CN115438181A - Metadata label classification method and device - Google Patents

Abstract

Embodiments of the present invention provide a method and device for label classification of metadata, which are applied in the field of big data technology to solve the problem of low efficiency of label classification for determining metadata in the prior art. Including: for any metadata, obtain the Chinese semantic information of the metadata and the English semantic information of the metadata; according to the Chinese word segmentation set, determine the first feature vector corresponding to the Chinese semantic information of the metadata; the Chinese word segmentation set is obtained by analyzing each metadata According to the English word segmentation set, determine the second feature vector corresponding to the English semantic information of the metadata; the English word segmentation set is obtained by segmenting the English semantic information of each metadata; the first feature vector The vector is concatenated with the second feature vector to obtain a feature encoding vector of the metadata; the feature encoding vector of the metadata is respectively input into each label classifier to determine the label category of the metadata.

Description

Method and device for label classification of metadata

technical field

Embodiments of the present invention relate to the technical field of big data, and in particular to a tag classification method and device for metadata.

Background technique

With the continuous development of science and technology, in the work of big data governance, one of the tasks is to label and classify metadata. By determining the label category of metadata, it can effectively help users better understand data.

At present, for each piece of metadata, the tag category to which it belongs is manually judged, which will result in low efficiency in determining the tag classification of the metadata.

To sum up, how to solve the low efficiency of label classification for determining metadata is a technical problem that needs to be solved urgently.

Contents of the invention

An embodiment of the present invention provides a method for label classification of metadata, which is used to solve the problem of low efficiency of label classification for determining metadata in the prior art.

In the first aspect, an embodiment of the present invention provides a tag classification method for metadata, including: for any metadata, acquiring Chinese semantic information of metadata and English semantic information of metadata; The first feature vector corresponding to the semantic information; the Chinese word segmentation set is obtained by segmenting the Chinese semantic information of each metadata; according to the English word segmentation set, determine the second feature vector corresponding to the English semantic information of the metadata; the English word segmentation set is It is obtained by segmenting the English semantic information of each metadata; splicing the first feature vector and the second feature vector to obtain the feature encoding vector of metadata; inputting the feature encoding vector of metadata into each label classifier respectively, and determining the element The label category for the data.

In the embodiment of the present invention, word segmentation is performed through the Chinese semantic information and English semantic information of the metadata, the feature encoding vector of the metadata is determined, and then the feature encoding vector of the metadata is input into each label classifier, and the maximum value of the predicted probability corresponds to The label category of the label classifier is used as the label category of the metadata, so as to realize more accurate determination of the label category of the metadata and improve the efficiency of determining the label category of the metadata.

Optionally, according to the Chinese word segmentation set, determine the first feature vector corresponding to the Chinese semantic information of the metadata, including: performing word segmentation on the Chinese semantic information of the metadata to obtain each first word; for any first word, if determined The first participle exists in the Chinese participle set, then set the sub-feature vector corresponding to the first participle as the first value; if it is determined that the first participle does not exist in the Chinese participle set, then set the sub-feature vector corresponding to the first participle as the first value Binary value: according to the position of each first participle in the Chinese semantic information of the metadata, the sub-feature vectors corresponding to each first participle are spliced to obtain the first feature vector corresponding to the Chinese semantic information of the metadata.

In the embodiment of the present invention, the first word segmentation is obtained by segmenting the Chinese semantic information of the metadata, and then the first feature vector of the metadata is determined according to the Chinese word segmentation set and the first word segmentation, so as to facilitate subsequent determination of the element based on the first feature vector. The feature encoding vector of the data, and then realize the input of the feature encoding vector of the metadata into each label classifier, and determine the label category of the metadata more accurately and quickly.

Optionally, according to the English word segmentation set, determine the second feature vector corresponding to the English semantic information of the metadata, including: performing word segmentation on the English semantic information of the metadata to obtain each second word; for any second word, if determined The second participle exists in the English participle set, then set the sub-feature vector corresponding to the second participle as the first value; if it is determined that the second participle does not exist in the English participle set, then set the sub-feature vector corresponding to the second participle as the first value Binary value: according to the position of each second participle in the English semantic information of the metadata, the sub-feature vectors corresponding to each second participle are spliced to obtain the second feature vector corresponding to the English semantic information of the metadata.

In the embodiment of the present invention, the second word segmentation is obtained by segmenting the English semantic information of the metadata, and then according to the second word segmentation and the English word segmentation set, the second feature vector of the metadata is determined, so as to facilitate subsequent determination of the metadata based on the second feature vector. The feature encoding vector of the data, and then realize the input of the feature encoding vector of the metadata into each label classifier, so that the label category of the metadata can be determined more accurately and quickly.

Optionally, the inputting the feature encoding vectors of the metadata into each label classifier to determine the label category of the metadata includes: inputting the feature encoding vectors of the metadata into each label classifier, and obtaining the The corresponding prediction probabilities in each of the above label classifiers; according to the corresponding prediction probabilities of each of the label classifiers, determine the label category of the metadata.

In the embodiment of the present invention, the label category of the label classifier corresponding to the maximum value of the predicted probability is used as the label category of the metadata, so as to achieve more accurate determination of the label category of the metadata and improve the efficiency of determining the label category of the metadata.

Optionally, each label classifier is obtained by training through training samples; each label classifier has an upper limit threshold and a lower limit threshold obtained through training; according to the corresponding prediction probability of each label classifier, determine the label category of the metadata, including : If the first predicted probability corresponding to the first label classifier is higher than the upper threshold of the first label classifier, and the second predicted probability corresponding to the second label classifier is lower than the lower threshold of the second label classifier, then the determination element The data has a label category corresponding to the first label classifier; the second label classifier is each label classifier except the first label classifier in each label classifier.

In the embodiment of the present invention, by determining qualified metadata as training samples, it is possible to increase the training samples, so that the upper threshold and lower threshold obtained by training each label classifier are more accurate, thereby realizing more accurate determination of metadata tag categories and improve the efficiency of determining tag categories for metadata.

Optionally, it also includes: for any label classifier, if the predicted probability of the label classifier is lower than the lower limit threshold of the label classifier, then determine that the metadata does not have the label category corresponding to the label classifier, and use the metadata as the Training samples for updating each label classifier.

In the embodiment of the present invention, by determining qualified metadata as training samples, it is possible to increase the number of training samples, so that in the training process, the upper limit threshold and lower limit threshold of each label classifier are optimized, thereby achieving more accurate Determining the label category of metadata and improving the efficiency of determining the label category of metadata.

