Financial information query method and device based on artificial intelligence

Abstract

The present application discloses a financial information query method and device based on artificial intelligence. The method includes: obtaining a query task text, wherein the task text is used to characterize the financial problem that the user plans to query; determining a first similarity between the query task text and the summary information of each knowledge vector in the financial knowledge base, and determining a candidate knowledge vector in the knowledge vector in the financial knowledge base based on the first similarity; determining a second similarity between the query task text and the candidate knowledge vector, and determining a target knowledge vector in the candidate knowledge vector based on the second similarity; determining a query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, wherein the query result is used to answer the financial problem corresponding to the query task text. The present application solves the technical problem that it is difficult for bank employees to effectively integrate and quickly query relevant financial information due to the increasing complexity of banking business.

Description

Technical Field

[0001] The present application relates to the field of smart financial technology, and more specifically, to a financial information query method and device based on artificial intelligence. Background Art

[0002] As banking business becomes increasingly complex, employees need to be familiar with a large number of rules and regulations, business operation guidelines and external regulatory policies. Information is scattered and updated frequently, which brings huge challenges to banks and their employees. Relevant technologies are difficult to effectively integrate and quickly retrieve relevant financial information, causing employees to spend a lot of time searching and learning, and unable to quickly and accurately query the financial information they want, affecting work efficiency and accuracy.

[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 financial information query method and device based on artificial intelligence to at least solve the technical problem that bank employees have difficulty in effectively integrating and quickly querying relevant financial information due to the increasing complexity of banking business.

[0005] According to one aspect of an embodiment of the present application, there is provided an artificial intelligence-based financial information query method, comprising: obtaining a query task text, wherein the query task text is used to characterize a financial question that a user plans to query; determining a first similarity between the query task text and summary information of each knowledge vector in a financial knowledge base, and determining a candidate knowledge vector in the knowledge vectors in the financial knowledge base based on the first similarity, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to characterize a semantic topic of the financial information corresponding to the knowledge vector; determining a second similarity between the query task text and the candidate knowledge vectors, and determining a target knowledge vector in the candidate knowledge vectors based on the second similarity; determining a query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, wherein the query result is used to answer the financial question corresponding to the query task text.

[0006] Optionally, the construction process of the financial knowledge base includes the following steps: according to a preset collection cycle, obtaining the original text of information in the information source, wherein the original text of the information includes at least one of the following: financial policy consultation, financial product introduction; processing the original text of the information into blocks to obtain multiple initial text blocks; using a semantic analysis model to determine whether the semantic information expressed by the initial text block is complete, and if the semantic information expressed by the initial text block is complete, determining the initial text block as the target text block; if the semantic information expressed by the initial text block is incomplete, determining other initial text blocks whose semantic similarity with the initial text block exceeds a preset similarity threshold, and based on the other initial text blocks, semantically completing the initial text block with incomplete semantic information to obtain the target text block; using a semantic analysis model to map the target text block to a preset vector space to obtain a knowledge vector, and determining summary information corresponding to the knowledge vector.

[0007] Optionally, the original text of the information is divided into blocks to obtain multiple initial text blocks, including: using regular expressions to identify delimiters in the original text of the information, and dividing the original text of the information into multiple sentence fragments based on the delimiters, wherein the delimiters include: punctuation marks, line breaks; using a sliding window to gradually slide backward from the beginning of the sentence fragment, and after each slide, judging whether to segment the sentence fragment at the position of the sliding window based on the sentence content in the sliding window; after the sliding window reaches the end of the sentence fragment, obtaining the initial text block corresponding to the sentence fragment, wherein, in the process of sliding the sliding window, there is a partial overlap between the framed area of ​​the sliding window after each slide and the framed area of ​​the sliding window before sliding.

[0008] Optionally, the training steps of the semantic analysis model include: obtaining a fine-tuning training data set, wherein the fine-tuning training data set contains multiple knowledge information related to the financial field, and category labels corresponding to each piece of knowledge information; obtaining a pre-trained model, and adding a low-rank bypass to the pre-trained model, wherein the input dimension and input dimension of the low-rank bypass are consistent with the original input dimension and output dimension of the pre-trained model, and the low-rank bypass is used to reduce the parameter adjustment amount of the pre-trained model during the training process of adapting to tasks in financial scenarios; based on the fine-tuning training data set, training the pre-trained model after adding the low-rank bypass to obtain a semantic analysis model, wherein the training process is used to adjust the matrix parameters in the low-rank bypass, while the original model parameters of the pre-trained model before adding the low-rank bypass remain unchanged.

[0009] Optionally, determining the second similarity between the query task text and the candidate knowledge vector includes: performing word segmentation on the query task text and determining the part of speech corresponding to each word obtained after the word segmentation; performing syntactic analysis on the query task text based on the part of speech corresponding to each word to obtain semantic structure information corresponding to the query task text, wherein the semantic structure information is used to characterize the contextual association relationship between each word in the query task text; using a semantic analysis model, based on the semantic structure information, determining keywords in the query task text, and mapping the keywords to a preset vector space to obtain a task semantic vector corresponding to the query task text; calculating the cosine similarity between the task semantic vector and each knowledge vector in the financial knowledge base, and determining the cosine similarity as the second similarity.

