Business question answering method and system based on large language model and scene label

Abstract

The invention belongs to the technical field of intelligent questioning and answering, and particularly discloses a business questioning and answering method and system based on a large language model and a scene label, corresponding knowledge base question sets are associated through business question expansion matching, and then answers are called from a knowledge base to determine corresponding data blocks and scene labels. Then, through scene label matching and scene consistency evaluation, a database with high scene consistency is selected to be combined with user questions to be input into a large language model to achieve intelligent question and answer. The business question and answer requirements of the user can be effectively met, and the user does not need to directly query professional knowledge; according to the method, questions proposed by users can be analyzed more comprehensively, questioning intentions and question scene information are fully covered, and question and answer feedback results are more targeted and have higher relevance with business questions of the users; and the question and answer result is simpler and more effective.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent question answering, and specifically relates to a business question answering method and system based on a large language model and scenario tags. Background Art

[0002] The intelligent question-answering system organizes the accumulated disordered corpus information in an orderly and scientific manner and establishes a knowledge-based classification model. These classification models can guide the newly added corpus consultation and service information, save human resources, and improve the automation of information processing. The intelligent question-answering system accurately locates the question knowledge required by users in the form of one question and one answer, and provides users with personalized information services by interacting with users. There are three main implementation methods of the existing intelligent question-answering system:

[0003] The first is intelligent question-answering technology based on traditional search technology. According to the string matching or semantic similarity of the user's input question, "question-answer" matching is directly performed from the knowledge base.

[0004] The second is the intelligent question-answering technology for knowledge graphs. Knowledge graphs are the main form of knowledge bases, and the intelligent question-answering technology system converts the user's natural language input conditions into query statements and reasoning formulas for graph databases.

[0005] The third is intelligent question-answering technology based on a large language model. Using the Retrieval Augmented Generation (RAG) technology framework, the system first queries the vector database for data blocks that are semantically related to the user's input question. Then, the data blocks are spliced ​​into the prompt word input conditions of the large language model. Finally, the large language model is used to combine the question and the data block content for dialogue answer feedback.

[0006] The existing intelligent question-answering technology still has the following main defects:

[0007] 1 Intelligent question-answering technology based on traditional search technology has a single form of expression, limited ability to answer questions, and low question matching degree.

[0008] 2 Intelligent question-answering technology based on knowledge graphs mainly solves the needs of factual questions, but it is difficult to solve personalized question-answering needs. The knowledge graph has limited ability to express complex knowledge.

[0009] 3 Intelligent question-answering technology based on large language models has flexible expression forms and can give reasonable responses regardless of whether relevant data blocks are matched. However, the answer results of large language models depend to a large extent on the query effect of data blocks. Directly querying data blocks based on user questions will result in low correlation between the data results and the user's question requirements, and cannot fully cover the scenario information of the user's question. Therefore, the answer content is too redundant or not very targeted.

[0010] 4 The existing intelligent question-answering system mainly adopts a technical architecture that targets knowledge-based questioning needs. In actual business scenarios, users generally do not directly ask questions about professional knowledge (such as "How to apply for a publication business license?" "What industrial and commercial qualifications should be required to sell books?" "What are the requirements for business premises for offline book sales?" etc.), but ask questions in the form of business needs (such as "How to run a bookstore?"). For business problems, the logical compatibility of existing question-answering technologies is poor, and the accuracy of returned results is insufficient. Summary of the invention

[0011] The purpose of the present invention is to provide a business question answering method and system based on a large language model and scenario tags to solve the above-mentioned problems existing in the prior art.

[0012] In order to achieve the above object, the present invention adopts the following technical solutions:

[0013] First, we provide a business question-answering method based on a large language model and scenario tags, including:

[0014] Obtain business questions and supplementary explanations of business questions input by users;

[0015] Import business problems into the preset large language model for problem expansion and obtain the corresponding expanded problem information;

[0016] Based on the expanded question information, a corresponding associated question set is matched in the knowledge base, where the associated question set includes a number of knowledge base associated questions;

[0017] In response to the user's operation instruction, a related question subset including a plurality of question items is selected from the related question set, and a knowledge base related answer corresponding to each question item in the related question subset is retrieved from the knowledge base as a reference answer to the question corresponding to the related question subset;

[0018] Determine the data blocks contained in the reference answer of each question in the knowledge base, as well as the set of scene labels associated with the data blocks;

[0019] For each question item in the associated question subset, the scenario label set of each corresponding data block and the supplementary description of the business question are evaluated for scenario consistency to obtain the scenario consistency score of each data block;

[0020] Several data blocks with the highest scenario consistency scores are selected as the target data blocks for each question item in the corresponding associated question subset, and the target data blocks, business questions, and supplementary explanations of the business questions corresponding to each question item in the associated question subset are combined and imported into the preset large language model for natural language processing to obtain the corresponding business answers respectively. Finally, multiple business answers are integrated to obtain a unified answer output result.

