CN113255357A

CN113255357A - Data processing method, target recognition model training method, target recognition method and device

Info

Publication number: CN113255357A
Application number: CN202110701378.4A
Authority: CN
Inventors: 王得贤; 李长亮; 毛璐
Original assignee: Beijing Kingsoft Software Co Ltd
Current assignee: Beijing Kingsoft Software Co Ltd; Beijing Kingsoft Digital Entertainment Co Ltd
Priority date: 2021-06-24
Filing date: 2021-06-24
Publication date: 2021-08-13

Abstract

The application provides a data processing method, a target recognition model training method and a target recognition method and device, wherein the data processing method comprises the following steps: training a plurality of different target recognition models based on the acquired training set, then respectively performing target recognition on each training text in the training set by using each target recognition model obtained by training to obtain a target recognition result of each training text, and if the target recognition result of the training text is inconsistent with the label information of the training text aiming at any training text, indicating that the label information of the training text is inaccurate, so that the label information of the training text in the training set is updated, and the data noise of the label information of the training text in the training set is removed.

Description

Data processing method, target recognition model training method, target recognition method and device

Technical Field

The present application relates to the field of computer technologies, and in particular, to a data processing method and apparatus, a target recognition model training method and apparatus, a target recognition method and apparatus, a computing device, and a computer-readable storage medium.

Background

In natural language processing tasks, the demand of target recognition tasks is increasing, such as entity recognition, which is also called "proper name recognition", refers to recognizing entities with specific meanings in texts, and mainly includes names of people, places, organizations, proper nouns, and the like.

In the current target recognition task, a deep learning method is usually adopted, specifically, a text to be recognized is input into a target recognition model obtained by pre-training, and the target recognition model is an end-to-end neural network model and can directly output a target in the text to be recognized.

Therefore, the accuracy of the target recognition model directly influences the accuracy of the recognition result, the target recognition model is trained based on massive training texts with labeled information in the training set, and the more the number of the training texts in the training set, the higher the accuracy of the trained target recognition model.

The labeling information of the training texts is generally labeled manually, however, as the number of the training texts in the training set is increased, the situations of manual label missing, label error and the like inevitably occur, so that certain data noise exists in the labeling information of the training texts in the training set, and the accuracy of the target recognition model is affected.

Disclosure of Invention

In view of the above, embodiments of the present application provide a data processing method and apparatus, a target recognition model training method and apparatus, a target recognition method and apparatus, a computing device, and a computer-readable storage medium, so as to solve technical defects in the prior art.

According to a first aspect of embodiments of the present application, there is provided a data processing method, including:

acquiring a training set, wherein the training set comprises a plurality of training texts;

training to obtain different target recognition models based on the training set;

respectively carrying out target recognition on each training text in the training set by using each target recognition model obtained by training to obtain a target recognition result of each training text;

and aiming at each training text, under the condition that the target recognition result of the training text is inconsistent with the labeling information of the training text, updating the labeling information of the training text in the training set.

Optionally, the step of training to obtain different target recognition models based on the training set includes:

and training at least one preset neural network by using the training set to obtain different target recognition models.

Optionally, the step of training at least one preset neural network by using a training set to obtain different target recognition models includes:

sequentially selecting a verification subset and a training subset from a training set;

and training the same preset neural network by using the selected training subset every time, and verifying the training result by using the selected verification subset to obtain different target recognition models.

and aiming at different preset neural networks, training the preset neural networks by using the selected training subsets each time, and verifying the training results by using the selected verification subsets to obtain different target recognition models.

Optionally, for each training text, in a case that the target recognition result of the training text is inconsistent with the label information of the training text, the step of updating the label information of the training text in the training set includes:

for each training text, if the target recognition result of the training text comprises a target and the labeling information of the training text does not exist in the training set, adding the labeling information of the training text in the training set, wherein the added labeling information of the training text is the recognized target in the training text;

and for each training text, if the target recognition result of the training text does not comprise a target and the labeling information of the training text exists in the training set, deleting the labeling information of the training text from the training set.

According to a second aspect of the embodiments of the present application, there is provided a target recognition model training method, including:

acquiring a training set, wherein the training set is obtained after data processing is performed by using the method provided by the first aspect of the embodiment of the application;

and training the preset neural network by using the training set to obtain a target recognition model.

According to a third aspect of the embodiments of the present application, there is provided a target identification method, including:

acquiring a text to be identified;

and inputting the text to be recognized into the target recognition model obtained by training by using the method provided by the second aspect of the embodiment of the application, so as to obtain the target recognition result of the text to be recognized.

According to a fourth aspect of the embodiments of the present application, there is provided a data processing apparatus including:

a first obtaining module configured to obtain a training set, wherein the training set comprises a plurality of training texts;

the first model training module is configured to train different target recognition models based on a training set;

the first target recognition module is configured to perform target recognition on each training text in the training set by using each target recognition model obtained through training to obtain a target recognition result of each training text;

and the updating module is configured to update the labeling information of the training texts in the training set under the condition that the target recognition result of each training text is inconsistent with the labeling information of the corresponding training text.

Optionally, the first model training module is further configured to: and training at least one preset neural network by using the training set to obtain different target recognition models.

