Batch instruction processing method, device and storage medium based on multi-modal input

Abstract

The application discloses a batch instruction processing method and device based on multi-modal input and a storage medium, and relates to the technical field of electronic digital data processing. The method comprises the following steps: converting a to-be-analyzed file into a horizontal two-dimensional matrix according to the file type of the to-be-analyzed file; calculating the intersection coverage of the horizontal two-dimensional matrix and a standard table header keyword, and determining a target two-dimensional matrix according to the intersection coverage; assembling the first row data of the target two-dimensional matrix through a key enumeration field and a preset configuration template to obtain a plurality of table header mapping prompt words; concurrently transmitting the plurality of table header mapping prompt words to a large model to obtain a standardized two-dimensional matrix; and calling a processing interface to process a standard table file corresponding to the standardized two-dimensional matrix to execute batch instruction ordering. The application overcomes the technical defects that a traditional string matching mode cannot understand business semantics and the recognition accuracy of synonymous fields is low, and improves the accuracy and processing efficiency of batch instruction conversion.

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, and in particular to a batch instruction processing method, device and storage medium based on multimodal input. Background Technology

[0002] In financial interbank trading platforms, batch order placement is a core business requirement for institutional users. The platform primarily provides batch order placement functionality via file upload. To ensure the standardization of transaction data and interface compatibility, the platform has established a fixed, standardized Excel template, requiring users to strictly adhere to the preset header fields and content enumeration specifications when filling in order data.

[0003] In related technologies, the platform's user template and standard template conversion functions are all based on string matching and fixed rules. This mechanism can only convert table headers and cell contents that are literally identical, and cannot understand the business semantics behind the data. For example, it cannot accurately map synonymous transaction type descriptions such as "buy" and "purchase", or synonymous field names such as "securities code" and "stock code". Summary of the Invention

[0004] The main objective of this application is to provide a batch instruction processing method, device, and storage medium based on multimodal input, aiming to solve the technical problem that the low accuracy of synonym field recognition in standardized batch instruction template scenarios leads to the failure of batch instruction processing.

[0005] To achieve the above objectives, this application provides a batch instruction processing method based on multimodal input, the batch instruction processing method based on multimodal input comprising: Convert the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed; Calculate the intersection coverage of the horizontal two-dimensional matrix and the standard header keywords, and determine the target two-dimensional matrix based on the intersection coverage. The first row of the target two-dimensional matrix is ​​assembled using key enumeration fields and a preset configuration template to obtain multiple header mapping prompts; The multiple header mapping prompts are concurrently transmitted to the large model to obtain a standardized two-dimensional matrix; The processing interface is invoked to process the standard table file corresponding to the standardized two-dimensional matrix in order to execute batch order placement.

[0006] In one embodiment, converting the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed includes: Receive at least one file to be parsed; The file type of the file to be parsed is determined based on the file name suffix and file byte prefix of the file to be parsed. Based on the parsing method corresponding to the file type, the file to be parsed is converted into the horizontal two-dimensional matrix.

[0007] In one embodiment, converting the file to be parsed into the horizontal two-dimensional matrix according to the parsing method corresponding to the file type includes: If the file type is a table format, read the worksheet data of the file to be parsed and extract the table header and content rows to generate the horizontal two-dimensional matrix; If the file type is a Word document or a PDF document, extract all table structures in the file to be parsed and generate a two-dimensional matrix, and filter the two-dimensional matrix with the highest matching degree based on the table matching degree algorithm and convert it into the horizontal two-dimensional matrix; If the file type is an image format, the text content of the file to be parsed is extracted by optical character recognition, and a large model is used to extract the table structure from the text content and generate a two-dimensional matrix, which is then converted into the horizontal two-dimensional matrix; if the table structure is not recognized after parsing, the full text is extracted and a two-dimensional matrix with a standard table header is generated by the large model, which is then adjusted into the horizontal two-dimensional matrix.

[0008] In one embodiment, calculating the intersection coverage of the horizontal two-dimensional matrix and the standard header keywords, and determining the target two-dimensional matrix based on the intersection coverage, includes: If the number of files in the file to be parsed is one, then calculate the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords. If the intersection coverage rate is greater than the coverage rate threshold, then use the horizontal two-dimensional matrix as the target two-dimensional matrix. If the number of files to be parsed is not one, the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords is calculated, and the horizontal two-dimensional matrix corresponding to the largest intersection coverage rate is taken as the target two-dimensional matrix.

[0009] In one embodiment, the preset configuration template includes a basic instruction template, header data, and enumeration restrictions. The step of assembling multiple header mapping prompts from the first row data of the target two-dimensional matrix using key enumeration fields and the preset configuration template includes: The source header list is determined based on the first row of the target two-dimensional matrix; After performing deduplication, null value removal, and space removal on the column data corresponding to each source table in sequence, the sample data corresponding to each source table is obtained by filtering the preset number of column data. The target table header corresponding to the source table is determined according to the conversion requirements in the configuration file, and the key enumeration fields are determined according to the enumeration constraints in the configuration file. Based on the basic instruction template corresponding to the key enumeration field, the table header data, the enumeration restrictions, the source table, and the example data, the table header mapping prompt words corresponding to each key enumeration field are assembled.

[0010] In one embodiment, the step of concurrently classifying and mapping the multiple header mapping prompts to a large model to obtain a standardized two-dimensional matrix includes: The multiple header mapping prompts are transmitted concurrently to the large model to obtain the header mapping scheme and confidence level output by the large model; The target header mapping scheme is determined based on the confidence level, and the target header mapping scheme includes transaction type values ​​and mapping relationships. Based on the header mapping dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value, the semantics of the target two-dimensional matrix are converted into the standardized two-dimensional matrix.

[0011] In one embodiment, before converting the semantics of the target two-dimensional matrix into the standardized two-dimensional matrix according to the header mapping dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value, the process includes: Based on the mapping relationship, determine the mapping logic corresponding to the unmapped fields in the target two-dimensional matrix; The header mapping relationship dictionary is obtained based on the mapping relationship and the mapping logic.

[0012] In one embodiment, the content mapping rule includes columns to be enumerated and their target enumerated value set, columns to be converted to units and their conversion coefficients, and columns to be formatted and their target format. The step of converting the semantics of the target two-dimensional matrix into the standardized two-dimensional matrix based on the header mapping dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value includes: The columns to be transformed in the target two-dimensional matrix are determined according to the content mapping rules; The target header name and enumeration constraints corresponding to the column to be converted are determined based on the header mapping relationship dictionary. Based on the source header name, original data, label header name, content mapping rules, and enumeration constraints of the column to be converted, construct the column hint words for the column to be converted; The column prompts for each column to be transformed are sent to the large model in terms of column dimension, so as to receive the standardized data output by the large model; The standardized two-dimensional matrix is ​​obtained by replacing the original data in the column to be transformed with the standardized data.

[0013] In addition, to achieve the above objectives, this application also provides a batch instruction processing device based on multimodal input, the batch instruction processing device based on multimodal input comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the batch instruction processing method based on multimodal input as described above.

[0014] In addition, to achieve the above objectives, this application also provides a storage medium, which is a computer-readable storage medium, storing a program that implements a batch instruction processing method based on multimodal input. The program that implements the batch instruction processing method based on multimodal input is executed by a processor to implement the steps of the batch instruction processing method based on multimodal input as described above.

[0015] This application provides a batch instruction processing method based on multimodal input. First, the file to be parsed is converted into a horizontal two-dimensional matrix according to its file type, unifying batch instruction data of different formats into a standard structured data form. This eliminates data structure differences caused by different file formats, providing a unified processing foundation for subsequent field semantic recognition. Next, the intersection coverage rate of the horizontal two-dimensional matrix and standard table header keywords is calculated, and a target two-dimensional matrix is ​​determined based on the intersection coverage rate. This accurately selects the core table containing batch instruction information from multiple candidate tables, eliminating interference from irrelevant data for field recognition, allowing subsequent semantic processing to focus on effective business data. Finally, the target two-dimensional matrix... The first row of data in the matrix is ​​assembled into multiple header mapping prompts using key enumeration fields and preset configuration templates. A differentiated semantic guidance framework is built around key business enumeration fields to provide clear business context for the large model, rather than relying solely on a single general instruction. These multiple header mapping prompts are then transmitted concurrently to the large model to obtain a standardized two-dimensional matrix. Through multi-dimensional semantic guidance, the large model can understand the actual business meaning of the headers from different business perspectives, accurately identifying synonymous fields such as "buy" and "purchase," "securities code" and "stock code," which have different literal expressions but consistent business semantics. Finally, the processing interface is called to process the standard table file corresponding to the standardized two-dimensional matrix to execute batch order placement.