Optionally, after determining the label category of the metadata, the method further includes: using the metadata as a training sample for updating each label classifier.

In the embodiment of the present invention, the metadata is used as the training sample updated by each label classifier, so that the upper limit threshold and the lower limit threshold of each label classifier can be optimized, and then the label category of the metadata can be determined more accurately and the determination of the metadata can be improved. The efficiency of labeling categories of data.

Optionally, it also includes: if there is no predicted probability corresponding to any label classifier higher than the upper threshold of the label classifier, after each label classifier is updated, continue to input the feature encoding vector of the metadata into the updated The individual label classifiers of , determine the label category of the metadata.

In the embodiment of the present invention, by inputting the feature encoding vector of the metadata into each updated label classifier, the efficiency of determining the label category of the metadata can be improved and the label category of the metadata can be determined more accurately.

In a second aspect, an embodiment of the present invention provides a tag classification device for metadata, including: an acquisition unit configured to acquire Chinese semantic information of metadata and English semantic information of metadata for any metadata. The processing unit is used to determine the first feature vector corresponding to the Chinese semantic information of the metadata according to the Chinese word segmentation set; the Chinese word segmentation set is obtained by segmenting the Chinese semantic information of each metadata; according to the English word segmentation set, determine the metadata The second feature vector corresponding to the English semantic information of the metadata; the English word segmentation set is obtained by segmenting the English semantic information of each metadata; the first feature vector and the second feature vector are spliced to obtain the feature encoding vector of the metadata; The feature encoding vectors of the metadata are respectively input into each label classifier to determine the label category of the metadata.

Optionally, the processing unit is specifically configured to: segment the Chinese semantic information of the metadata to obtain each first segment; for any first segment, if it is determined that the first segment exists in the Chinese segment set, set the first segment The corresponding sub-characteristic vector is the first value; if it is determined that the first participle does not exist in the Chinese word segmentation set, then set the sub-characteristic vector corresponding to the first participle as the second value; according to the Chinese semantic information of each first participle in the metadata In the position, the sub-feature vectors corresponding to each first word segmentation are concatenated to obtain the first feature vector corresponding to the Chinese semantic information of the metadata.

Optionally, the processing unit is specifically configured to: segment the English semantic information of the metadata to obtain each second participle; for any second participle, if it is determined that the second participle exists in the English participle set, set the second participle The corresponding sub-characteristic vector is the first value; if it is determined that the second participle does not exist in the English participle set, then set the sub-characteristic vector corresponding to the second participle as the second value; according to the English semantic information of each second participle in the metadata The position in , the sub-feature vectors corresponding to each second word segmentation are concatenated to obtain the second feature vector corresponding to the English semantic information of the metadata.

Optionally, the processing unit is specifically configured to: input the feature encoding vector of the metadata into each label classifier, and obtain the corresponding prediction probability in each label classifier; according to the corresponding prediction probability of each label classifier, Determines the tag category for the metadata.

Optionally, each label classifier is obtained by training through training samples; each label classifier has an upper threshold and a lower threshold obtained through training, and the processing unit is specifically used for: if the first predicted probability corresponding to the first label classifier is higher than the upper limit threshold of the first label classifier, and the second predicted probability corresponding to the second label classifier is lower than the lower limit threshold of the second label classifier, then it is determined that the metadata has a label category corresponding to the first label classifier; The two-label classifiers are all label classifiers except the first label classifier among the label classifiers.

Optionally, the processing unit is specifically configured to: for any label classifier, if the predicted probability of the label classifier is lower than the lower limit threshold of the label classifier, determine that the metadata does not have a label category corresponding to the label classifier, and send the metadata The data are used as training samples for updating each label classifier.

Optionally, the processing unit is specifically configured to: use the metadata as training samples for updating each label classifier.

Optionally, the processing unit is specifically configured to: if there is no predicted probability corresponding to any label classifier higher than the upper threshold of the label classifier, after each label classifier is updated, continue to separate the feature encoding vectors of the metadata into Input the updated label classifiers to determine the label categories of the metadata.

In the third aspect, the embodiment of the present invention also provides a computing device, including at least one processor and at least one memory, wherein the memory stores a computer program, and when the program is executed by the processor, the The processor executes a tag classification method for metadata according to the first aspect.

In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium, the storage medium stores a program, and when the program is run on a computer, the computer implements a meta-data implementation of the above-mentioned first aspect. Label classification method.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings that need to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For Those of ordinary skill in the art can also obtain other drawings based on these drawings without any creative effort.

FIG. 1 is a flowchart of a tag classification method for metadata provided by an embodiment of the present invention;

FIG. 2 is a flow chart of a method for improving the accuracy of the upper threshold and the lower threshold of each label classifier provided by an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of tag classification of metadata provided by an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of a computing device provided by an embodiment of the present invention.

detailed description

In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In the following, some terms used in this application are generally explained to facilitate the understanding of those skilled in the art, and the terms used in this application are not limited.

1. Metadata: Metadata, also known as intermediary data and relay data, is data describing data (data about data), mainly information describing data attributes (property).

In a possible scenario, when company A is performing data governance, company A needs to label and classify the metadata of the secondary directory in the data resource directory, where the label categories include institutional information, merchant information, user information, terminal information , transaction information, etc. Classifying metadata can effectively help users better understand data on the one hand, and on the other hand, labeling metadata can further help big data platforms establish data levels.

As described in the background technology, due to the large amount of metadata, it is too wasteful of manpower and material resources to use manual methods to determine the tag category to which metadata belongs, and because the error in using manual methods to determine the label category to which metadata belongs Therefore, using a manual method to determine the label category of metadata not only increases the cost of labor, but also reduces the efficiency of label classification of metadata and the accuracy of determining the label category of metadata.

In view of the above problems, the present application proposes a metadata label classification method, which can obtain the label category corresponding to the metadata by inputting the metadata into a label classifier, thereby improving the efficiency of metadata classification, Reduced labor costs.

Because the attribute information of metadata is various, for example: the collected metadata attribute information includes but not limited to English name, Chinese name, field English name, field type, field Chinese name, field value description, etc. By performing feature encoding vectors on the metadata according to different attribute information, and then inputting the feature encoding vectors into each label classifier, and outputting the label categories of the metadata, it is possible to more accurately determine the label categories of the metadata and improve the quality of the metadata. The efficiency of label classification. In order to facilitate the introduction of this solution, the attribute information of metadata is taken as the Chinese semantic information of metadata and the English semantic information of metadata as an example to facilitate the understanding of those skilled in the art, and does not limit the attribute information of metadata in this application.

