Method and system for analyzing index fluctuation attribution of large language model, electronic device and storage medium

Abstract

The application provides an index fluctuation attribution analysis method, system, electronic equipment and storage medium of a large language model, determines a corresponding to-be-attributed index according to a user question sent by a user, obtains a first text template and attribution dimension metadata corresponding to the to-be-attributed index, generates a semantic analysis request according to the user question, the first text template and the attribution dimension metadata, processes the semantic analysis request through a large language model to obtain a DSL object, generates a corresponding attribution analysis request according to the DSL object, determines at least one target attribution analysis tool according to the user question, maps the attribution analysis request to each target attribution analysis tool, performs attribution analysis according to the attribution analysis request through each target attribution analysis tool to obtain at least one attribution analysis result, and outputs an attribution result summary of the to-be-attributed index through a result summary request generated by a large language model according to a second text template, the user question and the at least one attribution analysis result.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a method, system, electronic device, and storage medium for attribution analysis of index fluctuations that integrates a large language model. Background Technology

[0002] As enterprises continue to develop, in order to achieve sustainable growth, they construct multiple business indicators to summarize key information related to business activities in a standardized manner. With the increase in business indicators, in order to avoid affecting the normal operation of the enterprise, real-time monitoring of business indicator changes has gradually become a key focus of the enterprise's daily operations.

[0003] Currently, attribution analysis for business metric anomalies typically involves methods such as drill-down attribution, topology-level attribution, and lineage attribution. However, due to objective issues such as planning, interaction, and logical differences, the entry points for different attribution analysis tools are scattered across different subsystems. This means that operations personnel need to be familiar with the product matrix to accurately deploy the corresponding attribution analysis tools to their respective subsystems. Furthermore, prior query configuration is required before using attribution analysis tools, and since the query configuration methods for each tool differ, operations personnel also need to have a certain understanding of each attribution method, requiring them to spend a significant amount of time configuring the queries for each tool. After obtaining the analysis results from the attribution analysis tools, operations personnel still need to synthesize the results from different tools to determine the final attribution conclusion. Therefore, existing methods for attribution analysis of business metric anomalies are not only difficult to use but also inefficient. Summary of the Invention

[0004] In view of this, this application provides a method, system, electronic device and storage medium for attribution analysis of index fluctuations that integrates a large language model, with the aim of reducing the user's usage threshold and improving the efficiency of index attribution analysis.

[0005] The first aspect of this application provides a method for attribution analysis of index fluctuations that integrates a large language model, the method comprising:

[0006] The corresponding attribution metric is determined based on the user's question, and the pre-set first text template and the attribution dimension metadata corresponding to the attribution metric are obtained.

[0007] A semantic analysis request is generated based on the user question, the first text template, and the attribution dimension metadata of the attribution metric to be attributed.

[0008] The semantic analysis request is processed by a large language model to obtain a DSL object, and a corresponding attribution analysis request is generated based on the DSL object.

[0009] Based on the user's question, at least one target attribution analysis tool is determined, and the attribution analysis request is mapped to each of the target attribution analysis tools;

[0010] Attribution analysis is performed on each of the target attribution analyses according to the attribution analysis request, and at least one attribution analysis result is obtained.

[0011] A result summary request is generated based on a pre-set second text template, the user question, and at least one of the attribution analysis results. The attribution result summary of the index to be attributed is then output by the large language model based on the result summary request.

[0012] Optionally, before generating the corresponding attribution analysis request based on the DSL object, the method further includes:

[0013] The DSL object is validated using pre-set specific business rules; wherein, the specific business rules include DSL definition information and preset required parameters and their dimension enumeration values;

[0014] If the DSL object passes the verification, the step of generating a corresponding attribution analysis request based on the DSL object is executed.

[0015] Optionally, determining at least one target attribution analysis tool based on the user query and mapping the attribution analysis request to the target attribution analysis tool includes:

[0016] Determine the user's intent based on the user's question;

[0017] Determine whether the user intent is empty;

[0018] If the user intent is not empty, at least one target attribution analysis tool is determined from the pre-set attribution analysis tools based on the intent enumeration and its indicator name indicated in the user intent.

[0019] The attribution analysis request is routed to the target attribution analysis tool.

[0020] Optionally, the method further includes:

[0021] If the user intent is empty, each pre-set attribution analysis tool will be identified as the target attribution analysis tool.

[0022] Optionally, the attribution analysis request includes at least the dimension filtering conditions for the attribution indicator, the current time interval of the attribution indicator, and the concurrent time interval.

[0023] The step of performing attribution analysis using each of the target attribution analysis tools according to the attribution analysis request to obtain the corresponding attribution analysis results includes:

[0024] Based on the dimension filtering conditions of the attributable indicator, the current time interval and the same period time interval of the attributable indicator, the attributable indicator data corresponding to the attributable indicator;

[0025] Based on the screening criteria of each of the target attribution analysis tools, target attribution index data that meet the requirements of each target attribution analysis tool are selected from the attribution index data.

[0026] The attribution analysis results are obtained by performing attribution analysis based on the corresponding target attribution index data using each of the aforementioned target attribution analysis tools.

[0027] Optionally, the screening criteria for the target attribution analysis tool include the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to the target attribution analysis tool;

[0028] The step of filtering the attribution index data from the attribution index data to select the target attribution index data that meets the requirements of each target attribution analysis tool includes:

[0029] For each of the target attribution analysis tools, obtain the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to the target attribution analysis tool;

[0030] Based on the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to each target attribution analysis tool, target attribution indicator data that satisfies each target attribution analysis tool is obtained from the attribution indicator data.