[0016] In summary, this application overcomes the technical shortcomings of traditional string matching methods, such as the inability to understand business semantics and the low accuracy of synonym field recognition, by using multimodal data normalization, precise filtering of target tables, assembly of multi-prompt words combined with key enumeration fields, and concurrent semantic mapping of large models. It fundamentally solves the problem of batch instruction processing failure caused by field mapping errors, and significantly improves the accuracy and processing efficiency of batch instruction conversion. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an embodiment of the batch instruction processing method based on multimodal input provided in this application. Figure 2 This is a schematic diagram of the processing flow for various file types provided in Embodiment 3 of the batch instruction processing method based on multimodal input of this application; Figure 3 This is a flowchart illustrating the verification process involved in Embodiment 9 of the batch instruction processing method based on multimodal input in this application. Figure 4 This is a schematic diagram of the hardware environment involved in the batch instruction processing device based on multimodal input in this application.

[0020] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not intended to limit this application.

[0022] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0023] Currently, the conversion function between user templates and standard templates on financial interbank trading platforms is based on string matching and fixed rules. This mechanism can only convert table headers and cell contents that are literally identical, and cannot understand the business semantics behind the data. For example, it cannot accurately map synonymous transaction type descriptions such as "buy" and "purchase", or synonymous field names such as "securities code" and "stock code".

[0024] The main solution of this application is as follows: convert the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed; calculate the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords, and determine the target two-dimensional matrix based on the intersection coverage rate; assemble the first row data of the target two-dimensional matrix into multiple header mapping prompts through key enumeration fields and preset configuration templates; concurrently transmit the multiple header mapping prompts to the large model to obtain a standardized two-dimensional matrix; and call the processing interface to process the standard table file corresponding to the standardized two-dimensional matrix to execute batch order placement.

[0025] This application overcomes the technical shortcomings of traditional string matching methods, such as the inability to understand business semantics and the low accuracy of synonym field recognition, by using multimodal data normalization, precise filtering of target tables, assembly of multiple prompt words combined with key enumeration fields, and concurrent semantic mapping of large models. It fundamentally solves the problem of batch instruction processing failure caused by field mapping errors, and significantly improves the accuracy and processing efficiency of batch instruction conversion.

[0026] It should be noted that the execution entity in this embodiment can be a batch instruction processing system based on multimodal input, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a batch instruction processing device based on multimodal input capable of performing the above functions. This embodiment does not specifically limit it in this way. The following uses a batch instruction processing system based on multimodal input as the execution entity as an example to describe this embodiment and the following embodiments.

[0027] Based on this, Embodiment 1 of this application proposes a batch instruction processing method based on multimodal input. Please refer to... Figure 1 The batch instruction processing method based on multimodal input includes steps S10 to S50: Step S10: Convert the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed.

[0028] In this embodiment, the file to be parsed refers to an electronic file containing batch order information. A horizontal two-dimensional matrix refers to a structured data set with rows as data records and columns as data fields.

[0029] As an optional implementation method, the file's metadata information is first obtained, the file type is determined based on the metadata information, the parsing engine corresponding to the file type is called to read the file content, the data records and field information in the file are extracted, the data records are arranged in row-major order, and a horizontal two-dimensional matrix is ​​generated.

[0030] As an alternative implementation, the binary data stream of the file is read, the actual file type is determined by byte feature matching, and the corresponding parsing engine is invoked to parse the file content. Table structures and non-table text content within the file are identified, and all valid order information is converted into a unified horizontal two-dimensional matrix format. This method solves the parsing failure problem caused by inconsistencies between the file extension and the actual file type, thus improving the accuracy of file parsing.

[0031] Step S20: Calculate the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords, and determine the target two-dimensional matrix based on the intersection coverage rate.

[0032] In this embodiment, standard header keywords refer to the set of keywords corresponding to all header fields in the predefined platform standard template. Intersection coverage refers to the ratio of the number of intersections between the headers of the horizontal two-dimensional matrix and the standard header keywords to the total number of standard header keywords.

[0033] As an optional implementation, the first row of data in the horizontal two-dimensional matrix is ​​extracted as the header to be matched. The header to be matched is compared with the keywords in the standard header one by one. The number of keywords that are successfully matched is counted, the intersection coverage rate is calculated, and the horizontal two-dimensional matrix with the highest intersection coverage rate is selected as the target two-dimensional matrix.

[0034] As an alternative implementation, a corresponding weight value is assigned to each standard header keyword, and a weighted intersection coverage rate is calculated based on the weight values. The horizontal two-dimensional matrix with the highest weighted intersection coverage rate is selected as the target two-dimensional matrix. This approach solves the filtering error problem caused by the different importance of different header fields and improves the accuracy of target table identification.

[0035] Step S30: Assemble multiple header mapping prompts by using key enumeration fields and preset configuration templates to assemble the first row data of the target two-dimensional matrix.

[0036] In this embodiment, the key enumeration field refers to the business field that has a decisive impact on the batch instruction processing result, and its value range is a predefined enumeration set. The preset configuration template refers to a pre-stored text template containing basic instructions, table header mapping rules, and enumeration constraints.

[0037] As an optional implementation, the first row of data of the target two-dimensional matrix is ​​extracted as the source header list, all enumeration values ​​of key enumeration fields are obtained from the preset configuration template, corresponding configuration parameters are generated for each enumeration value, the source header list and configuration parameters are filled into the preset configuration template, and multiple header mapping prompts are generated.

[0038] As an alternative implementation, sample data is extracted from each column of the target two-dimensional matrix. This sample data is then associated with the source header list. Configuration parameters containing sample data are generated for the enumeration values ​​of each key enumeration field. The source header list and the configuration parameters containing sample data are then filled into a preset configuration template to generate multiple header mapping prompts. This approach provides more semantic reference information for the large model, improving the accuracy of header mapping.

[0039] Step S40: The multiple header mapping prompts are concurrently transmitted to the large model to obtain a standardized two-dimensional matrix.

[0040] In this embodiment, the large model refers to a pre-trained language model with natural language understanding and generation capabilities. The standardized two-dimensional matrix refers to a horizontal two-dimensional matrix whose header and content both conform to the platform's standard template requirements.

[0041] As an optional implementation, multiple concurrent connections are established with the large model, and multiple header mapping prompts are transmitted to the large model through different concurrent connections. Multiple header mapping results returned by the large model are received, the optimal header mapping result is selected, and the target two-dimensional matrix is ​​transformed according to the optimal header mapping result to generate a standardized two-dimensional matrix.

[0042] As an alternative implementation, multiple header mapping prompts are packaged into a single request and transmitted in batches to a large model. The large model is then instructed to process all prompts in parallel, receive multiple header mapping results and their corresponding confidence scores returned by the large model, select the header mapping result with the highest confidence score, and perform content transformation and format adjustment on the target two-dimensional matrix based on this result to generate a standardized two-dimensional matrix. This approach reduces the number of network requests and improves processing efficiency.

[0043] Step S50: Call the processing interface to process the standard table file corresponding to the standardized two-dimensional matrix in order to execute batch order placement.

[0044] In this embodiment, the processing interface refers to the application programming interface provided by the platform for receiving standard form files and executing batch order placement operations. The standard form file refers to an electronic spreadsheet file generated according to the platform's standard template format.

[0045] As an optional implementation, a standardized two-dimensional matrix is ​​written into a preset standard table template to generate a standard table file. The platform's batch order processing interface is then called, and the standard table file is transmitted as a parameter to the processing interface to trigger a batch order placement operation.

[0046] As an alternative implementation, the standardized two-dimensional matrix is ​​converted into a structured data format supported by the platform's processing interface. The structured data is then directly transmitted via the platform's batch order processing interface, eliminating the need to generate intermediate standard table files. This approach reduces the overhead of file generation and transmission, thereby increasing the speed of order processing.

[0047] For example, a financial institution's customer uploads a PDF file containing bulk fund purchase instructions. The system first reads the binary data stream of the PDF file, determines its true file type as PDF through byte feature matching, and calls the PDF parsing engine to convert the file into multiple horizontal two-dimensional matrices. The system extracts the first row of data from each horizontal two-dimensional matrix as the header to be matched, compares it with predefined fund transaction standard header keywords, calculates the weighted intersection coverage of each horizontal two-dimensional matrix, and selects the horizontal two-dimensional matrix with the highest weighted intersection coverage as the target two-dimensional matrix. The system extracts the first row of data from the target two-dimensional matrix as the source header list, obtains four enumerated values ​​for the transaction type field from the preset configuration template, generates configuration parameters containing sample data of the corresponding column for each enumerated value, fills the source header list and configuration parameters into the preset configuration template, and generates four header mapping prompts. The system transmits the four header mapping prompts in batches to the large model, specifies that the large model processes all prompts in parallel, receives the four header mapping results returned by the large model and their corresponding confidence scores, selects the mapping result with the highest confidence score, and converts the target two-dimensional matrix into a standardized two-dimensional matrix based on this result. The system writes a standardized two-dimensional matrix into the platform's standard Excel template, generates a standard table file, calls the platform's batch order processing interface to transmit the standard table file, and executes batch fund purchase instructions.

