A method and system for identifying black and gray market text messages by integrating multiple information

Abstract

The present invention proposes a method for identifying black and gray market text messages that integrates multiple information, including: performing AI text restoration on the acquired original text to generate normal text, extracting features from the normal text, and then inputting the extracted feature vector into a pre-set AI text classification model to complete the normal feature classification task; performing multi-modal feature extraction on the acquired original text, and further performing multi-modal feature fusion to complete the multi-modal classification task; performing multi-modal fusion on the features of the acquired normal text and the features of the multiple information features to complete the multi-modal feature classification task; and performing weighted loss update to obtain the final multi-task learning model. By integrating multiple feature information to improve the model recognition effect of AI modeling technology and reduce the cumulative transmission of errors; at the same time, combining multi-modal fusion with multi-task learning to enhance the model effect; and then using data enhancement methods to improve the model robustness and generalization ability, and solve the pain points in traditional practices.

Description

Technical Field

[0001] The present invention belongs to the technical field of text message recognition, and in particular relates to a method and system for identifying black and gray market text messages by integrating multiple information. Background Art

[0002] The security challenges posed by the black and gray industries are becoming increasingly severe, with various crimes using internet technology occurring frequently. Chinese netizens are suffering enormous economic losses due to spam text messages, fraudulent information, and personal information leaks.

[0003] Malicious text messages have always been a key channel for black and gray industries to attract traffic. To evade detection, criminals are increasingly using variant characters to send malicious text messages, bypassing interception rules. Variant characters can be mutated in many ways, rapidly, and with varying complexity, making them difficult to identify.

[0004] Traditional industry practices involve using rules to identify variant characters and words, combined with manual review to intercept illegal content, or using illegal and harmful content detection engines to automatically intercept content on mobile devices. However, this approach, which relies on keywords and rules, is costly and time-consuming, with low recognition efficiency, poor adaptability and generalization, and inability to quickly iterate.

[0005] The latest industry practice uses AI modeling methods, such as deep learning, to restore variant text to normal text. Specifically, AI modeling technology is used to restore newly discovered variant characters in text messages to normal text without the variant characters or interfering characters, while ensuring that the normal text without the variant characters remains intact. Typically, the restored text contains common simplified Chinese characters and symbols. After the information is restored, common machine learning or deep learning methods are used for text classification to identify black and gray market information.

[0006] While AI modeling combined with text classification technology can address the pain points of traditional approaches, the results are often suboptimal, and many areas of information restoration are difficult. AI modeling alone has limited effectiveness and can reduce the effectiveness of subsequent classification models. Furthermore, the cumulative error between the two layers of the model can lead to poor overall recognition performance.

[0007] In view of this, it is very meaningful to propose a black and gray market text message identification method and system that integrates multiple information. Summary of the Invention

[0008] In order to solve the problems of low efficiency, poor adaptability and generalization, and inability to quickly iterate in the existing identification of illegal and gray market text messages, the present invention provides a method and system for identifying illegal and gray market text messages that integrates multiple information to solve the above-mentioned technical defects.

[0009] In a first aspect, the present invention proposes a method for identifying illegal and gray market text messages by integrating multiple information. The method comprises the following steps:

[0010] S1. Perform AI text restoration on the original text to generate normal text. After feature extraction of the normal text, the extracted feature vector is input into a pre-set AI text classification model to complete the normal feature classification task task 1.

[0011] S2. Extract multi-features from the original text, perform multi-feature fusion, and then input the obtained multi-feature vector into the pre-set AI text classification model to complete the multimodal classification task task2.

[0012] S3, performing multimodal fusion on the features of the normal text obtained in S1 and the multimodal information features obtained by multimodal feature fusion in S2, and then inputting the obtained multimodal vector into a pre-set AI text classification model to complete the multimodal feature classification task task3;

[0013] S4. Perform weighted loss update on the normal feature classification task task1, the multimodal classification task task2, and the multivariate feature classification task task3 to obtain a final multi-task learning model.

