Fund transaction form element extraction method and device based on large model and medium

Abstract

The invention discloses a fund transaction form element extraction method and device based on a large model and a medium, and the method comprises the steps: obtaining a to-be-processed fund transaction form, and converting the fund transaction form into form image data; performing classification processing on the form image data through a deep learning technology so as to remove invalid form image data, and storing valid form image data as first target image data; performing character recognition on the first target image data through an OCR recognition technology to obtain character information in the first target image data; inputting the character information into a large language model, and outputting corresponding structured data information by the large language model; and performing legality verification on the structured data information, and outputting an element extraction result of the fund transaction form. According to the method, the deep learning technology, the OCR recognition technology and the artificial intelligence technology are combined to perform element extraction on the fund transaction form, and the element extraction efficiency can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of computer software technology, and in particular to a method, device and medium for extracting elements from a fund transaction form based on a large model. Background Art

[0002] The fund transaction form is a form that investors need to fill out when conducting fund transactions. It is mainly used to record specific transaction information, including investor information, fund information, transaction amount, etc. Among them, investor information includes investor name, ID number, contact information, etc.; fund information includes fund name, fund code, transaction amount, transaction currency, etc.; transaction information includes transaction type (such as subscription, application, redemption, etc.), transaction amount, transaction time, etc. Risk warning means that investors need to confirm that they have understood the risks of the fund and promise that the information provided is true, accurate and complete.

[0003] Currently, fund transaction forms are presented in PDF files, and existing form recognition uses OCR technology to extract text information from PDF files. The drawback of this method is that the extracted text is unstructured information, which will cause the content to be disorganized, making it difficult for staff to quickly obtain the corresponding form information, thereby making the overall form information extraction efficiency low and unable to better meet transaction needs. Summary of the invention

[0004] The embodiments of the present invention provide a method, device, computer equipment and storage medium for extracting elements from a fund transaction form based on a large model, aiming to improve the efficiency of extracting elements from a fund transaction form.

[0005] In a first aspect, an embodiment of the present invention provides a method for extracting elements from a fund transaction form based on a large model, comprising:

[0006] Acquire a fund transaction form to be processed, and convert the fund transaction form into form image data;

[0007] Classify the form image data by deep learning technology to remove invalid form image data and save valid form image data as first target image data;

[0008] Performing text recognition on the first target image data by using OCR recognition technology to obtain text information therein;

[0009] Inputting the text information into a large language model, and having the large language model output corresponding structured data information;

[0010] The structured data information is subjected to a legality check, and based on the result of the legality check, an element extraction result of the fund transaction form is output.

[0011] In a second aspect, an embodiment of the present invention provides a fund transaction form element extraction device based on a large model, comprising:

[0012] An image conversion unit, used for acquiring a fund transaction form to be processed, and converting the fund transaction form into form image data;

[0013] An image classification unit, used for classifying the form image data by deep learning technology to remove invalid form image data and save valid form image data as first target image data;

[0014] A text recognition unit, used to perform text recognition on the first target image data by using OCR recognition technology to obtain text information therein;

[0015] An information extraction unit, used for inputting the text information into a large language model, and having the large language model output corresponding structured data information;

[0016] The information verification unit is used to perform a legality verification on the structured data information and output an element extraction result of the fund transaction form based on the result of the legality verification.

[0017] In a third aspect, an embodiment of the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for extracting fund transaction form elements based on a large model as described in the first aspect is implemented.

[0018] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for extracting fund transaction form elements based on a large model as in the first aspect is implemented.