[0010] Optionally, the keywords are mapped to a preset vector space to obtain a task semantic vector corresponding to the query task text, including: determining whether the keyword is a polysemous word, and if the keyword is a polysemous word, determining the meaning of the keyword based on the context information corresponding to the keyword in the query task text to ensure that the meaning of each keyword is clear and unique; determining the synonyms or near-synonyms corresponding to each keyword, and using the synonyms or near-synonyms as extended keywords; based on the keywords and extended keywords, vectorizing the query task text to obtain a task semantic vector.

[0011] Optionally, determining the query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector includes: determining the task intent corresponding to the query task text, wherein the task intent includes at least one of the following: policy interpretation, operational guidance, market analysis; obtaining an input prompt template corresponding to the task intent, and adjusting the input prompt template based on the query task text and the financial information corresponding to the target knowledge vector to obtain a model input text, wherein the input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the query task text, and to instruct the large language model on the knowledge data based on which the query task text is analyzed; and using the large language model to generate a query result corresponding to the query task text based on the model input text.

[0012] According to another aspect of an embodiment of the present application, there is also provided an artificial intelligence-based financial information query device, including: a task acquisition module, used to acquire a query task text, wherein the task text is used to characterize the financial problem that the user plans to query; a first matching module, used to determine a first similarity between the query task text and summary information of each knowledge vector in the financial knowledge base, and based on the first similarity, determine a candidate knowledge vector in the knowledge vector in the financial knowledge base, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to characterize the semantic theme of the financial information corresponding to the knowledge vector; a second matching module, used to determine a second similarity between the query task text and the candidate knowledge vector, and based on the second similarity, determine a target knowledge vector in the candidate knowledge vector; an answer generation module, used to determine a query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, wherein the query result is used to answer the financial problem corresponding to the query task text.

[0013] According to another aspect of an embodiment of the present application, a non-volatile storage medium is also provided, the non-volatile storage medium including a stored computer program, wherein the device where the non-volatile storage medium is located executes an artificial intelligence-based financial information query method by running the computer program.

[0014] According to another aspect of the embodiments of the present application, a computer program product is also provided, including a computer program, which implements the steps of the financial information query method based on artificial intelligence when executed by a processor.

[0015] In an embodiment of the present application, a query task text is obtained, wherein the task text is used to characterize the financial problem that the user plans to query; a first similarity between the query task text and the summary information of each knowledge vector in the financial knowledge base is determined, and based on the first similarity, a candidate knowledge vector is determined in the knowledge vector in the financial knowledge base, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to characterize the semantic theme of the financial information corresponding to the knowledge vector; a second similarity between the query task text and the candidate knowledge vector is determined, and based on the second similarity, a target knowledge vector is determined in the candidate knowledge vector; based on the financial information corresponding to the target knowledge vector, a query result corresponding to the query task text is determined, wherein the query result is used to answer the financial problem corresponding to the query task text. Through the financial knowledge vector database and the search enhancement technology, the purpose of realizing efficient retrieval and query of financial business information is achieved, thereby solving the problem that bank employees find it difficult to effectively integrate and quickly query related financial information technology due to the increasing complexity of banking business. 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 financial information query provided in an embodiment of the present application;

[0018] Figure 2 It is a schematic diagram of a method flow for financial information query provided according to an embodiment of the present application;

[0019] Figure 3 A schematic diagram of the structure of a financial information query device based on artificial intelligence 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] At present, with the accelerated development of a new round of scientific and technological revolution and industrial transformation, as well as the accelerated pace of global industrial transformation and adjustment, commercial banks are facing higher requirements for their operations. Globally, the progress and development of artificial intelligence increasingly rely on the improvement and expansion of large language models and large model application platforms. This trend is reflected in the breakthroughs and extensive applications of a series of core technologies, including but not limited to the construction and optimization of large-scale pre-trained models, innovations in deep learning frameworks, progress in model compression and acceleration technologies, improvement of model service deployment, and how to efficiently integrate various plug-ins and small models to achieve a wider range of integrated application scenarios.

[0023] Although the banking industry has begun to apply big data and artificial intelligence technologies to achieve technological upgrades, the relevant technologies still have significant limitations in the following aspects:

[0024] 1) Data silo problem. The data systems of commercial banks have serious data silo problems. The systems are scattered in different departments and lack a connection and sharing mechanism, which leads to inconsistent data and affects the accuracy of analytical conclusions. Although many banks have implemented a certain degree of knowledge management system, these systems usually rely on predefined rules and structured data. This approach is often difficult to achieve effective integration and management when faced with complex and unstructured information (such as free-text policy documents, emails, etc.).

[0025] 2) Limited intelligent search capabilities. Although some banks have deployed search engines and document management systems, these systems mostly rely on keyword matching and have difficulty understanding users’ actual needs or query intent, resulting in inaccurate search results and failure to meet employees’ instant information needs in complex business scenarios.

[0026] In order to solve the above problems, relevant solutions are provided in the embodiments of the present application, which are described in detail below.

[0027] According to an embodiment of the present application, a method embodiment of financial information query 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.

[0028] 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 The hardware structure block diagram of a computer terminal (or electronic device) for implementing a financial information query method based on artificial intelligence is shown. Figure 1As 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.

[0029] 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).