[0031] A second aspect of this application provides a system for attributing index fluctuations by integrating a large language model, the system comprising:

[0032] The user question receiving unit is used to determine the corresponding attribution metric based on the user question sent by the user, and to obtain the pre-set first text template and the attribution dimension metadata corresponding to the attribution metric.

[0033] The semantic analysis unit is configured to generate a semantic analysis request based on the user question, the first text template, and the attribution dimension metadata of the attribution metric to be attributed; process the semantic analysis request through a large language model to obtain a DSL object; generate a corresponding attribution analysis request based on the DSL object; determine at least one target attribution analysis tool based on the user question, and map the attribution analysis request to each target attribution analysis tool.

[0034] The attribution analysis unit is configured to perform attribution analysis according to the attribution analysis request for each of the target attribution analyses, and obtain at least one attribution analysis result.

[0035] The attribution result summary unit is used to generate a result summary request based on a pre-set second text template, the user question, and at least one of the attribution analysis results, and output the attribution result summary of the index to be attributed according to the result summary request through the large language model.

[0036] Optionally, the semantic analysis unit is further configured to:

[0037] The DSL object is validated using pre-set specific business rules; wherein, the specific business rules include DSL definition information and preset required parameters and their dimension enumeration values; if the DSL object passes the validation, the step of generating a corresponding attribution analysis request based on the DSL object is executed.

[0038] A third aspect of this application provides an electronic device, comprising: a processor and a memory, the processor and the memory being connected via a communication bus; wherein the processor is configured to call and execute a program stored in the memory; and the memory is configured to store the program, the program being configured to implement the index fluctuation attribution analysis method of the fusion large language model provided in the first aspect of this application.

[0039] The fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions for performing the index fluctuation attribution analysis method for fusion large language model as provided in the first aspect of this application.

[0040] This application provides a method, system, electronic device, and storage medium for attribution analysis of index fluctuations that integrates a large language model. Upon receiving a user query, the system can determine the attributable index to be analyzed based on the user query. Simultaneously, it acquires pre-set first text and the attribution dimension metadata corresponding to the attributable index. Based on the user query, the first text template, and the attribution dimension metadata of the attributable index, it generates a semantic analysis request that satisfies the input of the large language model. This allows the system to leverage the information extraction capabilities of the large language model to generate a DSL object usable by downstream attribution analysis tools based on the input semantic analysis request. This eliminates the need for operators to spend significant time configuring various attribution analysis tools, thus lowering the barrier to entry for using attribution analysis tools. The DSL object is encapsulated into an attribution analysis request, and at least one target attribution analysis tool is determined based on the user query, enabling selection based on user needs. By selecting appropriate attribution analysis tools, the system integrates entry points for attribution analysis of indicator fluctuations across different scenarios. The attribution analysis request is mapped to at least one designated target attribution analysis tool. This allows the target tool to perform attribution analysis based on the mapped request, yielding at least one attribution analysis result. Finally, leveraging the logical reasoning capabilities of a large language model, combined with the attribution analysis results from each target tool, user questions, and a pre-set second text template, further heuristic analysis is conducted to output a summary of the attribution results for the indicator to be attributed. In other words, by combining the summarization and text generation capabilities of the large language model, a final summary of the various attribution analysis results is provided. This eliminates the need for operations personnel to perform comprehensive analysis of results from different attribution analysis tools, improving the efficiency of attribution analysis, reducing reliance on operations personnel, and enhancing the accuracy of attribution analysis. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0042] Figure 1 A flowchart illustrating an index fluctuation attribution analysis method integrating a large language model, provided for an embodiment of this application;

[0043] Figure 2 A schematic diagram of the structure of an index fluctuation attribution analysis system that integrates a large language model, provided in an embodiment of this application;

[0044] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0047] As can be seen from the above background technology, existing attribution analysis methods for abnormal business indicators are not only difficult to use but also inefficient.

[0048] To address the aforementioned issues, existing technologies utilize numerous large language models (hereinafter referred to as large models) such as ChatGPT to perform attribution analysis on anomalies in business metrics. However, practical experience has revealed several drawbacks. First, due to inaccurate understanding of internal business terminology by these large models, it is difficult to obtain satisfactory attribution results when relying solely on them for attribution analysis. Second, precise numerical calculations are a weak point in large models, while attribution analysis typically involves extensive numerical computations, increasing the risk of errors in the attribution results obtained through large models.

[0049] Therefore, this application provides a method, system, electronic device, and storage medium for attribution analysis of index fluctuations that integrates a large language model. It determines the attributable index to be analyzed based on received user queries, simultaneously acquires pre-set first text and attribution dimension metadata corresponding to the attributable index, and generates a semantic analysis request that satisfies the input of a large language model based on the user query, the first text template, and the attribution dimension metadata of the attributable index. This allows the large language model to extract information and generate a DSL object usable by downstream attribution analysis tools based on the input semantic analysis request, eliminating the need for operators to spend significant time configuring various attribution analysis tools and thus lowering the barrier to entry for using attribution analysis tools. The DSL object is encapsulated into an attribution analysis request, and at least one target attribution analysis tool is determined based on the user query, enabling the selection of the appropriate attribution analysis tool according to user needs. This integrates entry points for different scenarios of index fluctuation attribution analysis and allows the attribution analysis request to be used effectively. Mapped to at least one designated target attribution analysis tool, this method enables the target attribution analysis tool to perform attribution analysis based on the mapped attribution analysis request, obtaining at least one attribution analysis result. Finally, leveraging the logical reasoning capabilities of a large language model combined with the attribution analysis results from each target attribution analysis tool, user questions, and a result summary request generated from a pre-set second text template, further heuristic analysis is performed, outputting a summary of the attribution results for the attributable indicators. In other words, by combining the summarization and text generation capabilities of the large language model, a final summary of the various attribution analysis results is provided, eliminating the need for operational personnel to perform comprehensive analysis of the results from different attribution analysis tools. This not only improves the efficiency of attribution analysis but also reduces reliance on operational personnel. Furthermore, by combining various target attribution analysis tools with a large language model to complete the corresponding attribution analysis, this application also addresses the risk of errors that may arise from relying solely on the large language model for attribution analysis in existing technologies, thus improving the accuracy of attribution analysis.