Optionally, the first model training module is further configured to: sequentially selecting a verification subset and a training subset from a training set; and training the same preset neural network by using the selected training subset every time, and verifying the training result by using the selected verification subset to obtain different target recognition models.

Optionally, the first model training module is further configured to: sequentially selecting a verification subset and a training subset from a training set; and aiming at different preset neural networks, training the preset neural networks by using the selected training subsets each time, and verifying the training results by using the selected verification subsets to obtain different target recognition models.

Optionally, the update module is further configured to: for each training text, if the target recognition result of the training text comprises a target and the labeling information of the training text does not exist in the training set, adding the labeling information of the training text in the training set, wherein the added labeling information of the training text is the recognized target in the training text; and for each training text, if the target recognition result of the training text does not comprise a target and the labeling information of the training text exists in the training set, deleting the labeling information of the training text from the training set.

According to a fifth aspect of the embodiments of the present application, there is provided a target recognition model training apparatus, including:

a second obtaining module, configured to obtain a training set, where the training set is a training set obtained by performing data processing by using the method provided in the first aspect of the embodiment of the present application;

and the second model training module is configured to train the preset neural network by using the training set to obtain the target recognition model.

According to a sixth aspect of embodiments of the present application, there is provided an object recognition apparatus, including:

the third acquisition module is configured to acquire a text to be recognized;

and the second target recognition module is configured to input the text to be recognized into the target recognition model obtained by training with the method provided by the second aspect of the embodiment of the application, so as to obtain the target recognition result of the text to be recognized.

According to a seventh aspect of embodiments of the present application, there is provided a computing device, including a memory, a processor, and computer instructions stored on the memory and executable on the processor, where the processor executes the computer instructions to implement the steps of the method provided in the first, second, or third aspects of embodiments of the present application.

According to an eighth aspect of embodiments of the present application, there is provided a computer-readable storage medium storing computer instructions which, when executed by a processor, implement the steps of the method provided by the first, second or third aspect of embodiments of the present application.

According to a ninth aspect of an embodiment of the present application, there is provided a chip storing computer instructions, which when executed by the chip, implement the steps of the method provided in the first, second or third aspect of the embodiment of the present application.

In the embodiment of the application, a plurality of different target recognition models are trained based on an acquired training set, then each target recognition model obtained by training is utilized to respectively perform target recognition on each training text in the training set, so as to obtain a target recognition result of each training text, if the target recognition result of each training text is inconsistent with the label information of each training text aiming at any training text, the label information of each training text is inaccurate, therefore, the label information of each training text in the training set is updated, and thus, the data noise of the label information of each training text in the training set is removed.

Drawings

Fig. 1 is a schematic flowchart of a data processing method provided in an embodiment of the present application;

FIG. 2 is a schematic flow chart diagram of another data processing method provided in the embodiments of the present application;

fig. 3 is a schematic flowchart of another data processing method provided in an embodiment of the present application;

FIG. 4 is a schematic flowchart of a method for training a target recognition model according to an embodiment of the present disclosure;

FIG. 5 is a schematic diagram of a target recognition model training process provided by an embodiment of the present application;

fig. 6 is a schematic flowchart of a target identification method according to an embodiment of the present application;

fig. 7 is a schematic structural diagram of a data processing apparatus according to an embodiment of the present application;

FIG. 8 is a schematic structural diagram of an apparatus for training a target recognition model according to an embodiment of the present disclosure;

fig. 9 is a schematic structural diagram of an object recognition apparatus according to an embodiment of the present application;

fig. 10 is a block diagram of a computing device according to an embodiment of the present disclosure.

Detailed Description

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. This application is capable of implementation in many different ways than those herein set forth and of similar import by those skilled in the art without departing from the spirit of this application and is therefore not limited to the specific implementations disclosed below.

The terminology used in the one or more embodiments of the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of the present application. As used in one or more embodiments of the present application and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used in one or more embodiments of the present application refers to and encompasses any and all possible combinations of one or more of the associated listed items.

It will be understood that, although the terms first, second, etc. may be used herein in one or more embodiments of the present application to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, a first aspect may be termed a second aspect, and, similarly, a second aspect may be termed a first aspect, without departing from the scope of one or more embodiments of the present application. The word "if," as used herein, may be interpreted as "responsive to a determination," depending on the context.

Aiming at the technical problem that labeling information of a training text in a training set has certain data noise, the denoising method mainly adopted at present comprises the following steps: artificial denoising and rule-based denoising. The manual denoising is to check the labeling information of each training text in a manual mode, judge whether the conditions of missing label, wrong label and wrong label exist one by one, and if yes, perform manual correction, and the mode needs to consume a large amount of manpower and material resources and has higher time cost; the rule-based denoising method is to design a judgment rule according to knowledge in each special field, judge the labeling information of the training text by using the judgment rule, judge whether the conditions of label missing, label missing and label missing exist or not, and manually correct if the conditions exist, and the method needs deeper professional field knowledge, needs professionals to design a large number of judgment rules, and has higher cost.

In order to solve the above problems, embodiments of the present application provide a data processing method and apparatus, a target recognition model training method and apparatus, a target recognition method and apparatus, a computing device, and a computer-readable storage medium, which will be described in detail one by one in the following embodiments.