[0021] In a possible design, matching and determining a corresponding set of related questions in a knowledge base based on the expanded question information includes:

[0022] Traversing a list of questions in a knowledge base, the list of questions comprising a number of candidate knowledge base associated questions;

[0023] If the trigger word label of a candidate knowledge base-related question in the question list is in the expanded question information, the candidate knowledge base-related question will be included in the related question set. In the knowledge base, each candidate knowledge base-related question contains one or more trigger word labels, which are confirmed in the process of building the knowledge base. When entering each candidate knowledge base-related question, the system automatically confirms 1 to N candidate trigger words through word segmentation, and the user selects and confirms them, or manually supplements and confirms them to obtain the corresponding trigger word label.

[0024] In a possible design, selecting a subset of related questions from a set of related questions in response to an operation instruction of a user includes:

[0025] Matching semantic similarity between each knowledge base related question and business question in the related question set, and sorting each knowledge base related question according to the semantic similarity between each knowledge base related question and business question, to obtain a sorted related question set;

[0026] The sorted related question set is output to the user, and in response to the user's operation instruction, a number of knowledge base related questions are selected from the sorted related question set to form a related question subset.

[0027] In a possible design, the data blocks included in the reference answer to the question and the scene label set associated with each data block include:

[0028] Determine the data blocks included in the reference answer to the question and several scene labels associated with each data block;

[0029] The scene labels associated with each data block are aggregated to obtain a scene label set for each data block.

[0030] In a possible design, the scenario consistency evaluation of the scenario tag set of each data block and the supplementary description of the business problem is performed to obtain the scenario consistency score of each data block, including:

[0031] Traversing the scene label set of each data block, and determining whether each scene label in the scene label set is included in the supplementary description of the business problem;

[0032] If a scene label in the scene label set of a data block is included in the supplementary description of the business problem, the scene consistency score of the data block will be increased by one point;

[0033] The final scene consistency score of each data block is obtained by statistics.

[0034] In a possible design, if a scene label in the scene label set of the data block is not included in the supplementary description of the business problem, the first semantic vector of the corresponding scene label and the second semantic vector of the supplementary description are determined, and the first semantic vector of the corresponding scene label and the second semantic vector of the supplementary description are substituted into a preset semantic similarity scoring formula for calculation to obtain a corresponding semantic similarity score, and the semantic similarity score is statistically added to the scene consistency score of the corresponding data block, and the semantic similarity scoring formula is:

[0035]

[0036] The first semantic vector represents the corresponding scene label, and the second semantic vector represents the supplementary description.

[0037] In a possible design, selecting several data blocks with the highest scene consistency scores as target data blocks includes: selecting three data blocks with the highest scene consistency scores as target data blocks.

[0038] Secondly, a business question-answering system based on a large language model and scene tags is provided, including a question acquisition unit, a question expansion unit, an association matching unit, a knowledge screening unit, a scene annotation unit, a scene matching unit and an intelligent question-answering unit, wherein:

[0039] A question acquisition unit, used to acquire the business question input by the user and the supplementary explanation of the business question;

[0040] The question expansion unit is used to import business questions into the preset large language model for question expansion and obtain the corresponding expanded question information;

[0041] An associated matching unit, used for matching and determining a corresponding associated question set in the knowledge base based on the expanded question information, wherein the associated question set includes a number of knowledge base associated questions;

[0042] The knowledge screening unit is used to select a related question subset containing a plurality of question items from the related question set in response to an operation instruction of a user, and retrieve a knowledge base related answer corresponding to each question item in the related question subset from the knowledge base as a reference answer to the question corresponding to the related question subset;

[0043] A scene annotation unit is used to determine the data blocks contained in the reference answer of each question in the knowledge base, as well as the scene label set associated with each data block;

[0044] A scenario matching unit is used to evaluate the scenario consistency of the scenario label set of each data block and the supplementary description of the business problem for each question item in the associated question subset, so as to obtain the scenario consistency score of each data block;

[0045] The intelligent question-answering unit is used to select several data blocks with the highest scenario consistency scores as the target data blocks for each question item in the corresponding associated question subset, and import the target data blocks, business questions and supplementary explanations of the business questions corresponding to each question item in the associated question subset into a preset large language model for natural language processing, to obtain the corresponding business answers respectively, and finally to integrate multiple business answers to obtain a unified answer output result.