[0050] See Figure 1 The diagram illustrates a flowchart of an index fluctuation attribution analysis method integrating a large language model, provided in an embodiment of this application. The method specifically includes the following steps:

[0051] S101: Determine the corresponding attribution metric based on the user's query, and obtain the pre-set first text template and the attribution dimension metadata corresponding to the attribution metric.

[0052] In this embodiment of the application, a corresponding first text template can be preset, wherein the first text template can be a predefined structured text; after receiving a user question, the corresponding attribution metric can be determined according to the user question, and the preset text template and the attribution dimension metadata corresponding to the attribution metric can be obtained, so that the user question and the attribution dimension metadata corresponding to the attribution metric can be filled into the first text template to assemble a semantic analysis request that meets the input of the large language model.

[0053] In some embodiments, the pre-set first text template can be a first prompt template, wherein the design structure of the first prompt template can be: "Return valid JSON in the corresponding format: ${DSL}, the recognizable dimension range includes ${dimensional set}, wherein the dimension range can include dimension A, dimension B, ... etc., and correspondingly, dimension A is ${dimensional definition}, that is, it can only be selected from ${dimensional value range}, dimension B, ... and must strictly follow the above rules and reference examples to handle user questions, and output the corresponding results in the format of JSON: ${DSL} at the beginning, and the output results are in correct JSON format." The reference example is shown below:

[0054] Q: ${Example Question 1}, A: ${Example Answer 1}

[0055] Q: ${Example Question 2}, A: ${Example Answer 2}

[0056] ......

[0057] It should be noted that adding relevant reference examples (example questions and their example answers (model output results)) to the first text template can enable the large language model to output results that meet the required format based on the reference examples, thereby improving the output quality of the large language model. Specifically, the few-shot prompting method can be used to add relevant reference examples (example questions and their example answers (model output results)) to the first prompt template (first text template).

[0058] In this embodiment of the application, a corresponding dimension set can be pre-set, wherein the dimension set includes a subset of all dimensions of each metric to be attributed, which is used to limit the range of each metric to be attributed; the dimension set includes multiple elements corresponding to each metric to be attributed, and each element consists of a metric business dimension and all its dimension enumeration values, that is, each element contains a metric business dimension and the optional range of its dimension values.

[0059] For example, if the metric to be attributed is order volume, then the corresponding business dimension of the metric to be attributed can be "city", "business line", etc., and the enumerated values ​​of all dimensions corresponding to "city (metric business dimension)" can be "Shenzhen, Guangzhou, Chongqing, etc.", and the enumerated values ​​of all dimensions corresponding to "business line (metric business dimension)" can be "large freight truck, small freight truck, etc."

[0060] In the specific execution step S101, after receiving a user's question, a pre-set first text template can be obtained, and the user's question can be analyzed to determine the attribution metric related to the user's question, so as to obtain the attribution dimension metadata related to the attribution metric from the pre-set dimension set; wherein, the attribution dimension metadata includes each indicator business dimension corresponding to the attribution metric and all its dimension enumeration values.

[0061] For example, if a user asks, "Why has the order volume of freight trucks in Shenzhen decreased compared to last week?", analyzing the user's question can determine that the corresponding attributable metric is "order volume". Then, the attribution dimension metadata related to the attributable metric can be obtained from the pre-set dimension set, that is, the business dimensions of each metric related to the attributable metric and all their dimension enumeration values ​​can be obtained.

[0062] S102: Generate a semantic analysis request based on the user's question, the first text template, and the attribution dimension metadata of the metric to be attributed.

[0063] In the specific execution step S102, after obtaining the user question, the first text template, and the attribution dimension metadata corresponding to the attribution metric, the user question, the first text template, and the attribution dimension metadata corresponding to the attribution metric can be concatenated to obtain a complete semantic analysis request that satisfies the large language template input. The semantic analysis request includes the user question, the first text template, and the attribution dimension metadata corresponding to the attribution metric.

[0064] Specifically, if the first text template is the first prompt template, the user's question and the metadata of the attribution dimension of the attribution indicator to be attributed can be filled into the first prompt template according to the design structure of the first prompt template to obtain a complete semantic analysis request that satisfies the input of the large language template.

[0065] S103: Process the semantic analysis request through the large language model to obtain the DSL object, and generate the corresponding attribution analysis request based on the DSL object.

[0066] In the specific execution step S103, after receiving the semantic analysis request, the request is input into the large language model. The large language model utilizes its information extraction capabilities to extract the dimension filtering conditions, current time interval, and concurrent time interval of the attribution metric based on the user question and the attribution dimension metadata corresponding to the attribution metric in the semantic analysis request. Furthermore, based on the example answers in the reference examples carried in the semantic analysis request, the model generates and outputs the dimension filtering conditions, current time interval, and concurrent time interval of the attribution metric in JSON format as a domain-specific language (DNS). A Language (DSL) object is generated and integrated with the AIP interface parameters provided by the large language model's indicator fluctuation attribution analysis system according to the convention, to obtain the corresponding attribution analysis request. The DSL object includes the dimension filtering conditions for the indicator to be attributed, the current time interval of the indicator, and the concurrent time interval. The corresponding attribution analysis request includes at least the dimension filtering conditions for the indicator to be attributed, the current time interval of the indicator, and the concurrent time interval. The dimension filtering conditions for the indicator to be attributed include the indicator to be attributed and indicator limiting conditions. The indicator limiting conditions may include the indicator business dimension corresponding to the indicator to be attributed and all its dimension enumeration values.

