Data processing method and device based on large model

Abstract

The invention discloses a data processing method and device based on a large language model. The method comprises the following steps: extracting a key entity from a natural language input by a user by adopting the large language model; calculating semantic similarity between the key entity and each index name in the industry index knowledge graph, and determining a preset number of index names as candidate indexes based on semantic similarity ranking; adopting a large language model to judge the index type of the key entity and extract the time corresponding to the key entity, retrieving the index name related to the key entity of the combination type in the candidate indexes and deducing, and retrieving the index name related to the key entity of the generic type in the candidate indexes and associating the index name with the key entity of the generic type; constructing a combined expression based on the time and the retrieval result corresponding to the key entity, and converting the combined expression into a database query statement; and executing the database query statement to obtain a query result. According to the method, data query can be accurately carried out according to the natural language of the user, and omission is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information retrieval, and in particular to a data processing method and device based on a large model. BACKGROUND

[0002] In the current digital era, data has become a key basis for enterprise decision-making. Traditional data analysis methods usually rely on professional data analysts to extract and analyze data through complex manual operations such as writing SQL, using BI tools drag-and-drop, and query language, which has made it difficult to meet the real-time analysis needs of non-technical business personnel.

[0003] To facilitate the real-time analysis needs of non-technical business personnel, NL2SQL technology has emerged in the prior art. NL2SQL technology can convert natural language into structured SQL statements, reducing the threshold for data queries, and allowing non-technical personnel to directly query data through everyday language.

[0004] However, NL2SQL technology still has problems with intent understanding, and cannot accurately match derived indicators and generalized indicators appearing in the proposed question, which may result in missing and incorrect queries. SUMMARY

[0005] The embodiments of the present application provide a data processing method and device based on a large model, which solves the technical problem of missing and incorrect queries in the prior art.

[0006] In a first aspect, the embodiments of the present application provide a data processing method based on a large model, which includes: extracting key entities from user input natural language using a large language model; calculating the semantic similarity between the key entities and each indicator name in an industry indicator knowledge graph, and determining a predetermined number of the indicator names as candidate indicators based on the semantic similarity ranking; using the large language model to determine the indicator type of the key entities and extract the time corresponding to the key entities, wherein the indicator type includes a combination type, a generic type, and an exact match type, and retrieving the indicator name related to the key entity of the combination type in the candidate indicators and performing deduction, and retrieving the indicator name related to the key entity of the generic type in the candidate indicators and performing association; constructing a combination expression based on the time corresponding to the key entities and the retrieval result, and converting the combination expression into a database query statement according to a predetermined rule; executing the database query statement to obtain a query result.

[0007] With reference to the first aspect, in a possible implementation manner, the method further includes: determining a data form of the query result; if the query result is single data, determining that an output format is single data overview and interpretation; if the query result is comparative data, determining that the output format is a comparative table and analysis; and if the query result is time series or multi-index data, determining that the output format is a trend chart and a deep report.

[0008] With reference to the first aspect, in a possible implementation manner, the method further includes: determining a visualization form based on the data form of the query result, and outputting a graph in the visualization form on a dialogue context; wherein the visualization form includes a column chart, a line chart, and a heat map.

[0009] With reference to the first aspect, in a possible implementation manner, the method further includes: performing inductive analysis and accurate data interpretation on the graph in the visualization form by using the large language model, to generate a language description conclusion.

[0010] With reference to the first aspect, in a possible implementation manner, after the key entity is extracted from the natural language input by the user by using the large language model, the method further includes: performing semantic normalization on the key entity in a form of a synonym, an abbreviation, or a combined calculation expression.

[0011] With reference to the first aspect, in a possible implementation manner, the deriving the key entity in the combination type includes: generating the key entity in the combination type according to a candidate index and a predefined business rule; wherein the business rule includes index operation logic and semantic constraint conditions.

[0012] With reference to the first aspect, in a possible implementation manner, the associating the key entity in the generic type includes: associating the key entity in the generic type to the candidate index by using semantic generalization and context mapping.