[0046] Third, we provide a business question-answering system based on a large language model and scenario tags, including:

[0047] A memory for storing instructions;

[0048] A processor is used to read the instructions stored in the memory and execute any one of the methods described in the first aspect according to the instructions.

[0049] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored on the computer-readable storage medium, and when the instructions are executed on a computer, the computer executes any one of the methods described in the first aspect. In addition, a computer program product is provided, and when the computer program product is executed on a computer, the computer executes any one of the methods described in the first aspect.

[0050] Beneficial effects: The present invention can effectively solve the business question and answer needs of users without the need for users to directly query professional knowledge; it can perform a more comprehensive analysis of the questions raised by users, fully covering the question intention and question scenario information; it makes the question and answer feedback results more relevant to the user's business questions, and provides the user with a better interactive experience; and, through scene label matching, it can improve the reach of business question and answer results, making the result content more targeted; at the same time, it can solve the knowledge base query task for long question and answer content, making the question and answer results more concise and effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0052] Figure 1 This is a schematic diagram of the steps of the method in Example 1 of the present invention;

[0053] Figure 2 This is a schematic diagram of the structure of the system in Example 2 of the present invention;

[0054] Figure 3 Schematic diagram of the system structure in Example 3 of the present invention. DETAILED DESCRIPTION

[0055] It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention. The specific structures and functional details disclosed herein are only used to describe the exemplary embodiments of the present invention. However, the present invention can be embodied in many alternative forms, and it should not be understood that the present invention is limited to the embodiments set forth herein.

[0056] It should be understood that, unless otherwise clearly specified and limited, the term "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments can be understood according to specific circumstances.

[0057] In the following description, certain details are provided to facilitate a complete understanding of the example embodiments. However, it will be appreciated by those of ordinary skill in the art that the example embodiments may be implemented without these certain details. For example, devices may be shown in block diagrams to avoid obscuring the examples with unnecessary details. In other embodiments, well-known processes, structures, and techniques may not be shown in unnecessary detail to avoid obscuring the embodiments.

[0058] Embodiment 1:

[0059] This embodiment provides a business question-answering method based on a large language model and scene tags, which can be applied to corresponding question-answering servers, such as Figure 1 As shown, the method comprises the following steps:

[0060] S1. Obtain the business problem input by the user and the supplementary explanation of the business problem.

[0061] In specific implementation, the user can input the corresponding business question and supplementary explanation for the business question to the question-answering server. The supplementary explanation is the relevant description / explanation text content for the business question, including the relevant scene / background limitation information of the question, etc. The question-answering server obtains the business question and supplementary explanation input by the user and performs subsequent analysis and processing.

[0062] S2. Import the business problem into the preset large language model for problem expansion, and obtain the corresponding expanded problem information.

[0063] In specific implementation, the question-answering server can import business questions into the preset large language model for question expansion query to obtain the question information after the business question is expanded. By expanding and strengthening the business questions, more question clues can be provided to the knowledge question-answering engine of the model in the future.

[0064] S3. Based on the expanded question information, a corresponding associated question set is matched in the knowledge base, where the associated question set includes several knowledge base associated questions.

[0065] In specific implementation, the question-answering server traverses and determines the trigger word tags (i.e., the corresponding keywords) contained in each expanded question in the expanded question information, and matches and determines the knowledge base associated questions corresponding to each trigger word tag in the knowledge base, and then uses the knowledge base associated questions corresponding to each trigger word tag to form an associated question set. The knowledge base is pre-configured with a number of candidate knowledge base associated questions and knowledge base associated answers corresponding to each candidate knowledge base associated question, and each knowledge base associated question is associated with a corresponding trigger word tag (i.e., the corresponding keyword). In the knowledge base, each candidate knowledge base associated question contains one or more trigger word tags, which are confirmed in the process of building the knowledge base. When entering each candidate knowledge base associated question, the system automatically confirms 1 to N candidate trigger words through word segmentation, which are selected and confirmed by the user, or manually supplemented and confirmed to obtain the corresponding trigger word tag. The trigger word tag is on the question, and the subsequent scenario tag is on the data block of the question answer.