[0019] The embodiment of the present invention provides a method, device, computer equipment and storage medium for extracting elements from a fund transaction form based on a large model. The method includes: obtaining a fund transaction form to be processed, and converting the fund transaction form into form image data; classifying the form image data through deep learning technology to eliminate invalid form image data, and saving valid form image data as first target image data; performing text recognition on the first target image data through OCR recognition technology to obtain text information therein; inputting the text information into a large language model, and the large language model outputs corresponding structured data information; performing a legitimacy check on the structured data information, and outputting the element extraction result of the fund transaction form based on the result of the legitimacy check. Before adopting the OCR recognition technology, the embodiment of the present invention pre-uses deep learning technology to perform image classification processing on the fund transaction form to eliminate invalid image data therein, which can reduce the workload in the OCR recognition process and improve the recognition effect. At the same time, after completing the OCR recognition, the embodiment uses a large language model to extract structured information from the OCR recognition result, which can solve the problem of the disorderly OCR recognition result. The embodiment further performs a legality check on the extracted structured information to output a more accurate and reliable element extraction result, thereby improving the element extraction efficiency for the fund transaction form. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0021] Figure 1 A schematic diagram of a flow chart of a method for extracting elements from a fund transaction form based on a large model provided by an embodiment of the present invention;

[0022] Figure 2 A schematic block diagram of a fund transaction form element extraction device based on a large model provided by an embodiment of the present invention;

[0023] Figure 3 A network architecture diagram of a fund transaction form element extraction method based on a large model provided by an embodiment of the present invention;

[0024] Figure 4 A schematic diagram of a sub-process in a method for extracting elements from a fund transaction form based on a large model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0027] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0028] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0029] See below Figure 1 The embodiment of the present invention provides a method for extracting elements from a fund transaction form based on a large model, which specifically includes steps S101 to S105.

[0030] Step S101, obtaining a fund transaction form to be processed, and converting the fund transaction form into form image data;

[0031] Step S102: classify the form image data using deep learning technology to remove invalid form image data and save valid form image data as first target image data;

[0032] Step S103, performing text recognition on the first target image data by using OCR recognition technology to obtain text information therein;

[0033] Step S104: input the text information into a large language model, and the large language model outputs corresponding structured data information;

[0034] Step S105: perform a validity check on the structured data information, and output a result of extracting elements of the fund transaction form based on the result of the validity check.

[0035] In this embodiment, when it is necessary to extract elements from the fund transaction form, since deep learning technology is more convenient and accurate in processing image data than processing PDF files, the PDF version of the fund transaction form is first converted into the corresponding form image data, and then the form image data is classified into two types: valid and invalid through deep learning technology. For invalid form image data, it can be directly eliminated. For valid form image data, text recognition is continued through OCR recognition technology, and the results of text recognition are output as structured data information through a large language model. Subsequently, the output results of the large language model are further verified for legitimacy to ensure the accuracy of the element extraction results, thereby determining the final element extraction results based on the results of the legitimacy verification. In practical applications, the element extraction results can be returned in Json format for subsequent review and editing.

[0036] Before adopting the OCR recognition technology, this embodiment uses deep learning technology to perform image classification processing on the fund transaction form in advance to eliminate invalid image data therein, which can reduce the workload in the OCR recognition process and improve the recognition effect. At the same time, after completing the OCR recognition, this embodiment uses a large language model to extract structured information from the OCR recognition results, which can solve the problem of disorderly OCR recognition results. In addition, this embodiment further performs a legality check on the extracted structured information to output a more accurate and reliable factor extraction result, thereby improving the factor extraction efficiency for the fund transaction form.

[0037] In one embodiment, the classifying process of the form image data by deep learning technology to remove invalid form image data and save valid form image data as first target image data includes:

[0038] Collect training images, and use the training images to perform classification training on the VGG16 model;

[0039] Perform transfer learning on the model parameters of the VGG16 model after classification training to build an image classification model;

[0040] The form image data is classified using the image classification model to obtain valid form image data or invalid form image data.

[0041] In the present embodiment, the VGG16 model refers to the Visual Geometry Group 16 model, which is a deep convolutional neural network model. Convolutional neural networks (CNN) are a type of feedforward neural networks (Feedforward Neural Networks) that include convolution calculations and have a deep structure, and are one of the representative algorithms of deep learning. In the present embodiment, the VGG16 model is used by the collected training images, and the parameters of the trained VGG16 model are transferred to learn, so as to construct an image classification model suitable for the application scenario of the present embodiment. Transfer learning is a machine learning method that allows the model to apply the knowledge learned in a task (source task) to another related or different task (target task). This method is particularly useful in cases such as data scarcity, limited computing resources or domain migration, because it can significantly reduce the demand for a large amount of labeled data and accelerate the learning process of the model on the new task. After that, the form image data can be effectively and invalidly classified by the constructed image classification model.