[0067] For example, the output DSL object may include the following metrics to be attributed: order volume; city (metric business dimension): Shenzhen (dimension enumeration value); business line (metric business dimension): freight vehicle (dimension enumeration value); current period: time: 2024-08-05; same period: time: 2024-07-29.

[0068] It should be noted that DSL, also known as Domain-Specific Language, refers to a computer language that is limited to a specific application domain. A domain-specific language is a language with very specific goals in its design and implementation. It is often used for business modeling to achieve specific goals in specific scenarios and is used by a group of computer programs that share the same goal.

[0069] In this embodiment, after obtaining the corresponding DSL object, to avoid anomalies when performing attribution analysis using attribution analysis requests generated from invalid DSL objects, the DSL object can be validated before generating the corresponding attribution analysis request. If the DSL object passes the validation, the corresponding attribution analysis request can be generated based on the DSL object; if the DSL object fails the validation, the DSL object can be filtered out to avoid anomalies when performing attribution analysis using attribution analysis requests generated from invalid DSL objects.

[0070] Optionally, the process of validating the DSL object can be as follows: validate the DSL object using pre-set specific business rules; if the DSL validation fails, filter out the DSL object; if the DSL object validation passes, execute the step of generating a corresponding attribution analysis request based on the DSL object.

[0071] It should be noted that specific business rules indicate whether a DSL object meets the DSL definition information and whether the DSL object includes preset mandatory parameters and their dimension values. Specifically, after obtaining a DSL object, it can be determined whether the DSL object meets the DSL definition information and whether the DSL object includes preset mandatory parameters and their dimension enumeration values. The preset parameters can be a pre-set indicator business dimension, which is not limited in this embodiment.

[0072] S104: Identify at least one target attribution analysis tool based on user queries and map the attribution analysis request to the target attribution analysis tool.

[0073] During the specific execution of step S104, a corresponding attribution analysis module can be pre-set, wherein the attribution analysis module includes various pre-set attribution analysis tools; after receiving the corresponding attribution analysis request, the user's question can be analyzed to obtain the user's intent, and at least one target attribution analysis tool can be determined from the various pre-set attribution analysis tools according to the user intent, and finally the attribution analysis request is mapped to each target attribution analysis tool.

[0074] It should be noted that various attribution analysis tools may include indicator dimension drill-down attribution analysis tools, indicator topology hierarchical attribution analysis tools, indicator lineage attribution analysis tools, etc., and are not limited to these in the embodiments of this application.

[0075] Optionally, the process of determining at least one target attribution analysis tool based on the user's question and mapping the attribution analysis request to the target attribution analysis tool can be as follows: determine the user's intent based on the user's question; determine whether the user intent is empty; if the user intent is not empty, determine at least one target attribution analysis tool from the pre-set attribution analysis tools based on the intent enumeration and its indicator names indicated in the user intent; if the user intent is empty, determine each pre-set attribution analysis tool as the target attribution analysis tool; and route the attribution analysis request to the target attribution analysis tool.

[0076] It should be noted that the intent enumeration can be dimension, topology, lineage, etc., and the indicator name can be the number of matched orders, the number of drivers accepting orders, etc., which are not limited in this embodiment of the application.

[0077] It should also be noted that the pre-configured attribution analysis tools may include indicator dimension drill-down attribution analysis tools, indicator topology hierarchical attribution analysis tools, and indicator lineage attribution analysis tools.

[0078] In practical applications, users can express their intentions in user questions or not. Therefore, by analyzing user questions and determining whether the user intention obtained from the analysis is empty, if it is empty, it means that the user did not express their intention in the user question. In this case, a fallback logic is used to determine all the pre-set attribution analysis tools as the target attribution analysis tool. If it is not empty, it means that the user expressed their intention in the user question. Then, the attribution analysis tool corresponding to the intent enumeration indicated in the user intention and its indicator name can be determined as the target attribution analysis tool.

[0079] For example, if a user asks, "Why has the order volume of freight trucks in Shenzhen decreased compared to last week, according to dimensional analysis?", analyzing the user's question can determine that the user is more concerned with dimensional analysis. Therefore, the intent enumeration and indicator names indicated in the user's intent can be identified as dimensions and indicator names, thus determining the appropriate target attribution analysis tool as the indicator dimension drill-down attribution analysis tool. Alternatively, if the user asks, "Why has the order volume of freight trucks in Shenzhen decreased compared to last week? Use dimension drill-down?", analyzing the user's question can determine that the user has expressed their intent (dimensional drill-down). Therefore, the intent enumeration and indicator names indicated in the user's intent can be identified as dimensions and indicator names, thus determining the dimension drill-down attribution analysis tool as the target attribution analysis tool. Or, if the user asks, "Why has the order volume of freight trucks in Shenzhen decreased compared to last week?", analyzing the user's question can determine that the user has not expressed their intent, i.e., the user intent is empty, thus determining all pre-set attribution analysis tools as the target attribution analysis tools.

[0080] S105: Perform attribution analysis according to the attribution analysis request for each target attribution analysis to obtain at least one attribution analysis result.