As shown in FIG. 1 , it provides a flow chart of a tag classification method for metadata according to an embodiment of the present invention, and the method includes the following steps:

Step 101, for any metadata, obtain the Chinese semantic information of the metadata and the English semantic information of the metadata.

In the embodiment of the present invention, for any metadata, the Chinese semantic information of the metadata and the English semantic information of the metadata are obtained. For example, if the data A is 01000000, the Chinese semantic information A' of the obtained metadata is the identification code of the card issuer and the English semantic information A' of the metadata is ISS_INS_ID_CD. For another example, if the data B is Li Mingming, the Chinese semantic information B' of the metadata obtained is name and the English semantic information B' of the metadata is NAME. For another example, if the data C is 8020, the Chinese semantic information C' of the acquired metadata is the acquirer identification code and the English semantic information C' of the metadata is ACQ_INS_CD.

Step 102, according to the Chinese word segmentation set, determine the first feature vector corresponding to the Chinese semantic information of the metadata.

In the embodiment of the present invention, according to the Chinese semantic information of the metadata, the Chinese word segmentation set of the metadata is determined. For example, if there is metadata A', metadata B' and metadata C'. Among them, the Chinese semantic information A' of the metadata is the identification code of the card issuer, and the Chinese semantic information A' of the metadata is segmented, and the first participle of the Chinese semantic information A' of the metadata is respectively the card issuer, the institution, and the identification code. The Chinese semantic information B' of the metadata is the name, and the Chinese semantic information B' of the metadata is segmented to obtain the first participle of the Chinese semantic information B' of the metadata as the name. The Chinese semantic information C' of the metadata is the identification code of the acquiring institution, and the Chinese semantic information C' of the metadata is segmented, and the first participle of the Chinese semantic information C' of the metadata is obtained as the acquirer, the institution, and the identification code. Therefore, the Chinese word segmentation set of metadata is: {issuer, institution, identification code, acquirer, name}.

Then, according to the Chinese word segmentation set of the metadata and the Chinese semantic information of the metadata, the first feature vector of the metadata is determined. For example, if there are metadata A', metadata B', and metadata C', the Chinese word segmentation set of metadata is {issuing card, organization, identification code, receipt, name}, and the Chinese semantic information of metadata A' is the identification code of the card issuer, the Chinese semantic information B' of the metadata is the name, and the Chinese semantic information C' of the metadata is the identification code of the acquiring institution. According to any first participle, it is determined that the first participle exists in the Chinese participle set, then the sub-feature vector corresponding to the first participle is set as the first value; if it is determined that the first participle does not exist in the Chinese participle set, then Set the sub-feature vector corresponding to the first participle as the second value, where the first value is 1 and the second value is 0 as an example, and then according to the position of each first participle in the Chinese semantic information of the metadata, each second participle The sub-feature vectors corresponding to a participle are spliced to obtain the first feature vector corresponding to the Chinese semantic information of the metadata. Therefore, the first feature vector of the metadata A' is {1, 1, 1, 0, 0}, and the metadata The first feature vector of B' is {0, 0, 0, 0, 1}, and the first feature vector of metadata C' is {0, 1, 1, 1, 0}.

Step 103, according to the English word segmentation set, determine the second feature vector corresponding to the English semantic information of the metadata.

In the embodiment of the present invention, according to the English semantic information of the metadata, the English word segmentation set of the metadata is determined. For example, if there are metadata A', metadata B', and metadata C', where the English semantic information A' of the metadata is ISS_INS_ID_CD, the English semantic information A' of the metadata is segmented to obtain the The second participles of English semantic information A' are ISS, INS, ID and CD respectively. The English semantic information B' of the metadata is NAME, and the English semantic information B' of the metadata is segmented to obtain the second participle of the English semantic information B' of the metadata respectively being NAME, and the English semantic information C' of the metadata is ACQ_INS_CD , segment the English semantic information C' of the metadata into words, and obtain the second word segmentation of the English semantic information C' of the metadata as ACQ, INS, and CD, respectively. Therefore, the English word segmentation set of metadata is: {ISS, INS, ID, CD, ACQ, NAME}. Then, according to the English word segmentation set of the metadata and the English semantic information of the metadata, the second feature vector of the metadata is determined. For example, if there is metadata A', metadata B', and metadata C', the English word segmentation set of metadata is: {ISS, INS, ID, CD, ACQ, NAME}, and the English semantic information A of metadata ' is ISS_INS_ID_CD, metadata English semantic information B' is NAME, metadata English semantic information C' is ACQ_INS_CD, according to any second participle, if it is determined that the second participle exists in the English participle set, then set the second participle The corresponding sub-characteristic vector is the first value; if it is determined that the second participle does not exist in the English participle set, then the sub-characteristic vector corresponding to the second participle is set as the second value, wherein the first value is 1, and the second participle The value is 0 as an example, and then according to the position of each second participle in the English semantic information of the metadata, the sub-feature vectors corresponding to each second participle are spliced to obtain the second feature vector corresponding to the English semantic information of the metadata. Therefore, the second eigenvector of metadata A' is {1, 1, 1, 1, 0, 0}, and the second eigenvector of metadata B' is {0, 0, 0, 0, 0, 1}, The second feature vector of metadata C' is {0, 1, 1, 1, 1, 0}. In order to facilitate the understanding of this solution, the implementation of step 102 before step 103 is used as an example. It is a possible implementation of step 102 before step 103, and it is also a possible implementation of step 102 after step 103. Parallel execution of step 102 and step 103 is also a possible implementation manner, which is not limited here.

Step 104, concatenating the first feature vector and the second feature vector to obtain a feature encoding vector of metadata.

In the embodiment of the present invention, the first feature vector of the metadata and the second feature vector of the metadata are spliced into a feature encoding vector of the metadata. For example, if the first feature vector of metadata A' is {1, 1, 1, 0, 0}, the second feature vector of metadata A' is {1, 1, 1, 1, 0, 0} , concatenate the first feature vector of metadata A' and the second feature vector of metadata A', and obtain the feature encoding vector of metadata A' as {1, 1, 1, 0, 0, 1, 1, 1 , 1, 0, 0}.