[0014] Preferably, the normal feature classification task task1, the multimodal classification task task2 and the multivariate feature classification task task3 are performed simultaneously.

[0015] Preferably, the AI ​​text restoration uses an AI text restoration model based on the FairSeq tool library, and the AI ​​text classification model uses ensemble learning CastBoost.

[0016] Further preferably, the multivariate feature extraction includes:

[0017] Segmenting the original text according to the Char Tokenizer Rule, extracting encoding features for each character in the sentence after segmentation, and obtaining the multi-information features of each character, wherein the multi-information features include Bert features, pinyin features, stroke features, font structure features, four-corner encoding features, Wubi features, Zhengma features, radical features, Unicode encoding features, GBK encoding features, and TFIDF features;

[0018] Use linear mapping to expand the feature dimension of the obtained multivariate information features and ensure that the mapping is in the same dimension to avoid feature offset caused by different dimensions;

[0019] The features after linear mapping are concatenated using Feature Concat to complete the multivariate feature extraction.

[0020] Further preferably, the multi-feature fusion includes: inputting the multi-features after feature concatenation Feature Concat into the network structure of the multi-feature fusion Multi Feature Fusion to obtain a final multi-feature vector output that fuses the multi-feature information, and the network structure of the multi-feature fusion Multi Feature Fusion includes a network structure consisting of a fully connected neural network, a random dropout Dropout, and an activation function tanh three-layer network.

[0021] Further preferably, the multimodal fusion includes:

[0022] Input the multivariate information features and the normal text features into the feedforward neural network FeedForward, and set random inactivation Dropout at the same time;

[0023] Then they are input together into the layer normalization layer normalize for layerNorm normalization, and then go through a layer of linear mapping (Linear);

[0024] Finally, the activation function tanh is used for nonlinear mapping to obtain the final multimodal vector output.

[0025] Further preferably, the multi-task learning includes:

[0026] The normal feature classification task Task1, the multimodal classification task Task2, and the multivariate feature classification task Task3 are performed in parallel. The three tasks are trained and updated simultaneously, and the loss is minimized after weighting. The loss of the three tasks is also minimized simultaneously. The weighted loss formula is as follows:

[0027]

[0028] α+β+γ=1 (2)

[0029] 0≤α≤1,0≤β≤1,0≤γ≤1 (3)

[0030] Formula (1) is the weighted loss at the tth moment of gradient update, where represents the weighted loss of the three tasks at time t, represents the loss of the normal feature classification task task1 at time t-1, represents the loss of the multimodal classification task task2 at time t-1, represents the loss of the multivariate feature classification task task3 at time t-1, α, β, γ are the corresponding weight parameters, and 0≤α, β, γ≤1; Formula (2) and Formula (3) are the constraints of the weight parameters α, β, γ, and the loss function The binary cross entropy loss BCE Loss is used.

[0031] Preferably, the data enhancement step is also included:

[0032] Select a normal training sample without variants, randomly select 20% of the data from this sample, and make random variant errors for each sentence in the data, including controlling the number of errors in each sentence based on the sentence length, and selecting keywords in the text based on the number of errors counted;

[0033] The number of errors count is selected as follows: when the sentence length is less than 4, the number count is 1; when the sentence length is less than 6, the number count is 2; when the sentence length is less than 35, the number count is a random integer value in the range of 2 to 1 / 3 of the sentence length + 1; when the sentence length is greater than or equal to 35, the number count is a random integer value in the range of 2 to 12;

[0034] Variant forms include: replacing keywords with pinyin, keywords mixed with symbols, replacing keywords with homophones, changing keyword shapes, splitting keywords, combining keyword sound changes with pinyin, replacing keywords with synonyms, replacing keywords with antonyms, and replacing keywords with references.