[0030] 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 artificial intelligence-based financial information query method 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, realizing the above-mentioned artificial intelligence-based financial information query method. 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 a combination thereof.

[0031] 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.

[0032] 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).

[0033] In the above operating environment, the embodiment of the present application provides a financial information query method based on artificial intelligence. Figure 2 is a schematic diagram of a method flow for financial information query provided according to an embodiment of the present application, such as Figure 2 As shown, the method comprises the following steps:

[0034] Step S202, obtaining a query task text, wherein the task text is used to characterize the financial question that the user plans to query;

[0035] Step S204, determining a first similarity between the query task text and summary information of each knowledge vector in the financial knowledge base, and determining a candidate knowledge vector from the knowledge vectors in the financial knowledge base based on the first similarity, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to characterize the semantic theme of the financial information corresponding to the knowledge vector;

[0036] Step S206, determining a second similarity between the query task text and the candidate knowledge vectors, and determining a target knowledge vector from the candidate knowledge vectors based on the second similarity;

[0037] Step S208, determining a query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, wherein the query result is used to answer the financial question corresponding to the query task text.

[0038] Through the above steps, the goal of efficient retrieval and query of financial business information is achieved through the financial knowledge vector database and search enhancement technology, thereby solving the problem of bank employees having difficulty in effectively integrating and quickly querying related financial information technology due to the increasing complexity of banking business.

[0039] The following is a further introduction to the artificial intelligence-based financial information query method in steps S202 to S208 of the embodiment of the present application.

[0040] First, the construction process of the above financial knowledge base is introduced. The specific steps are as follows.

[0041] In some embodiments of the present application, the construction process of the financial knowledge base includes the following steps: according to a preset collection cycle, the original text of the information in the information source is obtained, wherein the original text of the information includes at least one of the following: financial policy advice, financial product introduction; the original text of the information is processed into blocks to obtain multiple initial text blocks; using a semantic analysis model to determine whether the semantic information expressed by the initial text block is complete, and if the semantic information expressed by the initial text block is complete, the initial text block is determined as a target text block; if the semantic information expressed by the initial text block is incomplete, other initial text blocks whose semantic similarity with the initial text block exceeds a preset similarity threshold are determined, and based on the other initial text blocks, the initial text block with incomplete semantic information is semantically completed to obtain a target text block; using a semantic analysis model, the target text block is mapped to a preset vector space to obtain a knowledge vector, and the summary information corresponding to the knowledge vector is determined.

[0042] Specifically, in order to ensure the timeliness and comprehensiveness of the financial knowledge base, a scheduled batch task automated entry mechanism can be adopted. This automated process can regularly capture fresh information from various information sources, automatically complete information screening, format conversion and standardization, and then seamlessly connect it to the knowledge base system, reducing the tediousness and delays of manual operations and ensuring the continuous updating and expansion of the knowledge base content.

[0043] For the original text of information obtained from the information source, this embodiment can process it by parsing and then dividing it into Chunks. The specific steps are as follows.

[0044] In some embodiments of the present application, the original text of the information is divided into blocks to obtain multiple initial text blocks, including the following steps: using regular expressions to identify delimiters in the original text of the information, and dividing the original text of the information into multiple sentence fragments based on the delimiters, wherein the delimiters include: punctuation marks, line breaks; using a sliding window, starting from the beginning of the sentence fragment and sliding backward step by step, and after each sliding, judging whether to segment the sentence fragment at the position of the sliding window based on the sentence content in the sliding window; after the sliding window reaches the end of the sentence fragment, obtaining the initial text block corresponding to the sentence fragment, wherein, in the process of sliding the sliding window, there is a partial overlap between the framed area of ​​the sliding window after each sliding and the framed area of ​​the sliding window before sliding.

[0045] Specifically, the content can be split according to fixed length and natural language boundaries (such as punctuation and paragraph breaks), and the sliding window strategy can be combined to ensure that the text segments have both length restrictions and retain semantic integrity. When splitting, regular expressions are used to automatically identify delimiters, and priority is given to splitting at semantically complete locations. For long texts, context overlap is introduced (that is, there is a partial overlap between the selected area of ​​the sliding window after each slide and the selected area of ​​the sliding window before the slide) to strengthen the semantic coherence between segments.

[0046] At the same time, to ensure the integrity of text semantics, context extension and semantic fusion strategies can also be used to attach contextual information through sliding window technology, and use semantic analysis models (such as Embedding models, such as BERT, Sentence-BERT (SBERT) and BGE models, etc.) to complete the semantics of the split text, that is, to determine whether the semantic information expressed by the initial text block is complete, and to complete the semantics of the initial text block with incomplete semantic information. In addition, the title and field information of special formatted content such as tables and lists are retained to ensure their semantic relevance and ensure that the split fragments have independent semantic expression capabilities during retrieval.

[0047] Through the above processing, the knowledge in the financial knowledge base needs to undergo strict sorting and structuring before being stored. This means that each piece of information will be carefully classified and classified into corresponding subject areas according to its content attributes, such as macroeconomic analysis, financial technology trends, personal financial management guides, etc. Such partitioning and classification not only helps to improve the systematic nature of the knowledge base, but also greatly facilitates users to quickly locate the required content in massive amounts of information.