For another example, if the first feature vector of metadata B' is {0, 0, 0, 0, 1}, the second feature vector of metadata B' is {0, 0, 0, 0, 0, 1 }, concatenate the first feature vector of metadata B' and the second feature vector of metadata B', and obtain the feature encoding vector of metadata B' as {0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1}.

For another example, if the first feature vector of metadata C' is {0, 1, 1, 1, 0}, the second feature vector of metadata C' is {0, 1, 1, 1, 1, 0 }, concatenate the first feature vector of metadata C' and the second feature vector of metadata C', and obtain the feature encoding vector of metadata C' as {0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 0}.

In step 105, the feature encoding vectors of the metadata are respectively input into each label classifier to determine the label category of the metadata.

In the embodiment of the present invention, N label categories are set, which may be preset N label categories, or N label categories set according to specific situations, which is not limited here. For example, if N is 5, the five tag categories are institution information category, merchant information category, user information category, terminal information category and transaction information category. Determine five label classifiers according to five label categories, among which the five label classifiers are institution information label classifier, merchant information label classifier, user information label classifier, terminal information label classifier and transaction information label classifier . The feature encoding vectors of the metadata are respectively input into the five label classifiers, and the corresponding prediction probabilities of the five label classifiers can be obtained.

For example, if the feature encoding vector of metadata A' is {1, 1, 1, 0, 0, 1, 1, 1, 1, 0, 0}, there are institutional information label classifier and merchant information label classifier , user information tag classifier, terminal information tag classifier and transaction information tag classifier, these five tag classifiers input the metadata A' into the above five tag classifiers respectively, and the output is the organization information tag classifier corresponding The prediction probability corresponding to the merchant information label classifier is 90%, the corresponding prediction probability of the merchant information label classifier is 20%, the corresponding prediction probability of the user information label classifier is 21%, the corresponding prediction probability of the terminal information label classifier is 22%, and the transaction information label classification The predictor corresponds to a predicted probability of 19%. Determine the label category corresponding to the label classifier corresponding to the maximum predicted probability as the label category of metadata A'. Since 90%>22%>21%>20%>19%, therefore, the label category of metadata A' for institutional information.

From the above steps 101 to 105, it can be seen that by segmenting the metadata, determining the feature encoding vector of the metadata, and then inputting the feature encoding vector of the metadata into each label classifier, the label corresponding to the maximum value of the predicted probability The label category of the classifier is used as the label category of the metadata, so as to realize more accurate determination of the label category of the metadata and improve the efficiency of determining the label category of the metadata.

Since the present application inputs the feature encoding vectors of the metadata into each label classifier, the label category of the metadata is quickly determined, thereby improving the classification efficiency of the metadata. Wherein, by training and optimizing each label classifier, the accuracy of determining the label category of the metadata can be improved. The following describes how to train and optimize each label classifier.

In order, the training samples can be input into each label classifier, so that each label classifier can be better trained and optimized, and the test metadata can be input into each optimized label classifier later, so that the metadata can be determined more accurately The label category of is only to improve the classification efficiency of metadata. The training samples are the label categories to which the manually judged metadata belongs. It can be understood that the label categories to which the training samples belong are known, and the training samples are used to train and optimize each label classifier.

Input the training samples to each label classifier, and determine the upper threshold and lower threshold of each label classifier. For example, if the training samples are metadata A', metadata B', metadata C', metadata D', metadata E', and metadata F', the label classifiers are organization information label classifier, Merchant information tag classifier, user information tag classifier, terminal information tag classifier and transaction information tag classifier. Among them, the label category of metadata A' is organization information, the label category of metadata B' is organization information, the label category of metadata C' is user information, the label category of metadata D' is merchant information, and the label category of metadata E The label category of 'is terminal information, the label category of metadata F' is transaction information, input metadata A' and metadata B' into the organization information label classifier, and iterate N times, wherein, when the number of iterations is 1 , the metadata A' and metadata B' will be input into the institution information label classifier, and then the upper limit threshold and the lower limit threshold of the institution information label classifier will be re-determined according to the accuracy of the results output by the institution information label classifier , and update the organization information label classifier according to the re-determined upper threshold and lower limit threshold. In the next iteration, metadata A' and metadata B' are input into the updated organization information label classifier, and then continue to be based on the organization information The accuracy of the results output by the label classifier is used to re-determine the upper threshold and lower threshold of the organization information label classifier, and update the organization information label classifier according to the re-determined upper threshold and lower threshold. After N iterations, the predicted probability that the label category of metadata A' is institutional information and the predicted probability that the label category of metadata B' is institutional information will be output, where the upper threshold and lower threshold of the institutional information label classifier are It is used to determine whether metadata belongs to the category of institutional information tags. If the upper limit threshold of the organization information label classifier is determined to be 85% and the lower limit threshold is 20% according to the training samples, then if the metadata M is subsequently input into the organization information label classifier, and after N iterations, the output organization information label The prediction probability corresponding to the classifier is 19%, and since 19% is less than the lower threshold of 20% of the institution information label classifier, it is determined that the metadata M does not belong to the institution information label category. If the output prediction probability is 90%, since 90% is greater than the upper threshold 85% of the institution information label classifier, it is determined that the metadata M belongs to the institution information label category.

Input the metadata C' and into the user information label classifier, and iterate N times, wherein, when the number of iterations is 1, the metadata C' will be input into the user information label classifier, and then classified according to the user information label The accuracy rate of the result output by the filter to re-determine the upper threshold and lower threshold of the user information label classifier, and update the user information label classifier according to the re-determined upper threshold and lower threshold. In the next iteration, the metadata C' Input to the updated user information label classifier, and then continue to re-determine the upper and lower thresholds of the user information label classifier according to the accuracy of the output results of the user information label classifier, and according to the re-determined upper threshold and The lower threshold updates the user information label classifier. After N iterations, the predicted probability that the label category of metadata C' is user information will be output, where the upper threshold and lower threshold of the user information label classifier are used to determine whether the metadata belongs to the user information label category. If the upper limit threshold of the user information label classifier is 90%, and the lower limit threshold is 25%, then when the subsequent test metadata N is input into the user information label classifier, if the predicted probability corresponding to the output user information label classifier is 19 %, since 19% is less than the lower limit threshold 25% of the user information label classifier, it is determined that the metadata N does not belong to the user information label category. If the output prediction probability is 91%, since 91% is greater than the upper threshold 90% of the user information label classifier, it is determined that the metadata N belongs to the user information label category.

