Large model and GBDT-based reimbursement document field prediction method and device, and medium

Abstract

The invention discloses a reimbursement document field prediction method and device based on a large model and GBDT and a medium, and relates to the technical field of natural language processing, and the method comprises the steps: obtaining reimbursement document data, carrying out the field prediction analysis of the reimbursement document data, and determining the unique identification data; performing related field extraction of prediction fields on the reimbursement bill data to obtain feature fields; performing training set division on the reimbursement bill data to obtain a target value; based on the feature field and the target value, normalizing through CatBoost coding to obtain normalized training data; performing classification task training on the gradient boosting decision tree model according to the unique identification data and the normalized training data, and determining a classification model; and through the classification model, performing filling prediction on the to-be-predicted field to obtain auxiliary filling data. Through the method, the technical problems of low parameter transmission efficiency and insufficient performance and generalization ability of a reimbursement document field prediction method are solved.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular, to a method, device, and medium for predicting reimbursement form fields based on large models and GBDT. Background Art

[0002] In recent years, the financial software industry has undergone significant technological changes and is moving towards a more intelligent, automated, and integrated direction. With the popularization of cloud computing technology, more and more enterprises choose cloud-based solutions to deploy financial management systems, which not only reduces the enterprise's IT costs but also improves the speed and flexibility of data processing. At the same time, the application of artificial intelligence and machine learning is changing the traditional financial workflow, greatly enhancing the efficiency and accuracy of financial management through automated data analysis, risk assessment, and prediction models.

[0003] Existing reimbursement form field prediction technologies have defects in parameter transfer setting, feature field selection, feature field encoding, and the algorithm part. For parameter transfer setting, the parameter transfer efficiency in each stage of the existing technology is low, and the interaction between tasks is blocked. For feature field selection, the existing technology relies on manual selection and cannot adapt to the changing business needs of enterprise reimbursement. For feature field encoding, the unique encoding methods in the existing technology are not applicable to encoding high-cardinality categorical features. Label encoding maps categories into consecutive integers, which may violate the assumption of feature disorder and reduce the possibility of prediction accuracy. For the algorithm of predicting reimbursement form fields, the existing technology is mainly based on algorithms such as single decision trees, linear models, and random forests. The splitting strategy of a single decision tree overly relies on local optimality rather than global optimality and is prone to overfitting. The linear model assumes a linear relationship between features and target variables, so it cannot capture the non-linear relationships or high-order interactions between reimbursement form data. In addition, the random forest model reduces the overfitting risk through random sampling and feature subset selection, but if the tree depth is too large or the number of trees is too large, it may still overfit the training data due to excessive model complexity. Summary of the Invention

[0004] The embodiments of this application provide a method, device, and medium for predicting reimbursement form fields based on large models and GBDT, which solve the technical problems of low parameter transfer efficiency, insufficient performance, and generalization ability of the reimbursement form field prediction method.

[0005] In a first aspect, an embodiment of the present application provides a method for predicting fields of a reimbursement form based on a large model and GBDT, characterized in that the method includes: obtaining reimbursement form data, performing field prediction analysis on the reimbursement form data to determine unique identification data; extracting relevant fields of the predicted fields from the reimbursement form data to obtain feature fields; dividing the reimbursement form data into a training set to obtain target values; based on the feature fields and target values, through CatBoost encoding normalization, obtaining normalized training data; according to the unique identification data and the normalized training data, training a gradient boosting decision tree model for a classification task to determine a classification model; through the classification model, performing fill-in prediction on the fields to be predicted to obtain auxiliary fill-in form data.

[0006] In an implementation manner of the present application, performing field prediction analysis on the reimbursement form data to determine unique identification data specifically includes: querying historical data of the reimbursement form data to obtain a two-dimensional format file of identification names; configuring unique identifications for the identification names, and setting the two-dimensional format file with the configured unique identifications as the unique identification data.

