A data analysis method and device based on large language model

Abstract

The present application proposes a data analysis method and device based on a large language model. Based on the basic portrait of the first user and the search portrait of the first user, multiple third users with high similarity to the first user are screened. In the screening process, the basic information of the user, the search question, and the style preference for the search answer are comprehensively considered. Based on the search history of multiple third users, multiple third mapping relationships used to characterize the search question and data analysis dimensions are determined, and the target mapping relationship is determined according to the data analysis needs of the first user, thereby generating the first recommendation result. In addition, different solutions can be given to the same problem according to the different data analysis capabilities of the users to meet the personalized needs of different users. The technical solution of the present application can further expand the application scenarios of the large language model and accelerate the convenience of data analysis.

Description

Technical Field

[0001] The present invention belongs to the technical field of behavior prediction, and in particular, relates to a data analysis method and device based on a large language model. Background Art

[0002] With the introduction of Google's Transformer model, large language models have become a new generation of emerging technology in the field of AI. Relying on their powerful computing power and massive sample data, large language models can provide satisfactory answers to user needs.

[0003] However, currently, users' use of large language models is often limited to document generation and knowledge search. It also has great potential in application scenarios such as data analysis. However, existing technologies are unable to complete data analysis tasks based on user characteristics and needs. Summary of the invention

[0004] The purpose of the present invention is to provide a data analysis method and device based on a large language model to solve the technical problem in the prior art that data analysis tasks cannot be completed according to the characteristics and needs of users.

[0005] A data analysis method based on a large language model, characterized in that the method comprises:

[0006] S1: Determine a basic profile of the first user according to the basic information of the first user;

[0007] S2: using the first basic user portrait, matching A second portraits in the first user basic portrait set, and obtaining A second users according to the A second portraits;

[0008] S3: Determine a search profile of the first user according to the first historical search information of the first user;

[0009] S4: using the first user search portrait to match A second portraits to obtain B third portraits, and obtaining B third users according to the B third portraits;

[0010] S5: Acquire C pieces of third historical search information of the B third users, cluster the C pieces of third historical search information, and determine D third mapping relationships according to the obtained D first clustering centers;

[0011] S6: Determine a first recommendation result from the D third mapping relationships according to the first analysis requirement issued by the first user;

[0012] S7: Obtain E first common questions according to the A second portraits and the first recommendation results, and push the E first common questions and corresponding first solutions to the first user;

[0013] Among them, A, B, C, D, and E are all natural numbers greater than 0.

[0014] Preferably, S1 comprises the following sub-steps:

[0015] S11: Acquire first basic information of the first user;

[0016] S12: Acquire multiple first historical search questions of the first user, and classify the multiple first historical search questions according to search topics to obtain multiple first search questions;

[0017] S13: extracting keywords from the first search questions to obtain a plurality of first search tags;

[0018] S14: Generate the first user basic portrait according to the first basic information and the first search tags.

[0019] Preferably, the step S14 further includes the following sub-steps:

[0020] S15: Determine a plurality of first search results of different levels according to each of the first search tags and the first basic information, and establish a first mapping relationship between the first search tags and the plurality of first search results of different levels.

[0021] Preferably, S3 includes the following sub-steps:

[0022] S31: for each of the first history search questions, obtain corresponding second history search information and first action information, and establish a plurality of second mapping relationships;

[0023] S32: for each piece of the first action information, when the matching degree with the target action information meets the preset requirement, determining the second historical search information that satisfies the second mapping relationship with the first action information as the first historical search information, otherwise continue to process the next piece of the first action information;

[0024] S33: Perform text clustering analysis on the plurality of first historical search information to obtain a plurality of first text styles, and determine the first user search portrait according to the plurality of first text styles.

[0025] Preferably, the S31 includes the following sub-steps:

[0026] S311: Obtaining a first archive file of the first user from a cloud server;

[0027] S312: Acquire some second historical search information and some first action information from the first archive file, and establish a plurality of second mapping relationships.

[0028] Preferably, the S33 includes the following sub-steps:

[0029] S331: performing keyword indexing on each piece of the first historical search information;

[0030] S332: Acquire multiple first keyword sets and multiple first keyword position relationships according to the indexing result;

[0031] S333: clustering the plurality of first keyword sets and the plurality of first keyword position relationships to obtain a plurality of the first text styles;

[0032] S334: Determine a first user search portrait based on the multiple first text styles.