[0048] This embodiment solves the problem of directly processing files of different formats by recognizing multimodal file types and unifying structured conversion, thus expanding the applicability of batch command processing. By filtering target tables based on intersection coverage, it automatically locates valid order tables from multiple candidate tables, reducing the workload of manual filtering. By combining multi-hint word assembly of key enumeration fields and concurrent semantic mapping of large models, it solves the problem of traditional string matching methods failing to identify synonymous fields, significantly improving the accuracy of table header mapping. By directly calling the platform's processing interface to execute batch orders, it achieves full automation from file upload to command execution, greatly improving the efficiency of batch command processing.

[0049] Based on any of the above embodiments, in Embodiment 2 of this application, converting the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed includes: Step S11: Receive at least one file to be parsed.

[0050] In this embodiment, the file to be parsed refers to an electronic file containing batch order information.

[0051] As an optional implementation, the file to be parsed is received from the client via the Hypertext Transfer Protocol, the file is stored in a temporary storage area, a unique file identifier is generated, and the corresponding storage path is associated with it.

[0052] As another optional implementation, the system receives batches of files to be parsed via a message queue, performs integrity checks on each file, generates a file identifier after the check passes, and stores it in a temporary storage area. This approach solves the problem of file upload loss in high-concurrency scenarios and improves the stability of the system.

[0053] Step S12: Determine the file type of the file to be parsed based on the file name suffix and file byte prefix of the file to be parsed.

[0054] In this embodiment, the filename suffix refers to the character sequence following the last period in the filename, used to identify the file's format type. The file byte prefix refers to the fixed-length byte sequence at the beginning of the file's binary data stream, used to uniquely identify the file's true type.

[0055] As an optional implementation, the file name suffix of the file to be parsed is extracted, the sixteen bytes of data at the beginning of the file are read as the file byte prefix, and the file name suffix and the file byte prefix are matched with a preset file type mapping table. When the matching results are consistent, the corresponding file type is determined.

[0056] As another optional implementation, the file name suffix and file byte prefix of the file to be parsed are extracted and matched with a preset file type mapping table. When the matching results of the two are inconsistent, the matching result of the file byte prefix shall prevail, and log information of inconsistent types shall be recorded. This method solves the problem of type identification error caused by file suffix tampering and improves the accuracy of file type identification.

[0057] Step S13: Convert the file to be parsed into the horizontal two-dimensional matrix according to the parsing method corresponding to the file type.

[0058] In this embodiment, the parsing method refers to a predefined data extraction and structured transformation method for different file formats. A horizontal two-dimensional matrix refers to a structured data set with rows as data records and columns as data fields.

[0059] As an optional implementation, the corresponding parsing engine is called according to the determined file type. The parsing engine reads the contents of the file to be parsed, extracts all data records and field information in the file, arranges the data records in row-major order, and generates a horizontal two-dimensional matrix.

[0060] As another optional implementation, the corresponding parsing engine is called according to the determined file type. The parsing engine identifies the table structure and non-table text content in the file, extracts the order information from the table data and the text respectively, and integrates all valid order information into a unified horizontal two-dimensional matrix format. This method solves the problem that order information cannot be extracted from non-table files and expands the applicability of the system.

[0061] For example, a financial institution's customer uploads three files containing batch transaction instructions via a platform webpage. The system receives these three files via Hypertext Transfer Protocol (HTTP) and stores them in a temporary storage area, generating a unique file identifier for each file. The system extracts the filename extension and the first 16 bytes of data for each file and matches them against a preset file type mapping table. The first file has an .xlsx extension and its byte prefix matches Excel format; the second file has a .doc extension but its byte prefix matches PDF format, so the system determines it to be PDF format based on the byte prefix result and logs it; the third file has a .png extension and its byte prefix matches image format. The system then calls the Excel parsing engine, PDF parsing engine, and image parsing engine to process the corresponding files. Each parsing engine extracts the valid order information from the files, arranges the data records in row-major order, and generates a corresponding horizontal two-dimensional matrix.

[0062] This embodiment uses dual verification of filename suffix and file byte prefix to determine file type, solving the problem of type errors caused by easy tampering of single suffix identification methods and improving the accuracy of file type identification. By configuring corresponding parsing methods for different file types, it achieves unified structured conversion of multiple file formats, eliminating data barriers between different file formats. By supporting the simultaneous reception and processing of multiple files to be parsed, it improves the system's batch processing capabilities and meets the large-scale transaction needs of institutional clients.

[0063] Based on any of the above embodiments, in Embodiment 3 of this application, converting the file to be parsed into the horizontal two-dimensional matrix according to the parsing method corresponding to the file type includes: Step S131: If the file type is a table format, read the worksheet data of the file to be parsed and extract the table header and content rows to generate the horizontal two-dimensional matrix.

[0064] In this embodiment, the table format refers to an electronic file format that stores data in a row-column structured manner.

[0065] As an optional implementation, the table parsing engine is invoked to open the file to be parsed, all worksheets in the file are traversed, the first row of data in each worksheet is extracted as the table header, all non-empty rows after the first row are extracted as content rows, and the table header and content rows are combined in row-major order to generate a horizontal two-dimensional matrix.

[0066] As another alternative implementation method, the table parsing engine is called to open the file to be parsed. First, the worksheets containing valid data are selected and blank worksheets without data are excluded. Then, the header and content rows of each valid worksheet are extracted to generate the corresponding horizontal two-dimensional matrix. This method reduces the amount of invalid data to be processed and improves the efficiency of table parsing.

[0067] Step S132: If the file type is a Word document or a PDF document, extract all table structures in the file to be parsed and generate a two-dimensional matrix, and filter the two-dimensional matrix with the highest matching degree based on the table matching degree algorithm and convert it into the horizontal two-dimensional matrix.

[0068] In this embodiment, the table matching degree algorithm refers to a numerical calculation method used to calculate the similarity between candidate tables and the platform's standard template.

[0069] As an optional implementation, the document parsing engine is called to extract all table structures in the file to be parsed, a corresponding two-dimensional matrix is ​​generated for each table, the matching degree value between each two-dimensional matrix and the standard template is calculated based on the table matching degree algorithm, and the two-dimensional matrix with the highest matching degree value is selected and converted into a horizontal two-dimensional matrix.

[0070] As another optional implementation, the document parsing engine is called to extract all table structures in the file to be parsed. Invalid tables with fewer than a preset threshold number of rows or columns are first filtered out. Then, a two-dimensional matrix is ​​generated for the remaining tables and the matching degree value is calculated. The two-dimensional matrix with the highest matching degree value is selected and converted into a horizontal two-dimensional matrix. This method eliminates the interference of irrelevant small tables and improves the accuracy of target table filtering.

[0071] Step S133: If the file type is an image format, extract the text content of the file to be parsed through optical character recognition, and use a large model to extract the table structure from the text content and generate a two-dimensional matrix, and then convert it into the horizontal two-dimensional matrix; if the table structure is not recognized after parsing, extract the full text and generate a two-dimensional matrix with a standard table header as the frame through the large model, and then adjust it into the horizontal two-dimensional matrix.

[0072] In this embodiment, optical character recognition refers to the technology of converting text information in an image into editable text.

[0073] As an optional implementation, an optical character recognition engine is called to extract all text content from the image. The text content is then input into a large model, which identifies the table boundaries and row and column information in the text and generates a two-dimensional matrix. This two-dimensional matrix is ​​then converted into a horizontal two-dimensional matrix. If the large model does not recognize the table structure, the full text is input into the large model, which extracts the order information according to the field order of the standard table header and generates a two-dimensional matrix, which is then adjusted to a horizontal two-dimensional matrix.

[0074] As another alternative implementation, an optical character recognition engine is invoked to extract the text content from the image. First, the text content is denoised and formatted, removing irrelevant whitespace characters and special symbols. Then, the processed text is input into a large model to extract the table structure and generate a two-dimensional matrix. If the table structure is not recognized, the formatted full text is input into the large model to generate a two-dimensional matrix with a standard table header as the frame, and then adjusted to a horizontal two-dimensional matrix. This method improves the accuracy of text recognition and reduces the impact of noise on the analysis results of the large model.