Fig. 1 shows a schematic flow chart of a data processing method provided in an embodiment of the present application, where the method specifically includes the following steps.

S101, a training set is obtained, wherein the training set comprises a plurality of training texts.

In the embodiment of the present application, the execution subject of the data processing method may be a data acquisition device for acquiring a training text, a database for storing a training set, an intelligent device for executing a target recognition function, and the like. Before data processing, a training set including a large amount of training texts needs to be acquired, and generally, the training set may be acquired by receiving manually input large amounts of training texts to form a training set, or by reading large amounts of training texts from other data acquisition devices or databases to form a training set.

The obtained training text in the training set is generally labeled manually, and the specifically labeled information is an object in the training text, such as an entity in the text, where the training text may be a sentence, a short sentence, a word, an article, and the like, and for example, the training text is "the total designer of the bridge of the wuhan changjiang river is majol", and the labeled information of the training text may include "the bridge of the wuhan changjiang river" and "majol". However, in some specific fields, a technician is required to have a very strong professional basis, otherwise, problems such as label missing, label error and the like easily occur during manual labeling, for example, in the medical field, the technician is required to have a very professional medical professional basis, and taking a training sample for irregular menstruation caused by blood stasis as an example, the technician is required to have a certain medical professional basis to be able to label labeling information of blood stasis and irregular menstruation.

And S102, training to obtain different target recognition models based on the training set.

After the training set is obtained, training of the target recognition model can be performed based on the training set, in the conventional training mode, training texts in the training set are sequentially input into a neural network to be trained, the neural network to be trained is a neural network commonly used for target recognition, such as a Long Short-Term Memory network (LSTM), a Conditional Random Field (CRF), a Bidirectional Long Short-Term Memory network (bilst), a Bi-directional Short-Term Memory network (Bi-directional Short-Term Memory), a Bidirectional code representation (BERT) based on a converter, and the like, when one training text is input, the neural network outputs a feature information representing target information of the training texts, the output of the neural network is compared with labeled information of the training texts to obtain a difference value, and network parameters of the neural network are adjusted based on the difference value by using methods such as gradient descent and the like, the process is called one-time iteration, a training text is input to the neural network after the parameters are adjusted, next iteration is carried out, after multiple iterations, until the difference value is smaller than a preset threshold value or the iteration times reach the preset times, the iteration is stopped, and the neural network after the last iteration is determined to be the trained target recognition model.

In this embodiment of the application, in order to denoise the labeled information of the training set, a plurality of different target recognition models are obtained by training based on the training set, each target recognition model has different target recognition performance, and in order to ensure that each trained target recognition model has different target recognition performance, the specific training mode may adopt the following mode: in the first mode, a training set is divided into a plurality of subsets, then one or more preset neural networks are respectively trained by using different subsets, a target recognition model can be trained based on each subset, and the trained target recognition models are different because the training texts in each subset are different and the marking information is different, and the more subsets are divided, the more obvious the difference of the target recognition performance of the trained target recognition model is; in the second mode, a training set is not split, a plurality of preset neural networks are trained by directly utilizing the training set, a target recognition model can be trained for each preset neural network, and the target recognition performance of the trained target recognition model has certain difference due to different adopted neural networks; and in the third mode, the training set is divided into a plurality of subsets, the subsets are divided into a training subset, a verification subset and the like, then one or more preset neural networks are trained by using a training and verification mode, different subsets can be selected as the training subset and the verification subset each time, and thus a plurality of target recognition models can be trained, and the target recognition performance of the target recognition models is different.

To sum up, in an implementation manner of the embodiment of the present application, S102 may specifically be: and training at least one preset neural network by using the training set to obtain a plurality of different target recognition models.

And S103, respectively carrying out target recognition on each training text in the training set by using each target recognition model obtained by training to obtain a target recognition result of each training text.

After a plurality of different target recognition models are trained, each target recognition model can be used for respectively carrying out target recognition on each training text in a training set, and due to the fact that the target recognition performance of each target recognition model is different, the obtained target recognition results for the same training text are possibly the same or different. If the target recognition results obtained by the same training text through all the target recognition models are the same, the marking information of the training text can be proved to be accurate; if the target recognition results obtained by the same training text through the target recognition models are different, especially different from the labeling information of the training text in the training set, it can be shown that the labeling information of the training text may be inaccurate.

And S104, aiming at each training text, under the condition that the target recognition result of the training text is inconsistent with the label information of the training text, updating the label information of the training text in the training set.

As described above, if the target recognition result obtained by each target recognition model of a training text is different from the labeling information of the training text in the training set, it can be shown that the labeling information of the training text may be inaccurate, and the labeling information of the training text in the training set needs to be updated.

In specific implementation, S104 can be implemented in the following ways:

the first method is that for a training text, original labeling information of the training text exists in a training set, and after a target of the training text is identified by each target identification model, a plurality of target identification results are obtained, wherein a part of the target identification results are the same as the original labeling information of the training text, and a part of the target identification results are different from the original labeling information of the training text, and if the number of the target identification results different from the original labeling information of the training text reaches a certain proportion (for example, reaches 80% of the total number of the target identification results), it indicates that the labeling information of the training text belongs to a wrong label, and needs to be corrected, and the original labeling information of the training text in the training set can be replaced by: and the target recognition result with the largest occurrence frequency in the target recognition results different from the original marking information of the training text. For example, the original tagging information of the training text is "a", a total of 20 target recognition models are provided, wherein 3 target recognition models perform target recognition on the training text to obtain a target recognition result "a", 10 target recognition models perform target recognition on the training text to obtain a target recognition result "B", 7 target recognition models perform target recognition on the training text to obtain a target recognition result "C", and since the ratio of the target recognition result to the original tagging information exceeds 80%, and the ratio of "B" is the largest, the tagging information of the training text needs to be replaced by "B".

