Method and system for automatically linking data resources based on pre-trained language model

Abstract

The present invention provides a method and system for automatically attaching data resources based on a pre-trained language model, and belongs to the field of data management. The method comprises the following steps: obtaining data resource information filled in by a user. Dividing attribute triples into three categories. Using a pre-trained language model to generate an attribute triple vector representation of new data resource information, and then using cosine similarity to calculate similarity with existing data resources. Using a sorting algorithm to generate a list of data resources ranked in the top 10 by similarity. Selecting the new data resource with the highest similarity after sorting for automatic attachment. The system comprises a data resource acquisition module, an attribute triple generation module, a similarity calculation module, a data screening module and a data attachment module. The method and system can effectively perform data cleaning and preprocessing, remove duplicate, erroneous or incomplete data items, and ensure data consistency and integrity.

Description

Technical Field

[0001] The present invention relates to a method and system for automatically linking data resources based on a pre-trained language model, and belongs to the field of computer technology. Background Art

[0002] With the advent of the information age, information exchange has become more and more frequent. Whether it is individuals, enterprises or various fields of society, data has become an indispensable basis for information exchange. Especially driven by emerging technologies such as big data, cloud computing, and artificial intelligence, the amount of data has shown explosive growth, and the types and sources of data have become increasingly diverse. How to effectively manage and utilize these massive amounts of data has become an important issue that needs to be solved urgently.

[0003] In all walks of life, the importance of data is becoming increasingly prominent. For enterprises, data is not only an asset, but also the foundation of project development and business operations. Whether it is market analysis, customer management, product design or operational decision-making, data support is indispensable. Effective data management and utilization can help enterprises gain insight into market trends, improve operational efficiency, optimize resource allocation, and thus occupy a favorable position in the fierce market competition. However, with the growth of data volume, data management also faces a series of challenges. First, the diversity of data sources leads to inconsistent data formats and structures, which increases the difficulty of data integration. Secondly, the quality of data varies, and there are a large number of repeated, erroneous or incomplete data items, which require data cleaning and preprocessing. Furthermore, there may be duplication or similarity between data items from different sources, which requires effective association and integration to improve the consistency and integrity of the data. However, many enterprises still rely on traditional methods in data management, which are inefficient and difficult to deal with the complex problems brought by massive data. Based on this, there is an urgent need for an efficient and flexible data management method that can provide all-round support in various links such as data acquisition, integration, analysis and utilization. Therefore, in the existing technology, it is necessary to catalog the data and then attach it to the existing data collection.

[0004] However, existing methods lack effective data cleaning and preprocessing mechanisms during the connection process, and are unable to remove noise and redundant information in the data. For large data sets, the speed of manual processing cannot meet the needs of rapidly growing data volumes. In addition, existing methods lack flexible data connection and conversion mechanisms and cannot adapt to the needs of different data sources. These shortcomings limit the effectiveness and efficiency of data management and cannot meet the needs of modern data processing. Summary of the invention

[0005] In response to the above problems, the present invention proposes a method and system for automatically linking data resources based on a pre-trained language model. The present invention generates attribute triples of data resources and uses a similarity algorithm to compare and integrate new data resources with existing data resources. It can more accurately identify and associate similar data items, avoid data redundancy and conflict, and thus effectively improve the efficiency of data management and data quality.

[0006] The technical solution of the present invention is: in the first aspect, the present invention provides a method for automatically attaching data resources based on a pre-trained language model, the method comprising the following steps:

[0007] Step 1: Obtain the basic information of the data resource, data item information and connection information filled in by the user;

[0008] Step 2: Generate attribute triples divided into three categories: basic information of data resources, data item information and attachment information according to different information modules;

[0009] Step 3: Use the pre-trained language model to generate attribute triple vector representations of new data resource information, and then use cosine similarity to calculate the similarity with the attribute triple vector representations of existing data resources;

[0010] Step 4: Use a sorting algorithm to generate a list of data resources ranked top in similarity;

[0011] Step 5: Automatically connect the new data resources with the highest similarity after selecting and sorting.

[0012] Furthermore, in step 1, obtaining the basic information of the data resource, data item information and connection information filled in by the user includes:

[0013] The data resource information is obtained from the data filled in by the user. The data resource information includes basic information of the data resource, data item information and connection information. The basic information of the data resource includes the name of the data resource, summary and the system to which it belongs.

