Data processing method and device, equipment, medium and program product

Abstract

The embodiment of the invention provides a data processing method and device, equipment, a medium and a program product, and the method comprises the steps: obtaining a first sample data set which comprises M pieces of first sample data and N pieces of unlabeled data; performing active learning processing on the N pieces of unlabeled data according to an active learning mode to obtain labeled K pieces of second sample data; performing sample enhancement processing on the M pieces of first sample data and the K pieces of second sample data to obtain a second sample data set, and constructing a model training data set based on the second sample data set, the M pieces of first sample data and the K pieces of second sample data; and training the text discrimination model by adopting the model training data set. According to the method, active learning and sample enhancement can be performed on a large amount of unlabeled data through a small amount of labeled risk control data, so that richer business sample data is provided for model training, and the accuracy of a text discrimination model is improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data processing method, a data processing apparatus, a computer device, a computer-readable storage medium, and a computer program product. Background Art

[0002] In any business scenario, such as finance, gaming, social networking, and live streaming, large amounts of business data are often involved. To ensure compliance with business regulations and data security, the large amount of data present in these scenarios often requires the identification of risk control text. Currently, manual screening or identification of risk control data from this vast amount of business data is the primary method used in these scenarios. This manual review process is time-consuming and labor-intensive, and is prone to misjudgment, resulting in a relatively low accuracy rate in identifying risk control text. Summary of the Invention

[0003] The embodiments of the present application propose a data processing method, apparatus, equipment, medium and program product, which can actively learn and enhance samples of a large amount of unlabeled data through a small amount of labeled risk control data, thereby providing richer business sample data for model training to improve the accuracy of the text discrimination model.

[0004] In one aspect, an embodiment of the present application provides a data processing method, the method comprising:

[0005] Obtain a first sample data set, where the first sample data set includes M first sample data and N unlabeled data, where the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N;

[0006] Active learning is performed on N unlabeled data in an active learning manner to select K unlabeled data with labeling value from the N unlabeled data, and business labeling is performed on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N;

[0007] Performing sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and constructing a model training data set based on the second sample data set, the M first sample data, and the K second sample data;

[0008] A text discrimination model is trained using a model training dataset; wherein the trained text discrimination model is used to recognize target text in at least one business scenario.

[0009] In one aspect, an embodiment of the present application provides a data processing method, the method comprising:

[0010] An acquisition unit is configured to acquire a first sample data set, wherein the first sample data set includes M first sample data and N unlabeled data, wherein the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N;

[0011] a processing unit, configured to perform active learning processing on the N unlabeled data in an active learning manner, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and perform business labeling processing on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N;

[0012] The processing unit is further configured to perform sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and construct a model training data set based on the second sample data set, the M first sample data, and the K second sample data;

[0013] The processing unit is further used to train the text discrimination model using a model training data set; wherein the trained text discrimination model is used to identify target text in at least one business scenario.

[0014] In one possible implementation, the active learning method includes a first active learning method and a second active learning method; the first active learning method is used to indicate active learning based on a standard paradigm, and the second active learning method is used to indicate active learning based on feedback data; the processing unit performs active learning processing on N unlabeled data according to the active learning method to screen out K unlabeled data with labeling value from the N unlabeled data, including any of the following methods:

[0015] Based on the M labeled first sample data, active learning processing is performed on the N unlabeled data according to the first active learning method to screen out K unlabeled data with labeling value from the N unlabeled data;

[0016] Obtain business feedback data in business scenarios, and perform data generalization processing on N unlabeled data based on the business feedback data according to the second active learning method, so as to screen out K unlabeled data with labeling value from the N unlabeled data.

[0017] In one possible implementation, the processing unit performs active learning processing on the N unlabeled data according to a first active learning method based on the M labeled first sample data, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and performs the following operations:

[0018] Based on the labeled M first sample data, train the active learner;

[0019] Call the trained active learner to perform prediction and scoring on N unlabeled data to obtain the scoring results of each unlabeled data;

[0020] Based on the scoring results of each unlabeled data, K unlabeled data with labeling value are screened from N unlabeled data according to the preset sampling strategy.

[0021] In one possible implementation, the processing unit performs data generalization processing on the N unlabeled data based on the business feedback data in accordance with the second active learning method, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and perform the following operations:

[0022] Perform multi-dimensional text expansion processing on the business feedback data to obtain P candidate sample data, where P is a positive integer;

[0023] Based on P candidate sample data, N unlabeled data are screened to select K unlabeled data with labeling value from the N unlabeled data.

[0024] In one possible implementation, the dimensions include: any one or more of a scenario dimension, a data type dimension, and a timeliness dimension; the processing unit performs multi-dimensional text expansion processing on the business feedback data in the financial scenario to obtain P candidate sample data for performing the following operations:

[0025] In multiple business scenarios, text expansion processing is performed on the business feedback data according to the scenario dimension to obtain first candidate sample data; wherein the business scenario includes: any one or more of: game scenario, live broadcast scenario, social scenario, and financial scenario;

[0026] Perform text expansion processing of multiple data types on the business feedback data according to the data type dimension to obtain second candidate sample data;

[0027] Perform text expansion processing on the business feedback data according to the timeliness dimension to obtain third candidate sample data;

[0028] Based on the first candidate sample data, the second candidate sample data, and the third candidate sample data, P candidate sample data are obtained.

[0029] In a possible implementation, the processing unit performs sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set for performing any one of the following operations:

[0030] Perform word replacement processing on the M first sample data and the K second sample data to obtain a second sample data set;

[0031] Using a pre-trained language model to perform masking on the M first sample data and the K second sample data to obtain a second sample data set;

[0032] Using a machine translation model to perform back-translation processing on the M first sample data and the K second sample data to obtain a second sample data set;

[0033] Based on the M first sample data and the K second sample data, a variational autoencoder is used to generate a second sample data set.

[0034] In one possible implementation, the second sample data set includes a plurality of enhanced sample data, and the model training data set includes a plurality of business sample data, where the business sample data refers to any one of the first sample data, the second sample data, and the enhanced sample data, and any one of the business sample data carries a business label; the processing unit uses the model training data set to train the text discrimination model to perform the following operations:

[0035] Call the text discrimination model to identify each business sample data in the model training data set and obtain the predicted label of each business sample data;

[0036] Adjust the model parameters of the text discrimination model based on the differences between the business labels and predicted labels of each business sample data;

[0037] When the adjusted text discrimination model meets the model convergence condition, the training of the text discrimination model is stopped; wherein the text discrimination model includes: any one of a text classification basic model, a large pre-trained language model, and a large pre-trained knowledge distillation model.

[0038] In one possible implementation, the text discrimination model includes Q different types of discrimination sub-models, where Q is a positive integer. The processing unit trains the text discrimination model using a model training dataset to perform the following operations:

[0039] Using the model training data set to train Q discriminant sub-models respectively, to obtain Q trained discriminant sub-models; wherein the training method of the Q discriminant sub-models includes: any one of parallel training and serial training;

[0040] A text discrimination model is constructed based on Q trained discriminant sub-models;

[0041] The Q trained discriminant sub-models are constructed in a manner including: a model cascade manner, a model selection manner, and a model combination manner.

[0042] In one possible implementation, the model training data set includes multiple business sample data, where the business sample data refers to any one of first sample data, second sample data, and enhanced sample data; the discriminant sub-model includes a large pre-trained knowledge distillation model, which includes a large pre-trained language model and a lightweight language model;

[0043] The processing unit trains a large pre-trained knowledge distillation model using a model training dataset, performing the following operations:

[0044] Use a lightweight language model to identify each business sample data in the model training data set and obtain a vectorized representation i of each business sample data;

[0045] Based on the large pre-trained language model, each business sample data in the model training data set is identified, and the vectorized representation j of each business sample data is obtained;

[0046] Based on the vectorized representation i and j of each business sample data, the inter-vector difference quantization loss function is used to train the large pre-trained knowledge distillation model.

[0047] In a possible implementation, the processing unit is further configured to perform the following operations:

[0048] Obtain data requests for target business scenarios; target business scenarios include: financial scenarios, social scenarios, gaming scenarios, and live broadcast scenarios;

[0049] Parse the data request to obtain the text to be recognized in the target business scenario;

[0050] Determine the target cache submodule that matches the target business scenario, and cache the parsed document to be identified into the target cache submodule;

[0051] The text discrimination model is called to perform recognition processing on the text to be recognized to obtain the recognition result of the text to be recognized; the recognition result is used to indicate whether the text to be recognized is the risk control text in the target business scenario.

[0052] In one possible implementation, the data processing method is applied to a text recognition engine, which includes an offline module and an online module. The offline module includes a discrimination module, an active learning module, and a sample enhancement module, and the online module includes a multi-level cache module.

[0053] The discrimination module is used to provide a text discrimination model and perform business discrimination processing on each business sample data in the model training data set based on the text discrimination model;

[0054] The active learning module is used to perform active learning processing on N unlabeled data in an active learning manner;

[0055] The sample enhancement module is used to perform sample enhancement processing on the M first sample data and the K second sample data;

[0056] The multi-level cache module includes a first-level cache module and a second-level cache module. The first-level cache module is a general cache module for various business scenarios, and the second-level cache module includes multiple cache sub-modules, and one cache sub-module corresponds to one business scenario.

[0057] In one possible implementation, the target business scenario is a financial scenario, the recognition result includes a scoring result of the text to be recognized; the online module further includes a post-scoring processing module; and the processing unit is further configured to perform the following operations:

[0058] If the text to be identified is a risk control text in a financial scenario, then the risk control strategy associated with the financial scenario is obtained; the risk control strategy is used to indicate the risk control method that matches different scoring results;

[0059] According to the risk control strategy in the financial scenario, based on the scoring results of the text to be recognized, the target risk control method corresponding to the text to be recognized is determined;

[0060] Use target risk control methods to perform data risk control on the identified text;

[0061] Among them, data risk control processing includes: any one of deletion processing, interception processing, recording processing, and warning processing.

[0062] In a possible implementation, the online module further includes a rule engine module; and the processing unit is further configured to perform the following operations:

[0063] Determine the text type of the text to be recognized in the target business scenario;

[0064] If the text type of the text to be recognized is a preset text type, the rule engine module is called to perform rule processing on the text to be recognized;

[0065] The rule processing includes filtering, intercepting and warning.

[0066] On the one hand, an embodiment of the present application provides a computer device, which includes a processor, an input device, an output device and a memory; a computer program is stored in the memory; when the computer program is executed by the processor, the above-mentioned data processing method is executed.

[0067] On the one hand, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned data processing method is executed.

[0068] On the one hand, an embodiment of the present application provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the above-mentioned data processing method is executed.