[0035] In a second aspect, the present invention further proposes a system for identifying illegal and gray market text messages that integrates multiple types of information. The system specifically includes:

[0036] Acquisition module: used to obtain original text;

[0037] Restoration module: used to restore the original text obtained by AI to generate normal text;

[0038] Feature extraction module: used to extract features from normal text and extract multi-features from the acquired original text;

[0039] Normal feature classification task module: used to input the extracted feature vector into the pre-set AI text classification model to complete the normal feature classification task task1;

[0040] Multimodal classification task module: used to input the obtained multivariate feature vector into the pre-set AI text classification model to complete the multimodal classification task task2;

[0041] Multimodal fusion module: used to perform multimodal fusion of features under these two modalities: the features of the acquired normal text and the multivariate information features obtained by fusion of multivariate features;

[0042] Multi-feature classification task module: used to input the obtained multimodal vector into the pre-set AI text classification model to complete the multi-feature classification task task3;

[0043] Weighted processing module: used to perform weighted loss update on the normal feature classification task task1, the multimodal classification task task2 and the multivariate feature classification task task3 to obtain the final multi-task learning model.

[0044] In a third aspect, an embodiment of the present invention provides an electronic device comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors implement the method described in any implementation manner in the first aspect.

[0045] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any implementation manner in the first aspect.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] (1) By using AI modeling methods such as deep learning, the variant information text is restored to normal information text, and then AI text classification is performed. On this basis, by fusing multiple information to extract the multiple features of the variant information text, and then combining multimodal fusion and multi-task learning methods, the training data is enhanced at the same time. This can not only reduce the error accumulation and transmission problem caused by simply using AI text restoration and then AI text classification; it can also improve the model recognition effect, enhance the overall recognition ability of the model, and improve the generalization ability and robustness of the model.

[0048] (2) The technical solution of the present invention not only adopts AI modeling technology, but also improves the model recognition effect of AI modeling technology by fusing multivariate feature information and reduces the cumulative transmission of errors; at the same time, it combines multimodal fusion and multi-task learning to enhance the model effect; and then improves the model robustness and generalization ability through data enhancement methods, and solves the pain points in traditional practices. It can not only achieve better technical effects, but also has good application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated into and constitute a part of this specification. The accompanying drawings illustrate the embodiments and, together with the description, serve to explain the principles of the present invention. Other embodiments and many of the expected advantages of the embodiments will be readily apparent as they become better understood by reference to the following detailed description. The elements of the drawings are not necessarily to scale with respect to each other. Like reference numerals designate corresponding similar parts.

[0050] Figure 1 is a diagram of an exemplary device architecture in which an embodiment of the present invention may be applied;

[0051] Figure 2 Schematic diagram of the flow of a method for identifying illegal and gray market text messages that integrates multiple information according to an embodiment of the present invention;

[0052] Figure 3 This is an overall flow chart of the method for identifying illegal and gray market text messages that integrates multiple information according to an embodiment of the present invention;

[0053] Figure 4 This is a flowchart of multi-feature extraction and fusion in the method for identifying illegal and gray market text messages integrating multi-information according to an embodiment of the present invention;

[0054] Figure 5 This is a structural diagram of multimodal fusion in a method for identifying illegal and gray market text messages that integrates multiple information according to an embodiment of the present invention;

[0055] Figure 6 This is a structural diagram of multi-task learning in the method for identifying illegal and gray market text messages that integrates multiple information according to an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram of the structure of the black and gray market SMS identification system integrating multiple information according to an embodiment of the present invention;

[0057] Figure 8 It is a schematic structural diagram of a computer device suitable for implementing an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In the following detailed description, reference is made to the accompanying drawings, which form a part of the detailed description and are illustrated by illustrative specific embodiments in which the present invention may be practiced. To this end, directional terms, such as "top," "bottom," "left," "right," "up," "down," etc., are used with reference to the orientation of the figures being described. Because the components of the embodiments may be positioned in several different orientations, directional terms are used for illustrative purposes and are in no way limiting. It should be understood that other embodiments may be utilized or logical changes may be made without departing from the scope of the present invention. Therefore, the following detailed description should not be adopted in a limiting sense, and the scope of the present invention is defined by the appended claims.