[0007] In an implementation manner of the present application, extracting relevant fields of the predicted fields from the reimbursement form data to obtain feature fields specifically includes: based on the unique identification data, determining the storage path of the reimbursement form data, and according to the storage path and the reimbursement form data, through large model feature extraction, obtaining feature prompt words; wherein, the feature prompt words include: a function description part and an instruction part; according to the feature prompt words, through executing the prompt word instructions, determining the feature fields.

[0008] In an implementation manner of the present application, dividing the reimbursement form data into a training set to obtain target values specifically includes: decomposing the fields of the reimbursement form data by time type to obtain a reimbursement form to be supplemented; filling in missing values with field distinctions for the reimbursement form to be supplemented to obtain a supplemented reimbursement form; based on the supplemented reimbursement form, converting the values of the predicted fields into target values.

[0009] In an implementation manner of the present application, based on the feature fields and target values, through CatBoost encoding normalization, obtaining normalized training data specifically includes: performing CatBoost encoding on the feature fields to obtain the encoded numerical values of the feature fields; wherein, the calculation formula for CatBoost encoding is:

[0010] Wherein, is the encoded numerical value, is the sum of the values of the specified feature on the target values, is the sum of the target values divided by the number of all the values of the feature fields, The number of times the value of a specified feature appears in the entire feature value list; perform normalization processing on the encoded numerical values to obtain normalized training data.

[0011] In one implementation manner of the present application, according to the unique identification data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine the classification model, specifically including: setting the feature field values in the normalized training data as features and setting the target value as the supervision signal; based on the features and the supervision signal, perform classification task training of the gradient boosting decision tree on the normalized training data until the model converges to determine the classification model.

[0012] In one implementation manner of the present application, through the classification model, perform fill-in form prediction on the field to be predicted to obtain auxiliary fill-in form data, specifically including: obtaining the feature field values of the current reimbursement form and inputting the feature field values into the classification model to obtain a prediction value group; performing prediction probability screening on the prediction value group to determine the output prediction value; performing fill-in form intention judgment on the output prediction value to obtain auxiliary fill-in form data.

[0013] In one implementation manner of the present application, after performing fill-in form prediction on the field to be predicted through the classification model to obtain auxiliary fill-in form data, the method further includes: manually filling in the form when the auxiliary fill-in form data does not have preset setting parameters.

[0014] In a second aspect, an embodiment of the present application further provides a reimbursement form field prediction device based on a large model and GBDT, characterized in that the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable 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: obtain reimbursement form data, perform field prediction analysis on the reimbursement form data to determine unique identification data; extract relevant fields of the prediction field from the reimbursement form data to obtain feature fields; divide the reimbursement form data into a training set to obtain a target value; based on the feature fields and the target value, obtain normalized training data through CatBoost encoding normalization; according to the unique identification data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine the classification model; through the classification model, perform fill-in form prediction on the field to be predicted to obtain auxiliary fill-in form data.

[0015] In a third aspect, an embodiment of the present application further provides a non-volatile computer storage medium for predicting reimbursement form fields based on a large model and GBDT, storing computer-executable instructions, characterized in that the computer-executable instructions are set as follows: obtaining reimbursement form data, performing field prediction analysis on the reimbursement form data to determine unique identification data; extracting relevant fields of the predicted fields from the reimbursement form data to obtain feature fields; dividing the reimbursement form data into a training set to obtain target values; based on the feature fields and target values, through CatBoost coding normalization, obtaining normalized training data; according to the unique identification data and the normalized training data, training a gradient boosting decision tree model for a classification task to determine a classification model; through the classification model, performing fill-in form prediction on the fields to be predicted to obtain auxiliary fill-in form data.

