A method, apparatus, storage medium, and computer equipment for automated text tagging in spreadsheets based on a large language model.

By using a large language model-based automated text tagging method for spreadsheets, the problems of low efficiency and lack of flexibility in unstructured text tagging in the fields of smart healthcare and fintech are solved, and an efficient and flexible automated tagging process is achieved.

CN123117017APending Publication Date: 2026-11-17CHINA PING AN LIFE INSURANCE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611275967.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the fields of smart healthcare and fintech, existing technologies for automated labeling of unstructured text are inefficient and struggle to flexibly meet diverse labeling needs. Traditional manual and rule-based labeling methods suffer from low efficiency and insufficient flexibility.

Method used

An automated text tagging method for spreadsheets based on a large language model is adopted. This method parses spreadsheet files into structured data and segments them according to a preset block strategy. It then uses a large language model to process large-scale spreadsheet data and dynamically constructs prompt words based on the tagging requirements of the user's natural language input, thereby achieving automated batch tagging.

Benefits of technology

It significantly improves tagging efficiency, can handle complex text situations, and has good flexibility and versatility. Users only need natural language input to adapt to different tagging task types without redesigning rules.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117017A_ABST
    Figure CN123117017A_ABST
Patent Text Reader

Abstract

This application relates to the field of data processing technology, specifically disclosing a method, apparatus, storage medium, and computer device for automated text tagging of spreadsheets based on a large language model. The method includes: acquiring a spreadsheet file to be processed and acquiring the tagging requirement text; parsing the spreadsheet file into structured data and dividing it into multiple data blocks; for each data block, constructing prompt words based on the tagging requirement text and multiple records contained in the data block, sending the prompt words to the large language model, receiving the returned tagging results, parsing the tagging results to obtain the tagging content corresponding to each record, and storing each record and its tagging content in association; after all data blocks have been processed, combining each record and its tagging content into result data containing a column of original text and a column of tagging content, and outputting it. This application can be applied to automatic text tagging scenarios in spreadsheets in the fields of fintech and smart healthcare.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus, storage medium, and computer equipment for automated text marking in spreadsheets based on a large language model. Background Technology

[0002] With the deepening of digital transformation, various industries are facing the common challenge of leaping from "business datafication" to "data intelligence." The automated processing of massive amounts of unstructured text data has become a key bottleneck restricting the release of data value. This problem is particularly prominent in the data-intensive fields of smart healthcare and fintech. In smart healthcare, unstructured texts such as electronic medical records, imaging reports, and clinical follow-up records constitute the data basis for clinical decision support and medical research. In fintech, massive amounts of text such as corporate financial statement notes, risk assessment reports, compliance review documents, and credit approval materials also carry core business information and decision-making basis. However, while these text data contain enormous value, they are limited by their unstructured form. Medical texts are filled with professional terminology, ambiguous abbreviations, and cross-departmental differences in expression, while financial texts involve complex industry contexts, regulatory clauses, and diverse disclosure formats. Text tagging is an indispensable key step in transforming these domain-specific unstructured texts into structured knowledge that can be accurately understood and computed by computers. However, the common dilemma faced by both fields is that tagging requirements are highly diverse and frequently changing. Against this backdrop, how to build an efficient, flexible, and cross-domain-adaptable automated text tagging solution has become a fundamental technical problem that urgently needs to be solved in the aforementioned fields.

[0003] Currently, industry and academia primarily rely on two approaches for text tagging, both of which have significant limitations. One approach is traditional manual annotation, where domain experts read and label text line by line. While this method boasts high accuracy, it is extremely inefficient. The other approach uses rule-based auxiliary annotation tools. These solutions can achieve pre-annotation or semi-automated processing to some extent, but their intelligence is limited by the scope of the rules. Because text data often contains complexities such as non-standard expressions, ambiguous professional abbreviations, and contextual dependencies—for example, the same symptom may be described differently in medical records from different departments, or the same financial indicator may be disclosed in completely different formats in the financial statement footnotes of different industries—rule-based annotation often struggles to flexibly address diverse tagging needs. When the tagging task changes, rules typically need to be redesigned, resulting in a severe lack of flexibility. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, storage medium, and computer device for automated text tagging of spreadsheets based on a large language model. By parsing spreadsheet files into structured data and segmenting them according to a preset segmentation strategy, the large language model can overcome the limitations of the context window and process large-scale spreadsheet data in batches. At the same time, during the segmentation process, prompt words are dynamically constructed based on the tagging requirements input by the user in natural language, enabling the system to flexibly adapt to different tagging task types. This solves the technical problems of low efficiency of manual tagging and the inability of rule-based tagging to flexibly cope with diverse tagging needs in traditional solutions.

[0005] According to one aspect of this application, a method for automated text markup in spreadsheets based on a large language model is provided, comprising: Obtain the spreadsheet file to be processed, and obtain the marked requirement text corresponding to the spreadsheet file, which is input by the user in natural language; The spreadsheet file is parsed into structured data, and the structured data is divided into multiple data blocks according to a preset block division strategy. Each row in the structured data corresponds to a record in the spreadsheet file, each record contains a text field to be labeled, and each data block includes multiple records. For each data block, a prompt word is constructed based on the tagging requirement text and multiple records contained in the data block. The prompt word is sent to the large language model, the tagging result returned by the large language model is received, the tagging result is parsed to obtain the tagging content corresponding to each record, and each record and its corresponding tagging content are associated and stored. After all data blocks have been processed, each record and its corresponding tag content are combined into result data containing the original text column and the tag content column, and then output.