[0059] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as required.

[0060] Figure 1 An exemplary system architecture 100 is shown to which a method for processing information or an apparatus for processing information according to an embodiment of the present invention can be applied.

[0061] like Figure 1 As shown, system architecture 100 may include terminal devices 101, 102, 103, a network 104, and a server 105. Network 104 is a medium for providing communication links between terminal devices 101, 102, 103 and server 105. Network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables.

[0062] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0063] The terminal devices 101 , 102 , and 103 may be various electronic devices with communication functions, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0064] The server 105 may be a server that provides various services, such as a background information processing server that processes verification request information sent by the terminal devices 101, 102, and 103. The background information processing server may analyze and perform other processing on the received verification request information and obtain a processing result (e.g., verification success information indicating that the verification request is a legitimate request).

[0065] It should be noted that the method for processing information provided in the embodiment of the present invention is generally executed by the server 105, and accordingly, the device for processing information is generally set in the server 105. In addition, the method for sending information provided in the embodiment of the present invention is generally executed by the terminal devices 101, 102, and 103, and accordingly, the device for sending information is generally set in the terminal devices 101, 102, and 103.

[0066] It should be noted that the server can be either hardware or software. When the server is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When the server is software, it can be implemented as multiple software programs or software modules (for example, to provide distributed services), or as a single software program or multiple software modules, without specific limitation herein.

[0067] Malicious text messages have always been a key channel for black and gray industries to attract traffic. To evade detection, criminals are increasingly using variant characters to send malicious text messages, bypassing interception rules. Variant characters can be mutated in many ways, rapidly, and with varying complexity, making them difficult to identify.

[0068] Manual keyword and rule matching methods suffer from pain points such as low recognition efficiency and an inability to quickly iterate and adapt to new changes. AI technology offers capabilities and efficiency that cannot be matched by humans. It can efficiently identify, quickly iterate, and adapt to new changes, resolving user pain points.

[0069] Therefore, the solution of the present invention, while adopting AI modeling technology, improves the model recognition effect of AI modeling technology by integrating multivariate feature information, reducing the cumulative transmission of errors; at the same time, it combines multimodal fusion and multi-task learning to enhance the model effect; and further improves the model robustness and generalization ability through data enhancement methods, and solves the pain points of traditional practices. The main improvements of the solution of the present invention are as follows:

[0070] 1) Multi-dimensional information fusion: Based on the original text -> AI text restoration -> AI text classification, a parallel feature fusion process is added, namely, original text -> multi-dimensional feature extraction -> multi-dimensional feature fusion -> AI text classification. By adding this process, the original text information is directly transferred to the AI ​​text classification model through multi-dimensional feature information, reducing the error transmission caused by AI text restoration;

[0071] 2) Multimodal fusion: This involves mapping the information from both modalities into a unified multimodal vector space. This involves mapping the multimodal feature fusion information and the normal text information restored by AI text into a unified vector space, which is then combined with the AI ​​text classification model for joint training to improve model performance.

[0072] 3) Multi-task learning: Build separate AI text classification model tasks for information in the two modalities, and then combine the multimodal fusion AI text classification task to perform weighted loss updates to enhance the model effect;

[0073] 4) Data Augmentation: We select a normal training sample without variants, randomly select 20% of the data from this sample, and perform random variant errors on each sentence in the data. We control the number of errors in each sentence based on the sentence length. This ensures that the model can learn as many errors as possible from the variant errors, improving the model's robustness and generalization performance.