[0066] S4. In response to the user's operation instruction, a related question subset including several question items is selected from the related question set, and the knowledge base associated answers corresponding to each question item in the related question subset are retrieved from the knowledge base as reference answers to the questions corresponding to the related question subset.

[0067] In specific implementation, the question-and-answer server may match the semantic similarity of each knowledge base-related question in the associated question set with the business question, and sort each knowledge base-related question according to the semantic similarity between each knowledge base-related question and the business question to obtain a sorted associated question set. The sorted associated question set is then output and displayed to the user so that the user can view which knowledge base-related questions match the knowledge questions of interest through human-computer interaction and select the corresponding knowledge base-related questions. The question-and-answer server may select a number of knowledge base-related questions from the sorted associated question set to form an associated question subset in response to the user's operation instruction, and retrieve the knowledge base-related answers corresponding to each knowledge base-related question in the associated question subset from the knowledge base as reference answers to the questions.

[0068] S5. Determine the data blocks contained in the reference answer of each question in the knowledge base, and the set of scene labels associated with the data blocks.

[0069] In specific implementation, the question-answering server first determines the data blocks contained in the reference answer to the question and the several scene tags associated with each data block. The reference answer to the question is usually a long answer content structure. Each reference answer to the question includes at least one data block, and each data block is associated with multiple scene tags (i.e., the corresponding scene keywords). A nested content structure is formed according to the number of scene tags. In actual business links, references to knowledge such as operating instructions, method suggestions, management systems, laws and regulations are closely related to specific segmented scene tags. For example, the corresponding question 1 in the knowledge base is recorded as Q1. The reference answer 1 to the question is divided into paragraphs and contains multiple data blocks. Each data block consists of several scene tags. The corresponding format example is as follows:

[0070] Q1:

[0071] A1-1: {T1}

[0072] A1-2: {T1, T2, T3}

[0073] A1-3: {T1, T2, T4}

[0074] A1-4: {T1, T5}

[0075] A1-5: {T1, T5, T6}

[0076] The data blocks included in the reference answer 1 are A1-1, A1-2, A1-3, A1-4 and A1-5, and each data block is associated with multiple scene tags such as T1, T2, T3, etc. The scene tags associated with each data block can be aggregated to obtain a scene tag set of each data block such as {T1, T2, T3}, etc.

[0077] S6. For each question item in the associated question subset, the scenario label set of each corresponding data block and the supplementary description of the business problem are evaluated for scenario consistency to obtain the scenario consistency score of each data block.

[0078] In specific implementation, the question-answering server can traverse the scene tag set of each data block to determine whether each scene tag (i.e., the corresponding scene keyword) in its scene tag set is included in the supplementary description of the business question. If a scene tag in the scene tag set of a data block is included in the supplementary description of the business question, the scene consistency score of the data block is increased by one point, and then the final scene consistency score of each data block is obtained by statistics. If a scene label in the scene label set of the data block is not included in the supplementary explanation of the business problem, the first semantic vector of the corresponding scene label and the second semantic vector of the supplementary explanation are determined, and the first semantic vector of the corresponding scene label and the second semantic vector of the supplementary explanation are substituted into the preset semantic similarity scoring formula for calculation (taking cosine similarity as an example, it can also be replaced by any similarity measurement function with a value in [0, 1]. For example, the semantic embedding models that can be used include Open AI's text-embedding-ada-002, embedding text2vec-base-chinese, BERT series models, FastText, etc.), and the corresponding semantic similarity score is obtained, and the semantic similarity score is counted into the scene consistency score of the corresponding data block, and then the final scene consistency score of each data block is obtained by statistics. The semantic similarity scoring formula is:

[0079]

[0080] The first semantic vector represents the corresponding scene label, and the second semantic vector represents the supplementary description.

[0081] S7. Select several data blocks with the highest scenario consistency scores as the target data blocks for each question item in the corresponding associated question subset, and import the target data blocks, business questions, and supplementary explanations of the business questions corresponding to each question item in the associated question subset into the preset large language model for natural language processing, and obtain the corresponding business answers respectively. Finally, merge multiple business answers to obtain a unified answer output result.

[0082] During the specific implementation, the question-and-answer server selects several data blocks with the highest scenario consistency scores, such as the three data blocks with the highest scenario consistency scores, as the target data blocks corresponding to each question item in the associated question subset, and then imports the target data blocks corresponding to each question item in the associated question subset, the business question, and the supplementary explanation of the business question into the preset large language model for natural language processing to obtain the corresponding business answers respectively, and finally integrates multiple business answers to obtain a unified answer output result.