[0014] Furthermore, in step 2, the generating of attribute triples divided into three categories of basic information, data item information and attachment information according to different information modules includes:

[0015] Generate attribute triples from the data resource name, summary, belonging system, data item and attachment information of the input data;

[0016] The generated attribute triples include:

[0017] Attribute triples are generated by cleaning and sorting data; different attribute triples in the same data item are given weights or priority information when generating attribute triples, and different weights are given according to different categories of data information; the different weights given according to different categories of data information include:

[0018] Data information is divided into three categories, namely basic information of data resources, data item information and attachment information, and different weight ratios are assigned according to the importance of each type of information.

[0019] Furthermore, the step 3 comprises:

[0020] Step 301: fine-tune the pre-trained language model according to data sets in different fields;

[0021] Step 302: Use the fine-tuned pre-trained language model to perform vector representation on the three types of attribute triplets of the data resource to obtain new data resources;

[0022] Step 303: perform weighted average of the vectors of the three attribute triplets according to the weights of different categories;

[0023] Step 304: Compare the vector similarity between the new data resource and the existing data resource.

[0024] Furthermore, the step 302 uses the fine-tuned pre-trained language model to perform vector representation on the three types of attribute triples of the data resource, including:

[0025] First, classify the attribute triples, and obtain the vector representation of each type of attribute triplet of the data resource through the fine-tuned pre-trained language model; use the LABSE pre-trained language model as the encoder of the word vector.

[0026] Furthermore, the step 301, fine-tuning the pre-trained language model according to data sets in different fields, includes:

[0027] First, we extract data sets from databases in a certain field, including financial databases. We clean the data in the financial databases to obtain trainable data sets. We use the data sets in the financial field to fine-tune the embedding layer of the pre-trained language model so that the word vectors can represent the semantics of terms and phrases in the financial field.

[0028] The cosine similarity loss is selected as the loss function, and its formula is as follows:

[0029] Loss( W t )=- 1 N ∑ i=1 N [ y i ∙ log cos H 1 , H 2 +(1- y i )∙ log 1- cos H 1 , H 2 ]

[0030] in, is the model parameter, representing the weight in the tth iteration, is the number of samples in the training data set, is the similarity label, =1 means similar, =0 means no similarity, is the vector representation of two financial texts; then the gradient descent method is used to add the L2 regularization term to the loss function. The L2 regularization term is the sum of the squares of all parameters. After adding L2 regularization, the formula for updating the parameters of the gradient descent method becomes:

[0031]

[0032] in, is the new weight in the t+1th iteration, which is the weight in the tth iteration Update, α is the learning rate, which controls the step size of each update. is the loss function over parameters The gradient of is the derivative of the L2 regularization term.

[0033] Furthermore, the step 302 and the step 303 include:

[0034] Each attribute triple of the data resource is represented by a vector through a fine-tuned pre-trained language model. The attribute triple is in the form of (h, a, v), where h represents the head node, a represents the attribute, and v represents the tail node.

[0035] First, concatenate attribute triplets of the same type together using the formula:

[0036]

[0037] represents the nth attribute, Indicates the nth attribute value;

[0038] is a sequence of attribute triplets of a data resource. The calculation process of the representation vector of the data resource aggregated by three attribute triplets is as follows:

[0039]

[0040]

[0041]

[0042] in, In order to finally get the embedded representation of data resource information, is the embedding representation of attribute triples of a category of data resources, express The transpose of for The variance of is the weight of the first type of data resource attribute triples, is the weight of the second type of data resource attribute triples, is the weight of the third type of data resource attribute triplet, is the weighted average sum of the three categories of data resource attribute triplets, It is a representation of data resources aggregated by three types of attribute triples. It is the representation of the data resource aggregated by the first type of attribute triples. It is the representation of the data resource aggregated by the second type of attribute triples. A representation of data resources aggregated by the third type of attribute triples.

[0043] Furthermore, in step 304, the vector similarity between the new data resource and the existing data resource is compared, and the formula is as follows:

[0044]

[0045] Among them, Cos() represents the cosine similarity calculation, A vector representation of the new data resource, A vector representation of an existing data resource.

[0046] Furthermore, in step 4, a ranking algorithm is used to generate a list of data resources ranked top in similarity, including:

[0047] The similarity between the data resource attribute triples in the database table to be attached and the attached data resource attribute triples is calculated, all attached data resources are sorted according to the similarity, and the top 10 data resources are displayed on the page for user selection.