[0069] In an embodiment of the present application, a first sample data set can be obtained, which includes M first sample data and N unlabeled data, and the first sample data includes labeled business texts in a business scenario; M and N are both positive integers and M<N; active learning processing is performed on the N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and business labeling processing is performed on these K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N; sample enhancement processing is performed on the M first sample data and the K second sample data to obtain a second sample data set, and a model training data set is constructed based on the second sample data set, the M first sample data and the K second sample data; the text discrimination model is trained using the model training data set; wherein the trained text discrimination model is used to identify target text (such as risk control text and non-risk control text) in at least one business scenario. It can be seen that in this application, a small amount of business sample data can be labeled in the business scenario, and then a large amount of unlabeled business data in the business scenario can be actively learned, and business data with labeling value can be screened out and labeled to obtain second sample data; and after the labeled first sample data and second sample data are enhanced, rich and diverse business sample data can be obtained. Using rich sample data to train the text discrimination model can improve the accuracy of model training, thereby improving the text discrimination model's recognition accuracy of risk control texts in various business scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present application. For ordinary technical objects in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0071] Figure 1 This is a schematic diagram of the architecture of a data processing system provided in an embodiment of the present application;

[0072] Figure 2a This is a structural diagram of a text recognition engine provided by an embodiment of the present application;

[0073] Figure 2b This is a schematic diagram of the structure of a multi-level cache module provided in an embodiment of the present application;

[0074] Figure 3 This is a flow chart of a data processing method provided in an embodiment of the present application;

[0075] Figure 4a This is a diagram of a data set for a supervised learning method provided in an embodiment of the present application;

[0076] Figure 4b This is a data set diagram of an active learning method provided in an embodiment of the present application;

[0077] Figure 5 This is a flowchart of a first active learning method provided in an embodiment of the present application;

[0078] Figure 6 This is a flow chart of another data processing method provided in an embodiment of the present application;

[0079] Figure 7a This is a structural diagram of a basic text classification model provided in an embodiment of the present application;

[0080] Figure 7b This is a schematic diagram of the structure of a large-scale pre-trained language model provided in an embodiment of the present application;

[0081] Figure 7c This is a schematic diagram of the structure of a large-scale pre-trained knowledge distillation model provided in an embodiment of the present application;

[0082] Figure 8 This is an online recognition flow chart of a text discrimination model provided in an embodiment of the present application;

[0083] Figure 9 is a structural diagram of a data processing device provided in an embodiment of the present application;

[0084] Figure 10 It is a structural diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0085] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0086] The present application provides a data processing solution that is suitable for the recognition of business texts (such as risk control texts) in any business scenarios such as financial scenarios, game scenarios, live broadcast scenarios, etc. Specifically, the present application can collect a small amount of labeled sample data and a large amount of unlabeled data from financial scenarios, and use active learning to automatically filter out data with labeling value from a large amount of unlabeled data for risk control labeling, thereby serving as samples for model training; then the labeled sample data is subjected to sample enhancement processing. The present application can provide rich training samples for the model, and then train a more accurate text discrimination model to improve the accuracy of risk control text recognition in various business scenarios. The following is a general explanation of the principles of the data processing solution provided by the present application:

[0087] ① Sample data collection. Specifically, a first sample data set can be collected in one or more business scenarios. The first sample data set includes M first sample data and N unlabeled data, and the first sample data includes labeled business texts (such as risk control texts) in one or more business scenarios (such as financial scenarios). M and N are both positive integers and M<N; that is, the first sample data set includes a small amount of labeled business sample data and a large amount of unlabeled business data. For example, the sample data collected from business scenarios include: reporting data, community type data, conversation data, and group message data.

[0088] ② Active learning of sample data. Specifically, active learning can be performed on N unlabeled data. In practice, active learning can be used to filter out K unlabeled data with annotation value from the N unlabeled data. Then, risk control annotation is performed on these K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K ≤ N.

[0089] ③ Sample data enhancement processing. Specifically, sample enhancement processing can be performed on the M first sample data and the K second sample data to obtain a second sample data set, and the second sample data set, the M first sample data and the K second sample data are constructed to obtain a model training data set.

[0090] ④ Train a text discrimination model for risk control identification. Specifically, the above-mentioned model training data set can be used to train a text discrimination model, wherein the trained text discrimination model is used to identify risk control texts in at least one business scenario (such as a financial scenario, a gaming scenario, a live broadcast scenario, etc.). Subsequently, the trained text discrimination model can be used to identify and process risk control texts in various business scenarios such as financial scenarios, gaming scenarios, and live broadcast scenarios, and further, risk control processing operations can be performed in the business scenarios (such as deleting, intercepting, recording, and alerting risk control texts).

[0091] From the above, it can be seen that, on the one hand, the present application can actively learn a large amount of unlabeled business data (N unlabeled data) in the business scenario in an active learning manner, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and then perform risk control labeling on these K unlabeled business data to obtain labeled second sample data. It can be seen that active learning can automatically perform risk control labeling on sample data; on the other hand, sample enhancement can be performed based on the labeled sample data to obtain richer sample data, and finally, the text discrimination model can be trained based on the rich sample data, which can improve the accuracy of the trained text discrimination model, so that the text discrimination model can be applicable in multiple business scenarios, that is, improve the text discrimination model's recognition accuracy of risk control texts in various business scenarios.

[0092] The key technical terms involved in this application are introduced in detail below.

[0093] 1. Business scenarios

[0094] A business scenario refers to a scenario that provides business needs for an object. For example, if the business scenario is a financial scenario, the financial scenario is used to provide financial needs (such as asset trading, asset settlement, etc.) for the object; if the business scenario is a gaming scenario, the gaming scenario is used to provide gaming needs (such as testing games, playing games, etc.) for the object; and if the business scenario is a live broadcast scenario, the live broadcast scenario is used to provide various types of live broadcast needs (such as entertainment, education, maternal and child care, etc.) for the object. The business scenarios mentioned in this application may include: any one or more of financial scenarios, gaming scenarios, and live broadcast scenarios.

[0095] 2. First sample data, unlabeled data, and second sample data.

[0096] The first sample data refers to the sample data obtained by manual annotation in the business scenario. The first sample data here may include the business data that has been annotated in the business scenario (such as business text, pictures, videos and other data in any format). The business data here may be risk control data or non-risk control data (such as game data, live broadcast data and other types of business data). If the first sample data is risk control data, then the risk control data may include: risk control text, risk control pictures, risk control videos, etc. For the sake of convenience, this application uses the first sample data as risk control text as an example. Different business scenarios correspond to different types of risk control texts. For example, in a financial scenario, the personal information of an object may belong to risk control text. In a game scenario, the internal information of the game may belong to risk control text. In a live broadcast scenario, the sensitive barrage of an object's comments may belong to risk control text.

[0097] As the name suggests, unlabeled data refers to business data that has not been manually labeled (i.e., unlabeled) in business scenarios. For example, unlabeled data includes: unlabeled financial data in financial scenarios, unlabeled game data in game scenarios, and any one or more of unlabeled live broadcast data in live broadcast scenarios.

[0098] The second sample data refers to the labeled sample data obtained by actively learning the unlabeled data through active learning. That is, the second sample data is the sample data with labeling value obtained by automatically screening from the unlabeled business data after active learning processing.

[0099] 3. Active learning processing, sample enhancement processing, and business labeling processing.

[0100] Active learning processing refers to a technology that automatically learns and labels unlabeled data. This allows the model to autonomously select the training data it uses. This allows the model to use less labeled data while achieving the same or even better results than models trained on larger datasets. Specifically, active learning processing can filter out data with labeling value from unlabeled data, thereby labeling it and obtaining sample data that is valuable for model training (i.e., providing business sample data for the model).

[0101] Sample enhancement processing refers to enhancing sample data, that is, enhancing the first sample data and the second sample data. Sample enhancement technology can provide more diverse sample enhancement data based on the existing sample data. In layman's terms, sample enhancement processing is a process of generalizing learning from existing sample data. In this application, by performing sample enhancement processing on M first sample data and K second sample data, more diverse sample data can be obtained, so that these sample data can be used to train the text discrimination model.

[0102] Business labeling processing refers to the process of labeling unlabeled data in business scenarios. During the labeling process, business labeling of unlabeled data can be performed according to business needs. For example, if the business needs are used to instruct risk control labeling, the unlabeled data can be processed for risk control labeling to obtain labeled risk control text; if the business needs are used to instruct game labeling, the unlabeled data can be processed for game labeling to obtain labeled game text.

[0103] 4. Artificial intelligence.

[0104] Artificial intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive field of computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making. AI technology is an interdisciplinary discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, pre-trained models, operating / interaction systems, and mechatronics. Pre-trained models, also known as large models or basic models, can be fine-tuned and widely applied to downstream tasks across various AI domains. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0105] The data processing scheme proposed in this application may involve machine learning technology and pre-training model technology in the field of artificial intelligence. On the one hand, this application can use pre-training model technology to obtain pre-training models in the text field such as BERT (Bidirectional Encoder Representation from Transformers) and LTD-BERT (LTD-Bidirectional Encoder Representation from Transformers). These pre-training models can then be fine-tuned to serve as text discrimination models and be quickly and widely applicable to downstream specific tasks (such as the discrimination of risk control texts). On the other hand, this application can use machine learning technology to train a basic text classification model, such as BoE, FastText, LSTM / CNN, etc., and use the trained text classification basic model as a text discrimination model. Subsequently, the text discrimination model can be used to identify risk control texts in at least one business scenario (such as a financial scenario or a live broadcast scenario).

[0106] 5. Cloud technology.

[0107] The data processing scheme proposed in this application involves a large number of data computing services and data storage services, so a large amount of computer operating costs are required. Therefore, cloud technology can be used to provide data computing services and data storage services for this scheme, so that data processing can be better performed. Specifically, based on the data computing service, active learning processing can be performed on N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and sample enhancement processing can be performed on M first sample data and K second sample data based on the data computing service; in addition, the first sample data set, the second sample data set and the model training data set can be stored based on the data storage service, so as to facilitate the subsequent training of the text discrimination model based on the model training data set. Among them, cloud technology is a general term for network technology, information technology, integration technology, management platform technology, application technology, etc. based on the cloud computing business model application, which can form a resource pool that is used on demand and is flexible and convenient. Among them, cloud technology can include cloud storage technology. Cloud storage is a new concept extended and developed from the concept of cloud computing. A distributed cloud storage system (hereinafter referred to as storage system) refers to a storage system that uses cluster applications, grid technology, and distributed storage file systems to bring together a large number of different types of storage devices (also called storage nodes) in the network through application software or application interfaces to work together and provide external data storage and business access functions.

[0108] 6. Blockchain.

[0109] Blockchain is a new application model for computer technologies, including distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a series of data blocks linked using cryptographic methods. Each block contains information about a batch of network transactions, used to verify the validity of this information (to prevent counterfeiting) and generate the next block. The following explains concepts related to blockchain systems, blockchain nodes, and block structures.

[0110] In this application, the data processing process involves a lot of data, such as: first sample data, second sample data, unlabeled data, second sample data set, and model training data set. Optionally, this application can send the above data to the blockchain for storage. Based on the blockchain's tamper-proof and traceable characteristics, the data can be prevented from being tampered with or leaked, thereby improving the data security and reliability of the data processing process.

[0111] It should be noted that many data are involved in the data processing process of this application, such as: first sample data, second sample data, unlabeled data, second sample data set, and model training data set, etc. When the above embodiments of this application are applied to specific products or technologies, user permission or consent must be obtained, and the relevant data collection, use and processing processes must comply with relevant local laws, regulations and standards, and comply with the principles of legality, legitimacy and necessity, and do not involve obtaining data types prohibited or restricted by laws and regulations. In some optional embodiments, the relevant data involved in the embodiments of this application are obtained after the object has been individually authorized. In addition, when obtaining the individual authorization of the object, the purpose of the relevant data involved must be indicated to the object.

[0112] The system architecture provided by this application is described in detail below with reference to the accompanying drawings.

[0113] See Figure 1 , Figure 1 This is a schematic diagram of the architecture of a data processing system provided in an embodiment of the present application. Figure 1 As shown, the architecture diagram of the data processing system may include at least: a terminal device cluster and a server 104. The terminal device cluster may include at least one terminal device, such as: terminal device 101, terminal device 102, terminal device 103, and so on. It should be understood that the embodiment of the present application does not specifically limit the number of terminal devices in the terminal device cluster. The number of devices can be flexibly changed according to the different needs of business scenarios, such as the number of financial devices in a financial scenario is 2, and the number of gaming devices in a gaming scenario is 3. It should be noted that any terminal device can be directly or indirectly connected to the server 104 through wired or wireless communication.

[0114] Any computer device (terminal device, or server) in the data processing system provided in this application can be a mobile phone, tablet computer, laptop computer, PDA, mobile internet device (MID), vehicle, vehicle-mounted equipment, roadside equipment, intelligent robot, aircraft, wearable device, such as smart watch, smart bracelet, pedometer and other smart devices, virtual reality device, etc.