[0048] It should be noted that the semantic analysis model used in the construction of the financial knowledge base is obtained after fine-tuning the model using banking scenario data and expanding the vocabulary of financial terms in order to enhance the model's adaptability to professional terms and industry contexts, thereby improving the accuracy and effectiveness of the embedding vector. The specific training process is as follows.

[0049] In some embodiments of the present application, the training step of the semantic analysis model includes the following steps: obtaining a fine-tuning training data set, wherein the fine-tuning training data set contains multiple knowledge information related to the financial field, and category labels corresponding to each piece of knowledge information; obtaining a pre-trained model, and adding a low-rank bypass to the pre-trained model, wherein the input dimension and input dimension of the low-rank bypass are consistent with the original input dimension and output dimension of the pre-trained model, and the low-rank bypass is used to reduce the parameter adjustment amount of the pre-trained model during the training process of adapting to tasks in financial scenarios; based on the fine-tuning training data set, training the pre-trained model after adding the low-rank bypass to obtain a semantic analysis model, wherein the training process is used to adjust the matrix parameters in the low-rank bypass, while the original model parameters of the pre-trained model before adding the low-rank bypass remain unchanged.

[0050] Specifically, in the embodiment of the present application, the rich knowledge resources accumulated internally can be fully utilized: the in-house knowledge base, as the core material for this customized training. This knowledge base covers a wide range of financial business knowledge, industry norms, market dynamics, and typical examples of customer communication, providing highly targeted training data for the deep learning of the model. Customized training of the in-house knowledge base needs to combine external and internal data resources. External data includes industry public data (such as central bank policy documents, industry research reports, academic papers) and Internet resources (such as industry forums, news websites, and professional content on social platforms). Internal data comes from banking business practices, etc. This internal data is domain-specific and is the key to customized training. Fine-tuning training data sets can be obtained by preprocessing these data, including: text parsing and cleaning, data classification and annotation, text splitting and chunking, data enhancement and vectorization processing. By removing invalid content, classifying and archiving, and annotating domain labels, the data is structured and high-quality before entering the model. At the same time, natural language demarcation points and sliding window strategies are used to split the text for semantic integrity, and the text is converted into vectors in combination with the Embedding model to provide support for subsequent semantic retrieval and model training.

[0051] Utilizing the above-mentioned fine-tuning training data set, the embodiment of the present application adopts the linear residual adjustment (LoRA, Linear Operator for Rank Adaptation) technology to perform the fine-tuning task. The essence of the LoRA mechanism lies in its design that emphasizes both high efficiency and accuracy: it does not touch the main structure and massive parameters of the large-scale pre-trained model, but cleverly introduces a set of small, learnable low-rank matrices to only fine-tune the key weights in the model. This approach is equivalent to adding a layer of knowledge enhancement filter specifically for financial scenarios on the basis of the original powerful model, which not only avoids the dilution of the original model's capabilities, but also significantly enhances the accuracy and response speed of the model when dealing with specific financial problems.

[0052] Specifically, efficient parameter updates are achieved through low-rank bypass matrices. When initializing bypass parameters, strategies such as random initialization, domain knowledge initialization, or transfer learning initialization can be used. Random initialization ensures the diversity of parameter value ranges, domain knowledge initialization gives parameter domain adaptability through small-scale pre-training, and transfer learning uses parameters in similar fields as initial values ​​to accelerate convergence and enhance initial semantic understanding capabilities.

[0053] During the training process of LoRA, most of the parameters of the pre-trained model are frozen, and only the weights of the bypass matrix are updated to optimize the loss function of the customized task. A separate learning rate strategy is adopted to enable the low-rank matrix to learn domain knowledge more quickly, and the customized performance can be further improved by gradually unfreezing some pre-trained model parameters. The entire process achieves reduced training costs and optimized model performance through efficient parameter adjustment.

[0054] LoRA technology inserts a low-rank bypass matrix into the attention module and adds the output of the bypass matrix to the original output of the pre-trained model to form the final representation. Parameter fusion uses parallel path fusion or weighted fusion to dynamically adjust the proportion of LoRA output in the total model. In different tasks, LoRA parameters can be dynamically activated according to data relevance to ensure adaptation to new domain tasks while retaining the generalization ability of the model.

[0055] Through the fine-tuning implemented by LoRA, the semantic analysis model can not only better understand and respond to complex financial-related inquiries, but also deepen its understanding of financial-specific terminology, business processes, and compliance requirements while maintaining the model's ability to respond to a wide range of topics. This precise and efficient fine-tuning strategy lays a solid foundation for building a conversational AI system that is both versatile and performs well in the financial field. It is expected to play an important role in multiple application scenarios such as customer service, risk assessment, and investment consulting in the future, and further promote the intelligent transformation of financial services.

[0056] After completing the construction of the financial knowledge base and fine-tuning the semantic analysis model, the query task text entered by the user can be processed. The processing flow is described in detail below.

[0057] First, the query task text entered by the user is obtained. When the user types a question in the front-end interactive interface, the front-end system responds immediately. It not only serves as a bridge between the user and the complex logic of the back-end, but also undertakes the task of initially understanding and preprocessing the user input, including removing redundancy, correcting spelling errors, etc., to ensure the accurate transmission of information. Subsequently, the processed query task text is securely transmitted to the back-end server through a secure and efficient communication protocol, such as JSON or RESTful API. This process ensures the integrity of the data and the real-time nature of the interaction.