Input the metadata D' and into the business information label classifier, and iterate N times, wherein, when the number of iterations is 1, the metadata D' will be input into the business information label classifier, and then classified according to the business information label The accuracy rate of the results output by the device to re-determine the upper threshold and lower threshold of the merchant information label classifier, and update the merchant information label classifier according to the re-determined upper threshold and lower threshold. In the next iteration, the metadata D' Input to the updated business information label classifier, and then continue to re-determine the upper limit threshold and lower limit threshold of the business information label classifier according to the accuracy of the results output by the business information label classifier, and according to the newly determined upper threshold and The lower threshold updates the business information label classifier. After N iterations, the predicted probability that the label category of the metadata D' is merchant information will be output, where the upper threshold and lower threshold of the merchant information label classifier are used to determine whether the metadata belongs to the merchant information label category. If the upper limit threshold of the merchant information label classifier is 80%, and the lower limit threshold is 15%, then when the subsequent metadata K is input into the merchant information label classifier, if the predicted probability corresponding to the output merchant information label classifier is 13% , since 13% is less than the lower threshold 15% of the merchant information label classifier, it is determined that the metadata K does not belong to the merchant information label category. If the output predicted probability is 85%, since 85% is greater than the upper threshold 80% of the merchant information label classifier, it is determined that the metadata K belongs to the merchant information label category.

Input the metadata E' into the terminal information label classifier, and iterate N times, wherein, when the number of iterations is 1, the metadata E' will be input into the terminal information label classifier, and then classified according to the terminal information label The accuracy rate of the results output by the device to re-determine the upper threshold and lower threshold of the terminal information label classifier, and update the terminal information label classifier according to the re-determined upper threshold and lower threshold. In the next iteration, the metadata E' Input to the updated terminal information label classifier, and then continue to re-determine the upper limit threshold and lower limit threshold of the terminal information label classifier according to the accuracy of the output result of the terminal information label classifier, and according to the re-determined upper limit threshold and The lower threshold updates the terminal information label classifier. After N iterations, the predicted probability that the label category of the metadata E' is terminal information will be output. Wherein, the upper threshold and the lower threshold of the terminal information label classifier are used to determine whether the metadata belongs to the terminal information label category. If the upper limit threshold of the terminal information label classifier is 87%, and the lower limit threshold is 23%, then when the subsequent metadata L is input into the terminal information label classifier, if the output terminal information label classifier corresponds to a prediction probability of 18% , since 18% is less than the lower limit threshold 23% of the terminal information label classifier, it is determined that the metadata L does not belong to the terminal information label category. If the output prediction probability is 90%, since 90% is greater than the upper threshold 87% of the terminal information label classifier, it is determined that the metadata L belongs to the terminal information label category.

Input the metadata F' and into the transaction information label classifier, and iterate N times, wherein, when the number of iterations is 1, the metadata F' will be input into the transaction information label classifier, and then classified according to the transaction information label The accuracy rate of the result output by the filter to re-determine the upper threshold and lower threshold of the transaction information label classifier, and update the transaction information label classifier according to the re-determined upper threshold and lower threshold. In the next iteration, the metadata F' Input to the updated transaction information label classifier, and then continue to re-determine the upper and lower thresholds of the transaction information label classifier according to the accuracy of the output results of the transaction information label classifier, and according to the re-determined upper threshold and The lower bound threshold updates the transaction information label classifier. After N iterations, the tag category of metadata F' will be output as the predicted probability of transaction information. According to the tag category of metadata F' is the predicted probability of transaction information, the upper threshold and lower threshold of the transaction information label classifier are determined. Wherein, the upper threshold and the lower threshold of the transaction information label classifier are used to determine whether the metadata belongs to the transaction information label category. If the upper limit threshold of the transaction information label classifier is 88%, and the lower limit threshold is 25%, then when the subsequent test metadata W is input into the transaction information label classifier, if the output transaction information label classifier corresponds to a prediction probability of 20 %, since 20% is less than the lower limit threshold 25% of the transaction information label classifier, it is determined that the metadata W does not belong to the transaction information label category. If the output prediction probability is 91%, since 91% is greater than the upper threshold 88% of the transaction information label classifier, it is determined that the test metadata W belongs to the transaction information label category.

After each tag classifier is trained and optimized, the tag category of the metadata can be determined by inputting the metadata into each tag classifier. The following describes how to determine the label category of metadata.

In a possible situation, since the number and types of training samples are large enough to meet the conditions for training each label classifier, the upper and lower thresholds of each label classifier can be determined more accurately through the training samples. Then the metadata is input into each trained label classifier, and each label classifier will output the corresponding prediction probability of each label classifier. For example, if there are 5 label classifiers in total, including the organization information label classifier, merchant information label classifier, user information label classifier, terminal information label classifier and transaction information label classifier, by matching the training samples to Input into these 5 label classifiers, so as to realize the determination of the upper limit threshold and lower limit threshold of these 5 label classifiers, among them, the upper limit threshold of the organization information label classifier is 85%, and the lower limit threshold of the organization information label classifier is 20%. The upper threshold of the merchant information tag classifier is 90%, and the lower threshold of the merchant information tag classifier is 25%. The upper threshold of the user information label classifier is 86%, and the lower threshold of the user information label classifier is 19%. The upper limit threshold of the terminal information label classifier is 91%, and the lower limit threshold of the terminal information label classifier is 23%. The upper threshold of the transaction information label classifier is 95%, and the lower threshold of the transaction information label classifier is 30%. If the metadata Z is input into the five trained label classifiers, five predicted probabilities will be output, which are 90% of the predicted probability corresponding to the organization information label classifier and 90% of the predicted probability corresponding to the merchant information label classifier. is 70%, the prediction probability corresponding to the user information label classifier is 60%, the prediction probability corresponding to the terminal information label classifier is 55%, and the prediction probability corresponding to the transaction information label classifier is 55%. The label category of the label classifier with the largest value among the five predicted probabilities is taken as the label category of metadata Z, therefore, the label category of metadata Z is determined as institutional information.

In another possible situation, since the number and types of training samples are not enough to meet the conditions for training each label classifier, then the accuracy rate of the upper threshold and lower threshold of each label classifier determined only by training samples lower. Therefore, it is necessary to improve the accuracy rate of the first upper threshold and the lower threshold of each label classifier, so as to achieve a more accurate determination of the label category of the metadata through each label classifier. The following describes how to improve the accuracy of the upper and lower thresholds for each label classifier.