[0115] Any computer device (terminal device, or server) in the data processing system provided in this application may also be a server. Specifically, a server may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0116] It is understood that the types of the various computer devices in the data processing system of this application can be the same or different. For example, terminal device 101 can be a mobile phone, terminal device 102 can be a gaming device, and terminal device 103 can be a live broadcast device. For another example, terminal device 101, terminal device 102, and terminal device 103 can all be live broadcast devices. This application does not limit the number and type of various computer devices in the data processing system. The following briefly describes the specific process of the data processing solution involved in the embodiment of this application, taking terminal device 101 and server 104 as an example.

[0117] ① The terminal device 101 can collect a first sample data set in a business scenario (such as a financial scenario), where the first sample data set includes M first sample data and N unlabeled data, and the first sample data includes labeled business text in the financial scenario; M and N are both positive integers and M<N.

[0118] ② The terminal device 101 sends the first sample data set to the server 104.

[0119] ③ The server 104 performs active learning processing on the N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and performs business labeling processing on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N.

[0120] ④ The server 104 performs sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and constructs a model training data set based on the second sample data set, the M first sample data and the K second sample data.

[0121] ⑤ The server 104 uses the model training data set to train the text discrimination model. Subsequently, the server 104 can send the trained text discrimination model to the terminal device 101.

[0122] ⑥ In real-time business scenarios (such as finance, live streaming, and gaming), terminal device 101 can use the trained text recognition model to identify risk control text in the corresponding business scenario. Furthermore, terminal device 101 can perform risk control processing operations (such as deletion, interception, recording, and warning) on ​​the identified risk control text in these business scenarios.

[0123] It should be noted that the above process is merely an example and does not specifically limit the execution steps of the terminal device 101 and the server 104. Optionally, the terminal device 101 may also perform active learning processing on the N unlabeled data and sample enhancement processing on the M first sample data and the K second sample data in an active learning manner; alternatively, the above process may be performed independently by the terminal device 101 or the server 104.

[0124] In one possible implementation, the data processing system provided in the present application can be deployed in a blockchain system, that is, the terminal device 101, the terminal device 101, the terminal device 103, and the server 104 can all be used as node devices in the blockchain system, and the relevant data involved in the above data processing process (such as the first sample data, the second sample data, the unlabeled data, the second sample data set, and the model training data set, etc.) are all stored on the blockchain, so that the specific processing flow of the data processing in the present application can be executed on the blockchain, which can not only ensure the fairness and justice of the data processing process, but also make the data processing process traceable, thereby improving the security and reliability of the data processing process.

[0125] It should be understood that Figure 1 Any computer device (terminal device or server) shown is integrated with a text recognition engine. The text recognition engine is used to execute the above-mentioned data processing solution of this application. The specific structure of the text recognition engine is described in detail below. Figure 2a , Figure 2a This is a structural diagram of a text recognition engine provided by an embodiment of the present application. Figure 2a As shown, the text recognition engine includes the following modules:

[0126] (1) Online modules. Specifically, the online modules may include: a pre-processing module, a multi-level cache module, a rule engine module, a discrimination module, and a scoring post-processing module.

[0127] ① Preprocessing module: responsible for data preprocessing of data requests initiated in real-time business scenarios, such as data parsing, data format conversion, data cleaning, etc.

[0128] ② Multi-level cache module: It mainly provides first-level cache and second-level cache. The first-level cache is a general cache for various business scenarios, and the second-level cache is a cache for specific business scenarios. Figure 2b , Figure 2b This is a schematic diagram of the structure of a multi-level cache module provided by an embodiment of the present application. Figure 2b As shown, within the first-level cache, it is further divided into: message type text cache, community type text cache, and global general text cache; for the second-level cache, for example, business scenario a can correspond to the business a cache submodule; and business scenario b can correspond to the business b cache submodule; and business scenario c can correspond to the business c cache submodule. In this application, the existence of a multi-level cache module greatly improves the service request processing capability. For example, in the scenario of expanding a temporary session, by constructing a cache of the top-5000 request texts (the texts of the first 5000 data requests), the average processing time of a single text of the expanded temporary session request can be reduced by about 43.6%. At the same time, due to the existence of a multi-level cache module, the speed of processing online cases fed back by the business party can also be improved, that is, the efficiency of message data can be improved.

[0129] ③Rule engine module: mainly responsible for processing some invalid request texts, thereby reducing the service pressure of the device, such as the request text of the operation activity text "XX rewarded XX yuan (100 million) red envelope".

[0130] ④ Discrimination Module: This module is primarily responsible for loading the text discrimination model (single model or combined model) trained in the offline module according to the configuration file to form a scoring unit. This unit scores the upstream input text and then passes it to the downstream module. This module may involve multiple businesses sharing a single scoring unit, or a single business corresponding to multiple scoring units.

[0131] ⑤ Scoring post-processing module: It mainly implements some heuristic post-processing strategies, which will be related to the application goals of specific business scenarios. That is, different business scenarios correspond to different post-processing strategies.

[0132] (2) Offline module. Specifically, the offline module may include: data collection module, active learning module, sample enhancement module, adversarial module, and discrimination module.

[0133] ① Data collection module: responsible for collecting business data in one or more business scenarios, and marking risk control text from the collected business data to obtain M first sample data, and then M first sample data and N unlabeled data constitute the first sample data set.

[0134] ② Active learning module: In this module, two types of active learning methods (the first active learning method and the second active learning method) are introduced. They are used to mine business data with annotation value from N unlabeled data according to the above active learning methods in specific business scenarios. On the one hand, it alleviates the problem of low proportion of risk control text in the sampled data, and on the other hand, it can also improve business results.

[0135] ③ Sample enhancement module: In this module, in addition to the basic sample enhancement method based on word replacement, a VAE-based text generation method can also be introduced. By generating sample data through VAE, it can bring more generalized content different from word replacement to risk control texts in business scenarios.

[0136] ④ Adversarial module: used for adversarial learning during the training process of the text discrimination model, which can improve the accuracy of the text discrimination model in identifying risk control texts.

[0137] ⑤Discrimination module: Mainly responsible for training to obtain a text discrimination model. The text discrimination model can include one or more types of text classification models, such as: basic text classification models such as BoE (FastText), multi-layer text classification models based on LSTM / CNN unit types, classification models based on pre-trained BERT for fine-tuning, and models based on LTD-BERT (BERT knowledge distillation).

[0138] In the system provided by the present application, a computer device (any terminal device or server) can obtain a first sample data set, which includes M first sample data and N unlabeled data, and the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N; active learning processing is performed on the N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and business labeling processing is performed on these K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N; sample enhancement processing is performed on the M first sample data and the K second sample data to obtain a second sample data set, and a model training data set is constructed based on the second sample data set, the M first sample data and the K second sample data; the text discrimination model is trained using the model training data set; wherein the trained text discrimination model is used to identify target text in at least one business scenario. It can be seen that in this application, a small amount of business sample data can be labeled in the business scenario, and then a large amount of unlabeled business data in the business scenario can be actively learned, and business data with labeling value can be screened out and labeled to obtain second sample data; and after the labeled first sample data and second sample data are enhanced, rich and diverse business sample data can be obtained. Using rich sample data to train the text discrimination model can improve the accuracy of model training, thereby improving the text discrimination model's recognition accuracy of risk control texts in various business scenarios.

[0139] It can be understood that the data processing system described in the embodiment of the present application is for the purpose of more clearly illustrating the technical solution of the embodiment of the present application, and does not constitute a limitation on the technical solution provided in the embodiment of the present application. It is known to those skilled in the art that with the evolution of the system architecture and the emergence of new business scenarios, the technical solution provided in the embodiment of the present application is also applicable to similar technical problems.

[0140] Specific embodiments of the data processing solution are described below with reference to the accompanying drawings.

[0141] See Figure 3 , Figure 3 This is a flow chart of a data processing method provided by an embodiment of the present application. Figure 1 The computer device (such as a terminal device or a server) in the data processing system shown in FIG. Figure 3 As shown, the data processing method mainly includes but is not limited to the following steps S301-S304:

[0142] S301: Obtain a first sample data set, where the first sample data set includes M first sample data and N unlabeled data, where the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N.

[0143] In this application, the marked business texts may include risk control texts and non-risk control texts (such as game type texts, live broadcast type texts, and other texts of any business type). For the sake of convenience, the subsequent embodiments of this application will use business texts as risk control texts as examples for illustration.

[0144] (1) The process of obtaining the first sample data.

[0145] In one possible implementation, ① first, corresponding business data can be collected in real time from one or more business scenarios. For example, financial data can be collected in a financial scenario (such as asset transaction data in an asset transaction scenario), game data can be collected in a game scenario (such as interaction data between game players in a real-time game), and live broadcast data can be collected in a live broadcast scenario (such as viewer comment data). ② Then, the above-mentioned collected business data is subjected to business labeling processing. The business labeling processing here can include risk control labeling processing, that is, risk control text in the corresponding business scenario is selected from the above-mentioned business data according to the business strategy for labeling processing. The labeling processing here is to add risk control tags (the tag form can be numbers, text, etc.); wherein, different business scenarios correspond to different risk control texts. For example, in a financial scenario, the personal information of the object can belong to risk control text, and in a game scenario, the internal information of the game can belong to risk control text. In a live broadcast scenario, sensitive barrage comments of the object belong to risk control text. ③ Finally, the risk control text in any business scenario after labeling can be used as a first sample data.

[0146] (2) The process of obtaining unlabeled data.

[0147] In one possible implementation, relevant business data can be collected from one or more business scenarios in real time. Unlabeled data is then obtained by preprocessing the collected business data (e.g., cleaning, format conversion, and normalization). In this implementation, real-time data collected from various business scenarios as unlabeled data ensures the timeliness of the business data and, to a certain extent, improves data accuracy.

[0148] In another possible implementation, a large amount of unlabeled business data can be directly obtained from a pre-built business database as unlabeled data. In this implementation, since the business database has been pre-built, it is possible to obtain unlabeled data more conveniently, thereby improving data acquisition efficiency.

[0149] Based on the data acquisition process shown in (1) and (2) above, the M first sample data and N unlabeled data obtained can be further constructed into a first sample data set, wherein the number M of labeled first sample data is less than the number N of unlabeled data, that is, the first sample data set obtained by the computer device includes a small amount of labeled business sample data and a large amount of unlabeled business data, which can reduce the workload of data labeling.

[0150] S302: Active learning is performed on the N unlabeled data in an active learning manner to select K unlabeled data with labeling value from the N unlabeled data, and business labeling is performed on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N.

[0151] Among them, the active learning method allows the model to independently select the training data to use, so the model can use less labeled data and achieve the same or even better results as the model trained with a larger data set. In the model training process, active learning and supervised learning methods are mainly involved. Figure 4a and Figure 4b For a comparative illustration, see Figure 4a , Figure 4a This is a diagram of a data set for a supervised learning method provided in an embodiment of the present application. Figure 4a As shown in , each data in the labeled dataset used in supervised learning has been labeled, that is, each data has a label. In addition, see Figure 4b , Figure 4b This is a data set diagram of an active learning method provided in an embodiment of the present application. Figure 4a As shown in the figure, the active learning method uses annotated data sets that usually include a part (small amount) of annotated data and a part (large amount) of unlabeled data. For example, triangle data and rectangle data are annotated data, and the rest are unlabeled data. The idea of ​​active learning is to train the model by selecting data close to the decision hyperplane for annotation, so as to achieve Figure 4a The training effect of using fully labeled data in the supervised learning method shown.