[0048] In a second aspect, the present invention also provides a data resource automatic mounting system based on a pre-trained language model, which is a module for executing the data resource automatic mounting method based on a pre-trained language model described in the first aspect.

[0049] The beneficial effects of the present invention are:

[0050] 1. Through attribute triple generation and similarity calculation, manual intervention is reduced and the efficiency of data management is significantly improved. It supports multiple data sources and can quickly acquire and process massive data to meet the needs of data growth.

[0051] 2. Perform data cleaning and preprocessing before data attachment to effectively remove duplicate, erroneous or incomplete data items, ensuring data accuracy and integrity. Using the cosine similarity calculation algorithm, similar data items can be accurately identified and associated to avoid data redundancy and conflict.

[0052] 3. The similarity calculation module can intelligently process complex data relationships. The data screening module automatically selects the most relevant data items based on similarity rankings, reducing the workload of manual selection. The data attachment module ensures the consistency and integrity of the data when associating and integrating data, and ensures the quality of the integrated data through consistency checks and integrity verification mechanisms.

[0053] 4. Friendly user interface, intuitively displays the list of the top data items in similarity ranking, convenient for users to select and operate. The automated data processing process reduces the user's operating burden and improves the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flowchart of the main steps of the method for automatically linking data resources based on a pre-trained language model according to an embodiment of the present invention;

[0055] Figure 2 It is a schematic diagram of the main steps of generating attribute triples of different categories in the implementation mode of the present invention;

[0056] Figure 3 This is a flow chart of the main steps of using a pre-trained language model to generate attribute triple vector representations of data resource information and then using cosine similarity to calculate similarity with existing data resources in an implementation of the present invention. DETAILED DESCRIPTION

[0057] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0058] In the implementation mode of the present invention, the data of the data provider is cataloged and linked. Figure 1 1 is a schematic diagram of the main steps of a method for automatically attaching data resources based on a pre-trained language model according to an embodiment of the present invention, the method comprising the following steps:

[0059] Step 101: Obtain basic information of data resources, data item information and connection information filled in by the user;

[0060] The data resource information is obtained from the data filled in by the user. The data resource information includes basic information of the data resource, data item information and connection information. The basic information of the data resource includes the name of the data resource, summary and the system to which it belongs.

[0061] Extract new data items from the connected data source. There is no restriction on the data content during the extraction process, as long as it is text information. The extracted data items can include multiple types of information. Perform preliminary preprocessing on the extracted data items to remove obvious duplicate, erroneous or incomplete data items. The preprocessing steps include but are not limited to: data deduplication, format standardization, error correction and missing value filling. Through preprocessing, the quality of data items is improved and the complexity of subsequent processing is reduced. The preprocessed data items are stored in the temporary storage area of ​​the system to facilitate access and use by subsequent processing modules. The format of data item storage can be customized according to system requirements to ensure the completeness and accuracy of data item information.

[0062] Step 102: Generate attribute triples divided into three categories: data resource basic information, data item information, and attachment information according to different information modules;

[0063] The attribute triples divided into three categories of basic information, data item information and attachment information are generated according to different information modules, including:

[0064] Generate attribute triples from the data resource name, summary, belonging system, data item and attachment information of the input data;

[0065] The generated attribute triples include:

[0066] Attribute triples are generated by cleaning and sorting data; different attribute triples in the same data item are given weights or priority information when generating attribute triples, and different weights are given according to different categories of data information; the different weights given according to different categories of data information include:

[0067] Data information is divided into three categories, namely basic information of data resources, data item information and attachment information, and different weight ratios are assigned according to the importance of each type of information.

[0068] Specifically, the preprocessed data resource information stored in step 101 is read, parsed, and key attributes therein are extracted, such as basic information of data items, data item information, and hooking information. The parsed data item information is used to generate attribute triples of different categories according to a predefined format. Weight information is added to the attribute triples according to the category of the data resource. The generated attribute triples are stored in a temporary storage area of ​​the system so that they can be accessed and used by subsequent similarity calculation modules. The storage format is JSON or XML to facilitate the storage and access of structured data.

[0069] More detailed, such as Figure 2 Schematic diagram of the main steps of generating attribute triples of different categories in the implementation mode of the present invention.

[0070] Step 201, extracting key attributes of data.