In order to better understand the solution, take the tag classifier as an example, which includes an organization information tag classifier, a merchant information tag classifier, a user information tag classifier, a terminal information tag classifier, and a transaction information tag classifier, to determine whether the metadata Belonging to the institutional information label classifier as an example to describe how to improve the accuracy of the upper threshold and lower threshold of each label classifier, so that those skilled in the art can understand, and does not limit whether the metadata in this application belongs to the institutional information label classifier .

As shown in Figure 2, it is a flow chart of a method for improving the accuracy of the upper threshold and lower threshold of each label classifier provided by the embodiment of the present invention, the method includes the following steps:

Step 201, input metadata into each label classifier.

In the embodiment of the present invention, by inputting the training sample into the corresponding label classifier, the upper limit threshold and the lower limit threshold of each label classifier are preliminarily determined, so as to obtain each trained label classifier, and then the metadata are respectively input into each in the label classifier.

Step 202, judging whether the number of iterations is greater than N times, if yes, execute step 203, if not, execute step 204.

In the embodiment of the present invention, after the metadata is input into each label classifier, multiple iterations are required, each iteration will output the predicted probability corresponding to the metadata in each label classifier, and by inputting the metadata into each label classifier The number of iterations in the middle is greater than N times, and the prediction probability corresponding to each label classifier of the metadata can be output more accurately.

In step 203, each tag classifier outputs the predicted probability corresponding to each tag classifier, and the tag category of the metadata is determined according to the predicted probability corresponding to each tag classifier.

In the embodiment of the present invention, when the number of iterations is greater than N times, the iteration ends, and each label classifier will output the predicted probability corresponding to each label classifier, and then determine the metadata according to the magnitude relationship of the predicted probability corresponding to each label classifier. label category.

For example, the tag classifiers are institution information tag classifier, merchant information tag classifier, user information tag classifier, terminal information tag classifier and transaction information tag classifier. Among them, the upper limit threshold of the institution information label classifier is 87%, and the lower limit threshold of the institution information label classifier is 21%. The upper threshold of the merchant information tag classifier is 91%, and the lower threshold of the merchant information tag classifier is 24%. The upper threshold of the user information label classifier is 85%, and the lower threshold of the user information label classifier is 18%. The upper threshold of the terminal information label classifier is 92%, and the lower threshold of the terminal information label classifier is 24%. The upper threshold of the transaction information label classifier is 96%, and the lower threshold of the transaction information label classifier is 26%. If the metadata U is respectively input into the five trained label classifiers, after more than N iterations, five predicted probabilities will be output correspondingly, which are respectively 90% of the predicted probability corresponding to the organization information label classifier and 90% of the merchant information label The prediction probability corresponding to the classifier is 70%, the prediction probability corresponding to the user information label classifier is 60%, the prediction probability corresponding to the terminal information label classifier is 55%, and the prediction probability corresponding to the transaction information label classifier is 55%. The label category of the label classifier with the largest value among the five predicted probabilities is used as the label category of metadata U, because the label category corresponding to the organization information label classifier is organization information, therefore, the label category of metadata U is determined to be organization information .

Step 204, judge whether the predicted probability corresponding to the organization information classifier is greater than the upper limit threshold of the organization information classifier, and judge whether the predicted probability corresponding to the other label classifiers is less than the lower limit threshold of the other label classifiers, if so, then perform step 205, if If not, go to step 206.

In the embodiment of the present invention, for example, if the upper limit threshold of the organization information tag classifier is 87%, the lower limit threshold of the organization information tag classifier is 21%. The upper threshold of the merchant information tag classifier is 91%, and the lower threshold of the merchant information tag classifier is 24%. The upper threshold of the user information label classifier is 85%, and the lower threshold of the user information label classifier is 18%. Input the metadata R into each label classifier respectively. After iterating M times, where M is less than N, the prediction probability corresponding to the organization information classifier is 98%, and the prediction probability corresponding to the user information classifier is 10%. The prediction probability corresponding to the merchant information classifier is 9%, the prediction probability corresponding to the terminal information classifier is 8%, and the prediction probability corresponding to the transaction information classifier is 5%. Since the predicted probability corresponding to the organization information classifier is 98% greater than the upper threshold of the organization information classifier, and the predicted probability corresponding to the user information classifier is 10% lower than the lower threshold of the user information classifier, the corresponding predicted probability of the merchant information classifier 10% is less than the lower limit threshold of the user information classifier, the corresponding prediction probability of the terminal information classifier is 10% less than the lower limit threshold of the user information classifier, and the corresponding prediction probability of the transaction information classifier is 10% less than the lower limit of the user information classifier threshold, then it can be determined that the label category of metadata R is institutional information.

Step 205, determine that the label category of the metadata is the category of institution information, and use the metadata as a training sample.

In the embodiment of the present invention, since the label category of the metadata has been determined before the number of iterations is less than N, the metadata can be used as a training sample to facilitate the subsequent updating of the upper threshold and lower threshold of each label classifier based on the training samples , to obtain the updated label classifiers, so as to improve the accuracy of the label category of the subsequent metadata.

Step 206 , judging whether the predicted probability corresponding to the institution information classifier is smaller than the lower limit threshold of the institution information classifier, if yes, execute step 207 , if not, execute step 201 .

In the embodiment of the present invention, since there are three possibilities for the label category of metadata, the first possibility is that the label category of metadata is organization information, the second possibility is that the label category of metadata must not be organization information, and the third possibility is Metadata may or may not be institutional information, and the tag category of metadata is temporarily uncertain. Through step 204, it can be determined that the metadata does not conform to the first possibility, then it needs to be determined according to step 206 whether the metadata conforms to the second possibility or the third possibility. For example, if the upper threshold of the institution information label classifier is 87%, the lower limit threshold of the institution information label classifier is 21%. The upper threshold of the merchant information tag classifier is 91%, and the lower threshold of the merchant information tag classifier is 24%. The upper threshold of the user information label classifier is 85%, and the lower threshold of the user information label classifier is 18%. Input the metadata H into each label classifier respectively. After M iterations, where M is less than N, the prediction probability corresponding to the organization information classifier is 13%, and the prediction probability corresponding to the user information classifier is 70%. The prediction probability corresponding to the merchant information classifier is 74%, the prediction probability corresponding to the terminal information classifier is 75%, and the prediction probability corresponding to the transaction information classifier is 80%. Since the predicted probability corresponding to the institution information classifier is 13% less than the lower limit threshold of the institution information classifier, it means that the label category of the metadata H is not institution information.