[0075] For example, a financial institution's client simultaneously uploads three batch transaction instruction files in different formats. The first file is an XLSX format spreadsheet file. The system uses a table parsing engine to open this file, filters out the worksheet containing transaction data, and extracts the table headers and content rows to generate a horizontal two-dimensional matrix. The second file is a DOCX format Word document. The system uses a document parsing engine to extract three table structures from the document, filters out invalid tables with two rows, generates two-dimensional matrices for the remaining two tables, calculates the matching degree, and selects the two-dimensional matrix with the higher matching degree to convert it into a horizontal two-dimensional matrix. The third file is a JPG format screenshot of the transaction instructions. The system uses an optical character recognition engine to extract the text content from the image, performs noise reduction on the text, and inputs it into a large model. The large model does not recognize a clear table structure, so it extracts all order information from the text according to the field order of the standard table header to generate a two-dimensional matrix, which is then adjusted to a horizontal two-dimensional matrix.

[0076] For example, refer to Figure 2For files with different input formats and content, a differentiated parsing strategy is adopted to facilitate efficient parsing. Specifically: XLSX files: The worksheet is read directly using the openpyxl library, extracting the header (first row) and content rows to generate a two-dimensional matrix list. XLLS files: First converted to XLLSX format using the pandas library, then incorporated into the XLLSX processing flow. Doc / DocX files: The spirit.doc library is used to search for table structures within the document, extracting a two-dimensional matrix for each table; if multiple tables exist, the best-matching table is selected using a table matching algorithm (based on header similarity calculation with a standard template). PNG / JPG files: Text information is extracted from the image using an OCR model (such as Tesseract), then input into an LLM (such as the GPT series) and combined with a Prompt (such as "Extract table data from the following text") to parse the table structure and generate a two-dimensional matrix; the optimal table is also selected using a table matching algorithm. PDF files: To improve accuracy, each page is first converted to image format and then incorporated into the PNG / JPG processing flow. Handling cases without tables: For doc / docx / png / jpg / pdf files, if no table structure is found after parsing, all text is extracted, and a two-dimensional matrix with a standard table header is directly generated using LLM+Prompt (e.g., "Identify order information from text and generate a table according to standard headers"). Unifying table orientation: Since the original table may contain vertical tables, to ensure consistency in subsequent processing, after extracting the tables using the above method, a vertical table recognition algorithm is used to filter out the vertical tables and transpose them to ensure that the final two-dimensional tables output by the current module are all horizontal tables.

[0077] This embodiment achieves unified structured conversion of table, document, and image files by employing differentiated parsing strategies for different file types, covering all instruction file formats commonly used by institutional clients. By automatically filtering target tables in documents through a table matching algorithm, it solves the problem of manual filtering required for documents with multiple tables, improving processing efficiency. Furthermore, by supporting the parsing of text without table structure in image formats, it enables the automatic extraction of scattered order information, further expanding the system's applicability and reducing operational costs for clients.

[0078] Based on any of the above embodiments, in Embodiment 4 of this application, calculating the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords, and determining the target two-dimensional matrix based on the intersection coverage rate, includes: Step S21: If the number of files in the file to be parsed is one, calculate the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords. If the intersection coverage rate is greater than the coverage rate threshold, then use the horizontal two-dimensional matrix as the target two-dimensional matrix.

[0079] In this embodiment, the coverage threshold refers to a pre-set numerical threshold used to determine whether the horizontal two-dimensional matrix is ​​a valid order form.

[0080] As an optional implementation, the number of all horizontal two-dimensional matrices generated by the file to be parsed is counted. When the number is one, the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords is calculated. The calculated intersection coverage rate is compared with a preset coverage threshold. When the intersection coverage rate is greater than the coverage threshold, the horizontal two-dimensional matrix is ​​determined as the target two-dimensional matrix.

[0081] As another optional implementation, the number of all horizontal two-dimensional matrices generated from the file to be parsed is counted. When the number is one, the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords is calculated. The calculated intersection coverage rate is compared with the dynamically adjusted coverage threshold. When the intersection coverage rate is greater than the coverage threshold, the horizontal two-dimensional matrix is ​​determined as the target two-dimensional matrix. The dynamically adjusted coverage threshold is automatically updated based on the average intersection coverage rate of historically successfully parsed table data. This method solves the problem that fixed thresholds cannot adapt to different business scenarios and improves the adaptability of target table recognition.

[0082] Step S22: If the number of files to be parsed is not one, calculate the intersection coverage of the horizontal two-dimensional matrix and the standard header keywords, and use the horizontal two-dimensional matrix corresponding to the largest intersection coverage as the target two-dimensional matrix.

[0083] As an optional implementation, the number of all horizontal two-dimensional matrices generated by the file to be parsed is counted. When the number is not one, the intersection coverage rate of each horizontal two-dimensional matrix with the standard header keyword is calculated in turn. The numerical values ​​of all intersection coverage rates are compared, and the horizontal two-dimensional matrix corresponding to the intersection coverage rate with the largest value is determined as the target two-dimensional matrix.

[0084] As another optional implementation, the number of all horizontal two-dimensional matrices generated by the file to be parsed is counted. When the number is not one, the maximum intersection coverage of the horizontal two-dimensional matrices in each file is calculated by grouping by file. Then, the maximum intersection coverage values ​​of different files are compared. The horizontal two-dimensional matrix corresponding to the global maximum intersection coverage is determined as the target two-dimensional matrix. This method reduces unnecessary comparison operations and improves the processing efficiency in multi-file scenarios.

[0085] For example, when a financial institution's customer uploads a PDF file containing batch transaction instructions for the first time, the system parses the file and generates a horizontal two-dimensional matrix. The intersection coverage of this matrix with the standard header keywords is calculated to be 0.85. The preset coverage threshold is 0.6. Since 0.85 is greater than 0.6, the system determines this horizontal two-dimensional matrix as the target two-dimensional matrix. The customer then uploads three batch transaction instruction files in different formats simultaneously. The system parses the three files, generating five horizontal two-dimensional matrices. The intersection coverage of each matrix is ​​calculated to be 0.72, 0.68, 0.81, 0.59, and 0.76, respectively. After comparing all values, the system determines the maximum intersection coverage to be 0.81, and the horizontal two-dimensional matrix corresponding to this value is determined as the target two-dimensional matrix.

[0086] This embodiment employs a differentiated target table determination strategy based on the number of horizontal two-dimensional matrices generated from the files to be parsed. In a single-matrix scenario, coverage threshold verification ensures the validity of the target table, preventing invalid tables from entering subsequent processing. In a multi-matrix scenario, global maximum intersection coverage filtering automatically locates the optimal order table, solving the problem of manual filtering required in multi-file or multi-table scenarios, and significantly improving the automation and processing efficiency of batch command processing.

[0087] Based on any of the above embodiments, in Embodiment 5 of this application, the preset configuration template includes a basic instruction template, header data, and enumeration restrictions. The step of assembling the first row data of the target two-dimensional matrix into multiple header mapping prompts using key enumeration fields and the preset configuration template includes: Step S31: Determine the source header list based on the first row of the target two-dimensional matrix.

[0088] In this embodiment, the source header list refers to an ordered list consisting of all field names extracted from the first row of the target two-dimensional matrix.

[0089] As an optional implementation, the contents of all cells in the first row of the target two-dimensional matrix are extracted, empty cells are filtered out, and the contents of the remaining cells are arranged in the original column order to generate the source header list.

[0090] As another alternative implementation method, the contents of all cells in the first row of the target two-dimensional matrix are extracted, empty cells and duplicate cells are filtered out, and the contents of the remaining cells are arranged in the original column order to generate a source header list. This method solves the mapping conflict problem caused by duplicate header fields and improves the accuracy of subsequent processing.

[0091] Step S32: After performing deduplication, null value removal, and space removal on the column data corresponding to each source table in sequence, filter the preset number of column data to obtain the sample data corresponding to each source table.

[0092] In this embodiment, sample data refers to a set of sample data that can represent the content and format of the corresponding column data.

[0093] As an optional implementation, each header field in the source header list is traversed, all data in the column corresponding to that field is extracted, duplicate data, null data and leading and trailing spaces are removed in sequence, and a preset number of data are selected from the processed data as the example data corresponding to that header.

[0094] As another optional implementation, each header field in the source header list is traversed, all data in the corresponding column are extracted, and duplicate data, null data, and leading and trailing spaces are removed in turn. A preset number of data are randomly selected from the processed data as the example data corresponding to the header. This method improves the representativeness of the example data and avoids the sample bias problem caused by sequential sampling.