Secondly, for a training text, original labeling information of the training text exists in a training set, and after the target recognition of the training text is performed by each target recognition model, a plurality of target recognition results are obtained, wherein a certain number (for example, greater than 90% of the total number of the target recognition results) of the target recognition results is that no target exists in the training text, or for the training text, more than a certain number of target recognition models have no output of the target recognition results (considered as no target), which indicates that the labeling information of the training text belongs to a wrong label and needs to be corrected, the labeling information of the training text in the training set can be deleted, or the original labeling information of the training text in the training set is modified to be "no target", or the original labeling information of the training text is set to be null.

Thirdly, for a training text, the training set does not have the labeling information of the training text, and after the target recognition of the training text is performed by each target recognition model, a plurality of target recognition results are obtained, wherein a certain number (for example, more than 60% of the total number of the target recognition results) of the target recognition results are targets in the training text, and the target recognition results are the same, which indicates that the labeling information of the training text belongs to a missing label and needs to be corrected, the labeling information can be added to the training text, and the added labeling information is the target recognition result.

The above three ways are only examples of the implementation process of S104, and in specific implementation, a technician may update the labeling information of the training text in the training set according to a different ratio between the target recognition result and the labeling information or other rules. For example, for a training text, if the target recognition result is different from the labeling information of the training text, the labeling is considered as wrong, and the labeling information of the training text in the training set is deleted; for another example, for a training text, each target recognition result is different from the labeling information of the training text, and it can also be considered that the labeling is correct, and the labeling information of the training text in the training set can be retained.

In an implementation manner of the embodiment of the present application, S104 may specifically be:

When the method is specifically implemented, the method mainly aims at two situations of label missing and label error.

The missing mark means that a target is in the training text, however, when the technician labels the training text, the technician omits the label of the training text, that is, the label information of the training text is absent in the training set, so that when the target recognition model is used to perform target recognition on the training text, the target recognition result includes the target with a high probability. If the target recognition result of one training text includes a target and the labeling information of the training text does not exist in the training set, it is described that a missing target occurs, the labeling information of the training text may be added in the training set (that is, a target recognized by a target recognition model is added as the labeling information of the training text), and if the targets recognized by the target recognition models are different for the training text, the target information with the largest number of the recognized targets may be selected as the labeling information of the training text.

The mislabeling means that no target is originally in the training text, however, because a technician mislabels the labeling information of one target when labeling the training text because of an understanding error, when the training text is subjected to target recognition by using each target recognition model, the target recognition result does not include the target with a high probability. Then, if the target recognition result of one training text does not include the target and the labeling information of the training text exists in the training set, it indicates that the mislabeling occurs, and the labeling information of the training text may be deleted from the training set. In specific implementation, for a training text, as long as one target recognition result does not include a target, the labeling information of the training text may be deleted, or when the ratio of the targets not included in the target recognition result reaches a certain value, the labeling information of the training text may be deleted.

Through the embodiment, the problems of label missing and label missing of the training texts in the training set can be well solved, and the purpose of removing the data noise of the labeling information of the training texts in the training set is achieved, wherein the data noise is caused by manual labeling and mainly comprises manual label missing and label missing.

By applying the scheme of the embodiment of the application, a plurality of different target recognition models are trained based on the obtained training set, then the target recognition models obtained by training are utilized to respectively perform target recognition on the training texts in the training set, so as to obtain the target recognition result of each training text, if the target recognition result of each training text is inconsistent with the label information of each training text aiming at any training text, the label information of each training text is inaccurate, therefore, the label information of each training text in the training set is updated, and the data noise of the label information of the training texts in the training set is removed.

Based on the embodiment shown in fig. 1, fig. 2 is a schematic flow chart of another data processing method provided in the embodiment of the present application, and the method specifically includes the following steps.

S201, a training set is obtained, wherein the training set comprises a plurality of training texts.

This step is the same as S101 in the embodiment shown in fig. 1, and is not described again here.

S202, a verification subset and a training subset are sequentially selected from the training set.

The training set may include a plurality of subsets, which may be divided into a validation subset and a training subset. Of course, after the training set is obtained, the training set may also be split to obtain a plurality of subsets, and specifically, the splitting may be performed on average or randomly during the splitting, and in order to ensure that the accuracy of each target recognition model is the same, the splitting may be performed on average, that is, the number of training texts in each subset is the same.

After the subsets are divided, the subsets can be divided into training subsets and verification subsets, one or more preset neural networks are trained in a training and verification mode, and one subset can be selected from the plurality of subsets as a verification subset and the other subsets can be selected as training subsets each time.

And S203, training the same preset neural network by using the selected training subset each time, and verifying the training result by using the selected verification subset to obtain a plurality of different target recognition models.