[0081] In the specific execution step S105, after identifying at least one target attribution analysis tool, attribution indicator data corresponding to the attribution indicator to be attributed can be obtained according to the attribution analysis request, and target attribution indicator data corresponding to each target attribution analysis tool can be filtered out from the attribution indicator data, so that the target attribution analysis tool with the attribution analysis request can perform attribution analysis according to the corresponding target attribution indicator data and obtain the corresponding attribution analysis results.

[0082] It should be noted that since different attribution analysis tools focus on different content, the corresponding filtering conditions for each attribution analysis tool can be pre-set according to the content that each attribution analysis tool focuses on.

[0083] It's worth noting that drill-down attribution analysis tools primarily focus on the set of dimensions and detailed information of the dimension items of the metric to be attributed. Therefore, the corresponding filtering conditions, such as metric waveform information, dimension influence coefficients, and dimension item fluctuation contribution rates, can be set based on the focus of the drill-down attribution analysis tool. Similarly, topological hierarchy attribution analysis tools primarily focus on metrics that have a topological hierarchical relationship with the metric to be attributed. Likewise, the corresponding filtering conditions, such as metric waveform information, dimension influence coefficients, and dimension item fluctuation contribution rates, can be set based on the focus of the topological hierarchy attribution analysis tool. Finally, kinship attribution analysis tools primarily focus on data tables, fields, and task information that are related to the metric to be attributed. Therefore, the corresponding filtering conditions, such as metric waveform information, dimension influence coefficients, and dimension item fluctuation contribution rates, can be set based on the focus of the kinship attribution analysis tool. Among them, the indicator waveform information includes the waveform information corresponding to each indicator to be attributed, the dimension influence coefficient includes the influence coefficient corresponding to each indicator business dimension corresponding to each indicator to be attributed, and the dimension item fluctuation contribution rate includes the contribution rate corresponding to each indicator business dimension corresponding to each indicator to be attributed.

[0084] Optionally, the attribution analysis request includes at least the dimension filtering conditions for the attributable indicator, the current time interval of the attributable indicator, and the concurrent time interval. The process of obtaining the corresponding attribution analysis results by performing attribution analysis according to the attribution analysis request by each target attribution analysis tool can be specifically as follows: based on the dimension filtering conditions for the attributable indicator, the current time interval of the attributable indicator, and the concurrent time interval, obtain the attribution indicator data corresponding to the attributable indicator; based on the filtering conditions of each target attribution analysis tool, filter out the target attribution indicator data that meets the requirements of each target attribution analysis tool; and perform attribution analysis by each target attribution analysis tool based on the corresponding target attribution indicator data to obtain the attribution analysis results.

[0085] In this embodiment of the application, the corresponding first attribution index data can be obtained according to the attribution index indicated in the dimension filtering conditions of the attribution index and the current time interval, the corresponding second attribution index data can be obtained according to the attribution index indicated in the dimension filtering conditions of the attribution index and the same time interval, and finally the attribution index data of the attribution index to be attributed can be generated according to the first attribution index data and the second attribution index data.

[0086] As a preferred embodiment of this application, the screening criteria for the target attribution analysis tool include the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to the target attribution analysis tool. The process of screening the target attribution indicator data that meets the requirements of each target attribution analysis tool from the attribution indicator data can be specifically as follows: for each target attribution analysis tool, obtain the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to the target attribution analysis tool; based on the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to each target attribution analysis tool, obtain the target attribution indicator data that meets the requirements of each target attribution analysis tool from the attribution indicator data.

[0087] In practical application, for each target attribution analysis tool, the fluctuation information of the corresponding indicator, the dimension influence coefficient, and the dimension item fluctuation contribution rate are obtained. From the fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate of the target attribution analysis tool, the waveform information of the indicator to be attributed and the influence coefficient and contribution rate of each business dimension of the indicator to be attributed, as indicated in the indicator constraints, are determined. This allows for the extraction of target attribution data that satisfies the target attribution analysis tool from the attribution indicator data corresponding to the indicator to be attributed, based on the waveform information, influence coefficient, and contribution rate of each business dimension of the indicator to be attributed. This enables the filtering of the attribution indicator data to obtain a topology analysis set (target attribution data) that has the same fluctuation direction as the indicator to be attributed and satisfies the target attribution analysis tool.

[0088] S106: Generate a result summary request based on the pre-set second text template, user questions, and at least one attribution analysis result, and output the attribution result summary of the index to be attributed according to the result summary request through the large language model.

[0089] In this application embodiment, the applicant discovered that the attribution results output by large models in the prior art are usually in the form of numerical values, vectors, and matrices. These semantic features cannot be understood by computer programs or humans. Therefore, after obtaining the attribution results, downstream users must design corresponding display tools based on the three semantic features rather than standardized computer speech, in order to convert the obtained attribution results into statements that computer programs and humans can understand. Therefore, in order to improve the accuracy of the obtained attribution results and to ensure that the obtained attribution results are statements that computer programs and humans can understand, after obtaining the attribution analysis results of each target attribution analysis tool, the logical reasoning ability of the large language model can be utilized, combined with the attribution analysis results of each target attribution analysis tool, for further heuristic analysis. Specifically, the attribution analysis results of each target attribution analysis tool can be combined with a pre-set second text template and user questions to generate a result summary request that satisfies the input of the large model, and this result summary request can be... The data is sent to a large language model, enabling it to utilize its logical reasoning capabilities to perform heuristic induction based on the attribution analysis results of various target attribution analysis tools, user queries, and a second text model. This yields a summary of the attribution results for the attribution indicators, eliminating the need for users to manually summarize and analyze each attribution analysis result. This not only improves the efficiency but also the accuracy of attribution analysis. Furthermore, upon receiving the result summary request, the large language model can perform attribution analysis based on the requirements of the second text template within the request, considering the attribution analysis results of various target attribution analysis tools and user queries. It then outputs the summary of attribution results according to the output format specified in the second text template. This addresses the problem in existing technologies where downstream users are required to design corresponding display tools based on three semantic features rather than standardized computer speech after obtaining the attribution results. This is necessary to convert the obtained attribution results into computer programs and human-understandable statements.