[0095] Step S33: Determine the target table header corresponding to the source table according to the conversion requirements in the configuration file, and determine the key enumeration fields according to the enumeration constraints in the configuration file.

[0096] In this embodiment, the configuration file refers to a pre-stored structured file that contains platform standard template conversion rules and business constraints.

[0097] As an optional implementation, a preset global configuration file is read to obtain the conversion requirements from the source table header to the target table header defined in the file, as well as a list of business fields containing enumeration constraints, and the list of business fields is determined as the key enumeration fields.

[0098] As another optional implementation, the corresponding scenario-based configuration file is read according to the current business scenario identifier, the conversion requirements from the source table header to the target table header defined in the file are obtained, and the list of business fields containing enumeration constraints is determined as the key enumeration fields. This method realizes flexible switching of conversion rules under different business scenarios and improves the versatility of the system.

[0099] Step S34: Based on the basic instruction template, the header data, and the enumeration restrictions corresponding to the key enumeration fields, and in conjunction with the source table and the example data, assemble the header mapping prompt words corresponding to each key enumeration field.

[0100] In this embodiment, the basic instruction template refers to a predefined text template that includes a general mapping task description and output format requirements.

[0101] As an optional implementation, each key enumeration field is traversed to obtain the basic instruction template, standard header data, and enumeration restriction rules corresponding to that field. The source header list and the example data corresponding to each header are inserted into the specified position of the basic instruction template to generate the header mapping prompt words corresponding to the key enumeration field.

[0102] As another optional implementation, each key enumeration field is traversed to obtain the corresponding basic instruction template, standard header data, and enumeration restriction rules. The content weight of the corresponding part in the basic instruction template is adjusted according to the business importance of the field. Then, the source header list and the example data corresponding to each header are inserted into the adjusted basic instruction template to generate the header mapping prompt words corresponding to the key enumeration field. This method strengthens the semantic guidance of key business fields and improves the accuracy of large model mapping results.

[0103] For example, a financial institution's customer uploads an Excel file containing batch redemption instructions. After parsing and generating a target two-dimensional matrix, the system extracts the content of the first row of non-empty cells to generate a source header list containing four fields: transaction type, product code, applied share, and whether full redemption is required. The system sequentially extracts the data from the corresponding column of each field, performs deduplication, denull value removal, and space removal, and randomly selects four data entries from the processing results of each field as sample data. The system reads the configuration file of the fund trading scenario to obtain the conversion requirements from the source header to the target header, as well as a list of key enumerated fields including the transaction type field. The system iterates through the transaction type fields to obtain the corresponding basic instruction template, standard header data, and enumeration restriction rules for purchase, redemption, conversion, and setting dividend methods. It inserts the source header list and the sample data of the four fields into the basic instruction template to generate header mapping prompts corresponding to the transaction type fields.

[0104] This embodiment provides rich semantic reference information for the large model by extracting the source header list and generating sample data for corresponding columns, enhancing the model's ability to understand the business meaning of the headers. Centralized configuration and flexible updates of business rules are achieved through unified management of conversion requirements and enumeration constraints via configuration files. By combining basic instruction templates, enumeration restrictions, and sample data for key enumeration fields to assemble prompt words, differentiated semantic guidance for different business fields is achieved, significantly improving the accuracy and reliability of header mapping.

[0105] Based on any of the above embodiments, in Embodiment Six of this application, the multiple header mapping prompts are concurrently classified and mapped to a large model to obtain a standardized two-dimensional matrix, including: Step S41: The multiple header mapping prompts are transmitted concurrently to the large model to obtain the header mapping scheme and confidence level output by the large model.

[0106] In this embodiment, confidence level refers to the quantitative evaluation value of the correctness of the header mapping scheme of the large model's output.

[0107] As an alternative implementation, multiple independent network connections are established with the large model, and each header mapping prompt word is transmitted to the large model separately through different network connections. At the same time, the header mapping scheme and confidence value corresponding to each prompt word returned by the large model are received.

[0108] As another alternative implementation, multiple header mapping hints are encapsulated into a batch request data packet, which is transmitted to the large model through a single network connection. The large model is then instructed to process all hints in parallel and receive a batch response data packet containing all header mapping schemes and corresponding confidence levels returned by the large model. This approach reduces the overhead of establishing and disconnecting network connections and improves processing efficiency in high-concurrency scenarios.

[0109] Step S42: Determine the target header mapping scheme based on the confidence level. The target header mapping scheme includes transaction type values ​​and mapping relationships.

[0110] In this embodiment, the target header mapping scheme refers to the set of optimal header mapping relationships determined through screening.

[0111] As an optional implementation, the confidence values ​​of all header mapping schemes are compared, and the header mapping scheme with the highest confidence value is selected as the target header mapping scheme. The transaction type value and the mapping relationship from the source header to the target header are then extracted from this scheme.

[0112] As another optional implementation, the confidence values ​​of all header mapping schemes are compared, and candidate schemes with confidence values ​​higher than a preset confidence threshold are selected. Business rules are validated for each candidate scheme, and the candidate scheme that passes the business rules validation and has the highest confidence value is selected as the target header mapping scheme. This approach solves the problem that high-confidence schemes do not conform to business logic and improves the reliability of the mapping results.

[0113] Step S43: Based on the header mapping relationship dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value, the semantics of the target two-dimensional matrix are converted into the standardized two-dimensional matrix.

[0114] In this embodiment, the content mapping rule refers to a predefined set of cell content conversion rules for different transaction types.

[0115] As an optional implementation, a header mapping relationship dictionary is generated based on the mapping relationship, and the corresponding content mapping rules are loaded according to the transaction type value. First, the columns of the target two-dimensional matrix are rearranged and mapped according to the header mapping relationship dictionary, and then the cell content of the corresponding column is semantically transformed according to the content mapping rules to generate a standardized two-dimensional matrix.

[0116] As another optional implementation, a header mapping dictionary is generated based on the mapping relationship, and the corresponding content mapping rules are loaded according to the transaction type value. At the same time, header column mapping and cell content semantic conversion operations are performed. The converted columns are arranged according to the standard header order to generate a standardized two-dimensional matrix. This method reduces the overall processing time and improves the conversion efficiency through parallel processing.

[0117] For example, a financial institution's customer uploads a Word document containing batch fund conversion instructions. After the system parses and generates a target two-dimensional matrix, it assembles four header mapping prompts corresponding to purchase, redemption, conversion, and setting dividend methods, respectively. The system encapsulates these four prompts into a batch request data packet and transmits it to the large model, which is then instructed to process them in parallel. The large model returns four header mapping schemes with corresponding confidence scores of 0.72, 0.81, 0.95, and 0.68. The system compares all confidence scores and selects three candidate schemes with scores higher than a preset threshold of 0.7. Each candidate scheme undergoes business rule verification. The conversion type scheme with a confidence score of 0.95 is confirmed to conform to the business rules and is thus identified as the target header mapping scheme. The transaction type value is extracted as conversion, along with the corresponding mapping relationship. The system generates a header mapping relationship dictionary based on the mapping relationship, loads the content mapping rules corresponding to the conversion type, and simultaneously performs header column mapping and cell content semantic conversion. The converted columns are then arranged according to the standard header order to generate a standardized two-dimensional matrix.

[0118] This embodiment shortens the overall processing time of header mapping by concurrently transmitting multiple header mapping prompts to a large model, fully utilizing the model's parallel processing capabilities. The optimal header mapping scheme is selected based on confidence scores, achieving automatic optimization of mapping results and avoiding manual intervention. Semantic conversion is performed by combining content mapping rules corresponding to transaction type values, ensuring the accuracy and business compliance of cell content conversion and significantly improving the overall quality and efficiency of batch instruction conversion.

[0119] Based on any of the above embodiments, in Embodiment Seven of this application, before converting the semantics of the target two-dimensional matrix into the standardized two-dimensional matrix according to the header mapping relationship dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value, the following steps are included: Step A10: Determine the mapping logic corresponding to the unmapped fields in the target two-dimensional matrix based on the mapping relationship.

[0120] In this embodiment, an unmapped field refers to a source header name in the target two-dimensional matrix for which no corresponding target header is found in the mapping relationship.

[0121] As an optional implementation, iterate through all source header names of the target two-dimensional matrix, filter out unmapped fields not included in the mapping relationship, and determine the corresponding mapping logic based on the field name of the unmapped field and the example data of the corresponding column.