[0071] Get the data resource of new data from the data filled in by the user. The data resource information that the user needs to fill in includes basic information of the data resource, data item information and hook information. Parse the data resource information and store the extracted key attributes in the temporary storage area of ​​the system according to the predefined categories for access and use in subsequent steps. The extracted key attributes are shown in Table 1:

[0072] Table 1

[0073]

[0074] Step 202: Construct attribute triples

[0075] The triplet includes data resource name, attribute, and attribute value. The data resource name is the field name of the data item, the attribute is the attribute possessed by the data resource, and the attribute value is the specific content of the data item. The constructed triplet is shown in Table 2:

[0076] Table 2

[0077]

[0078] Step 203: assign different weights to different attribute triples

[0079] Define the weight distribution rules according to the business requirements of the data items and the importance of the attributes. According to the defined weight distribution rules, assign corresponding weights to different categories of each attribute triple. The weight is a floating point number between 0 and 1, indicating the importance of the attribute. The weight value distribution is shown in Table 3:

[0080] Table 3

[0081]

[0082] Step 103: Generate attribute triple vector representation of new data resource information using the pre-trained language model, and then use cosine similarity to calculate similarity with the attribute triple vector representation of existing data resources;

[0083] Furthermore, the step 103 includes:

[0084] Step 301: fine-tune the pre-trained language model according to data sets in different fields;

[0085] First, we extract data sets from databases in a certain field, including financial databases. We clean the data in the financial databases to obtain trainable data sets. We use the data sets in the financial field to fine-tune the embedding layer of the pre-trained language model so that the word vectors can represent the semantics of terms and phrases in the financial field.

[0086] Since this task is mainly to optimize the vector representation of financial data text, the present invention selects cosine similarity loss as the loss function, and its formula is as follows:

[0087] Loss( W t )=- 1 N ∑ i=1 N [ y i ∙ log cos H 1 , H 2 +(1- y i )∙ log 1- cos H 1 , H 2 ]

[0088] in, is the model parameter, representing the weight in the tth iteration, is the number of samples in the training data set, is the similarity label, =1 means similar, =0 means no similarity, is the vector representation of two financial texts; then the gradient descent method is used to add the L2 regularization term to the loss function. The L2 regularization term is the sum of the squares of all parameters. After adding L2 regularization, the formula for updating the parameters of the gradient descent method becomes:

[0089]

[0090] in, is the new weight in the t+1th iteration, which is the weight in the tth iteration Update, α is the learning rate, which controls the step size of each update. is the loss function over parameters The gradient of is the derivative of the L2 regularization term.

[0091] Step 302: Use the fine-tuned pre-trained language model to perform vector representation on the three types of attribute triplets of the data resource to obtain new data resources;

[0092] First, the attribute triples are classified. Each type of attribute triple in the data resource obtains a vector representation through a fine-tuned pre-trained language model. Since the pre-trained language model can capture the semantics of words in different contexts and performs well in tasks that require context sensitivity, the present invention uses the LABSE pre-trained language model as the encoder of the word vector. The attribute triple is in the form of (h, a, v), where h represents the head node, a represents the attribute, and v represents the tail node (attribute value);

[0093] First, concatenate attribute triplets of the same type together using the formula:

[0094]

[0095] represents the nth attribute, Indicates the nth attribute value;

[0096] is a sequence of attribute triplets of a data resource. The calculation process of the representation vector of the data resource aggregated by three attribute triplets is as follows:

[0097]

[0098] in, is the embedding representation of attribute triples of a category of data resources, express The transpose of for The variance of It is a representation of data resources aggregated by three types of attribute triples.

[0099] Step 303: perform weighted average of the vectors of the three attribute triplets according to the weights of different categories;

[0100]

[0101]

[0102] In order to finally get the embedded representation of data resource information, is the weight of the first type of data resource attribute triples, is the weight of the second type of data resource attribute triples, is the weight of the third type of data resource attribute triplet, is the weighted average sum of the three categories of data resource attribute triplets, It is the representation of the data resource aggregated by the first type of attribute triples. It is the representation of the data resource aggregated by the second type of attribute triples. A representation of data resources aggregated by the third type of attribute triples.

[0103] Step 304: Compare the vector similarity between the new data resource and the existing data resource.

[0104] Furthermore, in step 304, the vector similarity between the new data resource and the existing data resource is compared, and the formula is as follows:

[0105]

[0106] Among them, Cos() represents the cosine similarity calculation, A vector representation of the new data resource, A vector representation of an existing data resource.