[0042] In actual application scenarios, the form image data can be divided into 7 categories through the image classification model, among which there are 4 categories of invalid form image data and 3 categories of valid form image data. For example, chat screenshots will be judged as invalid, etc. Figure 3As shown, the image classification model includes a three-layer structure of input layer, hidden layer and output layer, wherein the input layer inputs an RGB picture of 896*896 pixels, so the total number of pixels is 896*896*3; at the same time, the input layer is normalized, so the data obtained for each picture is 896*896*3 numbers in the range of 0-1. The hidden layer uses 6 groups of convolutional pooling layers, 2 groups of 64-dimensional output channel convolutional layers, 2 groups of 128-dimensional output channel convolutional layers, and 2 groups of 256-dimensional output channel convolutional layers. All pooling is the maximum pooling of 2*2 dimensions, the convolution kernel uses 3*3, and the step size is set to 1. The calculation process of the hidden layer is roughly as follows: the 3*3 numbers of the convolution kernel are multiplied with the 896*896 pixels of the input layer in sequence, and after multiplication, 894*894 numbers (896-3+1) can be obtained. Because it is 64-dimensional, 894*894*64 neurons will be output. The 2*2 maximum pooling layer is equivalent to selecting the largest number from the 4 numbers, so 447*447*64 numbers are obtained. This is the calculation operation of a group of neurons 4. The last of the convolution pooling layer is a 128-dimensional fully connected layer. For each of these 128 numbers = the number finally output by the previous convolution pooling layer multiplied by the parameter of the same dimension. The calculation method of the output layer is similar to that of the fully connected layer. In the end, an array determined by the number of categories will be obtained. The serial number corresponding to the largest value selected from the array is the category.

[0043] In one embodiment, the performing text recognition on the first target image data by using OCR recognition technology to obtain text information therein includes:

[0044] Performing text recognition on the first target image data by using OCR recognition technology to obtain candidate text information;

[0045] The candidate text information is detected for abnormal content, and the detected abnormal content is deleted to obtain the text information.

[0046] In this embodiment, the OCR recognition technology (Optical Character Recognition) refers to the process in which an electronic device (such as a scanner or a digital camera) examines characters printed on paper, determines their shapes by detecting dark and light patterns, and then translates the shapes into computer text using a character recognition method. That is, for printed characters, the text in a paper document is converted into a black and white dot matrix image file by an optical method, and the text in the image is converted into a text format by a recognition software for further editing and processing by a word processing software.

[0047] Here, after completing OCR recognition, this embodiment will perform anomaly detection on the recognized candidate text information, such as detecting whether it contains abnormal symbols such as "_", and delete these abnormal contents after detection, so that the subsequent large model reasoning can be more accurate.

[0048] In one embodiment, the step of inputting the text information into a large language model, and having the large language model output corresponding structured data information, includes:

[0049] Combining the text information with preset element extraction reference information to set a large model prompt;

[0050] The large model prompt is input into the Qwen2 large model, and the Qwen2 large model outputs the structured data information corresponding to the text information.

[0051] In this embodiment, the large model prompt is first set according to the text information obtained by OCR recognition, and then the large model prompt is input into the Qwen2 large model to obtain the corresponding structured data information. The Qwen2 large model is an open source large language model that can generate natural language text or understand the meaning of language text. The large language model can handle a variety of natural language tasks, such as text classification, question and answer, dialogue, etc., and is an important way to artificial intelligence. In addition, the large model prompt refers to a prompt used to stimulate or guide the AI ​​model to generate a specific type of text or answer. The large model prompt mainly consists of two parts, namely the text information corresponding to the original application form and the field information to be extracted, such as field content and field explanation.