[0090] It should be noted that the pre-set second text model can be a prompt template (for ease of distinction, the prompt template set here will be referred to as the second prompt template), and the design of the second prompt template can be as follows:

[0091] Summarize and analyze the user-raised questions (user questions) based on the provided attribution data (attribution analysis results from various target attribution analysis tools), and explain the reasons for data fluctuations in the attribution indicators that the user is concerned about, according to the user's intent. Note: 1. The main diagnostic dimensions include: each dimension corresponding to the attribution indicator, such as "Dimension A," "Dimension B," etc.; 2. The given data (attribution analysis results from various target attribution analysis tools) has been sorted from high to low according to "impact"; 3. Dimension data must be retained when providing diagnostic conclusions (summary of attribution results); 4. If the actual diagnostic results (summary of attribution results) do not match the user questions, analysis should be conducted based on the actual fluctuation results. Results should be returned according to the following format rules, which can be: 1. Format Rule A; 2. Format Rule B; 3. Format Rule C.

[0092] It should be noted that the formatting rules can generate text content (attribution result summary) that is easy for computer programs and humans to read and understand for standardized large language models, including but not limited to the following formatting rules: 1. Year-on-year fluctuation results of {indicator}; 2. Diagnostic results show that, ranked by influence, {dimension 1} is the most important dimension affecting the change of {indicator}; 3. In the influence ranking of {dimension 1}, it can be seen that {dimension value A}, {dimension value B}, and {dimension value C} have the greatest influence on the fluctuation of {indicator}. Here, the indicator can be the indicator to be attributed, and dimension 1 can be the most important indicator business dimension affecting the change of the indicator to be attributed among the various indicator business dimensions of the indicator to be attributed; dimension value A, dimension value B, and dimension value C can be the enumerated values ​​of each dimension corresponding to dimension 1 (indicator business dimension). The above is only a preferred content of a formatting rule provided by the embodiment of this application. The specific content of the formatting rules can be set according to actual applications, and is not limited in this embodiment of the application.

[0093] Therefore, after the large language model constructs the result summary request, it can perform attribution analysis based on the requirements of the second text template in the result summary request, according to the attribution analysis results of each target attribution analysis tool in the result summary request and the user's questions. It can also output the attribution result summary obtained from the attribution analysis according to the output format required in the second text template. This solves the problem in the prior art that after obtaining the attribution results, downstream users must design corresponding display tools based on three semantic features rather than standardized computer speech, in order to realize the transformation of the obtained attribution results into computer programs and human-understandable sentences.

[0094] This application provides a method for attribution analysis of index fluctuations that integrates a large language model. It determines the attributable index to be analyzed based on received user queries, and simultaneously acquires pre-set first text and attribution dimension metadata corresponding to the attributable index. Based on the user query, the first text template, and the attribution dimension metadata of the attributable index, it generates a semantic analysis request that satisfies the input of the large language model. This allows the large language model to leverage its information extraction capabilities to generate a DSL object usable by downstream attribution analysis tools based on the input semantic analysis request. This eliminates the need for operations personnel to spend significant time configuring various attribution analysis tools, thus lowering the barrier to entry for using attribution analysis tools. The DSL object is encapsulated into an attribution analysis request, and at least one target attribution analysis tool is determined based on the user query. This enables the selection of the appropriate attribution analysis tool according to user needs, integrating entry points for different scenarios of index fluctuation attribution analysis, and mapping the attribution analysis request to at least one determined attribution analysis tool. A target attribution analysis tool performs attribution analysis based on mapped attribution analysis requests, obtaining at least one attribution analysis result. Finally, leveraging the logical reasoning capabilities of a large language model, combined with the attribution analysis results from various target attribution analysis tools, user questions, and a pre-set second text template, a result summary request is generated for further heuristic analysis, outputting a summary of the attribution results for the attributable indicators. This combines the summarization and text generation capabilities of the large language model to provide a final summary of the various attribution analysis results, eliminating the need for operational personnel to perform comprehensive analysis of the results from different attribution analysis tools. This not only improves the efficiency of attribution analysis but also reduces reliance on operational personnel. Furthermore, by combining various target attribution analysis tools with a large language model to complete the corresponding attribution analysis, this application also addresses the risk of errors that may arise from relying solely on large language models for attribution analysis in existing technologies, thereby improving the accuracy of attribution analysis.

[0095] Based on the attribution analysis method for index fluctuations fused with a large language model provided in the above embodiments of this application, correspondingly, the embodiments of this application also provide an attribution analysis system for index fluctuations fused with a large language model, such as... Figure 2 As shown, the index fluctuation attribution analysis system integrating a large language model includes a user question receiving unit, a semantic analysis unit, an attribution analysis unit, an attribution result summarization unit, and a large language model.

[0096] The user question receiving unit is used to determine the corresponding attribution metric based on the user question sent by the user, and to obtain the pre-set first text template and the attribution dimension metadata corresponding to the attribution metric.

[0097] The semantic analysis unit is used to generate semantic analysis requests based on user queries, first text templates, and attribution dimension metadata of the metrics to be attributed; process the semantic analysis requests through a large language model to obtain DSL objects; generate corresponding attribution analysis requests based on DSL objects; determine at least one target attribution analysis tool based on user queries, and map the attribution analysis requests to each target attribution analysis tool.