[0016] An embodiment of the present application provides a method, device, and medium for predicting reimbursement form fields based on a large model and GBDT. Through CatBoost coding, classification model construction, etc., it solves the technical problems of low parameter transfer efficiency, insufficient performance and generalization ability in the reimbursement form field prediction method, realizes real-time communication between the reimbursement form field prediction in the fill-in form task and model training, improves the parameter transfer efficiency, reduces the overfitting probability of the prediction model, and enhances the accuracy and generalization ability of the prediction model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings: Figure 1 It is a flowchart of a method for predicting reimbursement form fields based on a large model and GBDT provided by an embodiment of the present application; Figure 2 It is a data organization format diagram of a two-dimensional format file provided by an embodiment of the present application; Figure 3 It is a flowchart of assisting a user to fill in a form in a method for predicting reimbursement form fields provided by an embodiment of the present application; Figure 4 It is a schematic internal structure diagram of a device for predicting reimbursement form fields based on a large model and GBDT provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments of this application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

[0019] The embodiments of this application provide a method, device, and medium for predicting reimbursement form fields based on a large model and GBDT. Through CatBoost encoding, classification model construction, etc., it solves the technical problems of low parameter transfer efficiency, insufficient performance and generalization ability in the reimbursement form field prediction method, realizes real-time communication between the form filling task and model training in the reimbursement form field prediction, improves the parameter transfer efficiency, reduces the overfitting probability of the prediction model, and enhances the accuracy and generalization ability of the prediction model.

[0020] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the drawings.

[0021] Figure 1 It is a flowchart of a method for parsing multi-version communication protocols provided by an embodiment of this application. As Figure 1 shown, a method for parsing multi-version communication protocols provided by an embodiment of this application specifically includes the following steps: Step 101: Obtain reimbursement form data, perform field prediction analysis on the reimbursement form data, and determine unique identification data.

[0022] Exemplarily, create a field prediction setting table for storing relevant settings. At the same time, for a certain type of reimbursement form, select the names of the fields on the reimbursement form that need to be predicted, and query the historical data of this reimbursement form from the database.

[0023] Specifically, performing field prediction analysis on the reimbursement form data to determine unique identification data specifically includes: querying historical data of the reimbursement form data to obtain a two-dimensional format file of the identification name; performing unique identification configuration on the identification name, and setting the two-dimensional format file with the unique identification configured as the unique identification data.

[0024] Figure 2 It is a data organization format diagram of a two-dimensional format file provided by an embodiment of this application.

[0025] In one embodiment, first, create a field prediction setting table. The database table includes the following fields: unique identification, reimbursement form data storage path, feature field name, CatBoost encoder storage file path, model status, classification model storage file path.

[0026] The field prediction settings table is used to store relevant settings. The unique identifier is the id of a setting, and the corresponding setting parameters can be queried in the field prediction settings table through the unique identifier; the reimbursement form data save path stores the save path of the reimbursement form, facilitating subsequent reading of the reimbursement form data according to the path; the feature field name stores the feature field names extracted from the large model, including multiple fields, and is stored in the form of a list.

[0027] The CatBoost encoder save file path stores the file path of the trained CatBoost encoder, facilitating subsequent reading and use; the model status stores the status of the classification model, which is an enumeration field, including two enumeration values: "training" and "training completed"; the classification model save file path stores the file path of the trained gradient boosting decision tree classification model, facilitating subsequent reading and use.

[0028] At the same time, the financial software may prefabricate various types of reimbursement form data for users to use. Select one type of reimbursement form and a field to be predicted on this type of reimbursement form. The selected reimbursement form type name and the selected field name are organized in the form of "{reimbursement form type name}_{field name}" for the unique identifier of one reimbursement type and one field.

[0029] Then, query the historical data of this type of reimbursement form from the database, and save the queried historical data as a two-dimensional format file with the name of "{unique identifier}.file".

[0030] The data organization format of the two-dimensional format file has two dimensions: rows and columns. The first row of the file is the name of each field on the reimbursement form, and each subsequent row corresponds to the data of one reimbursement form. The first column of the file is the index of the reimbursement form data, which can be used to identify the data of one reimbursement form, and each subsequent column corresponds to the value of each field. Using this data organization format facilitates data reading. For example, reading the data of the first reimbursement form can be recorded as data[1][~]; reading the data of field 1 can be recorded as data[~][field 1 name]; reading the data of field 1 on the first reimbursement form can be recorded as data[1][1].