[0013] In a second aspect, an embodiment of the present application provides a data processing apparatus based on a large model, the data processing apparatus based on the large model comprising: an entity extraction module configured to extract a key entity from a natural language input by a user using a large language model; calculate semantic similarity between the key entity and each index name in an industry index knowledge graph, and determine a preset number of index names as candidate indexes based on the semantic similarity ranking; an index retrieval module configured to determine an index type of the key entity and extract a time corresponding to the key entity using the large language model, wherein the index type comprises a combination type, a generic type, and an exact match type, and retrieve the index names related to the key entity of the combination type from the candidate indexes and perform deduction, and retrieve the index names related to the key entity of the generic type from the candidate indexes and perform association; a sentence generation module configured to construct a combination expression based on the time corresponding to the key entity and the retrieval result, and convert the combination expression into a database query statement according to a preset rule; and a data interpretation module configured to execute the database query statement to obtain a query result.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising: one or more processors; and a memory storing computer executable instructions that, when executed by the one or more processors, cause the one or more processors to perform the data processing method based on the large model as described in the first aspect or any possible implementation manner of the first aspect.

[0015] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium storing computer readable instructions, the computer readable instructions being executed by a computer to implement the data processing method based on the large model as described in the first aspect or any possible implementation manner of the first aspect.

[0016] The embodiment of the present application provides a data processing method based on a large model, which can extract a key entity from a natural language by using a large language model, and determine candidate indexes by calculating semantic similarity between the key entity and each index name in an industry index knowledge graph, and match the key entity of a combination type, a generic type, and an exact match type from the candidate indexes, so that the natural language is accurately understood and accurately matched with the index name in the industry index knowledge graph, avoiding the situation of missing due to being unable to understand, and then the query result obtained by using the generated query language is more accurate, and the data processing method based on the large model does not need to predefine all analysis paths, which significantly reduces the maintenance cost of the knowledge base. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the description of the embodiments of the present application will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 The flowchart of the data processing method based on a large model provided by the present application is shown in the figure. Figure 2 The line chart generated in the embodiments of the present application for the question "What is the cumulative tax and profit situation this year and last year?" is shown in the figure. Figure 3 The line chart generated in the embodiments of the present application for the question "Calculate the same ring ratio of the actual collection amount in February 2025" is shown in the figure. Figure 4 The schematic diagram of the data processing device based on a large model provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0020] The present application provides a data processing method based on a large model, which comprises steps S101 to S105.

[0021] S101, extracting key entities from the natural language input by the user using a large language model.

[0022] After the user inputs the natural language, the large language model extracts the key entities in the natural language according to the preset prompt word template. Exemplarily, the large language model is DeepSeek-R1-Distill-Qwen-32B.

[0023] For example, the natural language input by the user is "comparison of sales and profit of the company in the last quarter", after step S101 is executed, the large language model extracts the key entities "company", "last quarter", "sales", "profit" and "comparison" from the natural language.

[0024] S102, calculating the semantic similarity between the key entities and each index name in the industry index knowledge graph, and determining a preset number of index names as candidate indexes based on the semantic similarity ranking.

[0025] Specifically, each indicator name in the knowledge graph of key entities and industry indicators is embedded into the vector semantic space. The semantic similarity between the key entities and each indicator name in the knowledge graph of industry indicators is calculated, and the semantic similarity is ranked from high to low. The top N indicator names are selected as candidate indicators, where N is a preset positive integer.

[0026] E(·) represents the embedding function, which is used to convert indicator names and key entities into vectors; Cosine similarity measures the angle between two vectors, with values ​​ranging from -1 to 1. A larger value indicates that the vectors are more similar. `target` represents the indicator name, an element belonging to the indicator name set in the industry indicator knowledge graph; `target` represents the key entity. In the entity extraction module, the formula for identifying the N most similar indicator names as candidate indicators is as follows:

[0027] in, is the indicator name, which is an element of the indicator name set in the industry indicator knowledge graph; target is the key entity; S is the N candidate indicator names selected from the highest to the lowest similarity.

[0028] For example, N can be 10.

[0029] S103. Use a large language model to determine the indicator type of the key entity and extract the time corresponding to the key entity. The indicator type includes combination type, generic type and exact match type. Also, retrieve the indicator names related to the key entity of the combination type from the candidate indicators and derive them. Retrieve the indicator names related to the key entity of the generic type from the candidate indicators and associate them.

[0030] S104. Construct a combined expression based on the time and search results corresponding to the key entities, and convert the combined expression into a database query statement according to preset rules.