[0006] According to another aspect of this application, an automated text tagging device for spreadsheets based on a large language model is provided, comprising: The data acquisition module is used to acquire the spreadsheet file to be processed and to acquire the marked requirement text corresponding to the spreadsheet file, which is input by the user in natural language. The data partitioning module is used to parse the spreadsheet file into structured data and divide the structured data into multiple data blocks according to a preset partitioning strategy. Each row in the structured data corresponds to a record in the spreadsheet file, each record contains a text field to be labeled, and each data block includes multiple records. The tagging module is used to construct prompt words for each data block based on the tagging requirement text and multiple records contained in the data block, send the prompt words to the large language model, receive the tagging results returned by the large language model, parse the tagging results to obtain the tagging content corresponding to each record, and store each record and its corresponding tagging content in association. The results output module is used to combine each record and its corresponding tag content into result data containing the original text column and the tag content column after all data blocks have been processed, and then output the result data.

[0007] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described automatic text marking method for spreadsheets based on a large language model.

[0008] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described automatic text marking method for spreadsheets based on a large language model.

[0009] By employing the aforementioned technical solutions, this application provides a method, apparatus, storage medium, and computer device for automated text tagging of spreadsheets based on a large language model. By combining spreadsheet file parsing with block processing, it achieves fully automated batch tagging of large-scale text data, significantly improving tagging efficiency and enabling tasks that previously relied on manual, line-by-line processing to be completed automatically in a short time. Simultaneously, by utilizing the deep semantic understanding capabilities of the large language model to replace traditional keyword matching or rule-driven methods, it effectively addresses complex text situations such as non-standard expressions, ambiguous professional abbreviations, and contextual dependencies, overcoming the inherent limitations of rule-based tagging in covering diverse expressions. Furthermore, users only need to input tagging requirements in natural language, and the system can automatically adapt to different tagging task types. When tagging requirements change, there is no need to redesign rules, demonstrating good flexibility and versatility.

[0010] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0011] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1This illustration shows an application environment diagram of an automated text tagging method for spreadsheets based on a large language model according to an embodiment of this application; Figure 2 The illustration shows a flowchart of an automated text tagging method for spreadsheets based on a large language model, according to an embodiment of this application. Figure 3 This illustration shows a structural schematic diagram of an automated text marking device for spreadsheets based on a large language model, according to an embodiment of this application. Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0012] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0013] This application provides an automated text tagging method for spreadsheets based on a large language model, which can be applied to applications such as... Figure 1 The application environment shown. For example... Figure 1 As shown, users access the system's front-end interface through client devices (such as personal computers, laptops, or tablets), upload the spreadsheet file to be labeled to the back-end server, and input the labeling requirements text in natural language on the interface. Upon receiving the spreadsheet file, the back-end server parses it into structured data, then divides the structured data into multiple data blocks according to a preset segmentation strategy. For each data block, it constructs prompt words and calls a large language model to obtain the labeling results returned by the large language model. Finally, it combines each record and its corresponding labeling content into result data containing columns of original text and labeled content, and provides a download address for the client. The client device can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices; the back-end server can be an independently deployed physical server, a cloud server instance, or a containerized computing node; the large language model can be accessed by calling an external API interface, or an open-source large language model instance can be deployed locally on the back-end server. The technical solution of this application is described in detail below through specific embodiments.

[0014] This embodiment provides a method for automated text tagging in spreadsheets based on a large language model, such as... Figure 2 As shown, the method includes: Step 101: Obtain the spreadsheet file to be processed, and obtain the marked requirement text corresponding to the spreadsheet file, which is input by the user in natural language.

[0015] Step 102: Parse the spreadsheet file into structured data, and divide the structured data into multiple data blocks according to a preset block division strategy. Each row in the structured data corresponds to a record in the spreadsheet file, each record contains a text field to be labeled, and each data block includes multiple records.

[0016] Step 103: For each data block, construct prompt words based on the tagging requirement text and the multiple records contained in the data block, send the prompt words to the large language model, receive the tagging results returned by the large language model, parse the tagging results to obtain the tagging content corresponding to each record, and store each record and its corresponding tagging content in association.

[0017] Step 104: After all data blocks have been processed, combine each record and its corresponding tag content into result data containing the original text column and the tag content column, and output it.

[0018] This application provides an automated text tagging method for spreadsheets based on a large language model. First, the spreadsheet file to be processed can be obtained. This file can be user-uploaded or a local file, such as thousands of electronic medical record follow-up records accumulated by a hospital department, or qualitative descriptive text from corporate loan application materials collected by a financial institution. Furthermore, the tagging request text described by the user in natural language can be obtained. This request text can characterize what kind of processing the user wants to perform on these texts, such as "Please determine which body system each patient's chief complaint belongs to" or "Please assess whether the implicit credit risk level in the document is high, medium, or low." The purpose of this step is to simultaneously provide the system with the user's subjective intent and the spreadsheet file to be processed, serving as the starting point for the entire tagging process. In a specific embodiment, the spreadsheet file can be an Excel file.

[0019] Subsequently, the spreadsheet file is parsed and converted into a structured format that can be quickly read internally. Each row corresponds to an independent record in the original table, and each column corresponds to a different text field within that record. The text content to be labeled is located in a specific column. Simultaneously, all records are divided into multiple data blocks according to a preset chunking strategy, with each data block containing several records. This is because large language models have an upper limit on the amount of data they can process in a single input. If all records are filled with prompt words at once, it is easy to exceed the input length limit of the large language model and cause errors. Taking a medical scenario as an example, a detailed discharge summary is considered as one record, and each record may be thousands of words long. Given the limited context window, each data block may only be able to hold dozens of records.