[0052] In one embodiment, the step of performing a validity check on the structured data information and outputting a result of extracting elements of the fund transaction form based on the result of the validity check includes:

[0053] The structured data information that has passed the legality check is used as the first factor extraction result;

[0054] Acquire a second target image corresponding to the structured data information that fails the legality check, extract table information from the second target image, extract cell characters therein based on the table information, and use the cell characters as the second factor extraction result;

[0055] The first element extraction result and the second element extraction result are aggregated into the element extraction result of the fund transaction form and outputted.

[0056] In this embodiment, after obtaining the structured data information, a legality check is performed on it, for example, to check whether the amount or number in the structured data information is reasonable, so as to further improve the reliability of the extraction result. When the legality check passes, the corresponding structured data information can be directly returned as the element extraction result. When the legality check fails, the form image data of the corresponding structured data information is obtained, and the table information is extracted from it, for example, by identifying the four point coordinates of the table in the image data to realize table recognition, so as to complete the table information extraction, and then based on the extracted table information, the cell characters are extracted from it, for example, the vertex coordinates of each table can be obtained, and the internal content can be extracted through the vertex coordinates. After the cell characters are finally extracted, they are returned as the element extraction result. Of course, if part of the information in the structured data information passes the legality check and the other part of the information fails the legality check, the element extraction results obtained by the two can be summarized and returned together.

[0057] In one embodiment, after the step of performing a validity check on the structured data information and outputting a result of extracting elements of the fund transaction form based on the result of the validity check, the following steps are included:

[0058] Obtaining Chinese capital numbers in the structured data information, and converting the Chinese capital numbers into Arabic numerals;

[0059] Obtaining the fund code, fund name, large redemption mark, dividend method and business type in the structured data information;

[0060] The fund code, fund name, large-scale redemption mark, dividend method and business type are optimized respectively, and the optimized structured data is used as the final factor extraction result.

[0061] In this embodiment, combined with Figure 4 After completing the legality check, in order to further improve the overall extraction efficiency, the structured data information is optimized. For example, the Chinese capital numbers are converted into Arabic numbers for quick identification, and invalid characters in the fund code, fund name, large redemption mark, dividend method and business type are filtered out to complete information optimization.

[0062] Specifically, obtaining Chinese capital numbers in the structured data information and converting the Chinese capital numbers into Arabic numerals includes:

[0063] Identify and filter invalid characters in the Chinese uppercase numbers;

[0064] Determine whether the Chinese capital number contains a decimal part;

[0065] If it is determined that the Chinese uppercase number contains a decimal part, determine the decimal point position in the Chinese uppercase number, and obtain the Chinese character string before the decimal point position and the decimal part after the decimal point position respectively;

[0066] If it is determined that the Chinese uppercase number does not contain a decimal part, then determine whether the Chinese uppercase number contains an uppercase unit; or when the Chinese uppercase number contains a decimal part, then determine whether the Chinese character string contains an uppercase unit;

[0067] If it is determined that the Chinese uppercase numeral or Chinese character string does not contain an uppercase unit, converting the Chinese uppercase numeral or Chinese character string into Arabic numerals using a preset digital mapping table;

[0068] If it is determined that the Chinese uppercase numerals or Chinese character string contain uppercase units, the Chinese uppercase numerals or Chinese character string are converted into Arabic numerals one by one.

[0069] In this embodiment, Figure 4 As shown, in order to improve the conversion efficiency, the structured data information can be divided into 24 categories, namely 10 Arabic numerals, 20 Chinese uppercase characters and 4 special characters, and then the invalid characters in the Chinese uppercase numerals are filtered out according to the division structure, and the remaining characters are judged whether they contain a decimal part. If a decimal part is contained, the part after the decimal point is extracted, and the Chinese string before the decimal point is returned, and then the returned Chinese string is processed, that is, whether the Chinese string contains an uppercase unit. If no uppercase unit is contained, the Chinese string is directly converted into Arabic numerals through a pre-set Chinese-Arabic numeral mapping table; if an uppercase unit is contained, the Chinese string is matched one by one as Arabic numerals. Further, the matched Arabic numerals can be stored in a list s, and then the list s is traversed to obtain the character c therein, and it is judged whether the character c is a number or a unit. When the character c is a number, the corresponding value is obtained based on the character c, and when the character c is a unit, the specific unit is determined based on the character c. In addition, for Chinese uppercase numbers that do not contain decimals, it is possible to directly determine whether they contain uppercase units, and then map the Chinese uppercase numbers to Arabic numerals or match them one by one to Arabic numerals.

