A method of automatically inferring semantics of a data cleaning script

Abstract

The application discloses a method for automatically inferring semantics of a data cleaning script, comprising loading an original table to the data cleaning script and obtaining an input table and an output table; constructing a table change space, obtaining a table change characteristic set based on comparison between the input / output table and the table change space; taking data conversion operation types that all satisfy inevitable occurrence, impossible occurrence and possible occurrence as candidate data conversion operation types; converting and splicing multiple descriptive texts corresponding to function names, function parameters and table change characteristics to obtain a semantic vector, inputting the semantic vector into a twin convolutional neural network model to obtain candidate data conversion operation type ranking; constructing a parameter slot of each data conversion operation type, inferring the parameter slot of the candidate data conversion operation type by using data cleaning code, table content and table change information of a corresponding row, and obtaining the type and parameters of the data conversion operation based on the inference result. The method can simply and conveniently infer the type and parameters of the data conversion operation in the cleaning code.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of data cleaning, and relates to a method for automatically inferring the semantics of a data cleaning script. BACKGROUND

[0002] Data wrangling (Data Cleaning) is a process of cleaning and transforming complex messy data into an ideal data format, which is an important pre-step for data access, data modeling and data visualization analysis. Using programming languages such as Python and R to write specific cleaning scripts is a common means to complete data cleaning work. In real work, data workers usually need to understand the specific data cleaning process performed in the cleaning script, i.e. the semantics of the cleaning script, to understand how the data has changed. For example, in program reuse, data workers need to learn the ideas of data cleaning in other scripts to modify and apply to their own data; in double checking, data workers need to verify the process of data cleaning performed by others to ensure the accuracy of the cleaning result; in code maintenance, data workers need to organize the code without good documentation to avoid errors in maintenance tasks.

[0003] Completing a data cleaning task requires various data transformation operations, such as filling missing values, removing duplicate rows, adding columns according to formulas, etc. The semantics of a data cleaning script refers to the type of data transformation operation performed by the cleaning code and its parameters. For example, given a line of R language cleaning code "tb2=arrange(tb1,-num)", the cleaning semantics of this code is to sort the num column of the input table tb1 in descending order, and then get the output table tb2, where the sorting operation is a data transformation operation type, and the sorted column num and the descending sorting method are the parameters of the operation. However, for data workers, it is not easy to determine the data transformation operation type and related parameters in the cleaning code, which involves two difficulties:

[0004] First, there are many types of data transformation operations, some of which are very complex. Related research has divided data transformation operations into 15 major categories and 21 subcategories according to the differences in operation objects and operation types, and the subcategories can be further divided. Some operations such as Pivot, Unpivot, etc. involve the conversion of table structure, which is difficult to understand. Therefore, data workers are difficult to master all types of data transformation operations.

[0005] Secondly, there are many programming ways to implement data transformation operations. At present, there are many programming languages and their related packages (libraries) that can complete data cleaning tasks, such as the Pandas package in Python, the tidyr and dplyr packages in R, etc. However, for the same data transformation operation, there can be different code implementation ways; and the same function can perform different operations when using different parameters, which makes it difficult for data workers who are not proficient in these programming languages to understand the semantics of data cleaning code. In addition, it takes a lot of time and effort for beginners to master a new programming language or tool package for data cleaning tasks.

[0006] In summary, understanding the semantics of complex cleaning scripts requires data workers to have a certain data cleaning foundation and programming skills, and this process is often tedious, time-consuming and prone to errors. Therefore, how to help data workers understand the data transformation semantics in the data cleaning script is a challenging research problem.

[0007] There are many research works aimed at helping data workers understand the semantics of cleaning scripts, which can be roughly divided into two categories:

[0008] (1) One is for code debugging, such as tools that can well reveal the effect of program execution on a table after a line of data cleaning code through text summary or visualization, and support data workers to further explore and debug the code, but this kind of tools cannot directly provide data workers with the type of data transformation operation and its parameters that the line of code specifically performs;

[0009] (2) The other is to visualize the semantics of data cleaning scripts, and tools can display cleaning semantics in the form of carefully designed graphical symbols or animations, although this kind of tools have the function of inferring the type of data transformation operation, but they rely on a rule-based engine to parse the semantics of data cleaning scripts, and the rule-based method has low generality and scalability, making it difficult to extend to different programming languages and other various cleaning functions. SUMMARY

[0010] The present application provides a method for automatically inferring the semantics of a data cleaning script, which can simply and conveniently infer the type and parameters of data transformation operations in cleaning code.