[0020] Furthermore, data is extracted block by block. For multiple records in the current data block, the user's initial input text specifying the tagging requirement is combined with these records to form a structured prompt, which is then sent to the large language model, awaiting the model's response. The original result returned by the large language model is typically a tag structure containing the tagging content corresponding to multiple records. Therefore, it can be decomposed so that each original record can obtain its corresponding tagging content, and then the record and its tagging content are saved as a pair. In this way, users do not need to learn any complex command formats or programming syntax; they can express their needs entirely in everyday language. Subsequently, the system can convert the user's text specifying the tagging requirement into a prompt format that the large language model can strictly execute. For example, in a medical scenario, if a user says, "Classify these chief complaint descriptions according to respiratory system, digestive system, circulatory system, nervous system, and others," the system can assemble this instruction along with dozens of chief complaint text records in the current data block and send it to the large language model. It should be noted that a Large Language Model (LLM) is a deep learning-based AI model that learns the syntax, semantics, and contextual relationships of a language by training on massive amounts of text data, thereby gaining the ability to understand, generate, and process natural language text. Essentially, it is a deep neural network based on the Transformer architecture. After pre-training on a large amount of text corpus, it can complete corresponding tasks based on given input.

[0021] After all data blocks have completed the above calls and parsing, each original record and its corresponding tag content that were previously saved in pairs are recombined into a new two-dimensional table containing two columns: the first column stores the text field of the original record, and the second column stores the tag content generated by the large language model for the text field. Finally, the table is output to the user.

[0022] By applying the technical solution of this embodiment, and combining spreadsheet file parsing with chunked processing, fully automated batch labeling of large-scale text data is achieved, significantly improving labeling efficiency and enabling tasks that previously relied on manual, line-by-line processing to be completed automatically in a short time. Simultaneously, by utilizing the deep semantic understanding capabilities of a large language model to replace traditional keyword matching or rule-driven methods, it effectively addresses complex text situations such as non-standard expressions, ambiguous professional abbreviations, and contextual dependencies, overcoming the inherent limitations of rule-based labeling in covering diverse expressions. Furthermore, users only need to input their labeling requirements in natural language, and the system can automatically adapt to different labeling task types. When labeling requirements change, there is no need to redesign rules, demonstrating excellent flexibility and versatility.

[0023] Optionally, in this embodiment, step 103, "constructing prompt words based on the tagging requirement text and multiple records contained in the data block," includes: identifying the tagging task type and tagging dimension in the tagging requirement text; selecting a target template from multiple preset prompt word templates according to the tagging task type; mapping the tagging dimension to a tag option list in the target template; filling the tag option list into the tagging option area of ​​the target template; concatenating the text fields of multiple records in the data block according to the record order and filling them into the data input area of ​​the target template; and generating a complete prompt word containing the tagging task type, the tag option list, the text fields of the multiple records, and the output format constraints according to the preset output format constraints in the target template.

[0024] In this embodiment, upon receiving the tagging request text, a structured parsing process can be performed. Specifically, two key pieces of information can be identified: first, the tagging task type, i.e., whether the user wants to perform classification, rating, sentiment assessment, or text polishing; and second, the tagging dimension, i.e., the specific angle from which the user wants to evaluate the text, such as "classified by body system" or "rated by risk level." Then, based on the identified tagging task type, a matching target template can be selected from a set of preset prompt word templates, with different prompt word templates corresponding to different tagging task types.

[0025] Next, the labeled dimensions are transformed into a finite set of options that the large language model can choose from in its actual output, i.e., a list of labeled options. For example, in a medical scenario, if a user says "classify the chief complaint symptoms by body system," the system identifies the labeled dimension as "body system" and maps it to a list of mutually exclusive labeled options such as "respiratory system, digestive system, circulatory system, nervous system, and others." In a financial scenario, if a user says "assess the credit risk level of a company," the labeled dimension "credit risk" is mapped to a list of labeled options including "high, medium, and low." This is because if the large language model generates text freely without constraints, it may output various natural language expressions such as "high risk," "certain hidden dangers," and "not recommended for lending." These expressions have inconsistent formats, cannot be automatically written into the labeled columns of a spreadsheet, and lack comparability between outputs. By mapping the labeled dimensions to a list of labeled options in advance, the large language model can only choose from a given range of options, and the output is structured and can be directly filled into a table. After obtaining the list of label options, it is filled into the label option area in the target template, which is specifically used to limit the output range. At the same time, the text fields of multiple records in the current data block are concatenated into a continuous text block according to their order in the original table, and then filled into the data input area of ​​the target template. The concatenation according to the record order is to ensure that when the large language model returns results later, each label content can be accurately traced back to the corresponding original record.

[0026] At this point, all the content that needs to be filled in the target template is in place. Next, based on the preset output format constraints in the target template, the four elements—task type, tag option list, text fields of multiple records, and output format constraints—can be combined into a complete prompt word and sent to the large language model. The output format constraints require the large language model to return the result in a purely structured form, such as a JSON array, where each element of the array corresponds one-to-one with each record in the data block. This is because the tagging result returned by the large language model is itself a continuous text; without such format constraints, it would be impossible to determine which segment of the tagging content belongs to which original record, thus making subsequent association storage and result output impossible. By fixing the output to a parsable structure through format constraints, the subsequent parsing process can be executed smoothly.

[0027] The system in this application automatically completes all tasks of intent parsing, template matching, option mapping, and format constraints. When the requirements change, only the instruction needs to be changed, without the need for technical personnel to modify the underlying logic, making it simple and convenient.