During specific training, the selected training subset is used for training the same preset neural network every time, and the selected verification subset is used for verifying the training result. For example, the training set is divided into 5 subsets, the subset 1, the subset 2, the subset 3, and the subset 4 are used as training subsets for the first time, a preset neural network is trained, the subset 5 is used as a verification subset, and a training result is verified to obtain a target recognition model 1; secondly, training the preset neural network by using the subset 1, the subset 2, the subset 3 and the subset 5 as training subsets, and verifying the training result by using the subset 4 as a verification subset to obtain a target recognition model 2; thirdly, training the preset neural network by using the subset 1, the subset 2, the subset 4 and the subset 5 as training subsets, and verifying the training result by using the subset 3 as a verification subset to obtain a target recognition model 3; fourthly, training the preset neural network by using the subset 1, the subset 3, the subset 4 and the subset 5 as training subsets, and verifying the training result by using the subset 2 as a verification subset to obtain a target recognition model 4; and fifthly, training the preset neural network by using the subsets 2, 3, 4 and 5 as training subsets, and verifying the training result by using the subset 1 as a verification subset to obtain the target recognition model 5. Thus, five object recognition models, namely, an object recognition model 1, an object recognition model 2, an object recognition model 3, an object recognition model 4 and an object recognition model 5 are obtained.

The preset neural network may be LSTM, CRF, BiLSTM, BERT, or the like, or may be a network obtained by combining the above networks. The specific training process is the same as or similar to the conventional training process of the target recognition model, and is not described herein again.

And S204, respectively carrying out target recognition on each training text in the training set by using each target recognition model obtained by training to obtain a target recognition result of each training text.

This step is the same as S103 in the embodiment shown in fig. 1, and is not described again here.

S205, for each training text, if the target recognition result of the training text is inconsistent with the label information of the training text, updating the label information of the training text in the training set.

This step is the same as S104 in the embodiment shown in fig. 1, and is not described again here.

By applying the scheme of the embodiment of the application, the acquired training set is divided into a plurality of subsets, one subset is selected from the plurality of subsets as a verification subset, other subsets are selected as training subsets in turn, the selected training subset is used for training the same preset neural network each time, the selected verification subset is used for verifying the training result to obtain a plurality of different target recognition models, then, each training text in the training set is respectively subjected to target recognition by utilizing each target recognition model obtained by training to obtain a target recognition result of each training text, if any training text is targeted, if the target recognition result of the training text is inconsistent with the labeling information of the training text, it indicates that the labeling information of the training text is inaccurate, and therefore, and updating the labeling information of the training texts in the training set, thereby removing the data noise of the labeling information of the training texts in the training set.

Based on the embodiment shown in fig. 1, fig. 3 shows a schematic flow chart of another data processing method provided in the embodiment of the present application, and as shown in fig. 3, the method specifically includes the following steps.

S301, a training set is obtained, wherein the training set comprises a plurality of training texts.

S302, a verification subset and a training subset are sequentially selected from the training set.

And S303, aiming at different preset neural networks, training the preset neural networks by using the selected training subsets each time, and verifying the training results by using the selected verification subsets to obtain different target recognition models.

During specific training, a plurality of different preset neural networks can be selected, and for each preset neural network, the selected training subset is used for training the preset neural network each time, and the selected verification subset is used for verifying the training result. For example, the training set is divided into 5 subsets, the number of preset neural networks is 2, the 1 st preset neural network is trained by using the subset 1, the subset 2, the subset 3 and the subset 4 as training subsets for the first time, the training result is verified by using the subset 5 as a verification subset, and the target identification model 1 is obtained; training the 2 nd preset neural network by using the subset 1, the subset 2, the subset 3 and the subset 4 as training subsets for the second time, and verifying the training result by using the subset 5 as a verification subset to obtain a target recognition model 2; thirdly, training the 1 st preset neural network by using the subsets 1, 2, 3 and 5 as training subsets, and verifying the training result by using the subset 4 as a verification subset to obtain a target recognition model 3; fourthly, training the 2 nd preset neural network by using the subset 1, the subset 2, the subset 3 and the subset 5 as training subsets, and verifying a training result by using the subset 4 as a verification subset to obtain a target recognition model 4; by analogy, ten target recognition models, namely a target recognition model 1, a target recognition model 2, a target recognition model 3, a target recognition model 4, a target recognition model 5, a target recognition model 6, a target recognition model 7, a target recognition model 8, a target recognition model 9 and a target recognition model 10 can be obtained. The specific training process is the same as or similar to the conventional training process of the target recognition model, and is not described herein again.

And S304, respectively carrying out target recognition on each training text in the training set by using each target recognition model obtained by training to obtain a target recognition result of each training text.

S305, for each training text, if the target recognition result of the training text is not consistent with the label information of the training text, updating the label information of the training text in the training set.

By applying the scheme of the embodiment of the application, the acquired training set is divided into a plurality of subsets, one subset is selected from the plurality of subsets as a verification subset and the other subsets are selected as training subsets in turn, for each preset neural network in a plurality of different preset neural networks, the same preset neural network is trained by using the selected training subset each time, the training result is verified by using the selected verification subset, a plurality of different target recognition models are obtained, then the target recognition models obtained by training are used for respectively carrying out target recognition on each training text in the training set, the target recognition result of each training text is obtained, if the target recognition result of any training text is inconsistent with the labeling information of the training text, the labeling information of the training text is inaccurate, therefore, the labeling information of the training text in the training set is updated, therefore, data noise of the labeling information of the training texts in the training set is removed.