[0152] In one possible implementation, the active learning method includes a first active learning method and a second active learning method; wherein the first active learning method is used to indicate active learning based on a standard paradigm, and the second active learning method is used to indicate active learning based on feedback data. The computer device performs active learning processing on N unlabeled data according to the active learning method to screen out K unlabeled data with labeling value from the N unlabeled data, including the following two methods:

[0153] Method 1: Based on the M labeled first sample data, active learning processing is performed on the N unlabeled data according to the first active learning method to screen out K unlabeled data with labeling value from the N unlabeled data.

[0154] Method 2: Obtain business feedback data in business scenarios, and perform data generalization processing on N unlabeled data based on the business feedback data according to the second active learning method, so as to screen out K unlabeled data with labeling value from the N unlabeled data.

[0155] The specific process of active learning processing is described in detail below.

[0156] (1) The processing of the first active learning method.

[0157] See Figure 5 , Figure 5 This is a flow chart of a first active learning method provided by an embodiment of the present application, such as Figure 5 As shown, the computer device performs active learning processing on N unlabeled data according to the first active learning method based on the M labeled first sample data to screen out K unlabeled data with labeling value from the N unlabeled data. The process mainly includes the following steps ①-③:

[0158] ① Based on the M labeled first sample data, an active learner is trained. The learner here has the ability to recognize risk control text. That is, a learner that can recognize risk control text in business scenarios can be trained based on the labeled first sample data.

[0159] ② Call the trained active learner to perform prediction and scoring on N unlabeled data, obtaining the scoring results for each unlabeled data. The scoring results here can be scoring probabilities. Ideally, if the scoring probability of any data is 0, it indicates that the data is not risk control data; conversely, if the scoring probability is 1, it indicates that the data is risk control data. Typically, the scoring probability value is a score between (0, 1).

[0160] ③ Based on the scoring results of each unlabeled data, K unlabeled data with annotation value are screened from N unlabeled data according to the preset sampling strategy. Among them, the preset sampling strategy can be used to indicate that data with a scoring probability value greater than or equal to a preset threshold (such as 0.8) have annotation value, then, data with a scoring probability value less than the preset threshold (such as 0.8) do not have annotation value. Based on this sampling strategy, K unlabeled data with annotation value (K≤N) can be screened from N unlabeled data based on each scoring probability value.

[0161] Furthermore, after screening out K unlabeled data with labeling value, these K unlabeled data can be subjected to business labeling processing. The business labeling processing here can label the unlabeled data according to business needs. For example, if the business needs are used to indicate risk control labeling processing, then risk control labels can be added to the K unlabeled data with labeling value. The risk control labels here can be any format such as: numbers (if the label is 1, it indicates risk control data; if the label is 0, it indicates non-risk control data), English (if the label is "Y", it indicates risk control data; if the label is "N", it indicates non-risk control data), strings, etc. This application does not make specific restrictions on this.

[0162] As can be seen from the above, the first active learning method is an active learning method based on the standard paradigm. This method can use a small amount of labeled data to train a basic active learner, and then perform prediction and scoring on a large amount of unlabeled data in the business database. Then, the sampling strategy combined with the scoring results can be used to screen out the samples to be labeled for labeling. Finally, the labeled data can be used as sample data for subsequent model training.

[0163] (2) The processing of the second active learning method.

[0164] In specific implementation, the computer device performs data generalization processing on N unlabeled data based on business feedback data in accordance with the second active learning method, so as to screen out K unlabeled data with annotation value from the N unlabeled data. The main steps include: performing multi-dimensional text expansion processing on the business feedback data to obtain P candidate sample data, where P is a positive integer; and screening out K unlabeled data with annotation value from the P candidate sample data (for example, randomly selecting or selecting K data with higher frequency of use from the P candidate sample data as K unlabeled data with annotation value). For example, taking the business feedback data obtained in a social scenario (which can be used as the original text) as an example, the text expansion process is shown in Tables 1-1 and 1-2 below:

[0165] Table 1-1. Extended text

[0166]

[0167] Table 1-2. Extended text

[0168]

[0169] As shown in Tables 1-1 and 1-2 above, during the text expansion process, the original text can be semantically expanded (e.g., to synonyms), its data format can be expanded (e.g., converted to English), and its data type can be expanded (e.g., to emoticons or other characters). Using these expansion methods, a single original text can be expanded into multiple expanded texts, thereby enhancing data generalization capabilities.

[0170] In one possible implementation, the aforementioned dimensions include: any one or more of the scenario dimension, the data type dimension, and the timeliness dimension. The computer device performs multi-dimensional text expansion processing on the business feedback data in the financial scenario to obtain P candidate sample data, specifically including the following steps:

[0171] ① In multiple business scenarios, text expansion is performed on the business feedback data according to the scenario dimension to obtain first candidate sample data. The business scenarios include any one or more of gaming scenarios, live broadcast scenarios, and financial scenarios. For example, if the business feedback data is feedback data from a financial scenario, the feedback data can be text expanded in other business scenarios (such as gaming scenarios or live broadcast scenarios).

[0172] ② Perform text expansion on the business feedback data into multiple data types according to the data type dimension to obtain second candidate sample data. For example, if the data type of the business feedback data is Chinese, the business feedback data can be expanded into other data types such as English and French.

[0173] ③ Perform text expansion on the business feedback data based on timeliness to obtain third candidate sample data. For example, if the timeliness of business feedback data is one week, the business feedback data can be expanded to other data with a timeliness of less than one month.

[0174] ④ Based on the first candidate sample data, the second candidate sample data, and the third candidate sample data, P candidate sample data are obtained. For example, the first candidate sample data, the second candidate sample data, and the third candidate sample data can be directly combined to obtain the P candidate sample data; for another example, the P candidate sample data can be screened from the first candidate sample data, the second candidate sample data, and the third candidate sample data; for another example, the first candidate sample data, the second candidate sample data, and the third candidate sample data can be fused (e.g., concatenated) to obtain the P candidate sample data.

[0175] Based on the above steps ①-④, this application can expand the business feedback data in multiple dimensions, so that the business feedback data can be applied in multiple business scenarios, and can enhance the generalization effect of the manual sampling strategy. For the specific case data (business feedback data) reported by business colleagues, this application can combine the accumulated large amount of log data to perform the generalization process of the feedback case, and then perform risk control labeling on the batch generalized data. Finally, the labeled data can be used as sample data to improve the effect of the discrimination module.

[0176] S303: Perform sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and construct a model training data set based on the second sample data set, the M first sample data and the K second sample data.

[0177] The second sample data set includes multiple enhanced sample data after sample enhancement. In one possible implementation, the computer device constructs a model training data set based on the second sample data set, M first sample data, and K second sample data, which may include: directly combining the second sample data set, M first sample data, and K second sample data into the model training data set; or randomly selecting a specified number (e.g., 1000) of sample data from the second sample data set, M first sample data, and K second sample data to construct the model training data set.

[0178] The following describes several methods of sample enhancement processing in detail.

[0179] (1) Word replacement-based approach. Specifically, word replacement processing can be performed on M first sample data and K second sample data to obtain a second sample data set. This approach is mainly achieved by replacing words in the text with their synonyms, that is, performing synonym replacement on the text in the first sample data and the second sample data, and then adding the text after synonym replacement as enhanced sample data to the second sample data set, thereby improving the generalization ability of the model training data.

[0180] (2) Based on a pre-trained language model. Specifically, a pre-trained language model can be used to mask the M first sample data and the K second sample data to obtain a second sample data set. This method mainly draws on the mask method in BERT, that is, the BERT model is allowed to predict the masked text, thereby increasing text diversity.

[0181] (3) Translation and back-translation method. Specifically, a machine translation model can be used to translate and back-translate M first sample data and K second sample data to obtain a second sample data set. For example, if the first sample data and the second sample data are in language A, machine translation model 1 can be used to translate language A into language B, and then machine translation model 2 can be used to back-translate language B into language A'. This method mainly utilizes the differences in the generalization capabilities of different machine translation models and introduces them into the original text, thereby achieving diversity enhancement of the labeled sample data.

[0182] (4) A method for generating text based on VAE. Based on M first sample data and K second sample data, a variational autoencoder (VAE) is used to generate a second sample data set. Among them, the VAE variational autoencoder is a commonly used unsupervised learning method that can be used for data generation. In this application, based on the labeled first sample data and second sample data, more generalized enhanced sample data can be generated through the VAE generation method. For example, the process of VAE text generation can be shown in Table 2 below:

[0183] Table 2. Enhanced text generated based on VAE

[0184]

[0185] Based on the sample enhancement methods shown in (1)-(4) above, this application can provide richer and more diverse sample data for subsequent model training. Practice has shown that when conducting data evaluation in specific business scenarios, the discriminant module after sample enhancement processing can improve the accuracy (i.e., accurate recall rate) index by an average of 2 to 3 percentage points on a fixed test set.

[0186] S304: Using the model training data set to train the text discrimination model.

[0187] Specifically, the model training data set includes multiple business sample data, and the business sample data refers to any one of the first sample data, the second sample data, and the enhanced sample data, and any one of the business sample data carries a business label; for example, if the business sample data is risk control sample data, then the business label is a risk control label; if the business sample data is game sample data, then the business label is a game label. In one possible implementation, a computer device uses the model training data set to train a text discrimination model, mainly including the following steps: calling the text discrimination model to identify each business sample data in the model training data set to obtain a predicted label for each business sample data; adjusting the model parameters of the text discrimination model based on the difference between the business label and the predicted label of each business sample data; and stopping training the text discrimination model when the adjusted text discrimination model meets the model convergence condition. Among them, the text discrimination model includes: any one of the text classification basic models (such as the BoE (FastText) model based on Embedding, the RNN multi-layer model based on LSTM units, and the multi-layer model based on CNN structure), large pre-trained language models (such as the BERT model), and large pre-trained knowledge distillation models (such as the LTD-BERT model).

[0188] Specifically, the model convergence conditions here may include any of the following: when the number of training times of the text discrimination model reaches a preset threshold (such as 100 times), the text discrimination model meets the model convergence conditions; when the error between the business label and the predicted label of any business sample data is less than the error threshold, the text discrimination model meets the model convergence conditions; when the change between the predicted labels obtained from two adjacent trainings of the text discrimination model is less than the change threshold, the text discrimination model meets the model convergence conditions.

[0189] In one possible implementation, the text discrimination model includes Q different types of discriminant sub-models, where Q is a positive integer. For example, when Q = 3, the discriminant sub-models may include: a text classification base model, a large pre-trained language model, and a large pre-trained knowledge distillation model. The computer device then trains the text discrimination model using the model training dataset, specifically including the following steps:

[0190] ① Using the model training dataset, Q discriminant sub-models are trained separately to obtain Q trained discriminant sub-models. The training methods for the Q discriminant sub-models include either parallel training or serial training. Specifically, the model training dataset is used to train a text classification base model; the model training dataset is used to train a large pre-trained language model; and the model training dataset is used to train a large pre-trained knowledge distillation model. It should be understood that the training method for any discriminant sub-model can refer to the training method for the text discriminant model described above and will not be repeated here.

[0191] ② A text discriminant model is constructed based on the Q trained discriminant sub-models. The Q trained discriminant sub-models can be constructed using any of the following methods: model cascading, model selection, or model combination. For example, the Q trained discriminant sub-models can be cascaded into a text discriminant model. Alternatively, any discriminant sub-model (e.g., a large pre-trained language model) can be randomly selected from the Q trained discriminant sub-models as the text discriminant model.

[0192] In the above implementation, the text discrimination model can be composed of multiple different types of discrimination sub-models. Since different types of discrimination sub-models can be applicable to text recognition in different business scenarios, the text discrimination model obtained after training can be applicable to the recognition of risk control texts in multiple business scenarios, thereby improving the accuracy and efficiency of model recognition.