[0031] Finally, form a piece of data with the unique identifier and the save path of the reimbursement form data, and insert it into the field prediction settings table.

[0032] Step 102: Extract relevant fields for the prediction field from the reimbursement form data to obtain feature fields.

[0033] Exemplarily, the queried expense report data is organized and input into a large model, and prompt engineering is used to enable the large model to extract the fields related to the prediction field as feature fields. This application explores the relationship between the fields on the document by combining large model prompt engineering with historical expense report data, extracts the fields related to the prediction field as feature fields, and does not require manual configuration by implementers, improving openness and thus adapting to the complex business change requirements in enterprise expense reimbursement.

[0034] Specifically, relevant fields of the prediction field are extracted from the expense report data to obtain feature fields, which specifically include: based on the unique identifier data, determining the storage path of the expense report data, and through large model feature extraction according to the storage path and the expense report data, obtaining a feature prompt; wherein, the feature prompt includes: a function description part and an instruction part; according to the feature prompt, by executing the prompt instruction, the feature field is determined.

[0035] In one embodiment, the accuracy of field prediction depends on the selection of feature fields. In traditional technical solutions, the selection of feature fields relies on the judgment of implementers, which requires implementers to be particularly familiar with relevant operations, and be able to accurately identify which fields in the expense report are related to the prediction field. However, due to the diverse changes in enterprise expense reimbursement requirements and the unstable business requirements, implementers need to have a high level of domain knowledge and business level for configuration.

[0036] First, query the storage path of the expense report data from the field prediction setting table according to the unique identifier, and read the expense report data according to this path, denoted as data. Then, configure the prompt for the large model to extract feature fields in combination with the expense report data.

[0037] A better prompt includes a function description part and an instruction part. Among them, the description part at least includes: a description of the function required to be provided by the large model, and input variables; while the instruction part at least includes specific instructions for the function provided by the large model, as well as answer format constraint statements and answer content constraint statements, and may also include input variables. A complete prompt may include the following content: You are an assistant for exploring the relationship between expense report fields, capable of learning the relationships between various fields of the expense report based on the expense report data I provide to you, where the expense report data is {data}. Please, based on the above information, after learning the relationships between the expense report fields, extract the fields related to the field: {name of the prediction field}. Please return the extracted fields to me in JSON format and organize these fields in a list. Do not add other information that needs to be supplemented.

[0038] Among them, the content "You are an exploration assistant for the relationships of fields in an expense report form. The data in the expense report form is {data}" belongs to the description part, which is used to describe the requirements for extracting feature fields this time. And {data} among them belongs to the input variable, representing the expense report form data obtained by input query. The remaining statements belong to the format statements; The content "Please, based on the above information, after learning the relationships of the fields in the expense report form, extract the fields related to the field: {name of the predicted field}" belongs to the instruction part, which is used to send a command to the large model to extract feature fields. And {name of the predicted field} among them belongs to the input variable, representing the name of the predicted field; The content "Please return the extracted fields to me in JSON format and organize these fields in the form of a list." belongs to the format constraint statement, which constrains the format returned by the large model; and the content "Do not add other information that needs to be supplemented." belongs to the answer constraint statement, which constrains the content of the answer from the large model.

[0039] After using the large model to return the list of names of the extracted feature fields, form a piece of data with the unique identifier and the list of names of the feature fields, and save it into the field prediction setting table.

[0040] Step 103: Divide the expense report form data into a training set to obtain the target value.

[0041] Exemplarily, divide the expense report form data into a training set and a test set according to a certain ratio. The training set data is used for the field prediction model to learn the rules in the data. The test set data is not visible during the training process of the field prediction model. When the model training is completed, use the test set data to verify the ability of the model. The commonly used ratios for dividing the training set and the test set are 8:2 and 7:3.

[0042] Specifically, dividing the expense report form data into a training set to obtain the target value specifically includes: decomposing the fields of the time type in the expense report form data to obtain the expense report form to be supplemented; filling in the missing values with field differentiation for the expense report form to be supplemented to obtain the supplemented expense report form; based on the supplemented expense report form, converting the value of the predicted field into the target value.