[0028] Optionally, in this embodiment, the step of "mapping the marker dimension to a list of tag options in the target template" includes: obtaining a preset mapping rule table, which stores multiple standard dimension names and their corresponding list of tag options, as well as mutual exclusion and association relationships between the standard dimension names; when there is only one marker dimension, mapping the marker dimension to a corresponding list of tag options according to the mapping rule table; when there are multiple marker dimensions, determining whether there is an association or mutual exclusion relationship between the multiple marker dimensions according to the mutual exclusion and association relationships in the mapping rule table; if there is no mutual exclusion relationship between the multiple marker dimensions, mapping each marker dimension to a corresponding list of tag options, and merging the multiple list of tag options into a composite list of tag options according to the association relationship, wherein each tag value in the composite list of tag options is a Cartesian product of the tag values ​​in each list of tag options; if there is a mutual exclusion relationship between the multiple marker dimensions, selecting the marker dimension with the highest weight as the target marker dimension according to the semantic weight of each marker dimension in the marker requirement text, and mapping the target marker dimension to a corresponding list of tag options.

[0029] In this embodiment, before performing the mapping operation, a preset mapping rule table can be obtained. This table stores three types of information in advance: multiple standard dimension names and their corresponding label option lists, as well as the mutual exclusion and association relationships between the standard dimension names. Standard dimension names are standardized expressions defined internally by the system, such as "severity of illness" and "credit risk level." Each standard dimension is pre-bound to a set of label option lists that can be selected by the large language model. For example, "severity of illness" corresponds to the three options "mild, moderate, and severe." Mutual exclusion and association relationships define whether different standard dimensions can logically coexist. If two standard dimensions are mutually exclusive, it means they cannot appear simultaneously in the same labeling task. For example, "evaluate by cost-effectiveness" and "evaluate by price" are mutually exclusive because cost-effectiveness already embeds the price factor; using them simultaneously would cause duplication and confusion in the evaluation logic. If two standard dimensions are associated, it means they can be used in combination to provide richer evaluation information.

[0030] When only one label dimension is identified, the system directly searches the mapping rule table for a matching standard dimension name and uses the corresponding label option list as the mapping result. Specifically, a precise or fuzzy match can be performed in the mapping rule table, based on text similarity or semantic equivalence, ultimately ensuring that any expression entered by the user can be correctly routed to the same set of label option lists.

[0031] When multiple labeled dimensions are identified, it's crucial to first determine whether there are any correlations or mutual exclusions between them. Specifically, this can be done by comparing each dimension with pre-stored mutually exclusive and correlated relationships in the mapping rule table to determine the current relationship status of the labeled dimensions. Mutually exclusive relationships refer to two labeled dimensions that are logically contradictory or semantically overlapping and should not be used simultaneously for the same labeling task. Correlated relationships, on the other hand, indicate that two labeled dimensions can coexist semantically, and their joint evaluation can provide more comprehensive information. This is because real-world user needs are often irregular. For example, in a financial scenario, if a user requests "analysis of companies based on industry prosperity and individual stock valuation," it's necessary to first determine whether there are correlations or mutual exclusions between these two labeled dimensions before deciding on the subsequent processing path.

[0032] Assuming there are no mutually exclusive relationships between the multiple labeled dimensions, each labeled dimension is mapped to its corresponding list of label options. These lists are then merged into a composite list of label options based on their relationships. Each label value in this composite list can be a Cartesian product of the label values ​​from the original lists. Here, the Cartesian product is simply a permutation of elements from two or more sets. For example, if the label options for "price level" are "high, medium, low," and the label options for "comfort level" are "excellent, good, poor," their Cartesian product generates nine composite labels: "high-excellent, high-good, high-poor, medium-excellent, medium-good, medium-poor, low-excellent, low-good, low-poor." Through this fusion method, a system that could only perform single-dimensional evaluations can generate multi-dimensional composite evaluation results, making the labeled content output by the large language model contain richer information. It is important to note that if there is no correlation between the above multiple label dimensions, then each label dimension can be independently mapped to its corresponding label option list, and multiple sets of label option lists can be included in the subsequent prompt word construction, so that the large language model can output the label results corresponding to each label dimension for each record.

[0033] If multiple mutually exclusive labeling dimensions are detected, they cannot be simply merged; a conflict resolution mechanism must be activated. Specifically, based on the semantic weight of each labeling dimension within the user's input labeling request text, the labeling dimension with the highest weight is selected as the target labeling dimension. Then, only this target labeling dimension is mapped to the corresponding list of label options. Semantic weight reflects the importance of different labeling dimensions in the user's needs. For example, if a user says "evaluate based on cost-effectiveness but not price," semantic analysis reveals that "cost-effectiveness" has a significantly higher weight than "price" in the sentence, and the two are mutually exclusive. Therefore, "cost-effectiveness" is selected as the target labeling dimension, and "price" is discarded. This conflict resolution mechanism ensures that when logical contradictions exist in the user's needs, it can automatically make reasonable trade-offs, rather than forcibly merging them, leading to distorted output or inability to execute.

[0034] The mapping mechanism in this embodiment enables the system to flexibly respond to diverse labeling needs input by users. Whether it is a simple labeling of a single dimension, a composite evaluation of multi-dimensional fusion, or a complex instruction with logical conflicts, it can automatically adapt to the corresponding processing path. The whole process does not require manual adjustment or additional explanation from the user, which effectively improves the accuracy and usability of the labeling results.

[0035] Optionally, in this embodiment, "mapping the marker dimension to a list of tag options in the target template" includes: semantically annotating the marker dimension to identify the evaluation attributes and evaluation scales in the marker dimension; matching a corresponding candidate tag option list from a preset attribute-tag mapping library based on the evaluation attributes, wherein the tag option list contains a tag hierarchy structure and a set of tag values ​​for each level; if a unique candidate tag option list is matched, then the candidate tag option list is used as the tag option list; if multiple candidate tag option lists are matched, then selecting a candidate tag option list whose tag granularity matches the evaluation scale from the multiple candidate tag option lists based on the semantic features of the evaluation scale, wherein the tag granularity is characterized by the number of tag levels and the dispersion of tag values ​​at each level.