[0193] Furthermore, the trained text discrimination model is used to identify target texts in at least one business scenario, where the target texts include risk control texts and non-risk control texts (such as game type, entertainment type, live broadcast type, and any other type of non-risk control texts). That is, the trained text discrimination model has the ability to identify risk control in various business scenarios, so the trained text discrimination model can be used to identify risk control texts and non-risk control texts in various business scenarios. For example, the business scenarios here may include but are not limited to: financial scenarios, game scenarios, live broadcast scenarios, social scenarios, and any other business scenarios. Subsequently, risk control processing operations can be performed on the identified risk control texts in the above-mentioned business scenarios, such as: deleting, intercepting, recording, warning, and other operations on the risk control texts, so as to ensure the risk control security in the business scenarios and improve data security.

[0194] Optionally, since the first sample data used for model training can include both labeled business texts and labeled non-risk control texts (such as game texts), the text discrimination model trained by the scheme shown in this application is also applicable to the recognition and processing of non-risk control texts (i.e., game texts). Furthermore, the text discrimination model trained by the above scheme of this application can be applied to the recognition of risk control texts and non-risk control texts in various business scenarios. Since the text discrimination model trained by this application has high accuracy, the use of this text discrimination model can more accurately perform various types of text recognition processing (such as risk control text recognition processing) in various business scenarios.

[0195] In an embodiment of the present application, a first sample data set can be obtained, which includes M first sample data and N unlabeled data, and the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N; active learning processing is performed on the N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and business labeling processing is performed on these K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N; sample enhancement processing is performed on the M first sample data and the K second sample data to obtain a second sample data set, and a model training data set is constructed based on the second sample data set, the M first sample data and the K second sample data; the text discrimination model is trained using the model training data set; wherein the trained text discrimination model is used to identify the target text in at least one business scenario. It can be seen that in this application, a small amount of business sample data can be labeled in the business scenario, and then a large amount of unlabeled business data in the business scenario can be actively learned, and business data with labeling value can be screened out and labeled to obtain second sample data; and after the labeled first sample data and second sample data are enhanced, rich and diverse business sample data can be obtained. Using rich sample data to train the text discrimination model can improve the accuracy of model training, thereby improving the text discrimination model's recognition accuracy of business texts in various business scenarios.

[0196] See Figure 6 , Figure 6 This is a flow chart of another data processing method provided by the embodiment of the present application. Figure 1 The computer device (such as a terminal device or a server) in the data processing system shown in FIG. Figure 6 As shown, the data processing method mainly includes but is not limited to the following steps S601-S607:

[0197] S601: Acquire a first sample data set.

[0198] The first sample data set includes M first sample data and N unlabeled data. The first sample data includes labeled business text in the business scenario. M and N are both positive integers and M < N. It should be noted that the specific process of how to obtain the first sample data set can be found in detail. Figure 3 The process shown in S301 in the embodiment will not be repeated here in the embodiment of this application.

[0199] S6021: Based on the M labeled first sample data, perform active learning processing on the N unlabeled data according to a first active learning method.

[0200] In one possible implementation, a computer device performs active learning on N unlabeled data using a first active learning method based on M labeled first sample data, primarily including: ① training an active learner based on the M labeled first sample data; ② invoking the trained active learner to perform predictive scoring on the N unlabeled data, obtaining a scoring result for each unlabeled data item; and ③ based on the scoring results for each unlabeled data item, screening K unlabeled data items with labeling value from the N unlabeled data items according to a preset sampling strategy.

[0201] S6022: According to the second active learning method, data generalization processing is performed on the N unlabeled data based on the business feedback data.

[0202] In one possible implementation, a computer device performs data generalization processing on N unlabeled data based on business feedback data according to the second active learning method, primarily including: performing multi-dimensional text expansion processing on the business feedback data to obtain P candidate sample data, where P is a positive integer; and screening out K unlabeled data with annotation value from the P candidate sample data. For example, K data with high usage frequency can be randomly selected or selected from the P candidate sample data as the K unlabeled data with annotation value.

[0203] Optionally, the computer device performs multi-dimensional text expansion processing on the business feedback data to obtain P candidate sample data, which specifically includes the following steps: in multiple business scenarios, performing text expansion processing on the business feedback data according to the scenario dimension to obtain first candidate sample data; performing text expansion processing on the business feedback data of multiple data types according to the data type dimension to obtain second candidate sample data; performing text expansion processing on the business feedback data according to the timeliness dimension to obtain third candidate sample data; based on the first candidate sample data, the second candidate sample data, and the third candidate sample data, obtain P candidate sample data; and from the P candidate sample data, screen out K unlabeled data with labeling value.

[0204] S603: Perform sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set.

[0205] In one possible implementation, the sample enhancement process mainly includes any one or more of the following methods: a word replacement-based method, a pre-trained language model-based method, a translation-back translation method, and a VAE-based text generation method. Figure 3 The step S303 in the embodiment will not be repeated here in the embodiment of the present application.

[0206] S604: A model training data set is constructed based on the second sample data set, the M first sample data and the K second sample data.

[0207] In one possible implementation, the computer device can directly combine the second sample data set, M first sample data and K second sample data into a model training data set; or, the computer device can randomly select a specified number (such as 1,000) of sample data from the second sample data set, M first sample data and K second sample data to construct a model training data set.

[0208] In one possible implementation, the text discriminant model includes Q different types of discriminant sub-models, where Q is a positive integer. The following describes the training process of each discriminant sub-model in detail, using Q (Q = 3) discriminant sub-models as examples: a text classification base model, a large pre-trained language model, and a large pre-trained knowledge distillation model.

[0209] S6051: Use the model training dataset to train the basic text classification model.

[0210] Specifically, basic text classification models are widely used in various current text classification scenarios. Generally speaking, they are divided into three types: Embedding-based BoE (FastText) models, LSTM-based RNN multi-layer models, and CNN-based multi-layer models. Figure 7a , Figure 7a This is a structural diagram of a basic text classification model provided by the embodiment of the present application. Figure 7a As shown in the figure, ① the BoE (FastText) model based on Embedding is usually used as a basic benchmark, that is, for any input text, the input text can be first divided into multiple words (such as word 1, word 2, word 3, word 4, word 5), and then these words are input into the BoE model for recognition to obtain the recognition result of the input text. ② The R&D test process shows that the RNN multi-layer model based on LSTM (or GRU) units (hereinafter referred to as Bi-LSTM model) has an effect that is close to the multi-layer model based on CNN structure (hereinafter referred to as CNN model) in terms of index values ​​(such as scoring indicators) in most cases. Figure 7a As shown in the figure, for the same text (word 1 + word 2 + word 3 + word 4 + word 5), the Bi-LSTM model and the CNN model achieve similar scoring results. However, the Bi-LSTM model and the CNN model differ in how they capture textual information. For example, the CNN model is more sensitive to certain keywords in black market text. In practice, the CNN model is more effective in identifying black market text.

[0211] Taking the BoE (FastText) model as an example, the training process of the model is illustrated.

[0212] In one possible implementation, the model training data set includes multiple business sample data, where the business sample data here refers to any one of the first sample data, the second sample data, and the enhanced sample data, and any business sample data carries a business label. Specifically, the process of training the BoE model using the model training data set is as follows: ① For any business sample data, first perform word segmentation on the business sample data to obtain multiple sample words (such as word 1, word 2, word 3, word 4, word 5); ② Input each sample word into the BoE model for recognition processing to obtain the Embedding feature of each sample word (such as Embedding1, Embedding2, Embedding3, Embedding4, Embedding5); ③ Fuse the above-mentioned Embedding1, Embedding2, Embedding3, Embedding4, and Embedding5 to obtain Embedding fusion features; ④ Predict the Embedding fusion features to obtain the predicted label of the current business sample data; ⑤ Adjust the parameters of the BoE model based on the business label and predicted label of the business sample data. Similarly, the BoE model is iteratively trained until the model reaches the model convergence condition, and then the model training can be stopped.

[0213] S6052: Train a large pre-trained language model using a model training dataset.

[0214] Specifically, a large pre-trained language model can be, for example, a BERT model. The BERT model has been proven to be an effective tool for solving the semantic quantification of colloquial short texts, and it has been found that it can bring good recognition effects for downstream tasks through fine-tuning. Therefore, in the research and development of financial risk control, the industry's popular model training dataset based on a small amount of annotations is mainly used to fine-tune the BERT model, thereby obtaining an offline model with good precision (i.e., accurate recall). Among them, the above-mentioned fine-tuning process uses the BERT model trained with the model training dataset in this application, and collects a small amount of specific business annotation data for model adjustment. Please refer to Figure 7b , Figure 7b This is a schematic diagram of the structure of a large pre-trained language model provided in the embodiment of this application. Figure 7bAs shown, for any business sample data, the BERT model can be called to identify the business sample data and obtain a predicted label; then, based on the difference between the business label and the predicted label of the business sample data, the parameters of the BERT model are adjusted; and so on, the BERT model is iteratively adjusted. It should be noted that the large pre-trained language model here can also be replaced by other models. Specifically, some independently developed custom models can be used to replace the model. For example, the "Litchi" pre-training model independently developed by some teams introduced lexical knowledge, structural knowledge, and semantic knowledge during the pre-training process. The "Litchi" pre-training model has a good model effect on the CLUE evaluation indicator.

[0215] S6053: Train a large pre-trained knowledge distillation model using a model training dataset.

[0216] In one possible implementation, when the discriminant sub-model is a large pre-trained knowledge distillation model, the large pre-trained knowledge distillation model includes: a large pre-trained language model and a lightweight language model. The process of using a computer device to train the large pre-trained knowledge distillation model using a model training dataset primarily includes: using a lightweight language model to identify each piece of business sample data in the model training dataset to obtain a vectorized representation i of each piece of business sample data; identifying each piece of business sample data in the model training dataset based on the large pre-trained language model to obtain a vectorized representation j of each piece of business sample data; and using an inter-vector difference quantization loss function based on the vectorized representation i and the vectorized representation j of each piece of business sample data to train the large pre-trained knowledge distillation model.

[0217] Specifically, although the classification model obtained after fine-tuning a large pre-trained language model (such as the BERT model) can achieve good precision-recall scores in risk control text evaluation in business scenarios, considering factors such as the possible combination of multiple models and the storage occupation of other module logic in the service, as well as the high complexity and computational complexity of the BERT model itself, the service efficiency of the entire model will be very low. Therefore, this application can use a more lightweight model (such as the LTD-BERT knowledge distillation model) without reducing the model effect. For example, see Figure 7c , Figure 7c This is a schematic diagram of the structure of a large-scale pre-training knowledge distillation model provided in the embodiment of this application. Figure 7c As shown in the figure, the LTD-BERT model's concept is as follows: for a pre-trained large language model (i.e., the BERT model), the LTD-BERT model uses a relatively lightweight Bi-LSTM model to learn and extract BERT's vectorized representation for the same input text, by optimizing the loss function that quantifies the difference between vectors. For example, the loss function is:

[0218]

[0219] Based on the above-mentioned inter-vector difference quantization loss function loss, the knowledge extraction of large language models is realized. This application can use the model training data set to perform knowledge distillation of the LTD-BERT model. Experimental data show that the LTD-BERT model has a very small loss in performance indicators compared to the BERT model itself, and its storage occupancy is usually only 22M, which is 20 times lower than the storage overhead of the BERT model, and the prediction speed is nearly 80 times higher than that of the BERT model. Moreover, through the fine-tuning training model of the LTD-BERT model, for the same fixed test set, the effect of using the LTD-BERT model for text recognition is only 1 to 2 percentage points different from that of using the BERT model for text recognition. It can be seen that the LTD-BERT model can not only ensure a higher accuracy rate, but also reduce model parameters.

[0220] S606: Based on the trained text classification basic model, the large pre-trained language model, and the large pre-trained knowledge distillation model, a text discrimination model is obtained.

[0221] In specific implementation, the text discrimination model can be a single model or a combined model.