Fig. 4 is a flowchart illustrating a method for training a target recognition model according to an embodiment of the present application, where the method specifically includes the following steps.

S401, a training set is obtained.

The training set is obtained by processing data by the data processing method. The specific data processing process is as in the above method embodiment, and is not described herein again.

S402, training the preset neural network by using the training set to obtain a target recognition model.

The preset neural network can be LSTM, CRF, BilTM, BERT, etc., or a network obtained by mutually combining the networks. The specific training process is the same as or similar to the conventional training process of the target recognition model, and is not described herein again.

In the embodiment of the application, the execution subject of the target recognition model training method may be training equipment of a target recognition model, intelligent equipment for executing a target recognition function, and the like. By applying the scheme of the embodiment of the application, the obtained training set is the training set subjected to data processing by the data processing method, and the training set well removes data noise, so that the influence of the data noise of the training text on the accuracy of the target recognition model can be reduced, and the accuracy of the target recognition model is improved.

For convenience of understanding, the following describes a training process of the target recognition model by taking two preset neural networks as an example, and as shown in fig. 5, the training process mainly includes three steps of five-fold model training, data noise reduction and target recognition model training. Firstly, selecting two preset neural networks of BERT + CRF and BERT + Machine Reading Comprehension (MRC), and splitting an acquired training set into 5 equal parts; then, carrying out five-fold model training: and (3) respectively utilizing BERT + CRF and BERT + MRC to carry out five-fold model training on the 5 equally divided subsets, wherein the five-fold model training is to select one subset as a verification subset and the other four subsets as training subsets, and respectively carry out training and verification on the two neural networks of BERT + CRF and BERT + MRC, so that 10 trained target recognition models can be obtained. Of course, the splitting of the training set into 5 equal parts is only an example, and the training set may also be split into K equal parts (K ≧ 2), and then the K-folding model training is performed on the K equal parts subsets obtained by splitting using BERT + CRF and BERT + MRC, respectively. Then, data noise reduction is carried out: carrying out target recognition on the 10 trained target recognition models on the whole training set, if a certain training text is recognized as a target by the 10 target recognition models at the same time, and the training text is not marked as the target in the training set, judging that the target is missed, and marking the training text as the target to be added into the training set; if a certain training text is marked as a target in the training set, but the recognition results of 10 target recognition models have no target, the training text is considered to be a wrong target, and the marking information of the training text is deleted from the training set. And finally, training the target recognition model, and training the target recognition model on the training set with updated labeling information again, wherein the training set is subjected to data denoising, so that the precision of the target recognition model is improved, and the accuracy of target recognition is improved.

Fig. 6 shows a schematic flowchart of a target identification method provided in an embodiment of the present application, where the method specifically includes the following steps.

S601, acquiring a text to be recognized.

And S602, inputting the text to be recognized into the trained target recognition model to obtain a target recognition result of the text to be recognized.

The target recognition model is obtained by the target recognition model training method. The trained target recognition model is an end-to-end deep learning model, and the obtained text to be recognized is directly input into the trained target recognition model, so that the target recognition result of the text to be recognized can be directly obtained, and the target recognition result comprises all targets which can be recognized in the text to be recognized.

In the embodiment of the present application, the execution subject of the target identification method may be an intelligent device that executes a target identification function. By applying the scheme of the embodiment of the application, the obtained training set is the training set subjected to data processing by the data processing method, and the training set well removes data noise, so that the influence of the data noise of the training text on the accuracy of the target recognition model can be reduced, and the accuracy of the target recognition model is improved. Therefore, when the trained target recognition model is used for carrying out target recognition on the text to be recognized, the accuracy of the target recognition result can be improved.

Corresponding to the above data processing method embodiment, fig. 7 shows a schematic structural diagram of a data processing apparatus provided in an embodiment of the present application, where the data processing apparatus includes:

a first obtaining module 710 configured to obtain a training set, wherein the training set includes a plurality of training texts;

a first model training module 720 configured to train different target recognition models based on the training set;

the first target recognition module 730 is configured to perform target recognition on each training text in the training set by using each target recognition model obtained through training to obtain a target recognition result of each training text;

and an updating module 740 configured to update, for each training text, the label information of the training text in the training set if the target recognition result of the training text is inconsistent with the label information of the training text.

Optionally, the first model training module 720 is further configured to: and training at least one preset neural network by using the training set to obtain different target recognition models.

Optionally, the first model training module 720 is further configured to: sequentially selecting a verification subset and a training subset from a training set; and training the same preset neural network by using the selected training subset every time, and verifying the training result by using the selected verification subset to obtain different target recognition models.

Optionally, the first model training module 720 is further configured to: sequentially selecting a verification subset and a training subset from a training set; and aiming at different preset neural networks, training the preset neural networks by using the selected training subsets each time, and verifying the training results by using the selected verification subsets to obtain different target recognition models.