[0031] For example, the default rule can be: SELECT * FROM `indicator` WHERE (indicator_name IN (?)) AND month = ? AND year = ?; the database query statement can be an SQL statement.

[0032] S105. Execute the database query statement to obtain the query results.

[0033] The big model-based data processing method can extract key entities from natural language through a large language model, and determine candidate indicators by calculating the semantic similarity between the key entities and each indicator name in the industry indicator knowledge graph, and match the key entities of the combination type, the generic name type and the accurate matching type from the candidate indicators, so that the natural language is accurately understood and accurately matched with the indicator name in the industry indicator knowledge graph, avoiding the omission of situations that cannot be understood, and then the query result obtained by the generated query language is more accurate, and the big model-based data processing method does not need to predefine all analysis paths, which significantly reduces the maintenance cost of the knowledge base.

[0034] The big model-based data processing method further comprises: judging the data form of the query result; if the query result is a single data, determining that the output format is a single data overview and interpretation; if the query result is comparative data, determining that the output format is a comparison table and analysis; if the query result is time series or multi-indicator data, determining that the output format is a trend chart and a deep report.

[0035] By adopting different output formats for different data forms, the query result can be presented to the query personnel in a more reasonable display manner, and the query personnel can understand and absorb the data in the most efficient way.

[0036] Further, the big model-based data processing method further comprises: determining a visualization form based on the data form of the query result and outputting a graph in the visualization form in the dialogue context; wherein the visualization form includes a column chart, a line chart and a heat map.

[0037] Specifically, the generated chart can be converted into a format that can be directly displayed in the dialogue interface through Base64 encoding, and embedded in the dialogue context, so that the query personnel can obtain intuitive and in-depth data insight results in the dialogue process, improving the interactivity and user experience of data analysis.

[0038] For example, for comparative data, a column chart is preferred to intuitively display the differences between different categories; for time series data, a line chart is preferred to clearly present the trend of data change over time.

[0039] Further, the big model-based data processing method further comprises: using a large language model to conduct inductive analysis and accurate data interpretation on the graph in the visualization form, and generating a language description conclusion.

[0040] By conducting inductive analysis and accurate data interpretation on the image, the query personnel can quickly understand the data and rules displayed in the chart by reading the language description result while viewing the chart.

[0041] After step S101, the data processing method based on a large model further includes: performing semantic normalization on the key entities in the form of synonyms, abbreviations, and combined calculation expressions.

[0042] This step can map key entities in different expression forms to a unified semantic space, avoid understanding bias caused by expression differences, and facilitate subsequent calculation of semantic similarity between the key entities and each index name in the industry index knowledge graph in step S102, so that the semantic similarity of key entities in the same semantic but different expression forms is the same, so that the data processing method based on a large model can more flexibly process user input and cover a wider range of application scenarios.

[0043] For example, "NP" and "net profit margin" are normalized to "net profit margin".

[0044] The derivation of the key entity of the combined type in step S103 includes: generating the key entity of the combined type according to the candidate index and the pre-defined business rule; wherein the business rule includes index operation logic and semantic constraint conditions.

[0045] Illustratively, the key entity is "sales growth rate", and according to the "sales" and "time" indexes in the candidate index, the "sales growth rate" is derived according to the calculation formula of the growth rate.

[0046] The association of the key entity of the generic type in step S103 includes: associating the key entity of the generic type to the candidate index through semantic generalization and context mapping.

[0047] Illustratively, the key entity is "financial index", and after semantic generalization and context mapping, the specific financial related indexes such as "sales", "profit" and "cost" are associated with the key entity.

[0048] Next, taking a certain cigarette factory query system as an example, the data processing method based on a large model provided by the embodiments of the present application will be described in detail.

[0049] The cigarette factory is a typical enterprise in the tobacco industry, and it needs to process a large amount of production, sales, and financial index data on a daily basis. The traditional query mode relies on data analysts to manually write SQL, which has problems such as response delay, high operation threshold, and difficulty for non-technical personnel to use.

[0050] The construction process of the industry index knowledge graph of the cigarette factory query system is as follows.