[0036] In this embodiment, firstly, semantic role labeling is performed on the labeling dimensions identified from the labeling requirement text. This is a deep text analysis technique aimed at identifying two key semantic elements from the phrase "labeling dimension": evaluation attributes and evaluation scale. Evaluation attributes refer to the specific angle from which the user evaluates the text, such as "severity of illness," "credit risk level," or "product comfort." Evaluation scales reflect the user's desired level of labeling granularity; for example, expressions like "roughly categorize," "precisely judge," "coarse-grained evaluation," and "detailed classification" all imply different scale requirements. The significance of semantic role labeling lies in its ability to break down a simple phrase input by the user into structured semantic information, providing a more accurate basis for subsequent tag matching.

[0037] After obtaining the evaluation attributes, they are used as search keywords to match against a pre-defined attribute-tag mapping library to retrieve a list of all candidate tag options related to that evaluation attribute. The attribute-tag mapping library can be a pre-built storage structure where each evaluation attribute corresponds to one or more tag option lists. Each tag option list contains not only a specific set of tag values ​​but also its hierarchical structure information. The tag hierarchy refers to the hierarchical relationship between tag values. For example, a tag option list describing the severity of a disease could be a single-level list of "mild, moderate, severe" or a multi-level list of "Level 1 (mild, moderate, severe), Level 2 (early, middle, late)". More levels mean a more refined classification dimension, but also more subtle distinctions between tags.

[0038] If only one candidate label option list is matched in the attribute-label mapping library, it means that the evaluation attribute has a unique and definite mapping relationship in the library, and this unique candidate label option list is directly used as the final label option list.

[0039] If multiple candidate label options are matched, it indicates that the evaluation attribute has different granularity label systems in different scenarios. At this point, further filtering can be performed based on the previously identified evaluation scale. Specifically, the semantic features of the evaluation scale can be analyzed to extract whether the user expects coarse-grained or fine-grained labeling. Then, the label granularity that best matches the evaluation scale can be selected from the multiple candidate label options. The criteria for determining label granularity are the number of label levels and the dispersion of label values ​​at each level: the more levels, the more complex the label structure and the more detailed the distinction; the higher the dispersion of label values ​​at each level, the greater the distinguishability between labels. For example, coarse-grained candidates: [mild, moderate, severe], with a single-layer hierarchy, significant semantic distance between label values ​​(significant difference between mild and severe), and moderate to low dispersion; fine-grained candidates: [very mild, mild, mild to moderate, moderate, moderate to severe, severe, very severe], with one or two layers, more densely distributed label values ​​in the semantic space, subtle differences between adjacent values, and high dispersion. In a specific embodiment, such as in a medical scenario, if a user inputs "roughly assess the severity of the condition," where the evaluation scale is "roughly assess," then a coarse-grained label list with fewer layers and lower label value dispersion, such as "mild, moderate, severe," can be prioritized from the list of multiple candidate label options. However, if the user inputs "precisely grade the severity of the condition," the evaluation scale changes to "precise grading," and a fine-grained label list with more layers and higher label value dispersion, such as a detailed grading system containing multiple sub-levels, can be preferred. Similarly, in financial scenarios, when a user enters "rough assessment of corporate credit risk", they can select a coarse-grained list of three levels: "high", "medium", and "low"; when a user enters "refined assessment of corporate credit risk", they can select a refined list of tags that includes more subdivided levels or additional scoring dimensions.

[0040] This application embodiment, through semantic parsing of the evaluation scale and quantitative matching of label granularity, explicitly transforms the user's implicit expectations into specific label selection decisions. When the label granularity matches the evaluation scale well, the labeling results output by the large language model will not lose information discrimination due to excessively coarse granularity, nor will they exceed the actual discrimination ability of the data due to excessively fine granularity, thereby significantly improving the usability of the labeling results and user satisfaction.

[0041] In this embodiment of the application, optionally, step 102, "dividing the structured data into multiple data blocks according to a preset segmentation strategy," includes: obtaining the upper limit of the context window of the large language model; counting the number of characters in the text field contained in each record of the structured data, and calculating the average number of characters per record; calculating the theoretical maximum number of records for each data block based on the upper limit of the context window and the average number of characters, and weighting and correcting the theoretical maximum number of records according to a preset safety coefficient to obtain the actual segmentation upper limit; and dividing the structured data into multiple data blocks according to the actual segmentation upper limit.

[0042] In this embodiment, before starting chunking, the upper limit of the context window of the large language model can be obtained. The upper limit of the context window is the maximum length of input that the large language model can accept in a single input, usually measured in the number of characters or tokens; it is the physical upper limit of the model's processing capacity. Once the total number of characters in the input exceeds this limit, the large language model will be unable to process the request normally, and may directly report an error or truncate the input. Different large language models have different upper limits of the context window.

[0043] Subsequently, statistical analysis was performed on the parsed structured data to calculate the number of characters in the text fields of each record and to determine the average number of characters across all records. This is because the text length of each record often varies greatly; some records may contain only a dozen or so characters, while others may contain hundreds or even thousands. For example, in a medical setting, a brief outpatient complaint might only have 20 characters, while a detailed discharge summary could exceed 2000 words. A similar difference exists in financial settings: a concise risk summary might only have 30 characters, while a detailed financial statement footnote analysis could approach 1000 words. In such cases, using a strategy of dividing data into blocks based on a fixed number of records could lead to some data blocks exceeding the capacity limit of the large language model, while other data blocks remain far below the capacity limit, resulting in a waste of the large language model's processing power. Calculating the average number of characters is precisely to quantify the actual density of the data, enabling subsequent calculations to be based on real-world conditions rather than subjective experience.

