding to the user question text based on the first score and the second score.

[0089] Optionally, the initial answer results generated by large language models of different manufacturers for the user question text are integrated to obtain the target answer result, including: determining the common content between the initial answer results generated by each large language model, wherein the common content is a text segment whose similarity between the initial answer results exceeds a preset similarity threshold; determining the weighted weight corresponding to each initial answer result based on the first score and the second score corresponding to the initial answer result generated by each large language model; and integrating the content in each initial answer result into the common content based on the weighted weight corresponding to each initial answer result to obtain the target answer result.

[0090] It should be noted that the various modules in the above-mentioned financial knowledge intelligent question and answer device can be program modules (for example, a set of program instructions that implement a certain specific function) or hardware modules. For the latter, it can be expressed in the following forms, but is not limited to this: the expression form of each of the above-mentioned modules is a processor, or the functions of each of the above-mentioned modules are implemented by a processor.

[0091] It should be noted that the financial knowledge intelligent question-answering device provided in this embodiment can be used to execute Figure 2 The financial knowledge intelligent question and answer method shown, therefore, the relevant explanations and descriptions of the above-mentioned financial knowledge intelligent question and answer method are also applicable to the embodiments of this application and will not be repeated here.

[0092] The embodiment of the present application also provides a non-volatile storage medium, the non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the following financial knowledge intelligent question-answering method by running the computer program: obtaining a user question text, and converting the user question text into a vector representation by semantically parsing the user question text to obtain a question text vector; determining a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library, and determining the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as the target knowledge information, wherein the financial knowledge vector library contains multiple knowledge vectors, each of which corresponds to a piece of knowledge point information in the financial field; using large language models from multiple different manufacturers to generate an initial answer result corresponding to the user question text based on the target knowledge information, wherein each large language model generates an initial answer result; fusing the initial answer results generated by the large language models of different manufacturers for the user question text to obtain the target answer result.

[0093] The embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the financial knowledge intelligent question-answering method described in each embodiment of the present application: obtaining a user question text, and converting the user question text into a vector representation by semantically parsing the user question text to obtain a question text vector; determining a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library, and determining the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as the target knowledge information, wherein the financial knowledge vector library contains multiple knowledge vectors, each of which corresponds to a piece of knowledge point information in the financial field; using large language models of multiple different manufacturers to generate an initial answer result corresponding to the user question text based on the target knowledge information, wherein each large language model generates an initial answer result; fusing the initial answer results generated by the large language models of different manufacturers for the user question text to obtain the target answer result.

[0094] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0095] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0096] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, 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 units or modules, which can be electrical or other forms.

[0097] 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 units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0098] 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. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0099] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc., which can store program code.

[0100] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A financial knowledge intelligent question-answering method, characterized in that: include: Obtaining a user question text, and converting the user question text into a vector representation by performing semantic analysis on the user question text, thereby obtaining a question text vector; Determine a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library, and determine the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as the target knowledge information, wherein the financial knowledge vector library contains a plurality of the knowledge vectors, each of which corresponds to a piece of knowledge point information in the financial field; Using large language models from multiple different manufacturers, based on the target knowledge information, to generate an initial answer result corresponding to the user question text, wherein each of the large language models generates one initial answer result; The initial answer results generated by the large language models of different manufacturers for the user question text are integrated to obtain the target answer result.

2. The intelligent question-answering method for financial knowledge according to claim 1, characterized in that: By performing semantic analysis on the user question text, the user question text is converted into a vector representation form, and the question text vector obtained includes: Performing word segmentation on the user question text and determining the part of speech corresponding to each word obtained after word segmentation; According to the part of speech corresponding to each word, the user question text is subjected to syntactic analysis to obtain semantic structure information corresponding to the user question text, wherein the semantic structure information is used to characterize the contextual association relationship between the words in the user question text; A language representation model is adopted to determine key sentence segments in the user question text according to the semantic structure information, and the key sentence segments are mapped into a preset vector space to obtain the question text vector.