[0070] In an actual application scenario, when extracting elements from a specified fund transaction form through the large model-based fund transaction form element extraction method provided in this embodiment, the element extraction result in json format returned is as follows:

[0071] {

[0072] "title":"Transaction Application Form",

[0073] "apkind":"022",

[0074] "fundName":"×××Bond C",

[0075] "fundCode":"019623",

[0076] "targetFundName":"",

[0077] "targetFundCode":"",

[0078] "applicationAmount":"293460.0",

[0079] "originChineseFigures":"293,460 yuan, 0 cents, 0 cents",

[0080] "chineseToArabicFigures":"293460.0",

[0081] "figuresMatch":"1",

[0082] "largeFlag":"",

[0083] "melonMd":"0",

[0084] "oldAppNo":"",

[0085] "investorName":"xxx Asset Management Plan",

[0086] "custNo":"12345678",

[0087] "transferOutAccountId":"",

[0088] "transferInAccountId":""}

[0089] in:

[0090] apkind: business type, one of the 9 business types: "redemption, conversion, subscription, subscription, subscription (subscription), subscription / subscription, withdrawal, custody transfer, setting dividend method", 022 is the digital representation of subscription;

[0091] applicationAmount: lowercase Arabic numerals, representing the corresponding amount or share in fund trading business;

[0092] chineseToArabicFigures: Arabic numerals corresponding to Chinese capital letters;

[0093] custNo: transaction account number or trading account number;

[0094] figuresMatch: whether the numbers corresponding to Chinese uppercase characters and lowercase Arabic numerals are equal;

[0095] fundCode: fund code, for example, 019623 is the fund code;

[0096] fundName: fund name;

[0097] investorName: investor name;

[0098] largeFlag: How to handle the unsuccessful or unconfirmed portion of a large redemption. There are two ways to handle it: postponement and cancellation;

[0099] melonMd: dividend method;

[0100] oldAppNo: the application number to be revoked;

[0101] originChineseFigures: uppercase representation of Chinese;

[0102] targetFundCode: the transfer-in fund code (the transfer-in fund code for conversion business);

[0103] targetFundName: the name of the fund to be transferred (the name of the fund to be transferred in the conversion business);

[0104] transferInAccountId: transfer-in transaction account (this field is returned for custody transfer business, and an empty string is returned for non-custody transfer business type);

[0105] transferOutAccountId: transfer-out transaction account number (this field is returned for custody transfer business, and an empty string is returned for non-custody transfer business type).

[0106] Figure 2 A schematic block diagram of a fund transaction form element extraction device 200 based on a large model provided in an embodiment of the present invention, the device 200 includes:

[0107] An image conversion unit 201 is used to obtain a fund transaction form to be processed and convert the fund transaction form into form image data;

[0108] An image classification unit 202 is used to classify the form image data by deep learning technology to remove invalid form image data and save valid form image data as first target image data;

[0109] A text recognition unit 203 is used to perform text recognition on the first target image data by using OCR recognition technology to obtain text information therein;

[0110] An information extraction unit 204, configured to input the text information into a large language model, and have the large language model output corresponding structured data information;

[0111] The information verification unit 205 is used to perform a validity verification on the structured data information and output the element extraction result of the fund transaction form based on the result of the validity verification.

[0112] In one embodiment, the image classification unit 202 includes:

[0113] A classification training unit, used to collect training images and use the training images to perform classification training on the VGG16 model;

[0114] The transfer learning unit is used to transfer the model parameters of the VGG16 model after classification training to build an image classification model;

[0115] The model classification unit is used to classify the form image data using the image classification model to obtain valid form image data or invalid form image data.