[0222] 1) If the text discrimination model is a single model, any model can be selected from the trained text classification basic model, large pre-trained language model, and large pre-trained knowledge distillation model as the text discrimination model;

[0223] 2) If the text discrimination model is a combination model, the trained text classification basic model, large pre-trained language model, and large pre-trained knowledge distillation model can be cascaded into a text discrimination model; alternatively, any two models can be selected from the trained text classification basic model, large pre-trained language model, and large pre-trained knowledge distillation model to form a text discrimination model.

[0224] S607: Using a text discrimination model to identify target text in a business scenario.

[0225] The trained text discrimination model is used to identify target text in at least one business scenario. For example, the trained text discrimination model can identify risk control text in a business scenario. In one possible implementation, a computer device can obtain a data request for a target business scenario, which includes any of a financial scenario, a gaming scenario, and a live broadcast scenario. The computer device then parses the data request to obtain the text to be recognized in the target business scenario. The computer device determines a target cache submodule that matches the target business scenario and caches the parsed text to be recognized in the target cache submodule. The computer device then calls the text discrimination model to perform recognition processing on the text to be recognized, obtaining a recognition result for the text to be recognized. The recognition result indicates whether the text to be recognized is risk control text in the target business scenario.

[0226] See Figure 8 , Figure 8 This is an online recognition flow chart of a text discrimination model provided by the embodiment of this application. Figure 8 As shown, the online identification process mainly includes the following steps S1-S6:

[0227] S1. Receive a data request. Specifically, a data request in any target business scenario such as a financial scenario, a social scenario, a game scenario, or a live broadcast scenario may be received.

[0228] S2. Parsing the data request. Specifically, the parsing method here includes but is not limited to: data packet parsing, data format conversion, and data preprocessing, etc. For example, after parsing the data packet of the data request, the text to be recognized in the target business scenario can be obtained.

[0229] S3. Cache to cache module. Specifically, a target cache submodule matching the target business scenario can be determined, and the parsed document to be identified is cached in the target cache submodule. For example, the target cache submodule can be a remote file server, through which the corresponding data is stored.

[0230] S4. The rule coverage module performs recognition. Specifically, the rule coverage module can recognize the text to be recognized. If the current text to be recognized has been recognized and processed in the previous round of recognition (that is, it indicates that the rule can cover the current text to be recognized), then the recognition result of the current text to be recognized can be directly obtained from the system data at this time, without the need for re-recognition, thereby improving processing efficiency.

[0231] S5. The text recognition model is used for recognition to obtain a recognition result. Optionally, the computer device can determine the text type of the text to be recognized in the target business scenario; if the text type of the text to be recognized is a preset text type (such as text of an operational activity type), the rule engine module can be called to perform rule processing on the text to be recognized; wherein, rule processing includes: any one of filtering processing, interception processing, and warning processing. The rule engine module can be used to process some invalid request texts, thereby reducing the service pressure of the device.

[0232] S6. Perform data post-processing on the text to be recognized based on the recognition result.

[0233] For example, if the target business scenario is a financial scenario, the recognition result includes a scoring result for the text to be recognized. If the text to be recognized is a risk control text in a financial scenario, the risk control policy associated with the financial scenario is obtained, which is used to indicate the risk control method that matches different scoring results. According to the risk control policy in the financial scenario, based on the scoring result of the text to be recognized, the target risk control method corresponding to the text to be recognized is determined; and data risk control processing is performed on the text to be recognized using the target risk control method. The data risk control processing includes any of the following: deletion processing, interception processing, recording processing, and warning processing. Among them, the scoring result can be a scoring probability value, and the risk control strategy can indicate different risk control methods corresponding to different scoring probability values, or the risk control strategy can also be used to indicate different risk control methods corresponding to different risk levels, wherein the risk level is determined based on the scoring probability value. For example, if the scoring probability value is between [0, 1], for example, the scoring probability value is [0, 0.3], then the risk level is low risk, and the corresponding risk control method is: record processing; if the scoring probability value is [0.3, 0.8], then the risk level is medium risk, and the corresponding risk control method is: warning processing; if the scoring probability value is [0.8, 1.0], then the risk level is high risk, and the corresponding risk control method is: deletion processing or interception processing. It can be seen that in the target business scenario, the target risk control method corresponding to the scoring probability value can be determined based on the risk control strategy matching the target business scenario, so that the recognized text can be subjected to data risk control processing according to the corresponding target risk control method. In this way, different risk control methods can be adaptively adopted according to the different needs of the business scenario, thereby improving the risk control effect in the business scenario.

[0234] In the embodiment of the present application, the text recognition engine provided has the following beneficial effects: 1) the discrimination module is constructed by the method of classification model and large pre-trained language model, which can effectively perform offline discrimination; 2) the construction and training of active learning modules are carried out by means of standard paradigms and manual feedback, thereby improving the accuracy of text discrimination; 3) the business pressure is reduced and processing efficiency is improved by multi-level cache; 4) the engine performance and accuracy are improved by combining offline modules and online modules; 5) the manual review cost is reduced, business efficiency is improved, and business risks are reduced by engine intelligent learning. Based on this, the present application enables text risk control recognition to be adapted to multiple business scenarios, solves the data imbalance problem through the intelligent recognition of offline modules, enables the engine to obtain intelligent recognition of new words in real time through the recognition of online modules and the application of machine learning, solves the implicit expression of data by annotating sample data, and improves the risk control recognition efficiency, recognition accuracy, and reduces the risk of financial business engines through the intelligent risk text recognition engine.

[0235] The following describes the data processing methods provided in the embodiments of the present invention.

[0236] See Figure 9 , Figure 9 This is a structural diagram of a data processing device provided in an embodiment of the present application. Figure 9 As shown, the data processing device 900 can be applied to the computer device (such as a terminal device or a server) mentioned in the above embodiments. Specifically, the data processing device 900 can be a computer program (including program code) running on a computer device, for example, the data processing device 900 is an application software; the data processing device 900 can be used to execute the corresponding steps of the data processing method provided in the embodiment of the present application. In specific implementation, the data processing device 900 can specifically include:

[0237] An acquisition unit 901 is configured to acquire a first sample data set, where the first sample data set includes M first sample data and N unlabeled data, where the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N;

[0238] Processing unit 902 is configured to perform active learning processing on the N unlabeled data in an active learning manner to select K unlabeled data with labeling value from the N unlabeled data, and perform business labeling processing on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N;

[0239] The processing unit 902 is further configured to perform sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and construct a model training data set based on the second sample data set, the M first sample data, and the K second sample data;

[0240] The processing unit 902 is further configured to train the text discrimination model using the model training data set; wherein the trained text discrimination model is used to identify target text in at least one business scenario.

[0241] In one possible implementation, the active learning method includes a first active learning method and a second active learning method; the first active learning method is used to indicate active learning based on a standard paradigm, and the second active learning method is used to indicate active learning based on feedback data; the processing unit 902 performs active learning processing on the N unlabeled data according to the active learning method to screen out K unlabeled data with labeling value from the N unlabeled data, including any of the following methods:

[0242] Based on the M labeled first sample data, active learning processing is performed on the N unlabeled data according to the first active learning method to screen out K unlabeled data with labeling value from the N unlabeled data;

[0243] Obtain business feedback data in business scenarios, and perform data generalization processing on N unlabeled data based on the business feedback data according to the second active learning method, so as to screen out K unlabeled data with labeling value from the N unlabeled data.

[0244] In one possible implementation, the processing unit 902 performs active learning processing on the N unlabeled data according to a first active learning method based on the M labeled first sample data, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and performs the following operations:

[0245] Based on the labeled M first sample data, train the active learner;

[0246] Call the trained active learner to perform prediction and scoring on N unlabeled data to obtain the scoring results of each unlabeled data;

[0247] Based on the scoring results of each unlabeled data, K unlabeled data with labeling value are screened from N unlabeled data according to the preset sampling strategy.

[0248] In one possible implementation, the processing unit 902 performs data generalization processing on the N unlabeled data based on the business feedback data in accordance with the second active learning method, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and perform the following operations:

[0249] Perform multi-dimensional text expansion processing on the business feedback data to obtain P candidate sample data, where P is a positive integer;

[0250] Based on P candidate sample data, N unlabeled data are screened to select K unlabeled data with labeling value from the N unlabeled data.

[0251] In one possible implementation, the dimensions include: any one or more of a scenario dimension, a data type dimension, and a timeliness dimension; the processing unit 902 performs multi-dimensional text expansion processing on the business feedback data in the financial scenario to obtain P candidate sample data for performing the following operations:

[0252] In multiple business scenarios, text expansion processing is performed on the business feedback data according to the scenario dimension to obtain first candidate sample data; wherein the business scenario includes: any one or more of: game scenario, live broadcast scenario, social scenario, and financial scenario;

[0253] Perform text expansion processing of multiple data types on the business feedback data according to the data type dimension to obtain second candidate sample data;

[0254] Perform text expansion processing on the business feedback data according to the timeliness dimension to obtain third candidate sample data;

[0255] Based on the first candidate sample data, the second candidate sample data, and the third candidate sample data, P candidate sample data are obtained.

[0256] In a possible implementation, the processing unit 902 performs sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set for performing any one of the following operations:

[0257] Perform word replacement processing on the M first sample data and the K second sample data to obtain a second sample data set;

[0258] Using a pre-trained language model to perform masking on the M first sample data and the K second sample data to obtain a second sample data set;

[0259] Using a machine translation model to perform back-translation processing on the M first sample data and the K second sample data to obtain a second sample data set;

[0260] Based on the M first sample data and the K second sample data, a variational autoencoder is used to generate a second sample data set.

[0261] In one possible implementation, the second sample data set includes multiple enhanced sample data, and the model training data set includes multiple business sample data, where the business sample data refers to any one of the first sample data, the second sample data, and the enhanced sample data, and any one of the business sample data carries a business label; the processing unit 902 uses the model training data set to train the text discrimination model to perform the following operations:

[0262] Call the text discrimination model to identify each business sample data in the model training data set and obtain the predicted label of each business sample data;

[0263] Adjust the model parameters of the text discrimination model based on the differences between the business labels and predicted labels of each business sample data;

[0264] When the adjusted text discrimination model meets the model convergence condition, the training of the text discrimination model is stopped; wherein the text discrimination model includes: any one of a text classification basic model, a large pre-trained language model, and a large pre-trained knowledge distillation model.

[0265] In one possible implementation, the text discrimination model includes Q different types of discrimination sub-models, where Q is a positive integer; the processing unit 902 trains the text discrimination model using the model training data set to perform the following operations:

[0266] Using the model training data set to train Q discriminant sub-models respectively, to obtain Q trained discriminant sub-models; wherein the training method of the Q discriminant sub-models includes: any one of parallel training and serial training;

[0267] A text discrimination model is constructed based on Q trained discriminant sub-models;

[0268] The Q trained discriminant sub-models are constructed in a manner including: a model cascade manner, a model selection manner, and a model combination manner.

[0269] In one possible implementation, the model training data set includes multiple business sample data, where the business sample data refers to any one of first sample data, second sample data, and enhanced sample data; the discriminant sub-model includes a large pre-trained knowledge distillation model, which includes a large pre-trained language model and a lightweight language model;

[0270] The processing unit 902 uses the model training dataset to train a large pre-trained knowledge distillation model, and is used to perform the following operations:

[0271] Use a lightweight language model to identify each business sample data in the model training data set and obtain a vectorized representation i of each business sample data;

[0272] Based on the large pre-trained language model, each business sample data in the model training data set is identified, and the vectorized representation j of each business sample data is obtained;

[0273] Based on the vectorized representation i and j of each business sample data, the inter-vector difference quantization loss function is used to train the large pre-trained knowledge distillation model.