[0098] The attribution analysis unit is used to perform attribution analysis according to the attribution analysis request for each target attribution analysis and obtain at least one attribution analysis result.

[0099] The attribution result summary unit is used to generate a result summary request based on a pre-set second text template, user questions, and at least one attribution analysis result, and output the attribution result summary of the index to be attributed through a large language model according to the result summary request.

[0100] The specific principles and execution processes of each unit in the index fluctuation attribution analysis system of the fusion large language model disclosed in the above embodiments of this application are the same as the index fluctuation attribution analysis method of the fusion large language model disclosed in the above embodiments of this application. Please refer to the corresponding parts in the index fluctuation attribution analysis method of the fusion large language model disclosed in the above embodiments of this application, and they will not be repeated here.

[0101] This application provides an index fluctuation attribution analysis system integrating a large language model. It determines the attributable index to be analyzed based on received user queries, and simultaneously acquires pre-set first text and attribution dimension metadata corresponding to the attribution index. Based on the user query, the first text template, and the attribution dimension metadata of the attribution index, it generates a semantic analysis request that satisfies the input of the large language model. This allows the system to leverage the information extraction capabilities of the large language model to generate a DSL object usable by downstream attribution analysis tools based on the input semantic analysis request. This eliminates the need for operators to spend significant time configuring various attribution analysis tools, thus lowering the barrier to entry for using attribution analysis tools. The DSL object is encapsulated into an attribution analysis request, and at least one target attribution analysis tool is determined based on the user query. This enables the selection of the appropriate attribution analysis tool according to user needs, integrating entry points for different scenarios of index fluctuation attribution analysis, and mapping the attribution analysis request to at least one determined attribution analysis tool. A target attribution analysis tool performs attribution analysis based on mapped attribution analysis requests, obtaining at least one attribution analysis result. Finally, leveraging the logical reasoning capabilities of a large language model, combined with the attribution analysis results from various target attribution analysis tools, user questions, and a pre-set second text template, a result summary request is generated for further heuristic analysis, outputting a summary of the attribution results for the attributable indicators. This combines the summarization and text generation capabilities of the large language model to provide a final summary of the various attribution analysis results, eliminating the need for operational personnel to perform comprehensive analysis of the results from different attribution analysis tools. This not only improves the efficiency of attribution analysis but also reduces reliance on operational personnel. Furthermore, by combining various target attribution analysis tools with a large language model to complete the corresponding attribution analysis, this application also addresses the risk of errors that may arise from relying solely on large language models for attribution analysis in existing technologies, thereby improving the accuracy of attribution analysis.

[0102] Optionally, the semantic analysis unit provided in this application embodiment is further used for:

[0103] The DSL object is validated using pre-defined specific business rules, which include DSL definition information and preset required parameters and their dimension enumeration values. If the DSL object passes the validation, the step of generating a corresponding attribution analysis request based on the DSL object is executed.

[0104] Optionally, at least one target attribution analysis tool is determined based on the user's query, and the attribution analysis request is mapped to a semantic analysis unit on the target attribution analysis tool, specifically for:

[0105] Determine the user's intent based on the user's question; determine if the user intent is empty; if the user intent is not empty, determine at least one target attribution analysis tool from the pre-set attribution analysis tools based on the intent enumeration and its indicator names indicated in the user intent; route the attribution analysis request to the target attribution analysis tool.

[0106] Optionally, the semantic analysis unit provided in this application embodiment is further configured to: if the user intent is empty, determine each pre-set attribution analysis tool as the target attribution analysis tool.

[0107] Optionally, the attribution analysis request includes at least the dimension filtering conditions for the attributable indicator, the current time interval of the attributable indicator, and the concurrent time interval. Each target attribution analysis tool performs attribution analysis based on the request to obtain the corresponding attribution analysis results. This attribution analysis unit is specifically used for:

[0108] Based on the dimensional filtering conditions of the attributable indicator, the current time interval and the same period time interval of the attributable indicator, the attributable indicator data corresponding to the attributable indicator is obtained; based on the filtering conditions of each target attributive analysis tool, the target attributable indicator data that meets the requirements of each target attributive analysis tool is filtered from the attributable indicator data; attributive analysis is performed by each target attributive analysis tool based on the corresponding target attributable indicator data to obtain the attribution analysis results.

[0109] Optionally, the selection criteria for the target attribution analysis tool include the indicator fluctuation information, dimensional influence coefficient, and dimensional item fluctuation contribution rate corresponding to the target attribution analysis tool. Attribution analysis units that meet the target attribution indicator data for each target attribution analysis tool are selected from the attribution indicator data. Specifically, this is used for:

[0110] For each target attribution analysis tool, obtain the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to the target attribution analysis tool; based on the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to each target attribution analysis tool, obtain the target attribution indicator data that meets the requirements of each target attribution analysis tool from the attribution indicator data.

[0111] This application provides an electronic device, such as... Figure 3 As shown, the electronic device includes a processor 301 and a memory 302. The memory 302 is used to store program code and data for the attribution analysis of index fluctuations of the fused large language model of the device. The processor 301 is used to call the program instructions in the memory to execute the steps shown in the attribution analysis method for index fluctuations of the fused large language model in the above embodiment.

[0112] This application provides a storage medium including a stored program, wherein, during program execution, the device where the storage medium is located executes the index fluctuation attribution analysis method shown in the above embodiments for integrating large language models.