Optionally, the update module 740 is further configured to: for each training text, if the target recognition result of the training text comprises a target and the labeling information of the training text does not exist in the training set, adding the labeling information of the training text in the training set, wherein the added labeling information of the training text is the recognized target in the training text; and for each training text, if the target recognition result of the training text does not comprise a target and the labeling information of the training text exists in the training set, deleting the labeling information of the training text from the training set.

The above is a schematic configuration of a data processing apparatus of the present embodiment. It should be noted that the technical solution of the data processing apparatus and the technical solution of the data processing method belong to the same concept, and details that are not described in detail in the technical solution of the data processing apparatus can be referred to the description of the technical solution of the data processing method.

Corresponding to the above-mentioned embodiment of the target recognition model training method, fig. 8 is a schematic structural diagram of a target recognition model training apparatus provided in this embodiment of the present application, where the target recognition model training apparatus includes:

a second obtaining module 810, configured to obtain a training set, where the training set is a training set obtained by performing data processing by using the data processing method;

and a second model training module 820 configured to train the preset neural network by using the training set to obtain the target recognition model.

By applying the scheme of the embodiment of the application, the obtained training set is the training set subjected to data denoising by the data processing method, and the training set well removes data noise, so that the influence of the data noise of the training text on the accuracy of the target recognition model can be reduced, and the accuracy of the target recognition model is improved.

The above is a schematic scheme of a target recognition model training apparatus according to this embodiment. It should be noted that the technical solution of the target recognition model training apparatus and the technical solution of the target recognition model training method belong to the same concept, and details that are not described in detail in the technical solution of the target recognition model training apparatus can be referred to the description of the technical solution of the target recognition model training method.

Corresponding to the above-mentioned embodiment of the target identification method, fig. 9 shows a schematic structural diagram of a target identification device provided in the embodiment of the present application, where the target identification device includes:

a third obtaining module 910 configured to obtain a text to be recognized;

and the second target recognition module 920 is configured to input the text to be recognized into the target recognition model obtained by training through the target recognition model training method, so as to obtain a target recognition result of the text to be recognized.

By applying the scheme of the embodiment of the application, the obtained training set is the training set subjected to data denoising by the data processing method, and the training set well removes data noise, so that the influence of the data noise of the training text on the accuracy of the target recognition model can be reduced, and the accuracy of the target recognition model is improved. Therefore, when the trained target recognition model is used for carrying out target recognition on the text to be recognized, the accuracy of the target recognition result can be improved.

The above is a schematic scheme of an object recognition apparatus of the present embodiment. It should be noted that the technical solution of the object recognition apparatus belongs to the same concept as the technical solution of the object recognition method, and for details that are not described in detail in the technical solution of the object recognition apparatus, reference may be made to the description of the technical solution of the object recognition method.

It should be noted that the components in the apparatus should be understood as functional blocks that must be established to implement the steps of the program flow or the steps of the method, and each functional block is not actually defined by division or separation of functions. The means defined by such a set of functional modules should be understood as a functional module framework that mainly implements the solution by means of a computer program described in the specification, and should not be understood as a physical means that mainly implements the solution by means of hardware.

Fig. 10 shows a block diagram of a computing device 100 according to an embodiment of the present application. The components of the computing device 100 include, but are not limited to, memory 110 and processor 120. The processor 120 is coupled to the memory 110 via a bus 130 and a database 150 is used to store data.

Computing device 100 also includes access device 140, access device 140 enabling computing device 100 to communicate via one or more networks 160. Examples of such networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the internet. The Access device 140 may include one or more of any type of Network Interface (e.g., a Network Interface Card (NIC)) whether wired or Wireless, such as an IEEE802.11 Wireless Local Area Network (WLAN) Wireless Interface, a worldwide Interoperability for Microwave Access (Wi-MAX) Interface, an ethernet Interface, a Universal Serial Bus (USB) Interface, a cellular Network Interface, a bluetooth Interface, a Near Field Communication (NFC) Interface, and so forth.

In one embodiment of the present application, the above-described components of the computing device 100 and other components not shown in fig. 10 may also be connected to each other, for example, by a bus. It should be understood that the block diagram of the computing device architecture shown in FIG. 10 is for purposes of example only and is not limiting as to the scope of the present application. Those skilled in the art may add or replace other components as desired.

Computing device 100 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., tablet, personal digital assistant, laptop, notebook, netbook, etc.), a mobile phone (e.g., smartphone), a wearable computing device (e.g., smartwatch, smartglasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or PC. Computing device 100 may also be a mobile or stationary server.

The processor 120 is configured to execute computer-executable instructions, and when the processor 120 executes the computer-executable instructions, the steps of the data processing method, the target recognition model training method, or the target recognition method are implemented.

The above is an illustrative scheme of a computing device of the present embodiment. It should be noted that the technical solution of the computing device belongs to the same concept as the technical solutions of the data processing method, the target recognition model training method, and the target recognition method, and details of the technical solution of the computing device, which are not described in detail, can be referred to the descriptions of the technical solutions of the data processing method, the target recognition model training method, and the target recognition method.

An embodiment of the present application further provides a computer readable storage medium, which stores computer instructions, and the computer instructions, when executed by a processor, implement the steps of the data processing method or the target recognition model training method or the target recognition method as described above.