[0074] Firstly, Figure 2 The embodiment of the present invention discloses a method for identifying black and gray market short messages by integrating multiple information. Figure 2 As shown, the method includes the following steps:

[0075] S101, performing AI text restoration on the acquired original text to generate normal text, extracting features from the normal text, and inputting the extracted feature vector into a pre-set AI text classification model to complete normal feature classification task 1;

[0076] S102: extract multi-features from the original text, perform multi-feature fusion, and then input the obtained multi-feature vector into a pre-set AI text classification model to complete the multimodal classification task task 2;

[0077] S103: Perform multimodal fusion on the features of the normal text obtained in S1 and the multimodal information features obtained by multimodal feature fusion in S2, and then input the obtained multimodal vector into a pre-set AI text classification model to complete the multimodal feature classification task 3;

[0078] S104: performing weighted loss update on the normal feature classification task task1, the multimodal classification task task2, and the multivariate feature classification task task3 to obtain a final multi-task learning model.

[0079] It also includes a data enhancement step S105:

[0080] Select a normal training sample without variants, randomly select 20% of the data from this sample, and make random variant errors for each sentence in the data, including controlling the number of errors in each sentence based on the sentence length, and selecting keywords in the text based on the number of errors counted;

[0081] The number of errors count is selected as follows: when the sentence length is less than 4, the number count is 1; when the sentence length is less than 6, the number count is 2; when the sentence length is less than 35, the number count is a random integer value in the range of 2 to 1 / 3 of the sentence length + 1; when the sentence length is greater than or equal to 35, the number count is a random integer value in the range of 2 to 12;

[0082] Variant forms include: replacing keywords with pinyin, keywords mixed with symbols, replacing keywords with homophones, changing keyword shapes, splitting keywords, combining keyword sound changes with pinyin, replacing keywords with synonyms, replacing keywords with antonyms, and replacing keywords with references.

[0083] Specifically, in this embodiment, the overall process of the technical solution of the present invention includes three parallel multi-task learning processes, such as Figure 3 shown.

[0084] 1) The first step is to generate normal text through AI text restoration, extract features, and then feed the feature vector into the AI ​​text classification model;

[0085] 2) The second method is to extract multi-features from the original text, fuse the multi-features, and then feed the multi-feature vectors into the AI ​​text classification model;

[0086] 3) The third item is divided into two parallel sub-processes, including:

[0087] First, normal text is generated through AI text restoration, and normal text features are obtained after feature extraction; secondly, multi-feature extraction is performed on the original text, and multi-feature fusion is performed to obtain multi-information features; finally, the features under the two modes of normal text features and multi-information features are multimodally fused, and the obtained multimodal vector is sent to the AI ​​text classification model.

[0088] Finally, the three parallel AI text classification tasks are weighted and updated to obtain the final multi-task learning model.

[0089] The AI ​​text restoration model is essentially a sequence-to-sequence machine translation task, for which many open-source libraries exist, such as OpenNMT, tensor2tensor, and FairSeq. The technical solution of this invention recommends using the FairSeq tool library, a high-level library developed based on PyTorch that includes many classic models and facilitates development and debugging.

[0090] For AI text classification models, the technical solution of this invention recommends using ensemble learning CastBoost. Other text classification models such as logistic regression, support vector machine, or TextCNN can also be used.

[0091] Further, Figure 4 The flowchart of multi-feature extraction and fusion in the technical solution of the present invention is shown as follows: Figure 4 shown.

[0092] Among them, the extraction of multivariate features includes:

[0093] 1) The original text is first segmented according to the Char Tokenizer Rule. After segmentation, encoding features are extracted for each character in the sentence to obtain multi-dimensional information features of each character. These include Bert features, pinyin features, stroke features, font structure features, four-corner encoding features, Wubi features, Zhengma features, radical features, Unicode encoding features, GBK encoding features, and TFIDF features. This allows us to extract as many multi-dimensional information features of the original text as possible.