[0116] In one embodiment, the text recognition unit 203 includes:

[0117] An OCR recognition unit, configured to perform text recognition on the first target image data by using an OCR recognition technology to obtain candidate text information;

[0118] The abnormality detection unit is used to detect abnormal content of the candidate text information and delete the detected abnormal content to obtain the text information.

[0119] In one embodiment, the information extraction unit 204 includes:

[0120] A prompt setting unit, used to set a large model prompt in combination with the text information and preset element extraction reference information;

[0121] The information output unit is used to input the large model prompt into the Qwen2 large model, and the Qwen2 large model outputs the structured data information corresponding to the text information.

[0122] In one embodiment, the information verification unit 205 includes:

[0123] A first setting unit, used to use the structured data information that has passed the legality check as the first factor extraction result;

[0124] A second setting unit is used to obtain a second target image corresponding to the structured data information that has not passed the legality check, extract table information from the second target image, extract cell characters therein based on the table information, and use the cell characters as the second factor extraction result;

[0125] A result aggregation unit is used to aggregate the first element extraction result and the second element extraction result into the element extraction result of the fund transaction form and output it.

[0126] In one embodiment, the fund transaction form element extraction device 200 based on the large model includes:

[0127] A number conversion unit, used for obtaining Chinese capital numbers in the structured data information and converting the Chinese capital numbers into Arabic numbers;

[0128] A parameter acquisition unit, used to acquire the fund code, fund name, large redemption mark, dividend method and business type in the structured data information;

[0129] The parameter optimization unit is used to optimize the fund code, fund name, large-scale redemption mark, dividend method and business type respectively, and use the optimized structured data as the final factor extraction result.

[0130] In one embodiment, the digital conversion unit includes:

[0131] An invalid identification unit, used for identifying and filtering invalid characters in the Chinese uppercase numbers;

[0132] A decimal determination unit, used to determine whether the Chinese capital number contains a decimal part;

[0133] A first determination unit is configured to determine the position of a decimal point in the Chinese uppercase number if it is determined that the Chinese uppercase number contains a decimal part, and to obtain the Chinese character string before the decimal point and the decimal part after the decimal point respectively;

[0134] A second determination unit is configured to determine whether the Chinese uppercase number contains an uppercase unit if it is determined that the Chinese uppercase number does not contain a decimal part; or to determine whether the Chinese character string contains an uppercase unit when the Chinese uppercase number contains a decimal part;

[0135] A first conversion unit, configured to convert the Chinese uppercase numeral or Chinese character string into Arabic numerals by using a preset digital mapping table if it is determined that the Chinese uppercase numeral or Chinese character string does not contain an uppercase unit;

[0136] The second conversion unit is configured to convert the Chinese uppercase numerals or Chinese character strings into Arabic numerals one by one if it is determined that the Chinese uppercase numerals or Chinese character strings contain uppercase units.

[0137] Since the embodiments of the apparatus part correspond to the embodiments of the method part, please refer to the description of the embodiments of the method part for the embodiments of the apparatus part, which will not be repeated here.

[0138] The embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed, the steps provided in the above embodiment can be implemented. The storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0139] The embodiment of the present invention also provides a computer device, which may include a memory and a processor, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps provided in the above embodiment may be implemented. Of course, the computer device may also include various network interfaces, power supplies and other components.

[0140] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of this application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of this application.

[0141] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A method for extracting elements from fund transaction forms based on a large model, characterized in that: include: Acquire a fund transaction form to be processed, and convert the fund transaction form into form image data; Classify the form image data by deep learning technology to remove invalid form image data and save valid form image data as first target image data; Performing text recognition on the first target image data by using OCR recognition technology to obtain text information therein; Inputting the text information into a large language model, and having the large language model output corresponding structured data information; The structured data information is subjected to a legality check, and based on the result of the legality check, an element extraction result of the fund transaction form is output.