[0058] After arriving at the backend, the query task text is deeply semantically parsed and vectorized. This process is equivalent to converting natural language text into a mathematical vector that can be understood by the machine. Each vector represents the semantic features of the problem, allowing the machine to understand the essence of the problem beyond the vocabulary level. The task semantics of the query task text is efficiently matched with the knowledge vectors in the pre-built huge financial knowledge base.

[0059] Specifically, in order to improve the relevance and practicality of the query matching results, a two-stage ranking optimization strategy is introduced in the embodiment of the present application. A lightweight model is initially used for rough sorting to quickly eliminate most of the irrelevant items, that is, the summary information of each knowledge vector in the financial knowledge base is preliminarily matched with the query task text to screen out candidate knowledge vectors; then a more complex fine ranking model can be used, based on richer features and deep learning technology, to further screen out the target knowledge vector from the TopN candidate knowledge vectors preliminarily screened out, to ensure that every piece of knowledge returned to the user is highly relevant and high-quality. This strategy effectively balances system efficiency and retrieval quality while ensuring the accuracy of the results through phased processing, providing users with a nearly instant and highly personalized knowledge service experience. Among them, the specific steps for determining the target indication vector are as follows.

[0060] In some embodiments of the present application, determining the second similarity between the query task text and the candidate knowledge vector includes the following steps: performing word segmentation on the query task text and determining the part of speech corresponding to each word obtained after the word segmentation; performing syntactic analysis on the query task text based on the part of speech corresponding to each word to obtain semantic structure information corresponding to the query task text, wherein the semantic structure information is used to characterize the contextual association relationship between each word in the query task text; using a semantic analysis model, based on the semantic structure information, determining the keywords in the query task text, and mapping the keywords to a preset vector space to obtain a task semantic vector corresponding to the query task text; calculating the cosine similarity between the task semantic vector and each knowledge vector in the financial knowledge base, and determining the cosine similarity as the second similarity.

[0061] In some embodiments of the present application, keywords are mapped to a preset vector space to obtain a task semantic vector corresponding to the query task text, including the following steps: determining whether the keyword is a polysemous word, and if the keyword is a polysemous word, determining the meaning of the keyword based on the context information corresponding to the keyword in the query task text to ensure that the meaning of each keyword is clear and unique; determining the synonyms or near-synonyms corresponding to each keyword, and using the synonyms or near-synonyms as extended keywords; based on the keywords and extended keywords, vectorizing the query task text to obtain a task semantic vector.

[0062] Specifically, the semantic analysis model can be used to achieve semantic matching between the user's query task text and the knowledge vector content in the financial knowledge base. By extracting keywords from the query task text and using synonym expansion, polysemy disambiguation and context analysis, the accuracy and semantic consistency of the retrieval results can be ensured. At the same time, with the help of multi-path recall and relevance ranking models, combined with multi-dimensional result fusion such as vector recall and rule matching, content that highly matches the user's intention can be displayed first.

[0063] After determining the financial information corresponding to the query task text, in the answer generation stage, the large language model can be guided by a carefully designed input prompt template (Prompt). Based on the determined financial information, the large language model can be guided to generate query results corresponding to the query task text. The specific steps are as follows.

[0064] In some embodiments of the present application, determining a query result corresponding to a query task text based on the financial information corresponding to the target knowledge vector includes the following steps: determining the task intent corresponding to the query task text, wherein the task intent includes at least one of the following: policy interpretation, operational guidance, and market analysis; obtaining an input prompt template corresponding to the task intent, and adjusting the input prompt template based on the query task text and the financial information corresponding to the target knowledge vector to obtain a model input text, wherein the input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the query task text, and to instruct the large language model on the knowledge data based on which the query task text is analyzed; using the large language model, based on the model input text, generating a query result corresponding to the query task text.

[0065] Specifically, constructing an effective prompt is a bridge between user queries and high-quality answers. In this process, first, the search results obtained from the financial knowledge base (i.e., the financial information corresponding to the target knowledge vector) are integrated. These results represent the knowledge fragments most relevant to the user's question. Next, the core intent of the user's original question is analyzed, and key words and potential context clues are extracted to ensure that the full picture of the user's inquiry can be accurately reflected when constructing the prompt.

[0066] The search results (i.e., the financial information corresponding to the target knowledge vector) are then cleverly combined with the question to form a coherent, context-rich input instruction. This prompt should not only contain all the key elements of the question, but also guide the large language model to understand the key points of each search result (i.e., the financial information corresponding to the target knowledge vector) and its relevance to the question. For example, the prompt is like: "Based on the following information found in the knowledge base [summary of search result 1], [summary of search result 2]... Please answer the user's question about [query task text] in detail, requiring the answer to be both accurate and comprehensive, and to be comprehensively analyzed in combination with all relevant information.

[0067] When the large language model is called, the constructed prompt is used as a stimulus to stimulate the model's deep thinking potential. With its powerful semantic understanding and generation capabilities, the large language model can not only parse the complex logic and subtle meanings in the prompt, but also further reason, synthesize and innovate on this basis, and finally generate a high-quality answer that not only fits the core of the question but also incorporates new search knowledge. This process not only greatly improves the accuracy and richness of the answer, but also provides users with a more intelligent and humane interaction method that goes beyond the traditional search experience.