[0033] Preferably, S5 comprises the following sub-steps:

[0034] S51: For the B third users, obtain C pieces of third historical search information;

[0035] S52: for each piece of the third historical search information, obtain first analysis dimension information, and establish a fourth mapping relationship using the third historical search information and the corresponding first analysis dimension;

[0036] S53: Clustering the plurality of fourth mapping relationships to obtain D first cluster centers, and determining D third mapping relationships according to the D first cluster centers.

[0037] Preferably, S6 comprises the following sub-steps:

[0038] S61: Obtain D first question information and D first analysis dimension information according to the D third mapping relationships;

[0039] S62: performing similarity matching between the first analysis requirement and each of the third mapping relationships one by one, and taking the third mapping relationship with the highest similarity as the target mapping relationship;

[0040] S63: Acquire first target data according to the first analysis requirement, and obtain the first recommendation result according to the first analysis dimension in the target mapping relationship and the first target data.

[0041] Preferably, the S7 comprises the following sub-steps:

[0042] S71: Acquire multiple second search tags in the A second portraits;

[0043] S72: Obtain G first frequently asked questions according to the similarities between the first recommendation result and the plurality of second search tags;

[0044] S73: Determine the first solution based on the first mapping relationship and the first user basic portrait and push it to the first user.

[0045] The present application also proposes a data analysis device based on a large language model, which is used to implement the above-mentioned data analysis method based on a large language model.

[0046] The present application proposes a data analysis method and device based on a large language model. Based on the basic portrait of the first user and the search portrait of the first user, multiple third users with high similarity to the first user are screened. In the screening process, the basic information of the user, the search question, and the style preference for the search answer are comprehensively considered. Based on the search history of multiple third users, multiple third mapping relationships used to characterize the search question and data analysis dimensions are determined, and the target mapping relationship is determined according to the data analysis needs of the first user, thereby generating the first recommendation result. In addition, different solutions can be given to the same problem according to the different data analysis capabilities of the users to meet the personalized needs of different users. The technical solution of the present application can further expand the application scenarios of the large language model and accelerate the convenience of data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the implementation methods of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the implementation methods or the description of the prior art. Obviously, the drawings in the following description are only exemplary, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0048] Figure 1 It is an execution flow chart of a data analysis method based on a large language model of the present invention.

[0049] Figure 2 It is a schematic diagram of the user sending different action information in response to different problems in the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments, wherein the illustrative embodiments and descriptions are only used to explain the present invention but are not intended to limit the present invention.

[0052] The following is a detailed description of a data analysis method and device based on a large language model of the present invention.

[0053] This embodiment proposes a data analysis method based on a large language model. The method flow is as follows: Figure 1 As shown, the specific steps include:

[0054] S1: Determine a basic profile of the first user according to the basic information of the first user.

[0055] The first user basic portrait is used to characterize the basic information, data analysis capabilities and other information of the first user.

[0056] The S1 comprises the following sub-steps:

[0057] S11: Obtain first basic information of the first user.

[0058] The first basic information includes the age, gender, occupation, job, etc. of the first user, and is mainly used to characterize the basic capabilities of the first user.

[0059] S12: Acquire multiple first historical search questions of the first user, and classify the multiple first historical search questions according to search topics to obtain multiple first search questions.

[0060] The search topics include at least two categories: one is a help-seeking search topic, such as "how to use the VLOOKUP function in EXCEL" and other tool usage issues in data analysis, and the other is a task-asking search topic, such as "please help me analyze the weather trend in my city" and other specific analysis task issues.

[0061] The first search question is a help-seeking search topic, and the first search question can reflect the data analysis ability of the first user.

[0062] S13: Extract keywords from the multiple first search questions to obtain multiple first search tags.

[0063] In this step, keyword extraction is performed on the multiple first search questions, so that multiple first search tags for representing the search questions can be obtained.

[0064] For example, for the question "How to use the VLOOKUP function in EXCEL", "VLOOKUP" can be added as a search tag to the first user basic portrait.

[0065] S14: Generate the first user basic portrait according to the first basic information and the multiple first search tags.

[0066] The first basic user portrait includes the first basic information characterizing the basic capabilities of the first user and the first search tag used to characterize the data analysis capabilities of the user.

[0067] Preferably, in order to provide differential search result recommendations for different types of users, S15 may be further included after S14.