[0043] In one embodiment, first, process the fields of the time type. For the fields on the expense report form with the data type of time type, decompose their values into 6 numerical type values of year, month, day, hour, minute, and second. For example, "2025-04-16 08:30:30" can be decomposed into "2025", "04", "16", "08", "30", and "30".

[0044] Then, handle the missing values. For the missing values in the expense report data, if it is a numerical type field, fill it with the average value; if it is a non-numerical type, fill it with the mode. For example, if there are missing values in field 1 of the expense report data, if field 1 is a numerical type, the calculation of the filling value is explained by the following formula.

[0045] (1) Among them, is the sum of the values of field 1 in all expense report data, is the number of expense reports where the value of field 1 is not empty, is the filling value.

[0046] If the field type of field 1 is not a numerical type, then use the mode to fill the missing part, and the calculation formula of the filling value is explained by the following formula.

[0047] (2) Among them, is the value that appears most frequently for field 1 in the expense report, is the filling value.

[0048] The values of the prediction field are generally strings. To facilitate subsequent feature encoding and model training, they need to be converted into target values. First, extract the unique values of the prediction field values and convert them into a list, and convert the values of the prediction field into the index values of their corresponding unique values in the list. For example, if the prediction field is the department of the expense report, and there are six expense report data, and the values of their expense departments are: ["Finance Department", "Legal Department", "Discipline Inspection Department", "Legal Department", "Finance Department", "Discipline Inspection Department"], then the list of their unique values is: ["Finance Department", "Legal Department", "Discipline Inspection Department"], and the index values of each unique value in the list are: 0, 1, 2. Then the result of converting the values of the prediction field into target values is: [0, 1, 2, 1, 0, 2].

[0049] Since machine learning classification algorithms require data to be in numerical format, we need to encode and convert the values of the feature fields in the dataset into numerical format.

[0050] Step 104: Based on the feature fields and target values, through CatBoost encoding and normalization, obtain the normalized training data.

[0051] Exemplarily, traditional solutions generally use one-hot encoding to convert non-numerical features into numerical features. The dimension of the converted numerical features is equal to the number of unique values of the feature field before conversion. If the number of unique values of the feature field is large, the converted feature space will increase rapidly and the data will be very sparse, which will cause the converted features to occupy a large amount of memory resources and cannot capture the relationships or similarities between categories. This application uses a CatBoost encoder to encode the values of the feature field, and the converted dimension will not change, so the above problems will not occur.

[0052] Specifically, based on the feature field and the target value, through CatBoost encoding normalization, normalized training data is obtained, which specifically includes: performing CatBoost encoding on the feature field to obtain the encoded value of the feature field; wherein, the calculation formula of CatBoost encoding is:

[0053] Wherein, is the encoded value, is the sum of the values of the specified feature on the target value, is the sum of the target values divided by the number of all feature field values, is the number of times the value of the specified feature appears in the entire feature value list; performing normalization processing on the encoded value to obtain the normalized training data.

[0054] In one embodiment, the CatBoost encoder is trained to encode the feature field, and the feature encoded value is normalized. Based on the feature encoded value and the target value, a field prediction model is trained using gradient boosting decision trees.

[0055] First, for the value of a certain feature field, perform CatBoost encoding on the feature field to obtain the encoded value of the feature field; wherein, the CatBoost encoding is explained by the following formula.

[0056] (3) Wherein, is the encoded value, is the sum of the values of the specified feature on the target value, is the sum of the target values divided by the number of all feature field values, is the number of times the value of the specified feature appears in the entire feature value list.