[0011] A method for automatically inferring the semantics of a data cleaning script, comprising:

[0012] (1) obtaining an original table, a data cleaning script and a plurality of data transformation operations, loading the original table to the data cleaning script, and saving the input table and the output table corresponding to each line of data cleaning code based on the loaded data cleaning script using a program executor;

[0013] (2) constructing a table change space by a first table change characteristic of a table object corresponding to each data conversion operation on a table attribute, comparing the change of the table object from the input table to the output table with the table change space to obtain a second table change characteristic set corresponding to each row of data cleaning code;

[0014] (3) dividing the first table change characteristic corresponding to each data conversion operation type into inevitable occurrence, impossible occurrence and possible occurrence to construct an operation-characteristic mapping table, comparing the operation-characteristic mapping table with the second table change characteristic set, and taking the data conversion operation type that makes the second table change characteristic set satisfy inevitable occurrence, impossible occurrence and possible occurrence as a candidate data conversion operation type;

[0015] parsing the function and function parameters of each row of data cleaning code, converting each table change characteristic in the corresponding second table change characteristic set into descriptive text, converting and splicing the function parameters, function and corresponding multiple descriptive texts into a first semantic vector through a FastText embedding model, inputting the first semantic vector into a twin convolutional neural network model, and comparing with a second semantic vector set in the twin convolutional neural network model to realize the possibility ranking of the candidate data conversion operation type;

[0016] (4) taking the parameters that can realize each data conversion operation type as parameter slots, and based on the cleaning code, sequentially inferring the parameter slots of the candidate data conversion operation type according to the possibility ranking, taking the first inferred successful candidate data conversion operation type and parameters as the type and parameters of the data conversion operation corresponding to the row of data cleaning code.

[0017] Based on the loaded data cleaning script, the program executor saves the input table and output table corresponding to each row of data cleaning code, including:

[0018] The program executor includes a program processor and a program interpreter, the program processor is used for marking the table information data with input and output tables in the loaded data cleaning script, the program interpreter is used for detecting the table information data of the input and output tables, and saving the input and output tables in the form of files.

[0019] The change of the table object corresponding to each data conversion operation on the table attribute is taken as the first table change characteristic, and the first table change characteristic set constructs a table change space, wherein the table object is a table, a row, a column and a cell, and the table attribute is quantity, order, relationship, value and type;

[0020] The first table change characteristic is the change of the table object in quantity, order, relationship or type.

[0021] The training process of the twin convolutional neural network model comprises:

[0022] obtaining a data cleaning script, a second semantic vector corresponding to each line of data cleaning code in the data cleaning script, and a labeled data conversion operation type to which the second semantic vector belongs, and taking the second semantic vector set as a training sample set;

[0023] constructing a training model, the training model comprising two sub-networks, each sub-network comprising a plurality of convolutional layers and a Flatten layer in sequence;

[0024] training the training model based on the training sample set by using a contrast loss function to obtain the twin convolutional neural network model.

[0025] The twin convolutional neural network model comprises a second semantic vector set, a second semantic vector is selected from the second semantic vector set saved in the model, each second semantic vector corresponds to a data conversion operation type, then the first semantic vector and each second semantic vector are taken as inputs of the twin convolutional neural network model, a first feature vector and a second feature vector are obtained respectively, the similarity degree of the first feature vector and the second feature vector is calculated based on the Euclidean distance, and the possibility size of each candidate data conversion operation type is obtained according to the similarity degree.

[0026] The first table change characteristic corresponding to each data conversion operation type is divided into a certain characteristic group, a possible characteristic group and an impossible characteristic group according to the certain occurrence, the impossible occurrence and the possible occurrence, to construct an operation-characteristic mapping table, wherein the certain characteristic group is a table change characteristic that must occur in the input-output table when the data conversion operation type is operated, the possible characteristic group is a table change characteristic that may occur in the input-output table when the data conversion operation type is operated, and the impossible characteristic group is a table change characteristic that may not occur in the input-output table when the data conversion operation type is operated.