[0051] Data source and preprocessing: The index data in the production management system, enterprise resource planning system, and sales management system of the cigarette factory in the past five years were integrated, covering 3847 core indicators such as production (e.g., comprehensive energy consumption of power center, actual warehouse quantity), sales (e.g., monthly transportation quantity, finished product delivery quantity), and finance (e.g., total tax and profit, total profit). After removing duplicate and invalid index data, a standardized index dataset was formed.

[0052] Vector embedding and semantic space construction: The Qwen3-Embedding model was used as the embedding function E(*) to convert the index names into vectors, and a semantic correlation matrix was constructed based on the cosine similarity algorithm. For example, the cosine similarity between “comprehensive energy consumption of power center” and “monthly specific energy consumption of comprehensive energy consumption of power center” was 0.92, which was much higher than the preset threshold of 0.7, and was determined as a highly correlated index. Definition of index relationship: Three core relationships between indexes were defined: hierarchical relationship, synonymous relationship, and combination and calculation relationship.

[0053] Hierarchical relationship: For example, “total tax and profit” includes “tax” and “total profit”, and “tax” is further divided into “consumption tax”, “value-added tax”, and “urban construction tax”; Synonymous relationship: For example, “warehouse quantity”, “actual warehouse quantity”, and “finished product warehouse quantity” are unified to the standard index name “actual warehouse quantity”; Combination and calculation relationship: For example, the calculation rules of “completion rate = completion quantity / planned quantity” and “growth rate = (current value - base value) / base value” are defined. First, the DeepSeek-R1-Distill-Qwen-32B large language model was used to extract key entities from the user's natural language input. The user's natural language input is shown in Table 1.

[0054] Table 1 When performing this step, the prompt word template input to the large language model is as follows.

[0055] The following Table 2 shows the extraction results of the key entities.

[0056] Table 2 Then, the cosine_similarities formula was used to calculate the semantic similarity between the key entities and each index name in the industry index knowledge graph, and based on the semantic similarity, the top 10 index names were ranked from high to low to determine the candidate index names. The cosine_similarities formula is as follows:

[0057] E(·) represents the embedding function, Cosine similarity measures the cosine of the angle between two vectors. The values range from -1 to 1, with higher values indicating more similar vectors.

[0058] Table 3 below shows the corresponding candidate indicators for each question.

[0059] Table 3 Serial number Question Candidate indicators 1 What is the cumulative tax and profit situation this year and the same period last year? [Total tax and profit, tax, total tax and profit, urban construction tax, other taxes, consumption tax, assessment tax (excluding marketing 5 expenses), value-added tax, this period's payable value-added tax, total profit] 2 What is the transportation quantity in April 2025? ['Monthly transportation quantity'] 3 Calculate the same ring ratio of actual warehouse quantity in February 2025 [Actual warehouse quantity, planned warehouse quantity, warehouse quantity of cigarette finished products, total monthly finished product warehouse quantity, total annual finished product warehouse quantity, warehouse quantity of finished tobacco, total annual auxiliary material warehouse quality detection inspection number, monthly warehouse batch, total annual auxiliary material warehouse safety detection inspection number, annual warehouse quantity of finished tobacco] Secondly, the DeepSeek-R1-Distill-Qwen-32B large language model is used to determine the type of each key entity and extract the corresponding time of the key entity. The indicator name related to the key entity of the combination type is retrieved and deduced in the candidate indicators. The indicator name related to the key entity of the generic type is retrieved and associated in the candidate indicators.

[0060] Then, based on the time corresponding to the key entity and the retrieval results, a combination expression is constructed.

[0061] When performing the above steps, the prompt word template input to the large language model is as follows.

[0062] Table 4 below shows the combination expression corresponding to each question.

[0063] Table 4 Secondly, according to the preset rule: SELECT * FROM `indicator` WHERE (indicator_name IN(?)) AND month =? AND year =?, the combination expression is converted into a SQL statement.

[0064] Finally, each SQL statement is executed to obtain the query result. Table 5 below shows the SQL statement generated for each question and the query result.