For another example, if the upper limit threshold of the institution information label classifier is 87%, the lower limit threshold of the institution information label classifier is 21%. The upper threshold of the merchant information tag classifier is 91%, and the lower threshold of the merchant information tag classifier is 24%. The upper threshold of the user information label classifier is 85%, and the lower threshold of the user information label classifier is 18%. Input the metadata P into each label classifier respectively. After M iterations, where M is less than N, the prediction probability corresponding to the organization information classifier is 25%, and the prediction probability corresponding to the user information classifier is 70%. The prediction probability corresponding to the merchant information classifier is 74%, the prediction probability corresponding to the terminal information classifier is 75%, and the prediction probability corresponding to the transaction information classifier is 80%. Since the predicted probability corresponding to the institution information classifier is 25% greater than the lower limit threshold of the institution information classifier and smaller than the upper limit threshold of the institution information classifier, the label category of metadata P cannot be determined, and the label category of metadata P may be institution The information may not be institutional information, and the label category of the metadata cannot be determined for the time being. Since the number of iterations is less than N, the metadata is re-input into each label classifier, so as to achieve a more accurate subsequent determination of the label category of the metadata.

In step 207, it is determined that the label category of the metadata is not an institution information category, and the metadata is used as a training sample.

In the embodiment of the present invention, since it has been determined that the label category of the metadata must not be institutional information before the number of iterations is less than N, the metadata can be used as a training sample, which is convenient for the subsequent upper limit threshold and The lower limit threshold is updated to obtain updated label classifiers, so as to improve the accuracy of subsequent determination of label categories of metadata.

From the above steps 201 to 207, it can be seen that in the iterative process of inputting metadata into each label classifier, by determining part of the metadata as training samples according to the above judgment conditions, the training samples of each label classifier are increased, thereby Realize updating the upper limit threshold and lower limit threshold of each label classifier, and then realize improving the accuracy rate of the label category of subsequent determination of metadata.

Based on the same technical idea as above, the embodiment of the present invention also provides a metadata label classification device, which can execute the method in the above method embodiment of the present invention. The structure of a tag classification device for metadata provided by an embodiment of the present invention can be seen in FIG. 3 . The device 300 includes: an acquisition unit 301 for acquiring Chinese semantic information of metadata and English semantic information of metadata for any metadata. information. The processing unit 302 is used to determine the first feature vector corresponding to the Chinese semantic information of the metadata according to the Chinese word segmentation set; the Chinese word segmentation set is obtained by segmenting the Chinese semantic information of each metadata; according to the English word segmentation set, determine the element The second feature vector corresponding to the English semantic information of the data; the English word segmentation set is obtained by segmenting the English semantic information of each metadata; splicing the first feature vector and the second feature vector to obtain the feature encoding vector of the metadata; The feature encoding vectors of metadata are respectively input into each label classifier to determine the label category of metadata.

Optionally, the processing unit 302 is specifically configured to: segment the Chinese semantic information of the metadata to obtain each first segment; for any first segment, if it is determined that the first segment exists in the Chinese segment set, set the first The sub-feature vector corresponding to the word segmentation is the first value; if it is determined that the first word segmentation does not exist in the Chinese word segmentation set, then set the sub-feature vector corresponding to the first word segmentation as the second value; according to the Chinese semantics of each first segmentation word in the metadata position in the information, the sub-feature vectors corresponding to each first participle are concatenated to obtain the first feature vector corresponding to the Chinese semantic information of the metadata.

Optionally, the processing unit 302 is specifically configured to: segment the English semantic information of the metadata to obtain each second segment; for any second segment, if it is determined that the second segment exists in the English segment set, set the second segment The sub-feature vector corresponding to the word segmentation is the first value; if it is determined that the second word segmentation does not exist in the English word segmentation set, then set the sub-feature vector corresponding to the second word segmentation as the second value; according to the English semantics of each second word segmentation in the metadata position in the information, the sub-feature vectors corresponding to the second participle are concatenated to obtain the second feature vector corresponding to the English semantic information of the metadata.

Optionally, the processing unit 302 is specifically configured to: input the feature encoding vector of the metadata into each label classifier, and obtain the corresponding prediction probability in each label classifier; according to the corresponding prediction probability of each label classifier , to determine the label category of the metadata.

Optionally, each label classifier is obtained by training through training samples; each label classifier has an upper threshold and a lower threshold obtained through training, and the processing unit 302 is specifically configured to: if the first prediction corresponding to the first label classifier The probability is higher than the upper limit threshold of the first label classifier, and the second predicted probability corresponding to the second label classifier is lower than the lower limit threshold of the second label classifier, then it is determined that the metadata has a label category corresponding to the first label classifier; The second label classifier is each label classifier except the first label classifier among the various label classifiers.

Optionally, the processing unit 302 is specifically configured to: for any label classifier, if the predicted probability of the label classifier is lower than the lower limit threshold of the label classifier, determine that the metadata does not have a label category corresponding to the label classifier, and set Metadata serve as training samples for updating classifiers for each label.

Optionally, the processing unit 302 is specifically configured to: use the metadata as a training sample for updating each label classifier.

Optionally, the processing unit 302 is specifically configured to: if there is no predicted probability corresponding to any label classifier higher than the upper threshold of the label classifier, after each label classifier is updated, continue to convert the feature encoding vector of the metadata to Input the updated label classifiers respectively to determine the label category of the metadata.

Based on the same technical concept, the embodiment of the present application also provides a computing device 400, as shown in FIG. As for the specific connection medium between the processor 401 and the memory 402, the bus connection between the processor 401 and the memory 402 in FIG. 4 is taken as an example. The bus can be divided into address bus, data bus, control bus and so on.

In the embodiment of the present application, the memory 402 stores instructions executable by at least one processor 401, and at least one processor 401 executes the instructions stored in the memory 402 to perform the steps included in the aforementioned metadata label classification method .

Among them, the processor 401 is the control center of the computing device, which can use various interfaces and lines to connect various parts of the computing device, by running or executing instructions stored in the memory 402 and calling data stored in the memory 402, thereby realizing data deal with. Optionally, the processor 801 may include one or more processing units, and the processor 401 may integrate an application processor and a modem processor. The call processor mainly handles issuing instructions. It can be understood that the foregoing modem processor may not be integrated into the processor 401 . In some embodiments, the processor 401 and the memory 402 can be implemented on the same chip, and in some embodiments, they can also be implemented on independent chips.