[0027] The parameter slot of the candidate data conversion operation type is inferred in sequence based on the cleaning code according to the possibility, comprising:

[0028] First, based on the parameter slot of each data conversion operation type constructed, the candidate data conversion operation type parameter slot with the maximum possibility is obtained, and the parameter slot of the candidate data conversion operation type is inferred using the corresponding line of data cleaning code, table content (data content in the table, i.e. the value of a cell in the table) and table change information, if the inference is successful, the candidate data conversion operation type with the maximum possibility and the parameter are taken as the type and parameter of the data conversion operation corresponding to the line of data cleaning code, if at least one parameter slot is not successfully inferred, the inference is performed in sequence according to the possibility of the candidate data conversion operation type from large to small until the inference is successful, and the inference is ended.

[0029] Compared with the prior art, the present application has the beneficial effects that:

[0030] Compared with the prior art, the present application has the beneficial effects that:

[0031] In addition, the output of the present application (i.e., the type of data conversion operation performed in the data cleaning script and its parameters) can also be applied to various existing tools to increase their practicality. For example, the present application can be integrated into the Somnus visualization system to enhance the robustness of its code parsing and improve its semantic inference capability; the present application can also be used as a plug-in for Jupyter Notebook to automatically generate text annotations for cleaning code. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A method flowchart for automatically inferring the semantics of a data cleaning script is provided for the specific embodiment;

[0033] Figure 2 A comparison diagram of the operation-characteristics mapping table and the second set of table change characteristics is provided for the specific embodiment;

[0034] Figure 3 An architecture diagram of the twin convolutional neural network model is provided for the specific embodiment;

[0035] Figure 4 An application diagram of the twin convolutional neural network model is provided for the specific embodiment;

[0036] Figure 5 An effect diagram of the method for automatically inferring the semantics of a data cleaning script integrated into the Somnus visualization system is provided for the specific embodiment;

[0037] Figure 6 An effect diagram of the method for automatically inferring the semantics of a data cleaning script integrated into the Jupyter Notebook interactive programming tool is provided for the specific embodiment. DETAILED DESCRIPTION

[0038] The present application provides a method for automatically inferring the semantics of a data cleaning script, as shown in Figure 1 which includes:

[0039] S1: generate intermediate input table and output table, as shown in A of Figure 1 obtain original table, data cleaning script and various data conversion operations, load original table to data cleaning script, save input table and output table corresponding to each row of data cleaning code based on loaded data cleaning script using program executor; program executor includes program processor and program interpreter, program processor is used for marking table information data with input and output table in loaded data cleaning script, program interpreter is used for detecting table information data of input and output table, and saving input and output table in file form, as shown in B of Figure 1 .

[0040] S2: detect table change characteristics and build operation-characteristics mapping table: as shown in C of Figure 1 , change of table object in table attribute through each data conversion operation corresponding table object is taken as first table change characteristic, first table change characteristic set is used for constructing table change space, wherein, table object is table, row, column and cell, and table attribute is number, order, relationship, value and type; for example, change characteristic of "column" and "number" is whether number of column of input table and output table is changed (such as increased, decreased, or unchanged) after executing data conversion operation. Based on this table change design space, changes of four table objects of input and output table in five attributes are compared and detected respectively, changes of table object from input table to output table are compared with table change space to obtain second table change characteristic set corresponding to each row of data cleaning code;

[0041] first table change characteristics corresponding to each data conversion operation type are divided into certain characteristics group, possible characteristics group and impossible characteristics group according to certain occurrence, impossible occurrence and possible occurrence, to build operation-characteristics mapping table, wherein, certain characteristics group is table change characteristic of input and output table which must occur when data conversion operation type is operated, possible characteristics group is table change characteristic of input and output table which may occur when data conversion operation type is operated, and impossible characteristics group is table change characteristic of input and output table which may not occur when data conversion operation type is operated;

[0042] For example, given a line of R language code "df = df[df.num > 1]", the data conversion operation type performed by the line of code is filtering operation, and the specific semantics is "filtering out the data in the num column of the df table whose value is greater than 1". In this operation, the number of columns, the position order of columns, etc. will not change, which are certain characteristics, and the number of rows cannot increase, and the data type of columns cannot change, which are impossible characteristics. In addition, since the num column in the df table may have cell data greater than 1, it may not exist, so the number of rows may decrease or remain unchanged, which are possible characteristics.