[0068] According to an embodiment of the present application, an embodiment of a financial information query device based on artificial intelligence is also provided. Figure 3 is a schematic diagram of the structure of a financial information query device based on artificial intelligence provided according to an embodiment of the present application. Figure 3 The device comprises:

[0069] The task acquisition module 30 is used to acquire a query task text, wherein the task text is used to represent the financial problem that the user plans to query;

[0070] The first matching module 32 is used to determine a first similarity between the query task text and the summary information of each knowledge vector in the financial knowledge base, and determine a candidate knowledge vector from the knowledge vectors in the financial knowledge base based on the first similarity, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to represent the semantic theme of the financial information corresponding to the knowledge vector;

[0071] The second matching module 34 is used to determine a second similarity between the query task text and the candidate knowledge vectors, and determine a target knowledge vector from the candidate knowledge vectors based on the second similarity;

[0072] The answer generation module 36 determines the query result corresponding to the query task text according to the financial information corresponding to the target knowledge vector, wherein the query result is used to answer the financial question corresponding to the query task text.

[0073] Optionally, the construction process of the financial knowledge base includes the following steps: according to a preset collection cycle, obtaining the original text of information in the information source, wherein the original text of the information includes at least one of the following: financial policy consultation, financial product introduction; processing the original text of the information into blocks to obtain multiple initial text blocks; using a semantic analysis model to determine whether the semantic information expressed by the initial text block is complete, and if the semantic information expressed by the initial text block is complete, determining the initial text block as the target text block; if the semantic information expressed by the initial text block is incomplete, determining other initial text blocks whose semantic similarity with the initial text block exceeds a preset similarity threshold, and based on the other initial text blocks, semantically completing the initial text block with incomplete semantic information to obtain the target text block; using a semantic analysis model to map the target text block to a preset vector space to obtain a knowledge vector, and determining summary information corresponding to the knowledge vector.

[0074] Optionally, the original text of the information is divided into blocks to obtain multiple initial text blocks, including: using regular expressions to identify delimiters in the original text of the information, and dividing the original text of the information into multiple sentence fragments based on the delimiters, wherein the delimiters include: punctuation marks, line breaks; using a sliding window to gradually slide backward from the beginning of the sentence fragment, and after each slide, judging whether to segment the sentence fragment at the position of the sliding window based on the sentence content in the sliding window; after the sliding window reaches the end of the sentence fragment, obtaining the initial text block corresponding to the sentence fragment, wherein, in the process of sliding the sliding window, there is a partial overlap between the framed area of ​​the sliding window after each slide and the framed area of ​​the sliding window before sliding.

[0075] Optionally, the training steps of the semantic analysis model include: obtaining a fine-tuning training data set, wherein the fine-tuning training data set contains multiple knowledge information related to the financial field, and category labels corresponding to each piece of knowledge information; obtaining a pre-trained model, and adding a low-rank bypass to the pre-trained model, wherein the input dimension and input dimension of the low-rank bypass are consistent with the original input dimension and output dimension of the pre-trained model, and the low-rank bypass is used to reduce the parameter adjustment amount of the pre-trained model during the training process of adapting to tasks in financial scenarios; based on the fine-tuning training data set, training the pre-trained model after adding the low-rank bypass to obtain a semantic analysis model, wherein the training process is used to adjust the matrix parameters in the low-rank bypass, while the original model parameters of the pre-trained model before adding the low-rank bypass remain unchanged.

[0076] Optionally, determining the second similarity between the query task text and the candidate knowledge vector includes: performing word segmentation on the query task text and determining the part of speech corresponding to each word obtained after the word segmentation; performing syntactic analysis on the query task text based on the part of speech corresponding to each word to obtain semantic structure information corresponding to the query task text, wherein the semantic structure information is used to characterize the contextual association relationship between each word in the query task text; using a semantic analysis model, based on the semantic structure information, determining keywords in the query task text, and mapping the keywords to a preset vector space to obtain a task semantic vector corresponding to the query task text; calculating the cosine similarity between the task semantic vector and each knowledge vector in the financial knowledge base, and determining the cosine similarity as the second similarity.

[0077] Optionally, the keywords are mapped to a preset vector space to obtain a task semantic vector corresponding to the query task text, including: determining whether the keyword is a polysemous word, and if the keyword is a polysemous word, determining the meaning of the keyword based on the context information corresponding to the keyword in the query task text to ensure that the meaning of each keyword is clear and unique; determining the synonyms or near-synonyms corresponding to each keyword, and using the synonyms or near-synonyms as extended keywords; based on the keywords and extended keywords, vectorizing the query task text to obtain a task semantic vector.

[0078] Optionally, determining the query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector includes: determining the task intent corresponding to the query task text, wherein the task intent includes at least one of the following: policy interpretation, operational guidance, market analysis; obtaining an input prompt template corresponding to the task intent, and adjusting the input prompt template based on the query task text and the financial information corresponding to the target knowledge vector to obtain a model input text, wherein the input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the query task text, and to instruct the large language model on the knowledge data based on which the query task text is analyzed; and using the large language model to generate a query result corresponding to the query task text based on the model input text.