The processor 401 may be a general processor, such as a central processing unit (CPU), a digital signal processor, an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array or other programmable logic devices, discrete gates or transistors Logic devices and discrete hardware components can implement or execute the methods, steps and logic block diagrams disclosed in the embodiments of the present application. A general purpose processor may be a microprocessor or any conventional processor or the like. The steps of the method disclosed in the embodiment of the label classification method combined with metadata can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

The memory 402, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs and modules. The memory 402 may include at least one type of storage medium, for example, may include flash memory, hard disk, multimedia card, card-type memory, random access memory (Random Access Memory, RAM), static random access memory (Static Random Access Memory, SRAM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Magnetic Memory, Disk, discs and more. Memory 402 is, but is not limited to, any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. The memory 402 in the embodiment of the present application may also be a circuit or any other device capable of implementing a storage function, and is used for storing program instructions and/or data.

Based on the same technical idea, an embodiment of the present application also provides a computer-readable storage medium, which stores a computer program executable by a computing device, and when the program is run on the computing device, the computing device The steps of the tag classification method for metadata described above are performed.

Those skilled in the art should understand that the embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to the present application. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow chart or blocks of the flowchart and/or the block or blocks of the block diagrams.

While preferred embodiments of the present application have been described, additional changes and modifications to these embodiments can be made by those skilled in the art once the basic inventive concept is appreciated. Therefore, the appended claims are intended to be construed to cover the preferred embodiment and all changes and modifications which fall within the scope of the application.

Obviously, those skilled in the art can make various changes and modifications to the application without departing from the spirit and scope of the application. In this way, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims (11)

1. A tag classification method for metadata, comprising: For any metadata, obtain the Chinese semantic information of the metadata and the English semantic information of the metadata; According to the Chinese word segmentation set, determine the first feature vector corresponding to the Chinese semantic information of the metadata; the Chinese word segmentation set is obtained by segmenting the Chinese semantic information of each metadata; Determine the second feature vector corresponding to the English semantic information of the metadata according to the English word segmentation set; the English word segmentation set is obtained by segmenting the English semantic information of each metadata; splicing the first feature vector and the second feature vector to obtain a feature encoding vector of the metadata; The feature encoding vectors of the metadata are respectively input into each label classifier to determine the label category of the metadata. 2. The method according to claim 1, wherein the first feature vector corresponding to the Chinese semantic information of the metadata is determined according to the Chinese word segmentation set, comprising: Segmenting the Chinese semantic information of the metadata to obtain each first participle; For any first participle, if it is determined that the first participle exists in the Chinese participle set, then the sub-feature vector corresponding to the first participle is set as the first value; if it is determined that the first participle does not exist in In the Chinese word segmentation set, the sub-feature vector corresponding to the first word segmentation is set as the second value; According to the position of each first participle in the Chinese semantic information of the metadata, the sub-feature vectors corresponding to each first participle are spliced to obtain the first feature vector corresponding to the Chinese semantic information of the metadata. 3. The method according to claim 1, wherein the second feature vector corresponding to the English semantic information of the metadata is determined according to the English word segmentation set, comprising: Segmenting the English semantic information of the metadata to obtain each second participle; For any second participle, if it is determined that the second participle exists in the English participle set, then the sub-feature vector corresponding to the second participle is set as the first value; if it is determined that the second participle does not exist in In the English participle set, the sub-feature vector corresponding to the second participle is set as the second value; According to the position of each second participle in the English semantic information of the metadata, the sub-feature vectors corresponding to each second participle are spliced to obtain the second feature vector corresponding to the English semantic information of the metadata. 4. The method according to any one of claims 1 to 3, wherein said inputting the feature encoding vectors of said metadata into respective label classifiers to determine the label category of said metadata comprises: Input the feature encoding vector of the metadata into each label classifier, and obtain the corresponding prediction probability in each label classifier; The label category of the metadata is determined according to the prediction probabilities corresponding to the respective label classifiers. 5. The method according to claim 4, wherein each label classifier is trained by training samples; each label classifier has an upper threshold and a lower threshold obtained through training; The determining the label category of the metadata according to the prediction probability corresponding to each label classifier includes: If the first predicted probability corresponding to the first label classifier is higher than the upper threshold of the first label classifier, and the second predicted probability corresponding to the second label classifier is lower than the lower threshold of the second label classifier, Then it is determined that the metadata has a tag category corresponding to the first tag classifier; and the second tag classifier is each tag classifier except the first tag classifier among the tag classifiers. 6. The method according to claim 5, further comprising: for any label classifier, if the predicted probability of the label classifier is lower than the lower limit threshold of the label classifier, then determining the The data does not have a label category corresponding to the label classifier, and the metadata is used as a training sample for updating each label classifier. 7. The method according to claim 5, further comprising: The metadata are used as training samples for updating the respective label classifiers. 8. The method according to claim 5, further comprising: if there is no predicted probability corresponding to any label classifier higher than the upper threshold of the label classifier, performing After the update, continue to input the feature encoding vectors of the metadata into the updated label classifiers respectively to determine the label category of the metadata. 9. A tag classification device for metadata, comprising: An acquisition unit, configured to acquire the Chinese semantic information of the metadata and the English semantic information of the metadata for any metadata; The processing unit is used to determine the first feature vector corresponding to the Chinese semantic information of the metadata according to the Chinese word segmentation set; the Chinese word segmentation set is obtained by segmenting the Chinese semantic information of each metadata; according to the English word segmentation set , determine the second feature vector corresponding to the English semantic information of the metadata; the English word segmentation set is obtained by segmenting the English semantic information of each metadata; combine the first feature vector with the second feature Concatenating the vectors to obtain the feature encoding vectors of the metadata; inputting the feature encoding vectors of the metadata into respective label classifiers to determine the label categories of the metadata. 10. A computing device, characterized in that it comprises at least one processor and at least one memory, wherein the memory stores a computer program that, when the program is executed by the processor, causes the processor to execute the rights The method according to any one of claims 1 to 8. 11. A computer-readable storage medium, wherein the storage medium stores a program, and when the program is run on a computer, the computer is enabled to implement the method according to any one of claims 1 to 8.

Images