[0057] Further, taking the feature fields of expense items as an example, there are six reimbursement data records, and the expense items are respectively: ["mobile phone fee", "business travel expense", "conference fee", "business travel expense", "mobile phone fee", "conference fee"], and the target values of these reimbursement data records are respectively [1, 2, 1, 2, 1, 2]. For the specified feature value: "mobile phone fee", the sum of the target values is: 1 + 1 = 2, so TargetCount("mobile phone fee") = 2; for the entire feature list, the sum of the target values divided by the number of values of all feature fields is: (1 + 2 + 1 + 2 + 1 + 2) / 6 = 1.5, so prior = 1.5; for the specified feature value: "mobile phone fee", the number of occurrences in the entire feature value list is 2 times, so FeatureCount("mobile phone fee") = 2. Based on the above calculation results, the value of the specified feature value: "mobile phone fee" after being encoded by the CatBoost encoder is: (2 + 1.5) / (2 + 1) = 1.167.

[0058] By training a CatBoost encoder in the above manner by combining the values of all feature fields with the target values, the values of the feature fields can be converted into numerical features for subsequent model training. Save the CatBoost encoder as a file, and form a piece of data with the unique identifier and the file path, and save it to the field prediction setting data table.

[0059] After encoding the values of the feature fields, it is necessary to perform normalization processing on the encoded values. Specifically, this application uses Z-Score standardization to normalize the data of each field of the reimbursement form, and the normalization processing is explained by the following formula.

[0060] (4) Among them, represents the average value of the data in field 1, represents the standard deviation of the data in field 1, is the normalized training data.

[0061] Step 105: According to the unique identifier data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine the classification model.

[0062] Exemplarily, this application performs classification task training on the gradient boosting decision tree model according to the unique identifier data and the normalized training data to determine the classification model, realizes real-time communication between the fill-in task and the model training of the reimbursement form field prediction, improves the parameter transfer efficiency, reduces the overfitting probability of the prediction model, and enhances the accuracy and generalization ability of the prediction model prediction.

[0063] Specifically, according to the unique identification data and the normalized training data, the gradient boosting decision tree model is trained for the classification task to determine the classification model, which specifically includes: setting the numerical values of the feature fields in the normalized training data as features and setting the target value as the supervision signal; based on the features and the supervision signal, performing the classification task training of the gradient boosting decision tree on the normalized training data until the model converges to determine the classification model.

[0064] In one embodiment, the classification model is trained and tested. The gradient boosting decision tree model is used as the learning model for the classification task, the values of the feature fields are used as features, and the target value is used as the supervision signal. The training set is trained in multiple rounds, and after each round of training, the test set is used for testing. When the model starts training, the model status in the field prediction setting data table is modified through the unique identification and set to "training". When the model finishes training, the model of the round with the best test effect is saved as the finally used field prediction model. At the same time, through the unique identification, the model status in the field prediction setting data table is modified to "training completed", and the classification model save file path is modified to the actual save path of the model.

[0065] Step 106: Through the classification model, perform fill-in form prediction on the field to be predicted to obtain auxiliary fill-in form data.

[0066] Specifically, through the classification model, perform fill-in form prediction on the field to be predicted to obtain auxiliary fill-in form data, which specifically includes: obtaining the numerical values of the feature fields of the current reimbursement form and inputting the numerical values of the feature fields into the classification model to obtain a prediction value group; performing prediction probability screening on the prediction value group to determine the output prediction value; performing fill-in form intention judgment on the output prediction value to obtain auxiliary fill-in form data.

[0067] Further, after performing fill-in form prediction on the field to be predicted through the classification model to obtain auxiliary fill-in form data, the method further includes: manually filling in the form when the preset setting parameters do not exist in the auxiliary fill-in form data.

[0068] Figure 3 This is a flowchart of an expense report field prediction method provided by an embodiment of the present application to assist users in filling in forms.

[0069] In one embodiment, when a user fills in a certain form field on a certain expense report, use "{expense report type name}_{field name}" as the unique identification to query in the field prediction setting table whether there is a field prediction model setting corresponding to this expense report type and this field. If there is no setting for this type and field, field prediction will not be performed, and the user manually fills in the field value; If the settings corresponding to this type and field can be queried, identify the name of the corresponding feature field from the settings, then collect the values of the feature fields on the document, and perform data cleaning on the values of the feature fields, including processing time-type fields and handling missing values.