2. The method for extracting elements from fund transaction forms based on a large model according to claim 1, characterized in that: The classifying and processing the form image data by deep learning technology to remove invalid form image data and save valid form image data as first target image data includes: Collect training images, and use the training images to perform classification training on the VGG16 model; Perform transfer learning on the model parameters of the VGG16 model after classification training to build an image classification model; The form image data is classified using the image classification model to obtain valid form image data or invalid form image data.

3. The method for extracting elements from fund transaction forms based on a large model according to claim 1, characterized in that: The performing text recognition on the first target image data by using the OCR recognition technology to obtain text information therein includes: Performing text recognition on the first target image data by using OCR recognition technology to obtain candidate text information; The candidate text information is detected for abnormal content, and the detected abnormal content is deleted to obtain the text information.

4. The method for extracting elements from fund transaction forms based on a large model according to claim 1, characterized in that: The step of inputting the text information into a large language model, and having the large language model output corresponding structured data information, includes: Combining the text information with preset element extraction reference information to set a large model prompt; The large model prompt is input into the Qwen2 large model, and the Qwen2 large model outputs the structured data information corresponding to the text information.

5. The method for extracting elements from fund transaction forms based on a large model according to claim 1, characterized in that: The step of performing a validity check on the structured data information and outputting a result of extracting elements of the fund transaction form based on the result of the validity check includes: The structured data information that has passed the legality check is used as the first factor extraction result; Acquire a second target image corresponding to the structured data information that fails the legality check, extract table information from the second target image, extract cell characters therein based on the table information, and use the cell characters as the second factor extraction result; The first element extraction result and the second element extraction result are aggregated into the element extraction result of the fund transaction form and outputted.

6. The method for extracting elements from fund transaction forms based on a large model according to claim 1, characterized in that: After the step of performing a legality check on the structured data information and outputting the element extraction result of the fund transaction form based on the result of the legality check, the method further includes: Obtaining Chinese capital numbers in the structured data information, and converting the Chinese capital numbers into Arabic numerals; Obtaining the fund code, fund name, large redemption mark, dividend method and business type in the structured data information; The fund code, fund name, large-scale redemption mark, dividend method and business type are optimized respectively, and the optimized structured data is used as the final factor extraction result.

7. The method for extracting elements from fund transaction forms based on a large model according to claim 6, characterized in that: The obtaining of Chinese capital numbers in the structured data information and converting the Chinese capital numbers into Arabic numbers includes: Identify and filter invalid characters in the Chinese uppercase numbers; Determine whether the Chinese capital number contains a decimal part; If it is determined that the Chinese uppercase number contains a decimal part, determine the decimal point position in the Chinese uppercase number, and obtain the Chinese character string before the decimal point position and the decimal part after the decimal point position respectively; If it is determined that the Chinese uppercase number does not contain a decimal part, then determine whether the Chinese uppercase number contains an uppercase unit; or when the Chinese uppercase number contains a decimal part, then determine whether the Chinese character string contains an uppercase unit; If it is determined that the Chinese uppercase numeral or Chinese character string does not contain an uppercase unit, converting the Chinese uppercase numeral or Chinese character string into Arabic numerals using a preset digital mapping table; If it is determined that the Chinese uppercase numerals or Chinese character string contain uppercase units, the Chinese uppercase numerals or Chinese character string are converted into Arabic numerals one by one.

8. A fund transaction form element extraction device based on a large model, characterized in that: include: An image conversion unit, used for acquiring a fund transaction form to be processed and converting the fund transaction form into form image data; An image classification unit, used for classifying the form image data by deep learning technology to remove invalid form image data and save valid form image data as first target image data; A text recognition unit, used to perform text recognition on the first target image data by using OCR recognition technology to obtain text information therein; An information extraction unit, used for inputting the text information into a large language model, and having the large language model output corresponding structured data information; The information verification unit is used to perform a legality verification on the structured data information and output an element extraction result of the fund transaction form based on the result of the legality verification.

9. A computer device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for extracting fund transaction form elements based on a large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for extracting elements from a fund transaction form based on a large model as described in any one of claims 1 to 7 is implemented.

Images