[0083] The method of this embodiment can effectively solve the user's business question and answer needs without the user directly querying professional knowledge; it can perform a more comprehensive analysis of the questions raised by the user, fully covering the question intention and question scenario information; it makes the question and answer feedback results more relevant to the user's business questions, and provides the user with a better interactive experience; and, through scene label matching, it can improve the reach of business question and answer results and make the result content more targeted; at the same time, it can solve the knowledge base query task for long question and answer content, making the question and answer results more concise and effective.

[0084] Embodiment 2:

[0085] This embodiment provides a business question answering system based on a large language model and scene tags, such as Figure 2 As shown, it includes a question acquisition unit, a question expansion unit, an association matching unit, a knowledge screening unit, a scene annotation unit, a scene matching unit and an intelligent question answering unit, wherein:

[0086] A question acquisition unit, used to acquire the business question input by the user and the supplementary explanation of the business question;

[0087] The question expansion unit is used to import business questions into the preset large language model for question expansion and obtain the corresponding expanded question information;

[0088] An associated matching unit, used for matching and determining a corresponding associated question set in the knowledge base based on the expanded question information, wherein the associated question set includes a number of knowledge base associated questions;

[0089] The knowledge screening unit is used to select a related question subset containing a plurality of question items from the related question set in response to an operation instruction of a user, and retrieve a knowledge base related answer corresponding to each question item in the related question subset from the knowledge base as a reference answer to the question corresponding to the related question subset;

[0090] A scene annotation unit is used to determine the data blocks contained in the reference answer of each question in the knowledge base, as well as the scene label set associated with each data block;

[0091] A scenario matching unit is used to evaluate the scenario consistency of the scenario label set of each data block and the supplementary description of the business problem for each question item in the associated question subset, so as to obtain the scenario consistency score of each data block;

[0092] The intelligent question-answering unit is used to select several data blocks with the highest scenario consistency scores as the target data blocks for each question item in the corresponding associated question subset, and import the target data blocks, business questions and supplementary explanations of the business questions corresponding to each question item in the associated question subset into a preset large language model for natural language processing, to obtain the corresponding business answers respectively, and finally to integrate multiple business answers to obtain a unified answer output result.

[0093] Embodiment 3:

[0094] This embodiment provides a business question answering system based on a large language model and scene tags, such as Figure 3 As shown, at the hardware level, it includes:

[0095] Data interface, used to establish data connection between the processor and the external data terminal;

[0096] A memory for storing instructions;

[0097] A processor is used to read the instructions stored in the memory and execute the business question answering method based on a large language model and scene labels in Example 1 according to the instructions.

[0098] Optionally, the system further includes an internal bus, through which the processor, the memory and the data interface can be interconnected, and the internal bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0099] The memory may include, but is not limited to, random access memory (RAM), read only memory (ROM), flash memory, first input first output (FIFO) and / or first in last out (FILO) memory, etc. The processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0100] Embodiment 4:

[0101] This embodiment provides a computer-readable storage medium, on which instructions are stored, and when the instructions are executed on a computer, the computer executes the business question answering method based on a large language model and scene labels in Embodiment 1. The computer-readable storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, an optical disk, a hard disk, a flash memory, a USB flash drive, and / or a memory stick, etc., and the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0102] This embodiment also provides a computer program product, which, when executed on a computer, executes the business question answering method based on a large language model and scene tags in Embodiment 1. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0103] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A business question answering method based on a large language model and scene labels, characterized in that: include: Obtain business questions and supplementary explanations of business questions input by users; Import business problems into the preset large language model for problem expansion and obtain the corresponding expanded problem information; Based on the expanded question information, a corresponding associated question set is matched in the knowledge base, where the associated question set includes a number of knowledge base associated questions; In response to the user's operation instruction, a related question subset including a plurality of question items is selected from the related question set, and a knowledge base related answer corresponding to each question item in the related question subset is retrieved from the knowledge base as a reference answer to the question corresponding to the related question subset; Determine the data blocks contained in the reference answer of each question in the knowledge base, as well as the set of scene labels associated with the data blocks; For each question item in the associated question subset, the scenario label set of each corresponding data block and the supplementary description of the business question are evaluated for scenario consistency to obtain the scenario consistency score of each data block; Several data blocks with the highest scenario consistency scores are selected as the target data blocks for each question item in the corresponding associated question subset, and the target data blocks, business questions, and supplementary explanations of the business questions corresponding to each question item in the associated question subset are combined and imported into the preset large language model for natural language processing to obtain the corresponding business answers respectively. Finally, multiple business answers are integrated to obtain a unified answer output result.