[0068] S15: Determine a plurality of first search results of different levels according to each of the first search tags and the first basic information, and establish a first mapping relationship between the first search tags and the plurality of first search results of different levels.

[0069] For the same search question, different types of users may need different search results. For example, for the same function in EXCEL, user 1 may need to know the basic usage, while user 2 may need to know the advanced usage. Therefore, in this step, for each of the first search tags, the type of the first user is determined based on the first basic information of the first user, and then several first search results of different levels are determined, so as to provide a basis for the subsequent targeted recommendation of search results to users.

[0070] S2: Using the first basic user portrait, match A second portraits in the first user basic portrait set, and obtain A second users based on the A second portraits.

[0071] The first user basic portrait set includes multiple historical user portraits, and each of the historical user portraits is constructed according to the method in S1.

[0072] In S2, the cosine similarity between the first user basic portrait and each of the historical user portraits in the first user basic portrait set is calculated one by one, and the historical user portraits whose cosine similarity is greater than a preset value are determined as second portraits, so that A second portraits can be obtained, and A second users can be determined according to the corresponding relationship between the second portraits and the second users. Each of the second users has relatively similar basic capabilities and search habits to the first user.

[0073] S3: Determine a search portrait of the first user according to the first historical search information of the first user.

[0074] The first historical search information refers to the most satisfactory answer information provided by the large language model for each question raised by the first user. Since the large language model provides answers based on the analysis of the questions raised by the user, when the user is not satisfied with the answer, the user may adjust the way of expressing the question, so as to make the large language model better understand the user's needs and give a satisfactory answer. Therefore, in this step, the first historical search information that the user is most satisfied with is analyzed to understand the characteristics of the user's expected answer.

[0075] The S3 comprises the following sub-steps:

[0076] S31: For each of the first history search questions, obtain corresponding second history search information and first action information, and establish a plurality of second mapping relationships.

[0077] The S31 includes the following sub-steps:

[0078] S311: Obtain a first archive file of the first user from a cloud server.

[0079] The first archive file stores all first historical search questions raised by the first user, and also stores second historical search information and first action information corresponding to each of the first historical search questions.

[0080] The second historical search information is obtained by analyzing the first historical search question with a large language model.

[0081] In addition, for each of the historical search questions, after the second historical search information is displayed on the first interface, the action information sent by the first user is the first action information. Figure 2 As shown, when the user is not satisfied with the answer currently displayed on the first interface, the user may adjust the expression of the previous question and ask the search question again. When the user is satisfied with the answer currently displayed on the first interface, the user may copy, forward, etc. the second historical search information currently displayed. For example, the user may be dissatisfied with the answers to questions 1 and 2, but satisfied with the answer to question 3, thus triggering the forwarding control.

[0082] S312: Acquire some second historical search information and some first action information from the first archive file, and establish a plurality of second mapping relationships.

[0083] Since each piece of the second historical search information corresponds to one piece of the first action information, a second mapping relationship may be established for each piece of the second historical search information.

[0084] S32: For each piece of the first action information, when the matching degree with the target action information meets the preset requirement, the second historical search information that satisfies the second mapping relationship with the first action information is determined as the first historical search information; otherwise, the next piece of the first action information is processed.

[0085] The target action information includes actions such as not issuing a search question again, copying or forwarding.

[0086] S33: Perform text clustering analysis on the plurality of first historical search information to obtain a plurality of first text styles, and determine the first user search portrait according to the plurality of first text styles.

[0087] The S33 comprises the following sub-steps:

[0088] S331: Perform keyword indexing on each piece of the first historical search information.

[0089] Each text type corresponds to a certain text style, each text style is composed of a number of keywords, and the keywords have a preset positional relationship.

[0090] S332: Acquire multiple first keyword sets and multiple first keyword position relationships according to the indexing result.

[0091] The first keyword set stores a number of extracted keywords, and the first keyword position relationship stores the position relationship between different keywords.

[0092] S333: Clustering the multiple first keyword sets and the multiple first keyword position relationships to obtain multiple first text styles.

[0093] In this step, after the clustering operation, the target keyword set and position relationship can be obtained, so that a number of the first text styles can be formed according to the obtained keyword set and position relationship, and the first text style can represent the first user's preference for the search answer.

[0094] S334: Determine a first user search portrait based on the multiple first text styles.

[0095] A number of tag information are obtained from the plurality of the first text styles, thereby forming the first user search portrait.