[0043] As shown in D of Figure 1 , comparing the operation-characteristics mapping table with the second table change characteristic set will make the second table change characteristic set satisfy the certain occurrence, impossible occurrence and possible occurrence data conversion operation types as candidate data conversion operation types, and multiple candidate data conversion operation types construct an unordered candidate data conversion operation type set, as shown in Figure 2 , the specific steps are as follows:

[0044] It contains two inference strategies: (1) If a data conversion operation type in the operation-characteristics mapping table has a change characteristic in the certain characteristic group that does not exist in the table change characteristic set, the data conversion operation type does not belong to the unordered candidate data conversion operation type set. As shown in T1 data conversion operation type in Figure 2 , the certain characteristic group has C1 change characteristic, but the change characteristic does not exist in the detected table change characteristic set, so T1 data conversion operation type is not in the unordered candidate data conversion operation type set; (2) If a data conversion operation type in the operation-characteristics mapping table has a change characteristic in the impossible characteristic group that already exists in the table change characteristic set, the data conversion operation type does not belong to the unordered candidate data conversion operation type set. As shown in T3 data conversion operation type in Figure 2 , the impossible characteristic group has C2 change characteristic, and the change characteristic also exists in the detected table change characteristic set, so T3 data conversion operation type is not in the unordered candidate data conversion operation type set. Since T2 is a data conversion operation type that satisfies the certain occurrence, impossible occurrence and possible occurrence, T2 data conversion operation type is in the unordered candidate data conversion operation type set.

[0045] S3: operation type inference based on CNN model: parsing the function and function parameters of each line of data cleaning code, and converting each table change characteristic in the corresponding second table change characteristic set into descriptive text, as shown in Figure 3As shown, the function parameters and the function are respectively converted into a 1x300-dimensional function name vector and a 1x300-dimensional function parameter vector by a FastText embedding model, a plurality of corresponding descriptive texts are converted into a 1x300-dimensional sentence vector, and the sentence vectors are averaged bit by bit to obtain a 1x300-dimensional change characteristic vector; the above three vectors are spliced to obtain a 3x300-dimensional first semantic vector, the first semantic vector is input into a twin convolutional neural network model, and the possibility of the candidate data conversion operation type is sorted by comparing with a second semantic vector set in the twin convolutional neural network model; the architecture of the twin convolutional neural network model is as shown in Figure 3 As shown, it contains two subnetworks of the same structure and shares the weight value. Each subnetwork is composed of 5 convolutional layers and 1 flatten layer.

[0046] The training process of the twin convolutional neural network model includes:

[0047] 921 rows of data cleaning codes are collected, and the corresponding input and output tables are generated for each row of code, and the data conversion operation type to which the cleaning code belongs is manually annotated, and the second semantic vector is obtained according to the above semantic vector obtaining method, and the second semantic vector set is used as a training sample set;

[0048] The training model is constructed, and the training model includes two subnetworks, and each subnetwork includes a plurality of convolutional layers and a Flatten layer in sequence;

[0049] The twin convolutional neural network model is obtained by training the training model based on the contrast loss function. After the model training is completed, the Top-1 accuracy rate of the inference operation type is as high as 92.2%. The model can calculate the possibility of each possible data conversion operation in the data conversion operation candidate set, and according to the size of the possibility, the operation candidate set is finally output in a descending order.

[0050] As shown in Figure 4 The twin convolutional neural network model includes a second semantic vector set, that is, T1-T5, a second semantic vector is selected from the second semantic vector set saved in the model, each second semantic vector corresponds to a data conversion operation type, then the first semantic vector and each second semantic vector are input into the twin convolutional neural network model to obtain a first feature vector and a second feature vector respectively, the similarity degree of the first feature vector and the second feature vector is calculated based on the Euclidean distance, and the possibility size of the candidate data conversion operation type T5 is obtained according to the similarity degree, as shown in D of Figure 1 .