[0122] As another alternative implementation, all source header names of the target two-dimensional matrix are traversed, and unmapped fields not included in the mapping relationship are filtered out. The names of the unmapped fields and the sample data of the corresponding columns are input into the large model, and the large model determines the corresponding mapping logic. This method solves the problem that manual methods cannot cover all synonyms and improves the recognition accuracy of unmapped fields.

[0123] Step A11: Obtain the header mapping relationship dictionary based on the mapping relationship and the mapping logic.

[0124] In this embodiment, the header mapping dictionary refers to a set of key-value pairs with the source header name as the key and the target header name as the value.

[0125] As an optional implementation, the existing mapping relationships and the mapping logic corresponding to the unmapped fields are merged to generate a dictionary of table header mapping relationships containing the corresponding relationships of all source table header names.

[0126] As another optional implementation, the existing mapping relationships and the mapping logic corresponding to the unmapped fields are merged. The merged mapping relationships are deduplicated and conflict-checked to remove duplicate key-value pairs and conflicting mapping relationships, generating the final header mapping relationship dictionary. This method avoids conversion errors caused by mapping relationship conflicts and improves the accuracy and reliability of the dictionary.

[0127] For example, a financial institution's customer uploads a PDF file containing bulk fund purchase instructions. After the system parses and generates a target two-dimensional matrix, it obtains a preliminary mapping relationship through concurrent mapping in a large model. This mapping relationship includes the correspondence between three fields: transaction type, product code, and application amount. The system iterates through all source header names of the target two-dimensional matrix, filters out the unmapped field "Whether there is a front-end fee" not included in the mapping relationship, and inputs the field name and sample data of the corresponding column into the large model. The large model determines that its corresponding mapping logic is mapped to the "Fee Method" field in the standard header. The system merges the existing mapping relationship with the mapping logic of the unmapped field, performs deduplication and conflict verification on the merged mapping relationship, and after confirming that there are no duplicates or conflicts, generates a header mapping relationship dictionary containing the correspondence between all source header names.

[0128] This embodiment solves the problem of field loss caused by incomplete initial mapping relationships by automatically identifying unmapped fields in the target two-dimensional matrix and determining their corresponding mapping logic, ensuring that all valid business fields can be correctly mapped. By merging existing mapping relationships and the mapping logic of unmapped fields and performing verification, a complete and accurate dictionary of table header mapping relationships is generated, providing a reliable basis for subsequent content semantic conversion and significantly improving the completeness and accuracy of batch instruction conversion.

[0129] Based on any of the above embodiments, in Embodiment 8 of this application, the content mapping rule includes columns to be enumerated and their target enumerated value set, columns to be converted to units and their conversion coefficients, and columns to be formatted and their target format. The step of converting the semantics of the target two-dimensional matrix into the standardized two-dimensional matrix according to the header mapping relationship dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value includes: Step A12: Determine the columns to be converted in the target two-dimensional matrix according to the content mapping rules.

[0130] In this embodiment, the column to be converted refers to the column in the target two-dimensional matrix that needs to undergo semantic conversion, unit conversion, or format adjustment.

[0131] As an optional implementation, all transformation items in the content mapping rules are traversed, the target header name corresponding to each transformation item is extracted, the source header column corresponding to each target header name is determined according to the header mapping relationship dictionary, and these source header columns are determined as columns to be transformed.

[0132] As another optional implementation, all columns of the target two-dimensional matrix are traversed, and the target header name corresponding to each column is determined according to the header mapping relationship dictionary. The target header name is checked to see if it exists in the content mapping rules. If it exists, the column is determined as the column to be converted. This method avoids missing the newly added conversion items in the content mapping rules and improves the completeness of the identification of the column to be converted.

[0133] Step A13: Determine the target header name and enumeration constraints corresponding to the column to be converted based on the header mapping relationship dictionary.

[0134] In this embodiment, the enumeration constraint refers to the finite set of values ​​that column data is allowed to take, as specified in the content mapping rules.

[0135] As an optional implementation, the target header name corresponding to each column to be converted is queried according to the header mapping relationship dictionary, and the enumeration constraint corresponding to the target header name is extracted from the content mapping rules.

[0136] As another optional implementation, the target header name corresponding to each column to be converted is queried according to the header mapping relationship dictionary. The enumeration constraints, unit conversion coefficients and format adjustment requirements corresponding to the target header name are extracted from the content mapping rules. This information is associated with the column to be converted and stored. This method centrally associates all conversion-related information, which facilitates the rapid assembly of subsequent prompt words.

[0137] Step A14: Construct column hints for the column to be converted based on the source header name, original data, target header name, content mapping rules, and enumeration constraints of the column to be converted.

[0138] In this embodiment, column prompts refer to text instructions specifically used to guide large models in performing semantic transformation and format adjustment on individual column data.

[0139] As an optional implementation, the source header name, all original data, corresponding target header name, enumeration constraints, unit conversion factor and format adjustment requirements of the column to be converted are extracted, and this information is assembled according to the preset column conversion instruction template to generate the column prompt words for the column to be converted.

[0140] As another optional implementation method, the source header name, all original data, corresponding target header name, enumeration constraints, unit conversion factor and format adjustment requirements of the column to be converted are extracted. After deduplication of the original data, a representative data subset is generated. The representative data subset is then assembled with other information according to a preset column conversion instruction template to generate the column prompt words for the column to be converted. This method reduces the length of the prompt words, reduces the processing overhead of large models, and ensures the accuracy of the conversion.

[0141] Step A15: Send the column prompts of each column to be converted to the large model by column dimension, so as to receive the standardized data output by the large model.

[0142] As an optional implementation, an independent large model call request is established for each column to be transformed. The column prompts for each column are sent to the large model concurrently through independent requests, while the standardized data corresponding to each column returned by the large model is received.

[0143] As another alternative implementation, the column hints of all columns to be converted are encapsulated into a batch request, sent to the large model, and the large model is specified to process in parallel on a column-by-column basis. The batch response containing the standardized data of all columns is received from the large model. This approach reduces the number of network requests and improves processing efficiency in high-concurrency scenarios.

[0144] Step A16: Replace the original data in the column to be transformed with the standardized data to obtain the standardized two-dimensional matrix.

[0145] As an optional implementation, each column to be transformed is traversed, and the corresponding standardized data returned by the large model is used to replace the original data of that column in row order. After all columns are replaced, a standardized two-dimensional matrix is ​​generated.

[0146] As another alternative implementation, a new two-dimensional matrix is ​​created, and the standardized data of each column is filled in sequentially according to the order of the standard header. For columns that do not require transformation, the original data is directly copied. After filling, a standardized two-dimensional matrix is ​​generated. This method avoids errors caused by modifying the original data and facilitates subsequent verification and backtracking.

[0147] For example, a financial institution's customer uploads an Excel file containing batch fund purchase instructions. The system parses and generates a target two-dimensional matrix. After determining the transaction type as "purchase," it loads the content mapping rules corresponding to the purchase type. These rules include enumerated conversions of transaction types, unit conversions of application amounts, and format adjustments for transaction dates. The system iterates through all columns of the target two-dimensional matrix and determines the transaction type, application amount, and transaction date columns as the columns to be converted based on the header mapping relationship dictionary. The system queries the target header corresponding to the transaction type, which is the operation type, with enumeration constraints of purchase, redemption, conversion, and setting dividend method; the target header corresponding to the application amount is the application amount, with a conversion factor of 10000; and the target header corresponding to the transaction date is the transaction date, with a target format of YYYY-MM-DD. The system extracts the source header name, original data, target header name, and corresponding conversion rules for each column, and constructs three column prompts accordingly. The system encapsulates the three column prompts into a batch request and sends it to the large model. It then receives standardized data returned by the large model. Specifically, "Buy" and "Purchase" in the transaction type column are converted to "Purchase," the values ​​in the application amount column are multiplied by 10000 to convert to yuan, and "2026 / 5 / 25" in the transaction date column is converted to "2026-05-25." The system creates a new two-dimensional matrix, filling it with the standardized data from each column and the columns that do not require conversion, according to the standard header order, generating the final standardized two-dimensional matrix.

[0148] This embodiment achieves precise column positioning by automatically identifying columns to be converted based on content mapping rules, avoiding the tedious manual specification of conversion columns. By constructing column hints using a dictionary of header mapping relationships and content mapping rules, clear conversion instructions and constraints are provided to the large model, ensuring the accuracy and business compliance of content conversion. By concurrently sending column hints to the large model along columns, the parallel processing capabilities of the large model are fully utilized, shortening the overall content conversion time. By replacing the original data with standardized data to generate a standardized two-dimensional matrix, a complete conversion from source data to standard data is achieved, providing a reliable data foundation for subsequent batch order placement.