The above is an illustrative scheme of a computer-readable storage medium of the present embodiment. It should be noted that the technical solution of the storage medium is the same as the technical solutions of the data processing method, the target recognition model training method, and the target recognition method, and details of the technical solution of the storage medium, which are not described in detail, can be referred to the descriptions of the technical solutions of the data processing method, the target recognition model training method, and the target recognition method.

The embodiment of the application discloses a chip, which stores computer instructions, and the computer instructions are executed by a processor to realize the steps of the data processing method or the target recognition model training method or the target recognition method.

The foregoing description of specific embodiments of the present application has been presented. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.

The computer instructions comprise computer program code which may be in the form of source code, object code, an executable file or some intermediate form, or the like. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, usb disk, removable hard disk, magnetic disk, optical disk, computer Memory, Read-Only Memory (ROM), Random Access Memory (RAM), electrical carrier wave signals, telecommunications signals, software distribution medium, and the like. It should be noted that the computer readable medium may contain content that is subject to appropriate increase or decrease as required by legislation and patent practice in jurisdictions, for example, in some jurisdictions, computer readable media does not include electrical carrier signals and telecommunications signals as is required by legislation and patent practice.

It should be noted that, for the sake of simplicity, the above-mentioned method embodiments are described as a series of acts or combinations, but those skilled in the art should understand that the present application is not limited by the described order of acts, as some steps may be performed in other orders or simultaneously according to the present application. Further, those skilled in the art should also appreciate that the embodiments described in the specification are preferred embodiments and that the acts and modules referred to are not necessarily required in this application.

In the above embodiments, the descriptions of the respective embodiments have respective emphasis, and for parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

The preferred embodiments of the present application disclosed above are intended only to aid in the explanation of the application. Alternative embodiments are not exhaustive and do not limit the invention to the precise embodiments described. Obviously, many modifications and variations are possible in light of the above teaching. The embodiments were chosen and described in order to best explain the principles of the application and its practical applications, to thereby enable others skilled in the art to best understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A data processing method, comprising:

and updating the labeling information of the training texts in the training set under the condition that the target recognition result of the training texts is inconsistent with the labeling information of the training texts aiming at each training text.

2. The method of claim 1, wherein the step of training different target recognition models based on the training set comprises:

3. The method according to claim 2, wherein the step of training at least one predetermined neural network by using the training set to obtain different target recognition models comprises:

sequentially selecting a verification subset and a training subset from the training set;

4. The method according to claim 2, wherein the step of training at least one predetermined neural network by using the training set to obtain different target recognition models comprises:

5. The method according to claim 1, wherein the step of updating the label information of the training text in the training set when the target recognition result of the training text is inconsistent with the label information of the training text for each training text comprises:

6. A method for training a target recognition model, comprising:

acquiring a training set, wherein the training set is obtained after data processing by using the method of any one of claims 1-5;

and training a preset neural network by using the training set to obtain a target recognition model.

7. A method of object recognition, comprising:

acquiring a text to be identified;

inputting the text to be recognized into the target recognition model obtained by training by using the method of claim 6, and obtaining the target recognition result of the text to be recognized.

8. A data processing apparatus, comprising:

a first model training module configured to train different target recognition models based on the training set;

and the updating module is configured to update the labeling information of the training text in the training set under the condition that the target recognition result of the training text is inconsistent with the labeling information of the training text for each training text.

9. The apparatus of claim 8, wherein the first model training module is further configured to: and training at least one preset neural network by using the training set to obtain different target recognition models.

10. The apparatus of claim 9, wherein the first model training module is further configured to: sequentially selecting a verification subset and a training subset from the training set; and training the same preset neural network by using the selected training subset every time, and verifying the training result by using the selected verification subset to obtain different target recognition models.

11. The apparatus of claim 9, wherein the first model training module is further configured to: sequentially selecting a verification subset and a training subset from the training set; and aiming at different preset neural networks, training the preset neural networks by using the selected training subsets each time, and verifying the training results by using the selected verification subsets to obtain different target recognition models.

12. The apparatus of claim 8, wherein the update module is further configured to: for each training text, if the target recognition result of the training text comprises a target and the labeling information of the training text does not exist in the training set, adding the labeling information of the training text in the training set, wherein the added labeling information of the training text is the recognized target in the training text; and for each training text, if the target recognition result of the training text does not comprise a target and the labeling information of the training text exists in the training set, deleting the labeling information of the training text from the training set.

13. An object recognition model training apparatus, comprising:

a second obtaining module configured to obtain a training set, wherein the training set is a training set obtained by performing data processing according to the method of any one of claims 1 to 5;

and the second model training module is configured to train a preset neural network by using the training set to obtain a target recognition model.

14. An object recognition apparatus, comprising:

the third acquisition module is configured to acquire a text to be recognized;

a second target recognition module, configured to input the text to be recognized into the target recognition model obtained by training according to the method of claim 6, and obtain a target recognition result of the text to be recognized.

15. A computing device comprising a memory, a processor and computer instructions stored on the memory and executable on the processor, wherein the steps of the method of any one of claims 1 to 5 or 6 or 7 are implemented when the computer instructions are executed by the processor.

16. A computer readable storage medium storing computer instructions, which when executed by a processor, perform the steps of the method of any one of claims 1 to 5 or 6 or 7.