[0113] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0114] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0115] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0116] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for attribution analysis of index fluctuations integrating a large language model, characterized in that, The method includes: The corresponding attribution metric is determined based on the user's question, and the pre-set first text template and the attribution dimension metadata corresponding to the attribution metric are obtained. A semantic analysis request is generated based on the user question, the first text template, and the attribution dimension metadata of the attribution metric to be attributed. The semantic analysis request is processed by a large language model to obtain a DSL object, and a corresponding attribution analysis request is generated based on the DSL object. The large language model uses its information extraction capabilities to extract the dimension filtering conditions, the current time interval and the concurrent time interval of the attribution indicator corresponding to the user question and the attribution dimension metadata in the semantic analysis request. Based on the example answer in the reference example carried by the semantic analysis request, the dimensional filtering conditions, the current time interval and the concurrent time interval of the attribution indicator are generated and output as a JSON format DSL object. The process includes: determining at least one target attribution analysis tool based on the user's question, and mapping the attribution analysis request to each target attribution analysis tool, including: determining the user's user intent based on the user's question; determining whether the user intent is empty; if the user intent is not empty, determining at least one target attribution analysis tool from a pre-set set of attribution analysis tools based on the intent enumeration and its metric names indicated in the user intent; routing the attribution analysis request to the target attribution analysis tool; and if the user intent is empty, determining each pre-set attribution analysis tool as a target attribution analysis tool. Attribution analysis is performed by each of the target attribution analysis tools according to the attribution analysis request, and at least one attribution analysis result is obtained. A result summary request is generated based on a pre-set second text template, the user question, and at least one of the attribution analysis results. The attribution result summary of the index to be attributed is then output by the large language model based on the result summary request.

2. The method according to claim 1, characterized in that, Before generating the corresponding attribution analysis request based on the DSL object, the method further includes: The DSL object is validated using pre-set specific business rules; wherein, the specific business rules include DSL definition information and preset required parameters and their dimension enumeration values; If the DSL object passes the verification, the step of generating a corresponding attribution analysis request based on the DSL object is executed.

3. The method according to claim 1, characterized in that, The attribution analysis request includes at least the dimension filtering conditions of the attribution indicator, the current time interval of the attribution indicator, and the same period time interval. Each of the target attribution analysis tools performs attribution analysis based on the attribution analysis request to obtain corresponding attribution analysis results, including: Based on the dimension filtering conditions of the attributable indicator, the current time interval and the same period time interval of the attributable indicator, the attributable indicator data corresponding to the attributable indicator; Based on the screening criteria of each of the target attribution analysis tools, target attribution index data that meet the requirements of each target attribution analysis tool are selected from the attribution index data. The attribution analysis results are obtained by performing attribution analysis based on the corresponding target attribution index data using each of the aforementioned target attribution analysis tools.

4. The method according to claim 3, characterized in that, The selection criteria for the target attribution analysis tool include the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to the target attribution analysis tool. The step of filtering the attribution index data from the attribution index data to select the target attribution index data that meets the requirements of each target attribution analysis tool includes: For each of the target attribution analysis tools, obtain the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to the target attribution analysis tool; Based on the indicator fluctuation information, dimension influence coefficient, and dimension item fluctuation contribution rate corresponding to each target attribution analysis tool, target attribution indicator data that satisfies each target attribution analysis tool is obtained from the attribution indicator data.

5. A system for attributing index fluctuations using a large language model, characterized in that, The system includes: The user question receiving unit is used to determine the corresponding attribution metric based on the user question sent by the user, and to obtain the pre-set first text template and the attribution dimension metadata corresponding to the attribution metric. The semantic analysis unit is used to generate a semantic analysis request based on the user question, the first text template, and the attribution dimension metadata of the metric to be attributed; and to process the semantic analysis request through a large language model to obtain a DSL object. The large language model utilizes its information extraction capabilities to extract the dimension filtering conditions, the current time interval, and the concurrent time interval of the metric to be attributed based on the user question and the attribution dimension metadata corresponding to the metric to be attributed in the semantic analysis request. Based on the example answer in the reference example carried in the semantic analysis request, it generates and outputs a JSON-formatted DSL object from the dimension filtering conditions, the current time interval, and the concurrent time interval of the metric to be attributed. The process involves: generating an SL object; generating a corresponding attribution analysis request based on the DSL object; determining at least one target attribution analysis tool based on the user's query, and mapping the attribution analysis request to each target attribution analysis tool, including: determining the user's user intent based on the user's query; determining whether the user intent is empty; if the user intent is not empty, determining at least one target attribution analysis tool from a pre-set set of attribution analysis tools based on the intent enumeration and its metric names indicated in the user intent; routing the attribution analysis request to the target attribution analysis tool; and if the user intent is empty, determining each pre-set attribution analysis tool as a target attribution analysis tool. The attribution analysis unit is used to perform attribution analysis according to the attribution analysis request using each of the target attribution analysis tools to obtain at least one attribution analysis result. The attribution result summary unit is used to generate a result summary request based on a pre-set second text template, the user question, and at least one of the attribution analysis results, and output the attribution result summary of the index to be attributed according to the result summary request through the large language model.

6. The system according to claim 5, characterized in that, The semantic analysis unit is also used for: The DSL object is validated using pre-set specific business rules; wherein, the specific business rules include DSL definition information and preset required parameters and their dimension enumeration values; if the DSL object passes the validation, the step of generating a corresponding attribution analysis request based on the DSL object is executed.

7. An electronic device, characterized in that, include: A processor and a memory are connected via a communication bus; wherein the processor is used to call and execute a program stored in the memory; The memory is used to store a program for implementing the index fluctuation attribution analysis method for fusion of large language models as described in any one of claims 1-4.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for performing the index fluctuation attribution analysis method according to any one of claims 1-4.

Images