[0149] Based on any of the above embodiments, in Embodiment Nine of this application, the aim is to establish an accurate correspondence between the customer table header and the platform standard table header. Its core innovation lies in using a small amount of sample data to enhance the understanding of the semantics of the large language model and to perform concurrent and accurate classification mapping for key fields.

[0150] Data Preparation and Extraction: Extract the first row from the input two-dimensional matrix as the "source header list" to be mapped. For each header field in the "source header list", perform the following operations in its corresponding data column to generate "sample data" for that field: Deduplication and Filtering: Traverse all data in the column, remove identical duplicate values, and filter out null values ​​or pure whitespace values.

[0151] Sampling: From the deduplicated and non-empty results, N rows of data (e.g., N=4 in a preferred embodiment) are randomly or sequentially selected as examples representing the content and format of the column of data.

[0152] Dynamic assembly of prompts: The system reads detailed rules about key mapping columns from a predefined configuration file. For example, for the key column "Transaction Type," the configuration file explicitly specifies: Conversion requirement: This column must be mapped to the "Operation Type" field in the platform's standard template. Enumeration constraint: The mapped value must strictly belong to a predefined enumeration set, such as {"Purchase," "Redemption," "Conversion," "Set Dividend Method"}.

[0153] The system dynamically assembles the following elements into the final header mapping Prompt: Basic instruction template: containing a task description ("Map the source header to the target header"), a list of target headers, general mapping rules (such as ignoring case sensitivity, handling common synonyms), and a few general examples. Current source header and its example data: formatted as "Header A: [Example value 1, Example value 2, ...]".

[0154] Special configuration for key columns: explicitly define the mapping target and enumeration restrictions for key columns such as "transaction type".

[0155] Concurrent Classification Mapping: Traditional mapping methods submit all table headers to the model at once. This embodiment employs a concurrent mapping strategy based on key enumeration fields (such as transaction type) as the classification dimension to improve the accuracy of key field mapping. Core Idea: Due to the significant differences in necessary fields and business context associated with different transaction types (such as "purchase" and "redemption"), separating and prioritizing the mapping of "transaction type" from other table header mappings can guide the model to obtain more accurate contextual information. Execution Process: The system first identifies the key enumeration fields (such as "transaction type") defined in the configuration, and then concurrently submits multiple mapping requests to the large language model. Each request's prompt emphasizes a candidate transaction type enumeration value and its associated field mapping tendency. The mapping results returned by the model include the mapping relationship of the "transaction type" field and mapping suggestions for other fields. By comparing and analyzing these concurrently returned results, the system can more reliably determine the final "transaction type" mapping relationship and accordingly determine the best-matching complete set of table header mapping schemes.

[0156] After completing the header mapping and establishing the correspondence between columns, this process performs semantic transformation on the specific data in the table, especially converting non-standardized descriptions into standardized enumeration values ​​or other formats required by the platform.

[0157] Initialization based on header mapping results: Based on the established header mapping relationships, especially the mapping result of the "Transaction Type" main field, the system loads the detailed content mapping rules corresponding to that transaction type. These rules define: which columns require enumeration conversion (e.g., mapping "Confirmation Method" to {"Yes", "No"}); which columns require unit conversion or format adjustment (e.g., converting the amount unit "ten thousand yuan" to "yuan"); and the field dependencies under different transaction types (e.g., under the "Redemption" type, "Full Redemption?" is a required field).

[0158] Prompt assembly and column-level concurrent processing: For each column requiring content transformation, the system constructs a dedicated content mapping prompt. This prompt includes: the source header name and target header name for the column; the specific transformation rules and enumeration constraints for the column; and a complete list of the original data for the column (or a sufficiently large representative subset). The system concurrently submits the above prompts to the large language model, column by column. The model's task is to transform each original value in the input list into a standardized value that meets the target requirements, according to the rules. For example, transforming ["buy", "purchase", "purchase"] into ["purchase", "purchase", "purchase"].

[0159] Reference Figure 3After intelligent mapping is completed, this stage refines the transformation results through code rules to ensure that the data absolutely meets the platform's order requirements. This includes, for example, validating and correcting enumerated values ​​and standardizing time formats. Through the precise collaboration of these three sub-steps—header mapping, content mapping, and data validation and post-processing—the high reliability and compliance of the output data in terms of business requirements are guaranteed, laying a solid foundation for the final standardized output.

[0160] Optionally, the header extraction operation is performed first, followed by a judgment step of "whether the example extraction of each header is completed". If the judgment result is no, four rows of deduplicated and non-empty example data are extracted from each header column, and then the judgment step is returned. If the judgment result is yes, the configuration information of the key column (transaction type) is obtained. Then, the large model is queried concurrently for each transaction type to obtain the corresponding mapping relationship. After summarizing, the main mapping relationship and its candidate relationship for each transaction type are obtained. Then, the operation of retrieving the first non-empty column in the candidate relationship of the content mapping relationship is performed concurrently. Then, other data format processing, including enumeration, time, .OF and other types, is performed. Finally, the strong enumeration field restriction processing is completed. For example, for fields such as whether to redeem in full, only the two values ​​"yes" or "no" can be recognized.

[0161] The data, after intelligent mapping and post-processing, is written into a new .xlsx file according to the platform's standard Excel template format. This file fully complies with the platform's batch order interface requirements and is automatically retrieved by the system via the order API.

[0162] Meanwhile, this embodiment aims to address the core problems of existing batch order technology on the platform, such as high customer adaptation costs, low operational efficiency, and susceptibility to errors due to the platform's single template, input format limitations, and lack of semantic understanding capabilities. Specifically, the technical problems to be solved in this embodiment include: how to automatically identify and understand the diverse batch order files (multimodal input) provided by customers and structure their content.

[0163] How to accurately map customer-defined table headers and semantic content to the platform's standard template fields and enumeration value system without pre-defined rules, especially accurately identifying transaction types (such as four standard transaction types: purchase, conversion, redemption, and setting dividend methods); how to design a unified and efficient multimodal input processing framework that is compatible with various input sources such as structured (Excel), semi-structured (Word, PDF tables), and unstructured (images, plain text), and outputs a standardized two-dimensional data matrix for subsequent processing.

[0164] Compared with existing technologies, what are the advantages of this embodiment? Multimodal compatibility: This embodiment not only supports traditional .xlsx / .xls table files, but also directly processes .doc / .docx documents, .pdf files, and image formats such as .png / .jpg, truly achieving "one-click conversion of any format," greatly expanding the application scenarios and user convenience. High-precision intelligent semantic mapping: By introducing a Large Language Model (LLM) combined with specific prompt word engineering, this embodiment can understand the business semantics of table headers and data content, achieving a mapping from "formal similarity" to "essential similarity." Even if the customer template and the standard template differ greatly in expression, high-accuracy conversion can be achieved, significantly reducing the need for manual intervention. Robustness and flexibility of the processing flow: Differentiated preprocessing flows (such as OCR recognition, document parsing, and table extraction) are designed for different input formats, and finally normalized into unified structured data (two-dimensional matrix). Meanwhile, the post-processing mechanism of "LLM understanding + code rule constraints" ensures that the conversion results are both intelligent and flexible, and strictly comply with the platform's order placement business rules (such as required field validation and enumeration value conversion); a robust table extraction mechanism: for documents or images containing multiple tables, a table matching degree algorithm (calculating the intersection coverage of keywords in the header fields of each candidate table and the platform's standard template) is introduced to automatically filter the most relevant order placement tables, improving parsing accuracy; end-to-end automated process: from uploading the original file to generating the standard template, no manual intervention is required, significantly shortening the order preparation time and improving platform usability and customer stickiness; scalable architecture design: the multimodal parsing module and the LLM table mapping module are decoupled, facilitating rapid iteration when adding new file formats or adjusting the standard template in the future.

[0165] This embodiment provides a batch transaction instruction conversion method based on multimodal data processing and intelligent mapping. Its core lies in constructing an automated system capable of handling multiple input formats and intelligently converting them into a standard output format. The technical solution mainly includes a multimodal input processing module, an intelligent mapping conversion engine, and a standardized output module. The overall system processing flow includes: Input preprocessing: The true type of the file is obtained by double verification using the file name suffix and the file byte representation prefix, and then classified into different parsing submodules.

[0166] Two-dimensional matrix generation: Convert various inputs into a unified horizontal two-dimensional matrix List (i.e., structured data of the source file header row and content row).

[0167] Intelligent semantic mapping: Utilizing the LLM+Prompt mechanism to perform header and content mapping on two-dimensional matrices, and then assembling the data.