[0065] Table 5 Serial number Question SQL statement Query result 1 What is the cumulative tax and profit situation this year and the same period last year? SQL template: SELECT * FROM `indicator` WHERE (indicator_name IN (?)) AND month IN (?,?,?,?,?,?,?,?) AND year IN (?,?) parameter: [total tax and profit, 01, 02, 03, 04, 05, 06, 07, 08, 2025, 2024] indicator_nameindicator_valueyearmonthTotal tax and profit 43,203 2024Total tax and profit 47,565 2024Total tax and profit 43,169 2024Total tax and profit 48,029 2024Total tax and profit 45,411 2024Total tax and profit 50,670 2024Total tax and profit 68,418 2025Total tax and profit 32,533 2025Total tax and profit 60,030 2025Total tax and profit 52,464 2025Total tax and profit 51,157 2025 2 What is the transportation quantity in April 2025? SQL template: SELECT * FROM `indicator` WHERE (indicator_name IN (?)) AND month =? AND year =? parameter: [monthly transportation quantity, 04, 2025] indicator_nameindicator_valueyearmonthMonthly transportation quantity 137636 2025 3 Calculate the same ring ratio of actual warehouse quantity in February 2025 SQL template: SELECT * FROM `indicator` WHERE (indicator_name IN (?)) AND month =? AND year IN (?,?) parameter: [actual warehouse quantity, 02, 2025, 2024] indicator_nameindicator_valueyearmonthactual_inbound_quantity12000020241actual_inbound_quantity13500020242actual_inbound_quantity15000020251actual_inbound_quantity16000020252 In addition, according to the following formula g(D), combined with the data characteristics of the query result of each question, the visualization form is dynamically selected. The query result of question 1 and question 3 generates a line chart. The line chart of question 1 is shown in Figure 2 , question 3 is shown in Figure 3 ; and Figure 2 and Figure 3 are automatically embedded in the conversation in Base64 encoding context and conclusion. Wherein, the formula g(D) is as follows: .

[0066] And, the large language model is adopted to perform inductive analysis and accurate data interpretation on the graph in the visual form, to generate a language description conclusion. Here, taking the question "Cumulative tax and benefit situation this year and last year?" as an example, the generated language description conclusion is as follows.

[0067] As shown in Figure 4 The embodiment of the application further provides a data processing apparatus 400 based on a large model, which comprises an entity extraction module 401, an index retrieval module 402, a sentence generation module 403 and a data interpretation module 404.

[0068] The entity extraction module 401 is configured to extract key entities from natural language input by a user by using a large language model; calculate semantic similarity between the key entities and each index name in an industry index knowledge graph, and determine a preset number of index names as candidate indexes based on the semantic similarity ranking; the index retrieval module 402 is configured to determine the index type of the key entities and extract the time corresponding to the key entities by using a large language model, wherein the index type comprises a combination type, a generic type and an accurate matching type, and the index retrieval module 402 is further configured to retrieve index names related to the key entities of the combination type in the candidate indexes and perform deduction, and retrieve index names related to the key entities of the generic type in the candidate indexes and perform association; the sentence generation module 403 is configured to construct a combination expression based on the time corresponding to the key entities and the retrieval result, and convert the combination expression into a database query statement according to a preset rule; and the data interpretation module 404 is configured to execute the database query statement to obtain a query result.

[0069] Further, the entity extraction module 401 is further configured to perform semantic normalization on the key entities in the form of synonyms, abbreviations and combination calculation expressions.

[0070] Further, the data interpretation module 404 is further configured to determine the data form of the query result; if the query result is a single data, the output format is determined to be a single data overview and interpretation; if the query result is comparative data, the output format is determined to be a comparison table and analysis; and if the query result is time series or multi-index data, the output format is determined to be a trend chart and a depth report.

[0071] The data processing apparatus 400 based on a large model further comprises an intelligent chart generation module: the intelligent chart generation module is configured to determine a visual form based on the data form of the query result, and output a graph in the visual form on a dialogue context; wherein the visual form comprises a column chart, a line chart and a heat map.

[0072] The data processing apparatus 400 based on a large model further comprises an index data interpretation module, which is configured to perform inductive analysis and accurate data interpretation on the graph in the visual form by using a large language model, to generate a language description conclusion.

[0073] The index retrieval module 402 derives the candidate index of the combination type, specifically comprising: generating the key entity of the combination type according to the candidate index and the predefined business rule; wherein the business rule comprises the index operation logic and the semantic constraint condition.

[0074] The index retrieval module 402 associates the candidate index of the generic type, specifically comprising: associating the key entity of the generic type to the candidate index through semantic generalization and context mapping.