[0079] It should be noted that the various modules in the above-mentioned artificial intelligence-based financial information query 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.

[0080] It should be noted that the financial information query device based on artificial intelligence provided in this embodiment can be used to execute Figure 2The artificial intelligence-based financial information query method shown, therefore, the relevant explanations and descriptions of the above-mentioned artificial intelligence-based financial information query method are also applicable to the embodiments of the present application and will not be repeated here.

[0081] An 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 artificial intelligence-based financial information query method by running the computer program: obtaining a query task text, wherein the task text is used to characterize the financial problem that the user plans to query; determining a first similarity between the query task text and summary information of each knowledge vector in the financial knowledge base, and determining a candidate knowledge vector in the knowledge vector in the financial knowledge base based on the first similarity, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to characterize the semantic topic of the financial information corresponding to the knowledge vector; determining a second similarity between the query task text and the candidate knowledge vector, and determining a target knowledge vector in the candidate knowledge vector based on the second similarity; determining a query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, wherein the query result is used to answer the financial problem corresponding to the query task text.

[0082] 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 artificial intelligence-based financial information query method described in each embodiment of the present application: obtaining a query task text, wherein the task text is used to characterize the financial problem that the user plans to query; determining a first similarity between the query task text and summary information of each knowledge vector in the financial knowledge base, and determining a candidate knowledge vector in the knowledge vector in the financial knowledge base based on the first similarity, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to characterize the semantic theme of the financial information corresponding to the knowledge vector; determining a second similarity between the query task text and the candidate knowledge vector, and determining a target knowledge vector in the candidate knowledge vector based on the second similarity; determining a query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, wherein the query result is used to answer the financial problem corresponding to the query task text.

[0083] 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.

[0084] 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.

[0085] 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.

[0086] 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.

[0087] 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.

[0088] 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.

[0089] 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 information query method based on artificial intelligence, characterized in that: include: Acquire a query task text, wherein the query task text is used to characterize the financial question that the user plans to query; Determine a first similarity between the query task text and summary information of each knowledge vector in the financial knowledge base, and determine a candidate knowledge vector from the knowledge vectors in the financial knowledge base based on the first similarity, wherein each knowledge vector corresponds to a piece of financial information, and the summary information is used to characterize a semantic topic of the financial information corresponding to the knowledge vector; The construction process of the financial knowledge base includes the following steps: obtaining the original text of information in the information source according to a preset collection cycle, wherein the original text of information includes at least one of the following: financial policy consultation, financial product introduction; processing the original text of information into blocks to obtain multiple initial text blocks; using a semantic analysis model to determine whether the semantic information expressed by the initial text block is complete, and if the semantic information expressed by the initial text block is complete, determining the initial text block as a target text block; if the semantic information expressed by the initial text block is incomplete, determining other initial text blocks whose semantic similarity with the initial text block exceeds a preset similarity threshold, and based on the other initial text blocks, semantically completing the initial text block with incomplete semantic information to obtain a target text block; using the semantic analysis model to map the target text block to a preset vector space to obtain the knowledge vector, and determining the summary information corresponding to the knowledge vector; Wherein, the original text of the information is divided into blocks to obtain a plurality of initial text blocks, including: using a regular expression to identify the separators in the original text of the information, and dividing the original text of the information into a plurality of sentence fragments according to the separators, wherein the separators include punctuation marks and line breaks; using a sliding window to gradually slide backward from the beginning of the sentence fragment, and after each sliding, judging whether to segment the sentence fragment at the position of the sliding window according to the sentence content in the sliding window; after the sliding window reaches the end of the sentence fragment, obtaining the initial text block corresponding to the sentence fragment, wherein, in the process of sliding the sliding window, there is a partial overlap between the framed area of ​​the sliding window after each sliding and the framed area of ​​the sliding window before sliding; wherein, The training steps of the semantic analysis model include: obtaining a fine-tuning training data set, wherein the fine-tuning training data set contains multiple knowledge information related to the financial field and category labels corresponding to each of the knowledge information; obtaining a pre-trained model and adding a low-rank bypass to the pre-trained model, wherein the input dimension and output dimension of the low-rank bypass are consistent with the original input dimension and output dimension of the pre-trained model, and the low-rank bypass is used to reduce the parameter adjustment amount of the pre-trained model during the training process of adapting to tasks in financial scenarios; based on the fine-tuning training data set, training the pre-trained model after adding the low-rank bypass to obtain the semantic analysis model, wherein the training process is used to adjust the matrix parameters in the low-rank bypass, while the original model parameters of the pre-trained model before adding the low-rank bypass remain unchanged; Determining a second similarity between the query task text and the candidate knowledge vectors, and determining a target knowledge vector among the candidate knowledge vectors based on the second similarity; Determine the query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, including: determine the task intent corresponding to the query task text, wherein the task intent includes at least one of the following: policy interpretation, operational guidance, market analysis; obtain an input prompt template corresponding to the task intent, and adjust the input prompt template based on the query task text and the financial information corresponding to the target knowledge vector to obtain a model input text, wherein the input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the query task text, and to instruct the large language model to analyze the query task text based on the knowledge data; use the large language model, based on the model input text, to generate the query result corresponding to the query task text, wherein the query result is used to answer the financial question corresponding to the query task text.