[0107] Step 104: using a sorting algorithm to generate a list of data resources ranked top in similarity;

[0108] Furthermore, in step 104, a ranking algorithm is used to generate a list of data resources ranked top in similarity, including:

[0109] Calculate the similarity between the data resource attribute triples in the database table of the information to be attached and the attached data resource attribute triples, and sort the read similarity results in descending order according to the similarity score to ensure that the most similar data items are ranked first. The sorting process can use efficient sorting algorithms such as quick sort and merge sort to ensure that the sorting can be completed quickly even in the case of large amounts of data. Filter out the top 10 existing data items in similarity ranking from the sorted similarity results. If there are less than 10 similar items in the similarity results, filter out all available similar data items. Generate a list of the top 10 similar data items filtered out and display them on the system interface for users to view and select.

[0110] Step 105: Automatically connect the new data resources with the highest similarity after sorting. The connection rules can be predefined or automatically generated by a machine learning model.

[0111] In a second aspect, the present invention also provides a data resource automatic mounting system based on a pre-trained language model, which is a module for executing the data resource automatic mounting method based on a pre-trained language model described in the first aspect.

[0112] Specifically include:

[0113] Data resource acquisition module: used to obtain basic information, data item information and connection information of data resources filled in by users;

[0114] Attribute triple generation module: used to generate attribute triples divided into three categories of data resource basic information, data item information and attachment information according to different information modules;

[0115] Similarity calculation module: used to generate attribute triple vector representation of new data resource information using the pre-trained language model, and then use cosine similarity to calculate the similarity with the attribute triple vector representation of existing data resources;

[0116] Data screening module: used to generate a list of data resources ranked top in similarity using a sorting algorithm;

[0117] Data attachment module: used to automatically attach the data resource with the highest similarity after selecting and sorting the new data resources.

[0118] In order to highlight the advantages of the method of the present invention, the data resource mounting method of the present invention is compared with the traditional data resource mounting method, and the results are shown in Table 4:

[0119] Table 4

[0120]

[0121] The automatic data resource connection method based on the pre-trained language model is characterized by high efficiency, intelligence, automated integration, high data quality and scalability.

[0122] Efficiency: The pre-trained language model is used to automatically extract and generate data attributes, which improves data management efficiency. There is no need to manually write rules or scripts, which greatly reduces human intervention.

[0123] Intelligence: By generating attribute triples, the semantic information of the data can be captured and the intelligence level of data integration and association can be improved.

[0124] Automated integration: Use similarity algorithms to automatically compare new data with existing data, reducing human errors and omissions in the data integration process.

[0125] Data quality: Improved data quality, as the system can match and verify data through intelligent algorithms rather than relying on manual matching methods with high error rates.

[0126] Scalability: Able to adapt to different types of data resources and changing business needs, supporting dynamic expansion and automatic updates.

[0127] Although traditional data resource attachment methods are simple to implement and low-cost, they have problems such as low efficiency, prone to errors, inadaptability and low data quality.

[0128] Low efficiency: For large-scale data or complex data structures, manual rules and simple algorithms are inefficient, and the data connection and integration process is cumbersome and time-consuming.

[0129] Error-proneness: Human intervention can easily introduce errors and omissions, and the formulation of matching rules may not be comprehensive enough, resulting in insufficient accuracy and completeness of data integration.

[0130] Inadaptability: It is difficult to handle changes in data formats or the introduction of new data, requires frequent rule updates and manual intervention, and has poor scalability.

[0131] Low data quality: The quality of data integration is limited by the accuracy of the rules and manual inspection, which makes it easy for data to be inconsistent and incomplete.

[0132] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for automatically linking data resources based on a pre-trained language model, characterized in that: The method comprises the following steps: Step 1: Obtain the basic information of the data resource, data item information and connection information filled in by the user; Step 2: Generate attribute triples divided into three categories: basic information of data resources, data item information and attachment information according to different information modules; Step 3: Use the pre-trained language model to generate attribute triple vector representations of new data resource information, and then use cosine similarity to calculate the similarity with the attribute triple vector representations of existing data resources; Step 4: Use a sorting algorithm to generate a list of data resources ranked top in similarity; Step 5: Automatically connect the new data resources with the highest similarity after selecting and sorting them; In step 2, the generating of attribute triples divided into three categories of basic information, data item information and attachment information according to different information modules includes: Generate attribute triples from the data resource name, summary, belonging system, data item and attachment information of the input data; The generated attribute triples include: Attribute triples are generated by cleaning and sorting data; different attribute triples in the same data item are given weights or priority information when generating attribute triples, and different weights are given according to different categories of data information; the different weights given according to different categories of data information include: Data information is divided into three categories, namely basic information of data resources, data item information and attachment information, and different weight ratios are assigned according to the importance of each type of information.