[0168] This application provides a batch instruction processing device based on multimodal input. The batch instruction processing device based on multimodal input includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the batch instruction processing method based on multimodal input in the above embodiment 1.

[0169] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a batch instruction processing device based on multimodal input suitable for implementing embodiments of this application. The batch instruction processing device based on multimodal input in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, personal digital assistants (PDAs), tablets, and in-vehicle terminals, as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The batch instruction processing device based on multimodal input shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0170] like Figure 4As shown, a batch instruction processing device based on multimodal input may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 1002 or a program loaded from storage device 1003 into random access memory (RAM) 1004. The random access memory 1004 also stores various programs and data required for the operation of the batch instruction processing device based on multimodal input. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the multimodal input-based batch instruction processing device to wirelessly or wiredly communicate with other devices to exchange data. Although a multimodal input-based batch instruction processing device with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems may be implemented alternatively.

[0171] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0172] The batch instruction processing device based on multimodal input provided in this application, employing the batch instruction processing method based on multimodal input in the above embodiments, can solve the technical problem of low accuracy in synonym field recognition leading to batch instruction processing failure in standardized batch instruction template scenarios. Compared with the prior art, the beneficial effects of the batch instruction processing device based on multimodal input provided in this application are the same as those of the batch instruction processing device based on multimodal input provided in the above embodiments, and other technical features in this batch instruction processing device based on multimodal input are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0173] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0174] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0175] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the batch instruction processing method based on multimodal input in the above embodiments.

[0176] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), or any suitable combination thereof.

[0177] The aforementioned computer-readable storage medium may be included in a batch instruction processing device based on multimodal input; or it may exist independently and not assembled into a batch instruction processing device based on multimodal input.

[0178] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by a batch instruction processing device based on multimodal input, cause the batch instruction processing device based on multimodal input to: convert the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed; Calculate the intersection coverage of the horizontal two-dimensional matrix and the standard header keywords, and determine the target two-dimensional matrix based on the intersection coverage. The first row of the target two-dimensional matrix is ​​assembled using key enumeration fields and a preset configuration template to obtain multiple header mapping prompts; The multiple header mapping prompts are concurrently transmitted to the large model to obtain a standardized two-dimensional matrix; The processing interface is invoked to process the standard table file corresponding to the standardized two-dimensional matrix in order to execute batch order placement.

[0179] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0180] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented in systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0181] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0182] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described batch instruction processing method based on multimodal input. This solves the technical problem of low accuracy in synonym field recognition, leading to batch instruction processing failure in standardized batch instruction template scenarios. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the batch instruction processing method based on multimodal input provided in the above embodiments, and will not be repeated here.

[0183] This application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the batch instruction processing method based on multimodal input as described above.

[0184] The computer program product provided in this application can solve the technical problem of low accuracy in synonym field recognition, which leads to the failure of batch instruction processing in standardized batch instruction template scenarios. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the batch instruction processing method based on multimodal input provided in the above embodiments, and will not be repeated here.

[0185] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. A batch instruction processing method based on multimodal input, characterized in that, The batch instruction processing method based on multimodal input includes: Convert the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed; Calculate the intersection coverage of the horizontal two-dimensional matrix and the standard header keywords, and determine the target two-dimensional matrix based on the intersection coverage. The first row of the target two-dimensional matrix is ​​assembled using key enumeration fields and a preset configuration template to obtain multiple header mapping prompts; The multiple header mapping prompts are concurrently transmitted to the large model to obtain a standardized two-dimensional matrix; The processing interface is invoked to process the standard table file corresponding to the standardized two-dimensional matrix in order to execute batch order placement.

2. The batch instruction processing method based on multimodal input as described in claim 1, characterized in that, The step of converting the file to be parsed into a horizontal two-dimensional matrix according to the file type of the file to be parsed includes: Receive at least one file to be parsed; The file type of the file to be parsed is determined based on the file name suffix and file byte prefix of the file to be parsed. Based on the parsing method corresponding to the file type, the file to be parsed is converted into the horizontal two-dimensional matrix.

3. The batch instruction processing method based on multimodal input as described in claim 2, characterized in that, The step of converting the file to be parsed into the horizontal two-dimensional matrix according to the parsing method corresponding to the file type includes: If the file type is a table format, read the worksheet data of the file to be parsed and extract the table header and content rows to generate the horizontal two-dimensional matrix; If the file type is a Word document or a PDF document, extract all table structures in the file to be parsed and generate a two-dimensional matrix, and filter the two-dimensional matrix with the highest matching degree based on the table matching degree algorithm and convert it into the horizontal two-dimensional matrix; If the file type is an image format, the text content of the file to be parsed is extracted by optical character recognition, and a large model is used to extract the table structure from the text content and generate a two-dimensional matrix, which is then converted into the horizontal two-dimensional matrix; if the table structure is not recognized after parsing, the full text is extracted and a two-dimensional matrix with a standard table header is generated by the large model, which is then adjusted into the horizontal two-dimensional matrix.

4. The batch instruction processing method based on multimodal input as described in claim 1, characterized in that, The calculation of the intersection coverage rate between the horizontal two-dimensional matrix and the standard header keywords, and the determination of the target two-dimensional matrix based on the intersection coverage rate, includes: If the number of files in the file to be parsed is one, then calculate the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords. If the intersection coverage rate is greater than the coverage rate threshold, then use the horizontal two-dimensional matrix as the target two-dimensional matrix. If the number of files to be parsed is not one, the intersection coverage rate of the horizontal two-dimensional matrix and the standard header keywords is calculated, and the horizontal two-dimensional matrix corresponding to the largest intersection coverage rate is taken as the target two-dimensional matrix.

5. The batch instruction processing method based on multimodal input as described in claim 1, characterized in that, The preset configuration template includes a basic instruction template, header data, and enumeration restrictions. The process of assembling the first row data of the target two-dimensional matrix using key enumeration fields and the preset configuration template to obtain multiple header mapping prompts includes: The source header list is determined based on the first row of the target two-dimensional matrix; After performing deduplication, null value removal, and space removal on the column data corresponding to each source table in sequence, the sample data corresponding to each source table is obtained by filtering the preset number of column data. The target table header corresponding to the source table is determined according to the conversion requirements in the configuration file, and the key enumeration fields are determined according to the enumeration constraints in the configuration file. Based on the basic instruction template corresponding to the key enumeration field, the table header data, the enumeration restrictions, the source table, and the example data, the table header mapping prompt words corresponding to each key enumeration field are assembled.

6. The batch instruction processing method based on multimodal input as described in claim 1, characterized in that, The step of concurrently classifying and mapping the multiple header mapping prompts to a large model to obtain a standardized two-dimensional matrix includes: The multiple header mapping prompts are transmitted concurrently to the large model to obtain the header mapping scheme and confidence level output by the large model; The target header mapping scheme is determined based on the confidence level, and the target header mapping scheme includes transaction type values ​​and mapping relationships. Based on the header mapping dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value, the semantics of the target two-dimensional matrix are converted into the standardized two-dimensional matrix.

7. The batch instruction processing method based on multimodal input as described in claim 6, characterized in that, Before converting the semantics of the target two-dimensional matrix into the standardized two-dimensional matrix based on the header mapping dictionary corresponding to the mapping relationship and the content mapping rule corresponding to the transaction type value, the process includes: Based on the mapping relationship, determine the mapping logic corresponding to the unmapped fields in the target two-dimensional matrix; The header mapping relationship dictionary is obtained based on the mapping relationship and the mapping logic.

8. The batch instruction processing method based on multimodal input as described in claim 6, characterized in that, The content mapping rules include columns to be enumerated and their target enumerated value set, columns to be converted to units and their conversion coefficients, and columns to be formatted and their target format. The step of converting the semantics of the target two-dimensional matrix into the standardized two-dimensional matrix based on the header mapping dictionary corresponding to the mapping relationship and the content mapping rules corresponding to the transaction type value includes: The columns to be transformed in the target two-dimensional matrix are determined according to the content mapping rules; The target header name and enumeration constraints corresponding to the column to be converted are determined based on the header mapping relationship dictionary. Based on the source header name, original data, label header name, content mapping rules, and enumeration constraints of the column to be converted, construct the column hint words for the column to be converted; The column prompts for each column to be transformed are sent to the large model in terms of column dimension, so as to receive the standardized data output by the large model; The standardized two-dimensional matrix is ​​obtained by replacing the original data in the column to be transformed with the standardized data.

9. A batch instruction processing device based on multimodal input, characterized in that, The batch instruction processing device based on multimodal input includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the batch instruction processing method based on multimodal input as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the batch instruction processing method based on multimodal input as described in any one of claims 1 to 8.

Images