2. The financial information query method based on artificial intelligence according to claim 1, characterized in that: Determining the second similarity between the query task text and the candidate knowledge vector includes: Performing word segmentation on the query task text and determining the part of speech corresponding to each word obtained after the word segmentation; According to the part of speech corresponding to each word, the query task text is subjected to syntactic analysis to obtain semantic structure information corresponding to the query task text, wherein the semantic structure information is used to characterize the contextual association relationship between the words in the query task text; Using a semantic analysis model, according to the semantic structure information, keywords in the query task text are determined, and the keywords are mapped to a preset vector space to obtain a task semantic vector corresponding to the query task text; The cosine similarity between the task semantic vector and each of the knowledge vectors in the financial knowledge base is calculated, and the cosine similarity is determined as the second similarity.

3. The financial information query method based on artificial intelligence according to claim 2 is characterized in that: Mapping the keywords to a preset vector space to obtain a task semantic vector corresponding to the query task text includes: Determine whether the keyword is a polysemous word, and if the keyword is a polysemous word, determine the meaning of the keyword according to the context information corresponding to the keyword in the query task text to ensure that the meaning of each keyword is clear and unique; Determine the synonyms or near-synonyms corresponding to each of the keywords, and use the synonyms or near-synonyms as extended keywords; The query task text is vectorized based on the keywords and the extended keywords to obtain the task semantic vector.

4. A financial information query device based on artificial intelligence, characterized in that: include: A task acquisition module, used to acquire a query task text, wherein the query task text is used to characterize the financial problem that the user plans to query; a first matching module, configured to determine a first similarity between the query task text and summary information of each knowledge vector in the financial knowledge base, and to determine a candidate knowledge vector from the knowledge vectors in the financial knowledge base based on the first similarity, wherein each of the knowledge vectors corresponds to a piece of financial information, and the summary information is used to characterize a semantic topic of the financial information corresponding to the knowledge vector; The construction process of the financial knowledge base includes the following steps: obtaining the original text of information in the information source according to a preset collection cycle, wherein the original text of information includes at least one of the following: financial policy consultation, financial product introduction; processing the original text of information into blocks to obtain multiple initial text blocks; using a semantic analysis model to determine whether the semantic information expressed by the initial text block is complete, and if the semantic information expressed by the initial text block is complete, determining the initial text block as a target text block; if the semantic information expressed by the initial text block is incomplete, determining other initial text blocks whose semantic similarity with the initial text block exceeds a preset similarity threshold, and based on the other initial text blocks, semantically completing the initial text block with incomplete semantic information to obtain a target text block; using the semantic analysis model to map the target text block to a preset vector space to obtain the knowledge vector, and determining the summary information corresponding to the knowledge vector; Wherein, the original text of the information is divided into blocks to obtain a plurality of initial text blocks, including: using regular expressions to identify separators in the original text of the information, and dividing the original text of the information into a plurality of sentence fragments according to the separators, wherein the separators include punctuation marks and line breaks; using a sliding window to gradually slide backward from the beginning of the sentence fragment, and after each sliding, judging whether to segment the sentence fragment at the position of the sliding window according to the sentence content in the sliding window; after the sliding window reaches the end of the sentence fragment, the initial text block corresponding to the sentence fragment is obtained, wherein, in the process of sliding the sliding window, there is a partial overlap between the framed area of ​​the sliding window after each sliding and the framed area of ​​the sliding window before sliding; wherein the training step of the semantic analysis model includes: obtaining a fine-tuning training data set, wherein the fine-tuning training data set The data set contains multiple knowledge information related to the financial field, and the category labels corresponding to each of the knowledge information; obtain a pre-trained model, and add a low-rank bypass to the pre-trained model, wherein the input dimension and output dimension of the low-rank bypass are consistent with the original input dimension and output dimension of the pre-trained model, and the low-rank bypass is used to reduce the parameter adjustment amount of the pre-trained model during the training process of adapting to tasks in financial scenarios; according to the fine-tuning training data set, the pre-trained model after adding the low-rank bypass is trained to obtain the semantic analysis model, wherein the training process is used to adjust the matrix parameters in the low-rank bypass, and the original model parameters of the pre-trained model before adding the low-rank bypass remain unchanged; a second matching module is used to determine the second similarity between the query task text and the candidate knowledge vector, and determine the target knowledge vector in the candidate knowledge vector based on the second similarity; The answer generation module is used to determine the query result corresponding to the query task text based on the financial information corresponding to the target knowledge vector, including: determining the task intent corresponding to the query task text, wherein the task intent includes at least one of the following: policy interpretation, operational guidance, and market analysis; obtaining an input prompt template corresponding to the task intent, and adjusting the input prompt template based on the query task text and the financial information corresponding to the target knowledge vector to obtain a model input text, wherein the input prompt template is used to guide the large language model to clarify the task to be completed corresponding to the query task text, and to instruct the large language model to analyze the query task text based on the knowledge data; using the large language model, based on the model input text, to generate the query result corresponding to the query task text, wherein the query result is used to answer the financial question corresponding to the query task text.

5. 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 artificial intelligence-based financial information query method described in any one of claims 1 to 3 by running the computer program.

6. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the artificial intelligence-based financial information query method described in any one of claims 1 to 3 are implemented.

Images