[0044] Furthermore, the theoretical maximum number of records that each data block can hold can be calculated by dividing the upper limit of the context window by the average number of characters. This theoretical value is then multiplied by a preset safety factor to obtain the actual block limit. The safety factor can be a decimal between 0 and 1, such as 0.7 or 0.8, and its purpose is to reserve sufficient space for fixed instruction text in prompts. Instructional content such as label option lists and output format constraints also require space in the context window. If the entire window capacity is allocated to the records contained in the data block, this instruction text will be pushed out of the window, resulting in incomplete input. By introducing a safety factor to correct the theoretical capacity, it is ensured that when each data block is fed into the large language model, the context window will neither overflow due to excessive data nor be truncated due to insufficient reserved space, thus guaranteeing the integrity and reliability of each request.

[0045] Next, based on the calculated upper limit of the actual block size, all records can be sequentially divided into multiple data blocks, with each data block containing no more than the upper limit. It's important to note that the splitting process is performed in record order, without altering the sequential relationship between records; the last block can contain fewer records than the upper limit.

[0046] This application embodiment implements dynamic adaptive calculation of block size, enabling the system to automatically find the optimal block scheme when facing different data features and different large language model capabilities, without the need for manual preset. This application embodiment can automatically increase the number of records per block to improve processing efficiency when dealing with short text data, and automatically decrease the number of records per block to avoid exceeding limits when dealing with long text data, thereby achieving a balance between security and efficiency and avoiding processing failures or wasted computing power due to uneven data length distribution.

[0047] Optionally, in this embodiment of the application, the method further includes: during the processing of each data block, recording the number of data blocks that have been processed, and generating progress information based on the number of data blocks that have been processed and the total number of the multiple data blocks; returning the progress information in response to a progress query request; and marking the progress information as completed when all the multiple data blocks have been processed, and returning the address for obtaining the result data.

[0048] In this embodiment, during the background data processing, the number of data blocks already processed can be recorded synchronously. Combined with the total number of all data blocks, a progress information that can be queried externally is generated. This progress information can be a status data containing two values: "number completed" and "total number," such as "3 blocks completed, 10 blocks total"; or a progress percentage. This mechanism is necessary because batch labeling using a large language model is a time-consuming operation, especially when the amount of data to be processed is large. The entire process can last for several minutes or even longer, and users cannot obtain results immediately after initiating a request. Without progress feedback, users will have no way of knowing whether the task is progressing normally, nearing completion, or already stuck, and can only passively wait or repeatedly refresh, resulting in a very poor user experience. The introduction of progress information allows users to understand the execution stage of the task in real time, thereby effectively alleviating the uncertainty during the waiting process.

[0049] Therefore, a progress query interface can be provided externally, capable of responding to progress query requests from outside and returning the current progress information upon receiving the request. This query request can be triggered by the user actively clicking the "Query Progress" button on the front-end interface, or it can be a polling request automatically initiated by the front-end page at fixed time intervals. Since the processing of the marking task is decoupled from the user's interface, after the user submits the marking task, it executes independently in the background without blocking the front-end interface. Therefore, the user can continue to operate other functions on the page or switch to other windows to handle other tasks during task execution, without needing to remain on the current page waiting for the result. The existence of the progress query interface makes this asynchronous execution mode possible, allowing users to return at any time to check the task progress.

[0050] Once all data blocks have been processed, the progress information is marked as complete, and the address for retrieving the result data is returned at this time. This ensures that the result data pointed to by the user's address is complete and usable, and avoids situations where the data is not yet written or is incomplete.

[0051] This application transforms a potentially time-consuming background batch processing task into an observable and interactive process, significantly enhancing the user's sense of control and trust in long-running tasks. Users no longer need to wait passively; they can handle other tasks in parallel during task execution, monitor progress at any time through a progress query interface, and obtain results immediately upon task completion. This design enables the system to provide a good user experience and task manageability when dealing with batch labeling of large-scale data.

[0052] In this embodiment of the application, optionally, after "obtaining the spreadsheet file to be processed" in step 101, the method further includes: converting the spreadsheet file into a Feather format binary file for storage, and generating a document identifier corresponding to the binary file; correspondingly, "parsing the spreadsheet file into structured data" in step 101 includes: in response to an asynchronous tag request, obtaining a corresponding target binary file based on the document identifier corresponding to the asynchronous tag request, and parsing the target binary file into the structured data.

[0053] In this embodiment, after receiving the spreadsheet file uploaded by the user, the spreadsheet file can be converted into a Feather format binary file for storage. Feather is a binary columnar storage format designed specifically for data frames, characterized by extremely fast read and write speeds and relatively small storage space requirements. Compared to spreadsheet files, the Feather format eliminates a large amount of redundant information related to the presentation layer, such as table styles, formulas, and macros, retaining only the pure data content. Therefore, it can significantly reduce I / O wait time in subsequent read and parsing operations. The fundamental reason for converting the original spreadsheet file to Feather format for storage is that the tagging task may involve repeatedly reading dozens or even hundreds of data blocks. Each data block needs to be retrieved from storage and parsed during processing. If data is read from the original spreadsheet file every time, the frequent format parsing and I / O overhead will seriously affect the overall processing efficiency. By converting the spreadsheet file to Feather format in one go, all subsequent read operations can be completed at the lowest cost. Simultaneously with the format conversion, a document identifier corresponding to the binary file can also be generated. In subsequent asynchronous tag requests, users do not need to upload the entire spreadsheet file again, nor do they need to remember the storage path of the spreadsheet file on the server; they only need to include this short document identifier. Generating and returning the document identifier can be a lightweight operation, and users can obtain the document identifier immediately after uploading the spreadsheet file.