[0096] S4: Using the first user search portrait, match and obtain B third portraits from the A second portraits, and obtain B third users based on the B third portraits.

[0097] In this step, first, a corresponding search portrait is generated for each of the second users according to the steps in S3, and then the cosine similarity between the search portrait of the first user and the search portrait of each second user is calculated one by one to obtain B third portraits whose cosine similarity meets the preset requirements, and B third users are obtained based on the B third portraits.

[0098] S5: Acquire C pieces of third historical search information of the B third users, cluster the C pieces of third historical search information, and determine D third mapping relationships according to the obtained D first clustering centers.

[0099] The S5 comprises the following sub-steps:

[0100] S51: For the B third users, obtain C pieces of third historical search information.

[0101] The third historical search information is search information about a data analysis table, and the third historical search information is a set of binary groups consisting of a question and an answer.

[0102] S52: For each piece of the third historical search information, obtain first analysis dimension information, and establish a fourth mapping relationship using the third historical search information and the corresponding first analysis dimension.

[0103] The first analysis dimension information is obtained by analyzing the answer information in the third historical search information, and includes analysis dimensions given by a large language model for the question raised by the third user.

[0104] S53: Clustering the plurality of fourth mapping relationships to obtain D first cluster centers, and determining D third mapping relationships according to the D first cluster centers.

[0105] Each of the first cluster centers includes a third mapping relationship consisting of a question and first analysis dimension information.

[0106] S6: Determine a first recommendation result from the D third mapping relationships according to the first analysis requirement issued by the first user.

[0107] Since the third mapping relationship stores the third mapping relationship consisting of the analysis question and the first analysis dimension information, when similarity matching is performed among the D third mapping relationships using the first analysis requirement, the first recommendation result can be determined based on the third mapping relationship with the highest matching degree.

[0108] The S6 comprises the following sub-steps:

[0109] S61: Obtain D first question information and D first analysis dimension information according to the D third mapping relationships.

[0110] S62: Perform similarity matching between the first analysis requirement and each of the third mapping relationships one by one, and use the third mapping relationship with the highest similarity as the target mapping relationship.

[0111] In this step, the similarity between the first analysis requirement and the first question information and the first analysis dimension information in the same third mapping relationship is calculated respectively to obtain a first similarity and a second similarity, and the first similarity and the second similarity are weightedly summed as the similarity between the first analysis requirement and each of the third mapping relationships.

[0112] S63: Acquire first target data according to the first analysis requirement, and obtain the first recommendation result according to the first analysis dimension in the target mapping relationship and the first target data.

[0113] The first target data is data content obtained by the large language model according to the first analysis requirement, and the first analysis dimension is a data analysis dimension obtained from the target mapping relationship. The first recommendation result can be obtained by fusing the two.

[0114] S7: Obtain E first common questions based on the A second portraits and the first recommendation results, and push the E first common questions and corresponding first solutions to the first user.

[0115] The S7 comprises the following sub-steps:

[0116] S71: Obtain multiple second search tags in the A second portraits.

[0117] Each of the second search tags is associated with a search question of the user.

[0118] S72: Obtain E first frequently asked questions according to the similarity between the first recommendation result and the plurality of second search tags.

[0119] The second search tag whose similarity exceeds a threshold is used as a target search tag, so as to determine E first common problems according to the target search tag.

[0120] S73: Determine the first solution based on the first mapping relationship and the first user basic portrait and push it to the first user.

[0121] Among them, the first user basic portrait can represent the data analysis ability of the first user, so according to the first mapping relationship, a first solution suitable for the first user can be analyzed.

[0122] Among them, A, B, C, D, and E are all natural numbers greater than 0.

[0123] The present application also proposes a data analysis device based on a large language model, which is used to execute any of the data analysis methods based on a large language model described above.

[0124] The present application proposes a data analysis method and device based on a large language model. Based on the basic portrait of the first user and the search portrait of the first user, multiple third users with high similarity to the first user are screened. In the screening process, the basic information of the user, the search question, and the style preference for the search answer are comprehensively considered. Based on the search history of multiple third users, multiple third mapping relationships used to characterize the search question and data analysis dimensions are determined, and the target mapping relationship is determined according to the data analysis needs of the first user, thereby generating the first recommendation result. In addition, different solutions can be given to the same problem according to the different data analysis capabilities of the users to meet the personalized needs of different users. The technical solution of the present application can further expand the application scenarios of the large language model and accelerate the convenience of data analysis.