[0051] S4: Parameter inference of data transformation operation. The parameters of each data transformation operation type are defined as parameter slots. The parameter slots of the candidate data transformation operation types are inferred in order of possibility based on the cleaning code. The first inferred candidate data transformation operation type and parameters are determined as the data transformation operation type and parameters of the data cleaning code. Figure 1 The specific steps are as follows:

[0052] (1) For each data transformation operation, the parameters required by the operation are defined as slots. For example, the sorting operation has four parameter slots, which are the input table slot, the output table slot, the sorting column slot, and the sorting mode slot.(2) The data transformation operation with the highest possibility in the operation candidate set is inferred using the data cleaning code, the table content, and the table change information, i.e., the table change characteristic set. For example, if the data transformation operation type with the highest possibility in the operation candidate set is the sorting operation, the parameter slots of the sorting operation are filled with slot values according to the sorting operation parameters. In the above example, the slot values of the input table slot and the output table slot can be directly obtained from the processing result of step S1, which are tb1 and tb2, respectively. In order to continue to infer the slot values of the sorting column slot and the sorting mode slot, the column name num involved in the cleaning code is extracted. Then, whether the num column is in an ordered state is determined according to the table change characteristic set or directly from the output table. In this example, the num column of the output table is in a descending order state, which meets the execution result of the sorting operation. Therefore, the sorting column slot value of the sorting operation is the num column, and the sorting mode slot value is the descending sorting. For different data transformation operation types, the slot value inference methods are different due to different parameter slots. (3) If the parameters of the data transformation operation type with the highest possibility can be successfully inferred, the inference task is ended, and the final data transformation operation type and parameters of the code are obtained. If the parameters cannot be successfully inferred, the data transformation operation with the second highest possibility in the operation candidate set is selected to continue inferring its parameters. This step is repeated until the parameters of the operation type are successfully inferred. For example, in the above example, if the num column of the output table is not in an ordered state, it does not meet the execution result of the sorting operation, and therefore the parameters of the sorting operation cannot be successfully inferred. In this case, the data transformation operation with the second highest possibility in the operation candidate set is selected, and the slot values of the operation type are inferred according to the parameter slots. After the inference is completed, the data transformation operation type and parameters of the data cleaning code are determined and output, as shown in F of Figure 1

[0053] The application can be applied to many existing tools to increase their practicality. The following are two application cases of the application:

[0054] Integrated into Somnus visualization system of patent publication with application number 202210066324X. Somnus visualization system utilizes graphical icon-based directed acyclic graph to demonstrate the semantics of data cleaning code, but since Somnus adopts rule-based method to infer the data transformation operation type and its parameters of cleaning code, its scalability and generalization are limited. In order to improve the inference ability of Somnus, the technical method of automatic semantic inference of the present invention can be applied to the backend of Somnus, replacing its original program adapter module, so that Somnus can infer more abundant cleaning code. Figure 5 Three examples are demonstrated that Somnus system originally cannot support, but can successfully infer the code semantics after integrating the technical method of the present invention, which are respectively the cleaning semantics that can help Somnus infer Extract function equation, non-equation statement (i.e. statement without assignment equal sign) and non-function equation (i.e. assignment statement without calling function).

[0055] Integrated into Jupyter Notebook interactive programming tool. Nowadays, more and more data workers use web-based interactive narrative programming platforms such as Jupyter Notebook to complete code writing work, and also write code comments for easy sharing. However, manually writing comments is a time-consuming and tedious work. In order to reduce the burden of data workers to write comments for cleaning code, the present invention can be integrated into Jupyter Notebook as an application plug-in to automatically generate corresponding code comments for data cleaning code. The plug-in uses the semantic inference results of the present invention (i.e. data transformation operation type and its parameters) to automatically generate comment text according to the corresponding data transformation operation comment template, and calls Jupyter plug-in API to insert the comment text above the corresponding cleaning code. For example, Figure 6 A and B of are two cleaning code units, when the cleaning code is written, click the integrated plug-in at the top right corner such as Figure 6 C, which can generate comment text such as Figure 6 D for the cleaning code in the Notebook.