[0054] When a user initiates an asynchronous marking request, the system can locate and retrieve the corresponding target binary file (the previously converted Feather file) from storage based on the document identifier carried in the request, and then parse it into structured data. An asynchronous marking request means that after a user submits a marking task, the system can immediately return a response informing the user that the task has been received, without waiting for the task to complete, and then execute the entire marking process independently in the background. In this process, the document identifier links the actions of "user uploading a spreadsheet file" and "user initiating an asynchronous marking request," two actions that may be far apart in time. Users can initiate an asynchronous marking request at any time after uploading the spreadsheet file without needing to re-upload the file or worrying about lost content or formatting changes. The system uses document identifiers to ensure the uniqueness and traceability of each spreadsheet file within the system, accurately routing each request to the corresponding data source even when handling large-scale concurrent requests.

[0055] This application's embodiments convert the original spreadsheet file into Feather format, enabling subsequent processing of dozens or even hundreds of data blocks to be completed with extremely low I / O overhead for each read operation. This avoids the redundant computation and storage overhead caused by repeatedly parsing the original spreadsheet file, thus significantly improving the overall processing efficiency. Simultaneously, the document identifier design completely decouples file access from the execution of marking tasks in the time dimension. Users can trigger marking tasks at any time needed using the document identifier; the system executes asynchronously in the background without blocking or waiting in the foreground.

[0056] Furthermore, as Figure 1 In terms of specific implementation, this application provides an automated text tagging device for spreadsheets based on a large language model, such as... Figure 3 As shown, the device includes: The data acquisition module is used to acquire the spreadsheet file to be processed and to acquire the marked requirement text corresponding to the spreadsheet file, which is input by the user in natural language. The data partitioning module is used to parse the spreadsheet file into structured data and divide the structured data into multiple data blocks according to a preset partitioning strategy. Each row in the structured data corresponds to a record in the spreadsheet file, each record contains a text field to be labeled, and each data block includes multiple records. The tagging module is used to construct prompt words for each data block based on the tagging requirement text and multiple records contained in the data block, send the prompt words to the large language model, receive the tagging results returned by the large language model, parse the tagging results to obtain the tagging content corresponding to each record, and store each record and its corresponding tagging content in association. The results output module is used to combine each record and its corresponding tag content into result data containing the original text column and the tag content column after all data blocks have been processed, and then output the result data.

[0057] Optionally, the marking module is used for: Identify the marking task type and marking dimension in the marking requirement text, and select a target template from a set of preset prompt word templates based on the marking task type; The marker dimension is mapped to the tag option list in the target template, the tag option list is filled into the marker option area of ​​the target template, and the text fields of multiple records in the data block are concatenated in the record order and filled into the data input area of ​​the target template. Based on the preset output format constraints in the target template, a complete prompt word is generated, which includes the marked task type, the tag option list, the text fields of the multiple records, and the output format constraints.

[0058] Optionally, the marking module is further configured to: Obtain a preset mapping rule table, which stores multiple standard dimension names and their corresponding label option lists, as well as the mutual exclusion and association relationships between the standard dimension names; When there is only one label dimension, the label dimension is mapped to the corresponding list of label options according to the mapping rule table; When there are multiple labeling dimensions, the system determines whether there is a correlation or mutual exclusion relationship between the multiple labeling dimensions based on the mutual exclusion and correlation relationships in the mapping rule table. If there is no mutual exclusion relationship between the multiple labeling dimensions, each labeling dimension is mapped to a corresponding tag option list, and the multiple tag option lists are merged into a composite tag option list based on the correlation relationship. Each tag value in the composite tag option list is a Cartesian product of the tag values ​​in each tag option list. If there is a mutual exclusion relationship between the multiple labeling dimensions, the labeling dimension with the highest semantic weight in the labeling requirement text is selected as the target labeling dimension, and the target labeling dimension is mapped to a corresponding tag option list.

[0059] Optionally, the marking module is further configured to: Semantic role labeling is performed on the labeled dimensions to identify the evaluation attributes and evaluation scales in the labeled dimensions; Based on the evaluation attributes, a list of corresponding candidate tags is matched from a preset attribute-tag mapping library. The tag option list contains a tag hierarchy structure and a set of tag values ​​for each level. If a unique candidate tag option list is matched, then the candidate tag option list is used as the tag option list; If multiple candidate label option lists are matched, then the candidate label option list whose label granularity matches the evaluation scale is selected from the multiple candidate label option lists according to the semantic features of the evaluation scale as the label option list. The label granularity is characterized by the number of label levels and the degree of dispersion of the label values ​​at each level.

[0060] Optionally, the data partitioning module is used for: Obtain the upper limit of the context window of the large language model; Count the number of characters in the text field of each record in the structured data, and calculate the average number of characters per record; Based on the upper limit of the context window and the average number of characters, the theoretical maximum number of records for each data block is calculated, and the theoretical maximum number of records is weighted and corrected according to a preset safety factor to obtain the actual block upper limit. Based on the actual upper limit of the block size, the structured data is divided into multiple data blocks.

[0061] Optionally, the device further includes a progress information query module; the progress information query module is used for: During the processing of each data block, the number of data blocks that have been processed is recorded, and progress information is generated based on the number of data blocks that have been processed and the total number of data blocks. In response to a progress query request, the progress information is returned; When all the data blocks have been processed, the progress information is marked as completed, and the address for obtaining the result data is returned.

[0062] Optionally, the device further includes a format conversion module; the format conversion module is used for: After obtaining the spreadsheet file to be processed, the spreadsheet file is converted into a Feather format binary file for storage, and a document identifier corresponding to the binary file is generated; Accordingly, the data partitioning module is used for: In response to an asynchronous tag request, based on the document identifier corresponding to the asynchronous tag request, the corresponding target binary file is obtained, and the target binary file is parsed into the structured data.