[0070] Then identify the file path for saving the CatBoost encoder from the settings, use the CatBoost encoder to encode the feature fields, and finally identify the classification model saving path from the settings. Input the values of the encoded feature fields into the classification model, and the model outputs the three predicted values with the highest probabilities for the user to select.

[0071] The user determines whether there is a value that conforms to their form-filling intention among these three predicted values. If there is, select and fill it in; if not, the user manually fills in the field value.

[0072] The above is the method embodiment proposed in this application. Based on the same inventive concept, the embodiments of this application also provide a reimbursement form field prediction device based on a large model and GBDT, and its structure is as Figure 4 shown.

[0073] Figure 4 It is a schematic internal structure diagram of a reimbursement form field prediction device based on a large model and GBDT provided by the embodiments of this application. As Figure 4 shown, the device includes: At least one processor 401; And a memory 402 communicatively connected to the at least one processor; Wherein, the memory 402 stores instructions executable by the at least one processor. The instructions are executed by the at least one processor 401 so that the at least one processor 401 can: Obtain reimbursement form data, perform field prediction analysis on the reimbursement form data to determine unique identification data; extract relevant fields of the predicted fields from the reimbursement form data to obtain feature fields; divide the reimbursement form data into training sets to obtain target values; based on the feature fields and target values, through CatBoost encoding normalization, obtain normalized training data; according to the unique identification data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine the classification model; through the classification model, perform form-filling prediction on the fields to be predicted to obtain auxiliary form-filling data.

[0074] Some embodiments of this application provide a Figure 1 non-volatile computer storage medium for reimbursement form field prediction based on a large model and GBDT, storing computer-executable instructions, and the computer-executable instructions are set as: Obtain reimbursement form data, perform field prediction analysis on the reimbursement form data to determine unique identification data; extract relevant fields of the predicted fields from the reimbursement form data to obtain feature fields; divide the reimbursement form data into training sets to obtain target values; based on the feature fields and target values, through CatBoost encoding normalization, obtain normalized training data; according to the unique identification data and the normalized training data, train the gradient boosting decision tree model for classification tasks to determine the classification model; through the classification model, perform fill-in form prediction on the fields to be predicted to obtain auxiliary fill-in form data.

[0075] Each embodiment in this application is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the Internet of Things device and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiments.

[0076] The systems and media provided by the embodiments of this application correspond one-to-one with the methods. Therefore, the systems and media also have beneficial technical effects similar to those of their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be elaborated here.

[0077] Those skilled in the art should understand that the embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0078] This application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of this application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, so that the instructions executed by the processors of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0079] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to operate in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction device that implements the functions specified in one or more of the acts Figure 1 or acts and / or boxes Figure 1 specified in one or more of the boxes.

[0080] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one or more of the acts Figure 1 or acts and / or boxes Figure 1 specified in one or more of the boxes.

[0081] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0082] The memory may include non-permanent memory in the computer-readable medium, random access memory (RAM) and / or non-volatile memory such as read only memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

[0083] Computer-readable media includes both permanent and non-permanent, removable and non-removable media implemented by any method or technology for storage of information such as computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technologies, compact disc read only memory (CD-ROM), digital versatile discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0084] It should also be noted that the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising the element.

[0085] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A reimbursement form field prediction method based on large models and GBDT, characterized in that, The method includes: Obtain reimbursement form data, perform field prediction analysis on the reimbursement form data, and determine unique identification data; Extract relevant fields of the predicted fields from the reimbursement form data to obtain feature fields; Divide the reimbursement form data into training sets to obtain target values; Based on the feature fields and target values, through CatBoost encoding normalization, obtain normalized training data; According to the unique identification data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine the classification model; Through the classification model, perform fill-in form prediction on the fields to be predicted to obtain auxiliary fill-in form data.

2. The reimbursement form field prediction method based on a large model and GBDT according to claim 1, wherein Performing field prediction analysis on the reimbursement form data to determine unique identification data specifically includes: Query historical data of the reimbursement form data to obtain a two-dimensional format file of identification names; Configure unique identifiers for the identification names, and set the two-dimensional format file with configured unique identifiers as the unique identification data.