Claims

1. A method of automatically inferring semantics of a data cleaning script, the method comprising: Comprise: (1) obtain the original table, data cleaning script and various data conversion operations, load the original table to the data cleaning script, save the input table and output table corresponding to each row of data cleaning code based on the program executor after loading the data cleaning script; (2) the first table change characteristic of each data conversion operation corresponding to the table object on the table attribute is used to construct the table change space, and the change of the table object from the input table to the output table is compared with the table change space to obtain the second table change characteristic set corresponding to each row of data cleaning code; (3) the first table change characteristic corresponding to each data conversion operation type is divided into three parts according to the necessary occurrence, the impossible occurrence and the possible occurrence to construct the operation-characteristics mapping table, and the operation-characteristics mapping table is compared with the second table change characteristic set, and the data conversion operation type which makes the second table change characteristic set satisfy the necessary occurrence, the impossible occurrence and the possible occurrence is taken as the candidate data conversion operation type; The function and function parameter of each row of data cleaning code are parsed, each table change characteristic in the corresponding second table change characteristic set is converted into descriptive text, the function parameter, the function and the corresponding multiple descriptive texts are converted and spliced into the first semantic vector by the FastText embedding model, the first semantic vector is input into the twin convolutional neural network model, and the possibility of the candidate data conversion operation type is sorted by comparing with the second semantic vector set in the twin convolutional neural network model; (4) the parameters capable of realizing each data conversion operation type are taken as parameter slots, the parameter slots of the candidate data conversion operation type are inferred in turn based on the cleaning code according to the possibility, and the type and parameter of the first inferred successful candidate data conversion operation are taken as the type and parameter of the data conversion operation corresponding to the row of data cleaning code.

2. The method for automatically inferring semantics of a data cleansing script of claim 1, wherein, The program executor includes a program processor and a program interpreter, the program processor is used for marking the table information data with input and output tables in the loaded data cleaning script, the program interpreter is used for detecting the table information data of the input and output tables, and the input and output tables are saved in the form of files. The change of each data conversion operation corresponding to the table object on the table attribute is taken as the first table change characteristic, and the first table change characteristic set is used to construct the table change space, wherein the table object is table, row, column and cell, and the table attribute is number, order, relationship, value and type.

3. The method for automatically inferring semantics of data cleaning scripts of claim 1, wherein, The first table change characteristic is the change of the table object in number, order, relationship or type.

4. The method for automatically inferring semantics of data cleaning scripts of claim 3, wherein, The training process of the twin convolutional neural network model comprises:

5. The method for automatically inferring semantics of data cleaning scripts of claim 1, wherein, obtaining a data cleaning script, and a second semantic vector corresponding to each row of data cleaning code in the data cleaning script and a data conversion operation type belonging to it, taking the second semantic vector set as a training sample set; construct a training model, the training model includes two subnetworks, each subnetwork includes a plurality of convolutional layers and a Flatten layer in turn; ​ The training model is trained based on the training sample set through a contrast loss function to obtain the twin convolutional neural network model.

6. The method for automatically inferring semantics of data cleaning scripts of claim 5, wherein, The twin convolutional neural network model includes a second semantic vector set, one second semantic vector is selected from the second semantic vector set saved in the model, each second semantic vector corresponds to a data conversion operation type, then the first semantic vector and each second semantic vector are taken as inputs of the twin convolutional neural network model, first feature vectors and second feature vectors are obtained respectively, the similarity of the first feature vectors and the second feature vectors is calculated based on the Euclidean distance, and the possibility of each candidate data conversion operation type is obtained according to the similarity.

7. The method for automatically inferring semantics of data cleaning scripts of claim 1, wherein, The first table change characteristics corresponding to each data conversion operation type are divided into a certain characteristic group, a possible characteristic group and an impossible characteristic group according to the certain occurrence, the impossible occurrence and the possible occurrence, to construct an operation-characteristic mapping table, wherein the certain characteristic group is the table change characteristic that the input-output table must occur when the data conversion operation type is operated, the possible characteristic group is the table change characteristic that the input-output table may occur when the data conversion operation type is operated, and the impossible characteristic group is the table change characteristic that the input-output table may not occur when the data conversion operation type is operated.

8. The method for automatically inferring semantics of data cleaning scripts of claim 1, wherein, Based on the cleaning code, the parameter slot of the candidate data conversion operation type is inferred in order according to the possibility, including: First, based on the constructed parameter slot of each data conversion operation type, the candidate data conversion operation type parameter slot with the maximum possibility is obtained, the parameter slot of the candidate data conversion operation type is inferred using the corresponding row of data cleaning code, table content and table change information, if the inference is successful, the candidate data conversion operation type with the maximum possibility and the parameter are taken as the type and parameter of the data conversion operation corresponding to the row of data cleaning code, if at least one parameter slot is not inferred successfully, the inference is performed in order according to the possibility of the inferred candidate data conversion operation type from large to small, until the inference is successful, and the inference is ended.

Images