[0063] It should be noted that other corresponding descriptions of the functional units involved in the automatic text marking device for spreadsheets based on a large language model provided in this application embodiment can be found in the following references. Figures 1 to 2 The corresponding descriptions in the method will not be repeated here.

[0064] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 4 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0065] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0066] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0067] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0068] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0069] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0070] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0071] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for automated text tagging in spreadsheets based on a large language model, characterized in that, include: Obtain the spreadsheet file to be processed, and obtain the marked requirement text corresponding to the spreadsheet file, which is input by the user in natural language; The spreadsheet file is parsed into structured data, and the structured data is divided into multiple data blocks according to a preset block division strategy. Each row in the structured data corresponds to a record in the spreadsheet file, each record contains a text field to be labeled, and each data block includes multiple records. For each data block, a prompt word is constructed based on the tagging requirement text and multiple records contained in the data block. The prompt word is sent to the large language model, the tagging result returned by the large language model is received, the tagging result is parsed to obtain the tagging content corresponding to each record, and each record and its corresponding tagging content are associated and stored. After all data blocks have been processed, each record and its corresponding tag content are combined into result data containing the original text column and the tag content column, and then output.

2. The method according to claim 1, characterized in that, The construction of prompt words based on the marked requirement text and multiple records contained in the data block includes: Identify the marking task type and marking dimension in the marking requirement text, and select a target template from a set of preset prompt word templates based on the marking task type; The marker dimension is mapped to the tag option list in the target template, the tag option list is filled into the marker option area of ​​the target template, and the text fields of multiple records in the data block are concatenated in the record order and filled into the data input area of ​​the target template. Based on the preset output format constraints in the target template, a complete prompt word is generated, which includes the marked task type, the tag option list, the text fields of the multiple records, and the output format constraints.

3. The method according to claim 2, characterized in that, The step of mapping the marker dimension to the list of label options in the target template includes: Obtain a preset mapping rule table, which stores multiple standard dimension names and their corresponding label option lists, as well as the mutual exclusion and association relationships between the standard dimension names; When there is only one label dimension, the label dimension is mapped to the corresponding list of label options according to the mapping rule table; When there are multiple labeling dimensions, the system determines whether there is a correlation or mutual exclusion relationship between the multiple labeling dimensions based on the mutual exclusion and correlation relationships in the mapping rule table: If there is no mutual exclusion relationship between the multiple labeling dimensions, each labeling dimension is mapped to a corresponding tag option list, and the multiple tag option lists are merged into a composite tag option list according to the correlation relationship, where each tag value in the composite tag option list is a Cartesian product of the tag values ​​in each tag option list; If there is a mutual exclusion relationship between the multiple labeling dimensions, the labeling dimension with the highest weight is selected as the target labeling dimension according to the semantic weight of each labeling dimension in the labeling requirement text, and the target labeling dimension is mapped to a corresponding tag option list.

4. The method according to claim 2, characterized in that, The step of mapping the marker dimension to the list of label options in the target template includes: Semantic role labeling is performed on the labeled dimensions to identify the evaluation attributes and evaluation scales in the labeled dimensions; Based on the evaluation attributes, a list of corresponding candidate tags is matched from a preset attribute-tag mapping library. The tag option list contains a tag hierarchy structure and a set of tag values ​​for each level. If a unique candidate tag option list is matched, then the candidate tag option list is used as the tag option list; If multiple candidate label option lists are matched, then the candidate label option list whose label granularity matches the evaluation scale is selected from the multiple candidate label option lists according to the semantic features of the evaluation scale as the label option list. The label granularity is characterized by the number of label levels and the degree of dispersion of the label values ​​at each level.

5. The method according to claim 1, characterized in that, The step of dividing the structured data into multiple data blocks according to a preset block division strategy includes: Obtain the upper limit of the context window of the large language model; Count the number of characters in the text field of each record in the structured data, and calculate the average number of characters per record; Based on the upper limit of the context window and the average number of characters, the theoretical maximum number of records for each data block is calculated, and the theoretical maximum number of records is weighted and corrected according to a preset safety factor to obtain the actual block upper limit. Based on the actual upper limit of the block size, the structured data is divided into multiple data blocks.

6. The method according to claim 1, characterized in that, The method further includes: During the processing of each data block, the number of data blocks that have been processed is recorded, and progress information is generated based on the number of data blocks that have been processed and the total number of data blocks. In response to a progress query request, the progress information is returned; When all the data blocks have been processed, the progress information is marked as completed, and the address for obtaining the result data is returned.

7. The method according to claim 1, characterized in that, After obtaining the spreadsheet file to be processed, the method further includes: The spreadsheet file is converted into a Feather format binary file for storage, and a document identifier corresponding to the binary file is generated; Accordingly, parsing the spreadsheet file into structured data includes: In response to an asynchronous tag request, based on the document identifier corresponding to the asynchronous tag request, the corresponding target binary file is obtained, and the target binary file is parsed into the structured data.

8. A spreadsheet text auto-marking device based on a large language model, characterized in that, include: The data acquisition module is used to acquire the spreadsheet file to be processed and to acquire the marked requirement text corresponding to the spreadsheet file, which is input by the user in natural language. The data partitioning module is used to parse the spreadsheet file into structured data and divide the structured data into multiple data blocks according to a preset partitioning strategy. Each row in the structured data corresponds to a record in the spreadsheet file, each record contains a text field to be labeled, and each data block includes multiple records. The tagging module is used to construct prompt words for each data block based on the tagging requirement text and multiple records contained in the data block, send the prompt words to the large language model, receive the tagging results returned by the large language model, parse the tagging results to obtain the tagging content corresponding to each record, and store each record and its corresponding tagging content in association. The results output module is used to combine each record and its corresponding tag content into result data containing the original text column and the tag content column after all data blocks have been processed, and then output the result data.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

10. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.