2. The method for automatically linking data resources based on a pre-trained language model according to claim 1, characterized in that: In step 1, obtaining the basic information of the data resource, data item information and connection information filled in by the user includes: The data resource information is obtained from the data filled in by the user. The data resource information includes basic information of the data resource, data item information and connection information. The basic information of the data resource includes the name of the data resource, summary and the system to which it belongs.

3. The method for automatically linking data resources based on a pre-trained language model according to claim 1, characterized in that: The step 3 comprises: Step 301: fine-tune the pre-trained language model according to data sets in different fields; Step 302: Use the fine-tuned pre-trained language model to perform vector representation on the three types of attribute triplets of the data resource to obtain new data resources; Step 303: perform weighted average of the vectors of the three attribute triplets according to the weights of different categories; Step 304: Compare the vector similarity between the new data resource and the existing data resource.

4. The method for automatically linking data resources based on a pre-trained language model according to claim 3, characterized in that: The step 302 uses the fine-tuned pre-trained language model to perform vector representation on the three types of attribute triples of the data resource, including: First, classify the attribute triples, and obtain the vector representation of each type of attribute triplet of the data resource through the fine-tuned pre-trained language model; use the LABSE pre-trained language model as the encoder of the word vector.

5. The method for automatically linking data resources based on a pre-trained language model according to claim 3, characterized in that: The step 301, fine-tuning the pre-trained language model according to data sets in different fields, includes: First, we extract data sets from databases in a certain field, including financial databases. We clean the data in the financial databases to obtain trainable data sets. We use the data sets in the financial field to fine-tune the embedding layer of the pre-trained language model so that the word vectors can represent the semantics of terms and phrases in the financial field. The cosine similarity loss is selected as the loss function, and its formula is as follows: ； in, is the model parameter, representing the weight in the tth iteration, is the number of samples in the training data set, is the similarity label, =1 means similar, =0 means no similarity, is the vector representation of two financial texts; then the gradient descent method is used to add the L2 regularization term to the loss function. The L2 regularization term is the sum of the squares of all parameters. After adding L2 regularization, the formula for updating the parameters of the gradient descent method becomes: ； in, is the new weight in the t+1th iteration, which is the weight in the tth iteration Updated, α is the learning rate, which controls the step size of each update. is the loss function over parameters The gradient of is the derivative of the L2 regularization term.

6. The method for automatically linking data resources based on a pre-trained language model according to claim 3, characterized in that: The steps 302 and 303 include: Each attribute triple of the data resource is represented by a vector through a fine-tuned pre-trained language model. The attribute triple is in the form of (h, a, v), where h represents the head node, a represents the attribute, and v represents the tail node. First, concatenate attribute triplets of the same type together using the formula: ； represents the nth attribute, Indicates the nth attribute value; is a sequence of attribute triplets of a data resource. The calculation process of the representation vector of the data resource aggregated by three attribute triplets is as follows: ； ； ； in, In order to finally get the embedded representation of data resource information, is the embedding representation of attribute triples of a category of data resources, express The transpose of for The variance of is the weight of the first type of data resource attribute triples, is the weight of the second type of data resource attribute triples, is the weight of the third type of data resource attribute triplet, is the weighted average sum of the three categories of data resource attribute triplets, It is a representation of data resources aggregated by three types of attribute triples. It is the representation of the data resource aggregated by the first type of attribute triples. It is the representation of the data resource aggregated by the second type of attribute triples. A representation of data resources aggregated by the third type of attribute triples.

7. The method for automatically linking data resources based on a pre-trained language model according to claim 3, characterized in that: In step 304, the vector similarity between the new data resource and the existing data resource is compared, and the formula is as follows: ； Among them, Cos() represents the cosine similarity calculation, A vector representation of the new data resource, A vector representation of an existing data resource.

8. The method for automatically linking data resources based on a pre-trained language model according to claim 1, characterized in that: In step 4, a sorting algorithm is used to generate a list of data resources ranked top in similarity, including: The similarity between the data resource attribute triples in the database table to be attached and the attached data resource attribute triples is calculated, all attached data resources are sorted according to the similarity, and the top 10 data resources are displayed on the page for user selection.

9. The data resource automatic connection system based on the pre-trained language model is characterized by: A module for executing the method for automatically attaching data resources based on a pre-trained language model as described in any one of claims 1 to 8.

Images