[0075] The apparatuses or modules illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. For the convenience of description, the above apparatuses are described in various modules with functions respectively. In the implementation of the embodiments of the present application, the functions of the modules can be implemented in one or more software and / or hardware. Of course, the modules for implementing certain functions can also be combined to implement the modules.

[0076] The embodiments of the present application also provide an electronic device, which comprises: one or more processors, and a memory storing computer executable instructions, the computer executable instructions causing the one or more processors to execute the big model based data processing method as described above when executed by the one or more processors.

[0077] The embodiments of the present application also provide a computer readable storage medium, which stores computer readable instructions, the computer readable instructions being executed by a computer to implement the big model based data processing method as described above.

[0078] The storage medium described above includes but is not limited to random access memory (English: Random Access Memory; abbreviation: RAM), read only memory (English: Read-Only Memory; abbreviation: ROM), cache (English: Cache), hard disk (English: Hard Disk Drive; abbreviation: HDD) or memory card (English: Memory Card).

[0079] The various embodiments in the specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment mainly explains the difference from other embodiments.

[0080] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can still be modified, or some or all of the technical features thereof can be replaced by equivalents; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present application.

Claims

1. A data processing method based on a large model, characterized in that, include: Key entities are extracted from the natural language input by the user using a large language model; Calculate the semantic similarity between the key entity and each indicator name in the industry indicator knowledge graph, and determine a preset number of indicator names as candidate indicators based on the semantic similarity ranking; The large language model is used to determine the indicator type of the key entity and extract the time corresponding to the key entity. The indicator type includes combination type, generic type and exact match type. The indicator name related to the key entity of combination type is retrieved from the candidate indicators and deduced. The indicator name related to the key entity of generic type is retrieved from the candidate indicators and associated. Based on the time and search results corresponding to the key entities, a combined expression is constructed, and the combined expression is converted into a database query statement according to preset rules; Execute the database query statement to obtain the query results.

2. The method according to claim 1, characterized in that, The method further includes: Determine the data format of the query results; If the query result is a single data entry, the output format is determined to be a single data entry overview and interpretation. If the query results are comparison data, determine the output format as a comparison table and analysis; The query results are time-series or multi-indicator data, and the output format is determined to be a trend chart and an in-depth report.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Based on the data format of the query results, a visualization format is determined and output as a graphic in the dialog context; wherein, the visualization format includes bar charts, line charts, and heatmaps.

4. The method according to claim 3, characterized in that, The method further includes: The large language model is used to perform inductive analysis and precise data interpretation on the visualized graphics, generating linguistic descriptions and conclusions.

5. The method according to claim 1, characterized in that, After extracting key entities from the user-input natural language using a large language model, the method further includes: Semantic normalization is performed on the key entities in the form of synonyms, abbreviations, and combined computational expressions.

6. The method according to claim 1, characterized in that, The derivation of the key entity of the combined type includes: Based on candidate indicators, the key entities of the combined type are generated according to predefined business rules; wherein the business rules include indicator calculation logic and semantic constraints.

7. The method according to claim 1 or 6, characterized in that, The association of the key entities of the generic type includes: Through semantic generalization and context mapping, the key entities of the generic type are associated with the candidate indicators.

8. A data processing device based on a large model, characterized in that, include: The entity extraction module is used to extract key entities from the natural language input by the user using a large language model; Calculate the semantic similarity between the key entity and each indicator name in the industry indicator knowledge graph, and determine a preset number of indicator names as candidate indicators based on the semantic similarity ranking; The indicator retrieval module is used to determine the indicator type of the key entity using the large language model and extract the time corresponding to the key entity, wherein the indicator type includes a combination type, a generic type and an exact match type, and to retrieve the indicator name related to the key entity of the combination type from the candidate indicators and derive it, and to retrieve the indicator name related to the key entity of the generic type from the candidate indicators and associate it. The statement generation module is used to construct a combined expression based on the time and search results corresponding to the key entity, and convert the combined expression into a database query statement according to preset rules. The data interpretation module is used to execute the database query statement and obtain the query results.

9. An electronic device, characterized in that, include: One or more processors; as well as A memory storing computer-executable instructions, which, when executed by the one or more processors, cause the one or more processors to perform the data processing method based on a large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a computer, implement the data processing method based on a large model as described in any one of claims 1 to 7.

Images