[0125] The above description is only a preferred embodiment of the present invention, so all equivalent changes or modifications made according to the structure, characteristics and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims

1. A data analysis method based on a large language model, characterized in that: The method includes: S1: Determine a basic profile of the first user according to the basic information of the first user; S2: using the first basic user portrait, matching A second portraits in the first user basic portrait set, and obtaining A second users according to the A second portraits; S3: Determine a search profile of the first user according to the first historical search information of the first user; S4: using the first user search portrait to match A second portraits to obtain B third portraits, and obtaining B third users according to the B third portraits; S5: Acquire C pieces of third historical search information of the B third users, cluster the C pieces of third historical search information, and determine D third mapping relationships according to the obtained D first clustering centers; S6: Determine a first recommendation result from the D third mapping relationships according to the first analysis requirement issued by the first user; S7: Obtain E first common questions according to the A second portraits and the first recommendation results, and push the E first common questions and corresponding first solutions to the first user; Among them, A, B, C, D, and E are all natural numbers greater than 0; The S6 comprises the following sub-steps: S61: Obtain D first question information and D first analysis dimension information according to the D third mapping relationships; S62: performing similarity matching between the first analysis requirement and each of the third mapping relationships one by one, and taking the third mapping relationship with the highest similarity as the target mapping relationship; S63: Acquire first target data according to the first analysis requirement, and obtain the first recommendation result according to the first analysis dimension in the target mapping relationship and the first target data.

2. A data analysis method based on a large language model according to claim 1, characterized in that: The S1 comprises the following sub-steps: S11: Acquire first basic information of the first user; S12: Acquire multiple first historical search questions of the first user, and classify the multiple first historical search questions according to search topics to obtain multiple first search questions; S13: extracting keywords from the first search questions to obtain a plurality of first search tags; S14: Generate the first user basic portrait according to the first basic information and the first search tags.

3. A data analysis method based on a large language model according to claim 2, characterized in that: The S14 further includes the following sub-steps: S15: Determine a plurality of first search results of different levels according to each of the first search tags and the first basic information, and establish a first mapping relationship between the first search tags and the plurality of first search results of different levels.

4. The data analysis method based on a large language model according to claim 3, characterized in that: The S3 comprises the following sub-steps: S31: for each of the first history search questions, obtain corresponding second history search information and first action information, and establish a plurality of second mapping relationships; S32: for each piece of the first action information, when the matching degree with the target action information meets the preset requirement, determining the second historical search information that satisfies the second mapping relationship with the first action information as the first historical search information, otherwise continue to process the next piece of the first action information; S33: Perform text clustering analysis on the plurality of first historical search information to obtain a plurality of first text styles, and determine the first user search portrait according to the plurality of first text styles.

5. The data analysis method based on a large language model according to claim 4, characterized in that: The S31 includes the following sub-steps: S311: Obtaining a first archive file of the first user from a cloud server; S312: Acquire some second historical search information and some first action information from the first archive file, and establish a plurality of second mapping relationships.

6. The data analysis method based on a large language model according to claim 5, characterized in that: The S33 comprises the following sub-steps: S331: performing keyword indexing on each piece of the first historical search information; S332: Acquire multiple first keyword sets and multiple first keyword position relationships according to the indexing result; S333: clustering the plurality of first keyword sets and the plurality of first keyword position relationships to obtain a plurality of the first text styles; S334: Determine a first user search portrait based on the multiple first text styles.

7. The data analysis method based on a large language model according to claim 6, characterized in that: The S5 comprises the following sub-steps: S51: For the B third users, obtaining C pieces of third historical search information; S52: for each piece of the third historical search information, obtain first analysis dimension information, and establish a fourth mapping relationship using the third historical search information and the corresponding first analysis dimension; S53: Clustering the plurality of fourth mapping relationships to obtain D first cluster centers, and determining D third mapping relationships according to the D first cluster centers.

8. The data analysis method based on a large language model according to claim 7, characterized in that: The S7 comprises the following sub-steps: S71: Acquire multiple second search tags in the A second portraits; S72: Obtain G first frequently asked questions according to the similarities between the first recommendation result and the plurality of second search tags; S73: Determine the first solution based on the first mapping relationship and the first user basic portrait and push it to the first user.

9. A data analysis device based on a large language model, used to implement a data analysis method based on a large language model as described in any one of claims 1 to 8.

Images