3. The intelligent question-answering method for financial knowledge according to claim 1, characterized in that: Before using large language models of multiple different manufacturers to generate an initial answer result corresponding to the user question text based on the target knowledge information, the method further includes: Determining the question intention type corresponding to the user question text, wherein the question intention type includes at least one of the following: policy interpretation, operation guidance, and market analysis; Obtaining professional scores corresponding to the large language models of different manufacturers, wherein the professional scores are used to characterize the quality levels of historical answers given by various large language models to historical questions of different question intent types; A preset number of the large language models with the highest professional scores under the question intention type corresponding to the user question text are determined as the large language models used to answer the user question text.

4. The intelligent question-answering method for financial knowledge according to claim 3, characterized in that: Using large language models from multiple different vendors, generating initial answer results corresponding to the user's question text based on the target knowledge information includes: Determine an initial input prompt template corresponding to the question text vector, and adjust the initial input prompt template according to the question intention type and the model features of each of the large language models used to answer the user question text to obtain a target input prompt template, wherein the target input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the user question text; Integrate the key sentence segments in the user question text into the target input prompt template, and add the target knowledge information to the target field position of the target input prompt template to obtain a model input text, wherein the target field position is used to indicate the knowledge data based on which the large language model analyzes the user question text; The large language model is used, and the initial answer result is generated according to the model input text.

5. The intelligent question-answering method for financial knowledge according to claim 4, characterized in that: The method further comprises: Acquire historical question and answer record information of the user corresponding to the user question text, wherein the historical question and answer record information includes at least one of the following: historical user question text, historical interaction feedback data, and historical satisfaction evaluation data; Constructing a user feature profile based on the historical question and answer record information and the user's account registration information, wherein the user feature profile is used to characterize the user's behavioral tendencies and decision-making style; Using sentiment analysis technology to determine the emotional state corresponding to the user question text, wherein the emotional state is used to characterize the emotional tendency of the user when entering the user question text; The large language model is used to analyze and process the model input text based on the user feature portrait and the emotional state to obtain the initial answer result that can meet the personalized needs of the user.

6. The intelligent question-answering method for financial knowledge according to claim 3, characterized in that: After generating an initial answer result corresponding to the user question text, the method further includes: Determine the relevance parameter and accuracy parameter corresponding to the initial answer result generated by each of the large language models, wherein the relevance parameter is used to characterize the degree of relevance between the initial answer result and the user question text, and the accuracy parameter is used to characterize the degree of conformity between the initial answer result and factual information; Determining a first score corresponding to the initial answer result according to the relevance parameter and the accuracy parameter; Obtaining a second score corresponding to each of the initial answer results generated by the large language model of the user feedback; Based on the first score and the second score, the professional score corresponding to each of the large language models under the question intent type corresponding to the user question text is updated.

7. The intelligent question-answering method for financial knowledge according to claim 6, characterized in that: The initial answer results generated by the large language models of different manufacturers for the user question text are integrated to obtain the target answer results including: Determine the common content between the initial answer results generated by the large language models, wherein the common content is a text segment whose similarity between the initial answer results exceeds a preset similarity threshold; Determine a weighted weight corresponding to each of the initial answer results generated by each of the large language models according to the first score and the second score corresponding to the initial answer results; According to the weighted weights corresponding to the respective initial answer results, the contents in the respective initial answer results are merged into the common content to obtain the target answer result.

8. A financial knowledge intelligent question-answering device, characterized in that: include: A vector conversion module is used to obtain a user question text, and convert the user question text into a vector representation by performing semantic analysis on the user question text to obtain a question text vector; A knowledge acquisition module, used to determine a similarity parameter between the question text vector and a knowledge vector in a financial knowledge vector library, and determine the knowledge point information corresponding to the knowledge vector whose similarity parameter is greater than a preset similarity threshold as target knowledge information, wherein the financial knowledge vector library contains a plurality of the knowledge vectors, each of which corresponds to a piece of knowledge point information in the financial field; A model answering module, used to determine to use a plurality of large language models of different manufacturers, and to generate an initial answer result corresponding to the user question text according to the target knowledge information, wherein each of the large language models generates one initial answer result; The answer fusion module is used to fuse the initial answer results generated by the large language models of different manufacturers for the user question text to obtain a target answer result.

9. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the financial knowledge intelligent question-answering method described in any one of claims 1 to 7 by running the computer program.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the financial knowledge intelligent question-answering method described in any one of claims 1 to 7 are implemented.

Images