3. The reimbursement form field prediction method based on a large model and GBDT according to claim 1, wherein, Extracting relevant fields of the predicted fields from the reimbursement form data to obtain feature fields specifically includes: Based on the unique identification data, determine the storage path of the reimbursement form data, and according to the storage path and the reimbursement form data, through large model feature extraction, obtain feature prompt words; wherein, the feature prompt words include: a function expression part and an instruction part; According to the feature prompt words, through the execution of the prompt word instructions, determine the feature fields.

4. The reimbursement form field prediction method based on a large model and GBDT according to claim 1, wherein, Dividing the reimbursement form data into training sets to obtain target values specifically includes: Decompose the fields of the reimbursement form data by time type to obtain a reimbursement form to be supplemented; Fill in the missing values of the reimbursement form to be supplemented by field distinction to obtain a supplemented reimbursement form; Based on the supplemented reimbursement form, convert the values of the predicted fields into the target values.

5. A reimbursement form field prediction method based on a large model and GBDT according to claim 1, characterized in that Based on the feature fields and target values, through CatBoost encoding normalization, obtain normalized training data specifically includes: Perform CatBoost encoding on the feature fields to obtain the encoded numerical values of the feature fields; wherein, the calculation formula of the CatBoost encoding is: Among them, is the encoded value, is the sum of the values of the specified feature on the target value, is the sum of the target values divided by the number of values of all feature fields, is the number of times the value of the specified feature appears in the entire list of feature values; Perform normalization processing on the encoded numerical values to obtain normalized training data.

6. The reimbursement form field prediction method based on a large model and GBDT according to claim 1, wherein According to the unique identification data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine the classification model specifically includes: Set the feature field numerical values in the normalized training data as features, and set the target values as supervision signals; Based on the features and the supervision signals, perform classification task training on the normalized training data using the gradient boosting decision tree until the model converges to determine the classification model.

7. A reimbursement form field prediction method based on a large model and GBDT according to claim 1, characterized in that Through the classification model, perform fill-in form prediction on the fields to be predicted to obtain auxiliary fill-in form data specifically includes: Obtain the feature field numerical values of the current reimbursement form, and input the feature field numerical values into the classification model to obtain a prediction value group; Perform prediction probability screening on the prediction value group to determine the output prediction value; Perform fill-in form intention judgment on the output prediction value to obtain the auxiliary fill-in form data.

8. A method for predicting reimbursement form fields based on a large model and GBDT according to claim 1, characterized in that After predicting the form filling for the field to be predicted through the classification model and obtaining auxiliary form filling data, the method further includes: When there is no preset setting parameter in the auxiliary form filling data, manually fill in the form.

9. A reimbursement form field prediction device based on a large model and GBDT, characterized in that, The device includes: At least one processor; And a memory communicatively connected to the at least one processor; Wherein, the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to: Obtain expense report data, perform field prediction analysis on the expense report data, and determine unique identification data; Extract related fields of the prediction field from the expense report data to obtain feature fields; Divide the expense report data into training sets to obtain target values; Based on the feature fields and target values, obtain normalized training data through CatBoost encoding normalization; According to the unique identification data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine a classification model; Through the classification model, predict the form filling for the field to be predicted to obtain auxiliary form filling data.

10. A non - volatile computer storage medium for predicting reimbursement form fields based on large models and GBDT, storing computer - executable instructions, characterized in that, The computer-executable instructions are set to: Obtain expense report data, perform field prediction analysis on the expense report data, and determine unique identification data; Extract related fields of the prediction field from the expense report data to obtain feature fields; Divide the expense report data into training sets to obtain target values; Based on the feature fields and target values, obtain normalized training data through CatBoost encoding normalization; According to the unique identification data and the normalized training data, perform classification task training on the gradient boosting decision tree model to determine a classification model; Through the classification model, predict the form filling for the field to be predicted to obtain auxiliary form filling data.

Images