[0274] In a possible implementation, the processing unit 902 is further configured to perform the following operations:

[0275] Obtain data requests for target business scenarios; target business scenarios include: financial scenarios, social scenarios, gaming scenarios, and live broadcast scenarios;

[0276] Parse the data request to obtain the text to be recognized in the target business scenario;

[0277] Determine the target cache submodule that matches the target business scenario, and cache the parsed document to be identified into the target cache submodule;

[0278] The text discrimination model is called to perform recognition processing on the text to be recognized to obtain the recognition result of the text to be recognized; the recognition result is used to indicate whether the text to be recognized is the risk control text in the target business scenario.

[0279] In one possible implementation, the data processing method is applied to a text recognition engine, which includes an offline module and an online module. The offline module includes a discrimination module, an active learning module, and a sample enhancement module, and the online module includes a multi-level cache module.

[0280] The discrimination module is used to provide a text discrimination model and perform business discrimination processing on each business sample data in the model training data set based on the text discrimination model;

[0281] The active learning module is used to perform active learning processing on N unlabeled data in an active learning manner;

[0282] The sample enhancement module is used to perform sample enhancement processing on the M first sample data and the K second sample data;

[0283] The multi-level cache module includes a first-level cache module and a second-level cache module. The first-level cache module is a general cache module for various business scenarios, and the second-level cache module includes multiple cache sub-modules, and one cache sub-module corresponds to one business scenario.

[0284] In one possible implementation, the target business scenario is a financial scenario, the recognition result includes a scoring result of the text to be recognized; the online module further includes a post-scoring processing module; and the processing unit 902 is further configured to perform the following operations:

[0285] If the text to be identified is a risk control text in a financial scenario, then the risk control strategy associated with the financial scenario is obtained; the risk control strategy is used to indicate the risk control method that matches different scoring results;

[0286] According to the risk control strategy in the financial scenario, based on the scoring results of the text to be recognized, the target risk control method corresponding to the text to be recognized is determined;

[0287] Use target risk control methods to perform data risk control on the identified text;

[0288] Among them, data risk control processing includes: any one of deletion processing, interception processing, recording processing, and warning processing.

[0289] In a possible implementation, the online module further includes a rule engine module; and the processing unit 902 is further configured to perform the following operations:

[0290] Determine the text type of the text to be recognized in the target business scenario;

[0291] If the text type of the text to be recognized is a preset text type, the rule engine module is called to perform rule processing on the text to be recognized;

[0292] The rule processing includes filtering, intercepting and warning.

[0293] In an embodiment of the present application, a first sample data set can be obtained, which includes M first sample data and N unlabeled data, and the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N; active learning processing is performed on the N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and business labeling processing is performed on these K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N; sample enhancement processing is performed on the M first sample data and the K second sample data to obtain a second sample data set, and a model training data set is constructed based on the second sample data set, the M first sample data and the K second sample data; the text discrimination model is trained using the model training data set; wherein the trained text discrimination model is used to identify the target text in at least one business scenario. It can be seen that in this application, a small amount of business sample data can be labeled in the business scenario, and then a large amount of unlabeled business data in the business scenario can be actively learned, and business data with labeling value can be screened out and labeled to obtain second sample data; and after the labeled first sample data and second sample data are enhanced, rich and diverse business sample data can be obtained. Using rich sample data to train the text discrimination model can improve the accuracy of model training, thereby improving the text discrimination model's recognition accuracy of risk control texts in various business scenarios.

[0294] See Figure 10 , Figure 10 : is a structural diagram of a computer device provided in an embodiment of the present application. The computer device 1000 is used to execute the steps performed by the terminal device or server in the aforementioned method embodiment. The computer device 1000 may include an independent device (such as one or more of a server, a node, a terminal, etc.), or may include components inside an independent device (such as a chip, a software module or a hardware module, etc.). The computer device may include at least one processor 1001 and a communication interface 1002. Further, optionally, the computer device may also include at least one memory 1003 and a bus 1004. In addition, the processor 1001, the communication interface 1002 and the memory 1003 are connected via a bus 1004. Wherein:

[0295] 1) The processor 1001 is a module that performs arithmetic operations and / or logical operations, and can specifically be one or more combinations of processing modules such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor unit (MPU), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a complex programmable logic device (CPLD), a coprocessor (assisting the central processing unit in completing corresponding processing and applications), and a microcontroller unit (MCU).

[0296] 2) The communication interface 1002 can be used to provide information input or output for the at least one processor 1001. Furthermore, the communication interface 1002 can be used to receive data sent externally and / or send data externally. It can be a wired link interface such as an Ethernet cable, or a wireless link interface (Wi-Fi, Bluetooth, general wireless transmission, in-vehicle short-range communication technology, or other short-range wireless communication technology). The communication interface 1002 can serve as a network interface.

[0297] 3) Memory 1003 is used to provide storage space for storing data such as the operating system and computer programs. Memory 1003 can be one or a combination of random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), or compact disc read-only memory (CD-ROM).

[0298] In a specific implementation, the processor 1001 is used to call the program instructions stored in the memory 1003 to perform the following operations:

[0299] Obtain a first sample data set, where the first sample data set includes M first sample data and N unlabeled data, where the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N;

[0300] Active learning is performed on N unlabeled data in an active learning manner to select K unlabeled data with labeling value from the N unlabeled data, and business labeling is performed on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N;

[0301] Performing sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and constructing a model training data set based on the second sample data set, the M first sample data, and the K second sample data;

[0302] A text discrimination model is trained using a model training dataset; wherein the trained text discrimination model is used to recognize target text in at least one business scenario.

[0303] In one possible implementation, the active learning mode includes a first active learning mode and a second active learning mode; the first active learning mode is used to indicate active learning based on a standard paradigm, and the second active learning mode is used to indicate active learning based on feedback data; the processor 1001 performs active learning processing on N unlabeled data according to the active learning mode to screen out K unlabeled data with labeling value from the N unlabeled data, including any of the following modes:

[0304] Based on the M labeled first sample data, active learning processing is performed on the N unlabeled data according to the first active learning method to screen out K unlabeled data with labeling value from the N unlabeled data;

[0305] Obtain business feedback data in business scenarios, and perform data generalization processing on N unlabeled data based on the business feedback data according to the second active learning method, so as to screen out K unlabeled data with labeling value from the N unlabeled data.

[0306] In one possible implementation, the processor 1001 performs active learning processing on the N unlabeled data according to a first active learning method based on the M labeled first sample data, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and performs the following operations:

[0307] Based on the labeled M first sample data, train the active learner;

[0308] Call the trained active learner to perform prediction and scoring on N unlabeled data to obtain the scoring results of each unlabeled data;

[0309] Based on the scoring results of each unlabeled data, K unlabeled data with labeling value are screened from N unlabeled data according to the preset sampling strategy.

[0310] In one possible implementation, the processor 1001 performs data generalization processing on the N unlabeled data based on the business feedback data in accordance with the second active learning mode, so as to screen out K unlabeled data with labeling value from the N unlabeled data, and performs the following operations:

[0311] Perform multi-dimensional text expansion processing on the business feedback data to obtain P candidate sample data, where P is a positive integer;

[0312] Based on P candidate sample data, N unlabeled data are screened to select K unlabeled data with labeling value from the N unlabeled data.

[0313] In one possible implementation, the dimensions include: any one or more of a scenario dimension, a data type dimension, and a timeliness dimension; the processor 1001 performs multi-dimensional text expansion processing on the business feedback data in the financial scenario to obtain P candidate sample data for performing the following operations:

[0314] In multiple business scenarios, text expansion processing is performed on the business feedback data according to the scenario dimension to obtain first candidate sample data; wherein the business scenario includes: any one or more of: game scenario, live broadcast scenario, social scenario, and financial scenario;

[0315] Perform text expansion processing of multiple data types on the business feedback data according to the data type dimension to obtain second candidate sample data;

[0316] Perform text expansion processing on the business feedback data according to the timeliness dimension to obtain third candidate sample data;

[0317] Based on the first candidate sample data, the second candidate sample data, and the third candidate sample data, P candidate sample data are obtained.

[0318] In a possible implementation, the processor 1001 performs sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, which is used to perform any one of the following operations:

[0319] Perform word replacement processing on the M first sample data and the K second sample data to obtain a second sample data set;

[0320] Using a pre-trained language model to perform masking on the M first sample data and the K second sample data to obtain a second sample data set;

[0321] Using a machine translation model to perform back-translation processing on the M first sample data and the K second sample data to obtain a second sample data set;

[0322] Based on the M first sample data and the K second sample data, a variational autoencoder is used to generate a second sample data set.

[0323] In one possible implementation, the second sample data set includes multiple enhanced sample data, and the model training data set includes multiple business sample data, where the business sample data refers to any one of the first sample data, the second sample data, and the enhanced sample data, and any one of the business sample data carries a business label; the processor 1001 uses the model training data set to train the text discrimination model to perform the following operations:

[0324] Call the text discrimination model to identify each business sample data in the model training data set and obtain the predicted label of each business sample data;

[0325] Adjust the model parameters of the text discrimination model based on the differences between the business labels and predicted labels of each business sample data;

[0326] When the adjusted text discrimination model meets the model convergence condition, the training of the text discrimination model is stopped; wherein the text discrimination model includes: any one of a text classification basic model, a large pre-trained language model, and a large pre-trained knowledge distillation model.

[0327] In one possible implementation, the text discrimination model includes Q different types of discrimination sub-models, where Q is a positive integer; the processor 1001 trains the text discrimination model using a model training dataset to perform the following operations:

[0328] Using the model training data set to train Q discriminant sub-models respectively, to obtain Q trained discriminant sub-models; wherein the training method of the Q discriminant sub-models includes: any one of parallel training and serial training;

[0329] A text discrimination model is constructed based on Q trained discriminant sub-models;

[0330] The Q trained discriminant sub-models are constructed in a manner including: a model cascade manner, a model selection manner, and a model combination manner.

[0331] In one possible implementation, the model training data set includes multiple business sample data, where the business sample data refers to any one of first sample data, second sample data, and enhanced sample data; the discriminant sub-model includes a large pre-trained knowledge distillation model, which includes a large pre-trained language model and a lightweight language model;

[0332] Processor 1001 uses a model training dataset to train a large pre-trained knowledge distillation model, and is used to perform the following operations:

[0333] Use a lightweight language model to identify each business sample data in the model training data set and obtain a vectorized representation i of each business sample data;

[0334] Based on the large pre-trained language model, each business sample data in the model training data set is identified, and the vectorized representation j of each business sample data is obtained;

[0335] Based on the vectorized representation i and j of each business sample data, the inter-vector difference quantization loss function is used to train the large pre-trained knowledge distillation model.

[0336] In a possible implementation, the processor 1001 is further configured to perform the following operations:

[0337] Obtain data requests for target business scenarios; target business scenarios include: financial scenarios, social scenarios, gaming scenarios, and live broadcast scenarios;

[0338] Parse the data request to obtain the text to be recognized in the target business scenario;

[0339] Determine the target cache submodule that matches the target business scenario, and cache the parsed document to be identified into the target cache submodule;

[0340] The text discrimination model is called to perform recognition processing on the text to be recognized to obtain the recognition result of the text to be recognized; the recognition result is used to indicate whether the text to be recognized is the risk control text in the target business scenario.

[0341] In one possible implementation, the data processing method is applied to a text recognition engine, which includes an offline module and an online module. The offline module includes a discrimination module, an active learning module, and a sample enhancement module, and the online module includes a multi-level cache module.

[0342] The discrimination module is used to provide a text discrimination model and perform business discrimination processing on each business sample data in the model training data set based on the text discrimination model;

[0343] The active learning module is used to perform active learning processing on N unlabeled data in an active learning manner;

[0344] The sample enhancement module is used to perform sample enhancement processing on the M first sample data and the K second sample data;

[0345] The multi-level cache module includes a first-level cache module and a second-level cache module. The first-level cache module is a general cache module for various business scenarios, and the second-level cache module includes multiple cache sub-modules, and one cache sub-module corresponds to one business scenario.