2. The business question answering method based on a large language model and scene tags according to claim 1 is characterized in that: The step of matching and determining a corresponding set of related questions in the knowledge base based on the expanded question information includes: Traversing a list of questions in a knowledge base, the list of questions comprising a number of candidate knowledge base associated questions; If the trigger word label of a candidate knowledge base associated question in the question list is in the expanded question information, the candidate knowledge base associated question is compiled into the associated question set.

3. The business question answering method based on a large language model and scene tags according to claim 1 is characterized in that: The selecting a subset of related questions from the set of related questions in response to the user's operation instruction comprises: Matching semantic similarity between each knowledge base related question and business question in the related question set, and sorting each knowledge base related question according to the semantic similarity between each knowledge base related question and business question, to obtain a sorted related question set; The sorted related question set is output to the user, and in response to the user's operation instruction, a number of knowledge base related questions are selected from the sorted related question set to form a related question subset.

4. The business question answering method based on a large language model and scene tags according to claim 1 is characterized in that: The data blocks contained in the reference answer to the question and the scene tag set associated with each data block include: Determine the data blocks included in the reference answer to the question and several scene labels associated with each data block; The scene labels associated with each data block are aggregated to obtain a scene label set for each data block.

5. The business question answering method based on a large language model and scene tags according to claim 1 is characterized in that: The scenario consistency evaluation of the scenario tag set of each data block and the supplementary description of the business problem is performed to obtain the scenario consistency score of each data block, including: Traversing the scene label set of each data block, and determining whether each scene label in the scene label set is included in the supplementary description of the business problem; If a scene label in the scene label set of a data block is included in the supplementary description of the business problem, the scene consistency score of the data block will be increased by one point; The final scene consistency score of each data block is obtained by statistics.

6. The business question answering method based on a large language model and scene tags according to claim 5 is characterized in that: If a scene label in the scene label set of the data block is not included in the supplementary description of the business problem, the first semantic vector of the corresponding scene label and the second semantic vector of the supplementary description are determined, and the first semantic vector of the corresponding scene label and the second semantic vector of the supplementary description are substituted into the preset semantic similarity scoring formula for calculation to obtain the corresponding semantic similarity score, and the semantic similarity score is statistically added to the scene consistency score of the corresponding data block. The semantic similarity scoring formula is: The first semantic vector represents the corresponding scene label, and the second semantic vector represents the supplementary description.

7. The business question answering method based on a large language model and scene tags according to claim 1 is characterized in that: The selecting several data blocks with the highest scene consistency scores as target data blocks includes: selecting three data blocks with the highest scene consistency scores as target data blocks.

8. Business question answering system based on large language model and scene label, characterized by: It includes question acquisition unit, question expansion unit, association matching unit, knowledge screening unit, scene annotation unit, scene matching unit and intelligent question answering unit, among which: A question acquisition unit, used to acquire the business question input by the user and the supplementary explanation of the business question; The question expansion unit is used to import business questions into the preset large language model for question expansion and obtain the corresponding expanded question information; An associated matching unit, used for matching and determining a corresponding associated question set in the knowledge base based on the expanded question information, wherein the associated question set includes a number of knowledge base associated questions; The knowledge screening unit is used to select a related question subset containing a plurality of question items from the related question set in response to an operation instruction of a user, and retrieve a knowledge base related answer corresponding to each question item in the related question subset from the knowledge base as a reference answer to the question corresponding to the related question subset; A scene annotation unit is used to determine the data blocks contained in the reference answer of each question in the knowledge base, as well as the scene label set associated with each data block; A scenario matching unit is used to evaluate the scenario consistency of the scenario label set of each data block and the supplementary description of the business problem for each question item in the associated question subset, so as to obtain the scenario consistency score of each data block; The intelligent question-answering unit is used to select several data blocks with the highest scenario consistency scores as the target data blocks for each question item in the corresponding associated question subset, and import the target data blocks, business questions and supplementary explanations of the business questions corresponding to each question item in the associated question subset into a preset large language model for natural language processing, to obtain the corresponding business answers respectively, and finally to integrate multiple business answers to obtain a unified answer output result.

9. Business question answering system based on large language model and scene labeling, characterized by: include: A memory for storing instructions; A processor is used to read the instructions stored in the memory and execute the business question answering method based on a large language model and scene labels according to any one of claims 1 to 7.

Images