[0346] In one possible implementation, the target business scenario is a financial scenario, the recognition result includes a scoring result of the text to be recognized; the online module further includes a scoring post-processing module; and the processor 1001 is further configured to perform the following operations:

[0347] If the text to be identified is a risk control text in a financial scenario, then the risk control strategy associated with the financial scenario is obtained; the risk control strategy is used to indicate the risk control method that matches different scoring results;

[0348] According to the risk control strategy in the financial scenario, based on the scoring results of the text to be recognized, the target risk control method corresponding to the text to be recognized is determined;

[0349] Use target risk control methods to perform data risk control on the identified text;

[0350] Among them, data risk control processing includes: any one of deletion processing, interception processing, recording processing, and warning processing.

[0351] In a possible implementation, the online module further includes a rule engine module; and the processor 1001 is further configured to perform the following operations:

[0352] Determine the text type of the text to be recognized in the target business scenario;

[0353] If the text type of the text to be recognized is a preset text type, the rule engine module is called to perform rule processing on the text to be recognized;

[0354] The rule processing includes filtering, intercepting and warning.

[0355] In an embodiment of the present application, a first sample data set can be obtained, which includes M first sample data and N unlabeled data, and the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N; active learning processing is performed on the N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and business labeling processing is performed on these K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N; sample enhancement processing is performed on the M first sample data and the K second sample data to obtain a second sample data set, and a model training data set is constructed based on the second sample data set, the M first sample data and the K second sample data; the text discrimination model is trained using the model training data set; wherein the trained text discrimination model is used to identify the target text in at least one business scenario. It can be seen that in this application, a small amount of business sample data can be labeled in the business scenario, and then a large amount of unlabeled business data in the business scenario can be actively learned, and business data with labeling value can be screened out and labeled to obtain second sample data; and after the labeled first sample data and second sample data are enhanced, rich and diverse business sample data can be obtained. Using rich sample data to train the text discrimination model can improve the accuracy of model training, thereby improving the text discrimination model's recognition accuracy of risk control texts in various business scenarios.

[0356] In the above embodiments, the term "module" or "unit" refers to a computer program or part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories), or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the functions of the module or unit.

[0357] In addition, it should be noted here that: the embodiment of the present application also provides a computer storage medium, and a computer program is stored in the computer storage medium, and the computer program includes program instructions. When the processor executes the above program instructions, it can execute the method in the corresponding embodiment above, so it will not be described in detail here. For technical details not disclosed in the computer storage medium embodiment involved in this application, please refer to the description of the method embodiment of this application. As an example, the program instructions can be deployed on a computer device, or executed on multiple computer devices located in one place, or, executed on multiple computer devices distributed in multiple locations and interconnected by a communication network.

[0358] According to one aspect of the present application, embodiments of the present application further provide a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, so that the computer device can perform the methods described in the corresponding embodiments above. Therefore, these methods will not be described in detail here.

[0359] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted via a computer-readable storage medium. The computer instructions can be transmitted from a website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that includes one or more available media integrations. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD) or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0360] The above disclosure is only a preferred embodiment of the present application, and certainly cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope covered by the present application.

Claims

1. A data processing method, characterized in that: include: Obtain a first sample data set, where the first sample data set includes M first sample data and N unlabeled data, where the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N; Active learning is performed on the N unlabeled data in an active learning manner to select K unlabeled data with labeling value from the N unlabeled data, and business labeling is performed on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N; Performing sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and constructing a model training data set based on the second sample data set, the M first sample data, and the K second sample data; The model training data set is used to train a text discrimination model; wherein the trained text discrimination model is used to identify target text in at least one business scenario.

2. The method according to claim 1, wherein The active learning method includes a first active learning method and a second active learning method; the first active learning method is used to indicate active learning based on a standard paradigm, and the second active learning method is used to indicate active learning based on feedback data; the active learning processing of the N unlabeled data according to the active learning method to screen out K unlabeled data with labeling value from the N unlabeled data includes any of the following methods: Based on the M labeled first sample data, active learning processing is performed on the N unlabeled data according to the first active learning method to screen out K unlabeled data with labeling value from the N unlabeled data; Business feedback data in a business scenario is obtained, and according to the second active learning method, data generalization processing is performed on the N unlabeled data based on the business feedback data to screen out K unlabeled data with labeling value from the N unlabeled data.

3. The method according to claim 2, wherein The active learning process is performed on the N unlabeled data based on the M labeled first sample data according to the first active learning method to screen out K unlabeled data with labeling value from the N unlabeled data, including: Training an active learner based on the M labeled first sample data; Calling the trained active learner to perform prediction and scoring processing on the N unlabeled data to obtain scoring results for each unlabeled data; Based on the scoring results of each unlabeled data, K unlabeled data with labeling value are screened from the N unlabeled data according to a preset sampling strategy.

4. The method according to claim 2, wherein According to the second active learning method, performing data generalization processing on the N unlabeled data based on the business feedback data to screen out K unlabeled data with labeling value from the N unlabeled data includes: Performing multi-dimensional text expansion processing on the business feedback data to obtain P candidate sample data, where P is a positive integer; The N unlabeled data are screened based on the P candidate sample data to select K unlabeled data with labeling value from the N unlabeled data.

5. The method according to claim 4, wherein The dimension includes: any one or more of a scenario dimension, a data type dimension, and a timeliness dimension; the multi-dimensional text expansion processing is performed on the business feedback data to obtain P candidate sample data, including: In multiple business scenarios, text expansion processing is performed on the business feedback data according to the scenario dimension to obtain first candidate sample data; wherein the business scenario includes: any one or more of: a gaming scenario, a social scenario, a live broadcast scenario, and a financial scenario; Performing text expansion processing of multiple data types on the business feedback data according to the data type dimension to obtain second candidate sample data; Performing text expansion processing on the business feedback data according to the timeliness dimension to obtain third candidate sample data; Based on the first candidate sample data, the second candidate sample data, and the third candidate sample data, P candidate sample data are obtained.

6. The method according to claim 1, wherein The performing sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set includes any one of the following: Performing word replacement processing on the M first sample data and the K second sample data to obtain a second sample data set; Using a pre-trained language model to perform masking on the M first sample data and the K second sample data to obtain a second sample data set; Using a machine translation model to perform back-translation processing on the M first sample data and the K second sample data to obtain a second sample data set; Based on the M first sample data and the K second sample data, a second sample data set is generated using a variational autoencoder.

7. The method according to claim 1, wherein The second sample data set includes a plurality of enhanced sample data, and the model training data set includes a plurality of business sample data, wherein the business sample data refers to: any one of the first sample data, the second sample data, and the enhanced sample data, and any one of the business sample data carries a business label; The method of training the text discrimination model using the model training data set includes: Calling the text discrimination model to identify each business sample data in the model training data set to obtain a predicted label for each business sample data; Adjusting the model parameters of the text discrimination model based on the difference between the business labels and the predicted labels of each business sample data; When the adjusted text discrimination model meets the model convergence condition, the training of the text discrimination model is stopped; wherein the text discrimination model includes: any one of a text classification basic model, a large pre-trained language model, and a large pre-trained knowledge distillation model.

8. The method according to claim 7, wherein The text discrimination model includes Q different types of discrimination sub-models, where Q is a positive integer; The method of training the text discrimination model using the model training data set includes: Using the model training data set to train Q discriminant sub-models respectively, to obtain Q trained discriminant sub-models; wherein the training method of the Q discriminant sub-models includes: any one of parallel training and serial training; A text discrimination model is constructed based on the Q trained discriminant sub-models; The Q trained discriminant sub-models are constructed in a manner including: a model cascade manner, a model selection manner, and a model combination manner.

9. The method according to claim 8, wherein The model training data set includes a plurality of business sample data, wherein the business sample data refers to: any one of the first sample data, the second sample data and the enhanced sample data; The discriminant sub-model includes: a large pre-trained knowledge distillation model, and the large pre-trained knowledge distillation model includes: a large pre-trained language model and a lightweight language model; The process of training a large pre-trained knowledge distillation model using the model training dataset includes: Using a lightweight language model to identify each business sample data in the model training data set, and obtaining a vectorized representation i of each business sample data; Identify each business sample data in the model training data set based on a large pre-trained language model, and obtain a vectorized representation j of each business sample data; Based on the vectorized representation i and the vectorized representation j of each business sample data, the large-scale pre-trained knowledge distillation model is trained using an inter-vector difference quantization loss function.

10. The method according to claim 1, wherein The business text is a risk control text; The method further comprises: Obtaining data requests in a target business scenario; the target business scenario includes: any one of a financial scenario, a gaming scenario, a social scenario, and a live broadcast scenario; Parsing the data request to obtain the text to be recognized in the target business scenario; Determine a target cache submodule that matches the target business scenario, and cache the document to be identified obtained by parsing into the target cache submodule; The text discrimination model is called to perform recognition processing on the text to be recognized to obtain a recognition result of the text to be recognized; the recognition result is used to indicate whether the text to be recognized is a risk control text in the target business scenario.

11. The method according to any one of claims 1 to 10, wherein: The data processing method is applied to a text recognition engine, which includes an offline module and an online module. The offline module includes a discrimination module, an active learning module, and a sample enhancement module. The online module includes a multi-level cache module. The discrimination module is used to provide a text discrimination model and perform business discrimination processing on each business sample data in the model training data set based on the text discrimination model; The active learning module is used to perform active learning processing on the N unlabeled data in an active learning manner; The sample enhancement module is used to perform sample enhancement processing on the M first sample data and the K second sample data; The multi-level cache module includes a first-level cache module and a second-level cache module. The first-level cache module is a general cache module for various business scenarios. The second-level cache module includes multiple cache sub-modules, and one cache sub-module corresponds to one business scenario.

12. The method according to claim 11, wherein The target business scenario is a financial scenario, the recognition result includes a scoring result of the text to be recognized; the online module also includes a scoring post-processing module; the method further includes: If the text to be identified is a risk control text in the financial scenario, then obtaining a risk control strategy associated with the financial scenario; the risk control strategy is used to indicate a risk control method that matches different scoring results; According to the risk control strategy in the financial scenario, based on the scoring result of the text to be recognized, determining the target risk control method corresponding to the text to be recognized; Performing data risk control processing on the text to be identified using the target risk control method; The data risk control processing includes any one of deletion processing, interception processing, recording processing and warning processing.

13. The method according to claim 12, wherein: The online module further includes a rule engine module; the method further includes: Determine the text type of the text to be recognized in the target business scenario; If the text type of the text to be recognized is a preset text type, calling the rule engine module to perform rule processing on the text to be recognized; The rule processing includes any one of filtering processing, interception processing and warning processing.

14. A data processing device, characterized in that: include: An acquiring unit, configured to acquire a first sample data set, wherein the first sample data set includes M first sample data and N unlabeled data, wherein the first sample data includes labeled business text in a business scenario; M and N are both positive integers and M<N; a processing unit, configured to perform active learning processing on the N unlabeled data in an active learning manner to screen out K unlabeled data with labeling value from the N unlabeled data, and perform business labeling processing on the K unlabeled data to obtain K labeled second sample data, where K is a positive integer and K≤N; The processing unit is further configured to perform sample enhancement processing on the M first sample data and the K second sample data to obtain a second sample data set, and construct a model training data set based on the second sample data set, the M first sample data, and the K second sample data; The processing unit is further used to train a text discrimination model using the model training data set; wherein the trained text discrimination model is used to identify target text in at least one business scenario.

15. A computer device, characterized in that: include: storage devices and processors; a memory storing one or more computer programs; A processor, configured to load the one or more computer programs to implement the data processing method according to any one of claims 1 to 13.

16. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the data processing method according to any one of claims 1 to 13.

17. A computer program product, characterized in that The computer program product comprises a computer program, and the computer program is adapted to be loaded by a processor and execute the data processing method according to any one of claims 1 to 13.

Images