[0094] 2) Then, linear mapping is used to expand the feature dimensions of these features and ensure that they are mapped to the same dimension to avoid feature offset caused by different dimensions;

[0095] 3) Finally, the features after linear mapping (Linear) are concatenated (Feature Concat) to complete the extraction of multivariate features.

[0096] Among them, the character tokenizer rule (Char Tokenizer Rule): Chinese is segmented by characters; English is segmented by English words; other symbols are segmented by single characters.

[0097] Multi-feature fusion involves inputting the multi-features obtained after feature concatenation into a multi-feature fusion network structure, resulting in a final multi-feature vector output that fuses the multi-faceted information. The multi-feature fusion network consists of three layers: a fully connected neural network, dropout, and a tanh activation function. This multi-feature fusion network is trained and optimized simultaneously with the model.

[0098] Furthermore, the multimodal fusion structure diagram of the technical solution of the present invention is as follows: Figure 5 shown.

[0099] Specifically, the multivariate information features and normal text features are fed into a feedforward neural network (FeedForward) with dropout applied. They are then fed into a layer normalization (Layer Normalization) for layer norm normalization. Furthermore, they undergo a linear mapping layer, and finally, a tanh activation function is used for nonlinear mapping to produce the final multimodal vector output. This multimodal fusion structure is optimized during model training.

[0100] Multimodal fusion offers significant advantages, as it can leverage data from both modalities, enabling more robust model predictions. Furthermore, the complementary information between different modalities can effectively enhance model performance.

[0101] Furthermore, the multi-task learning structure diagram of the technical solution of the present invention is as follows: Figure 6 shown.

[0102] Specifically, multi-task learning involves three parallel classification tasks: normal feature classification (Task1), multimodal classification (Task2), and multivariate feature classification (Task3). These three related tasks are learned simultaneously, trained and updated simultaneously, and the weighted loss is minimized, minimizing the loss across all three tasks simultaneously.

[0103] Among them, the weighted loss formula is as follows:

[0104]

[0105] α+β+γ=1 (2)

[0106] 0≤α≤1,0≤β≤1,0≤γ≤1 (3)

[0107] Formula (1) is the weighted loss at the tth moment of gradient update, where represents the weighted loss of the three tasks at time t, represents the loss of the normal feature classification task task1 at time t-1, represents the loss of the multimodal classification task task2 at time t-1, represents the loss of the multi-feature classification task task3 at time t-1, α, β, γ are the corresponding weight parameters, and 0≤α, β, γ≤1; formula (2) and formula (3) are the constraints of the weight parameters α, β, γ. Since it is a binary classification, the loss function The binary cross entropy loss (BCE Loss) is used.

[0108] In this example, after multi-task learning, only the results of multimodal classification task 2 are output. The goal is to minimize the losses of all three tasks simultaneously and minimize the weighted loss. In other words, the three tasks complement and promote each other's learning, sharing loss values ​​and updating parameters to enhance model performance and improve generalization.

[0109] Furthermore, the data enhancement in the technical solution of the present invention adopts the following method:

[0110] This involves selecting a normal training sample without variants, randomly selecting 20% ​​of the data from this sample, and randomly making errors in each sentence. The number of errors in each sentence is controlled based on the sentence length, and the keywords in the text are selected based on the number of errors counted.

[0111] The number of errors count is selected as follows: when the sentence length is less than 4, the number count is 1; when the sentence length is less than 6, the number count is 2; when the sentence length is less than 35, the number count is a randomly selected integer value in the range of 2 to 1 / 3 of the sentence length + 1; when the sentence length is greater than or equal to 35, the number count is a randomly selected integer value in the range of 2 to 12.

[0112] Variant forms include: replacing keywords with pinyin, keywords mixed with symbols, replacing keywords with homophones, keyword deformation, keyword splitting, keyword sound change + pinyin combination, keyword synonym replacement, keyword antonym replacement, and keyword reference replacement.

[0113] The present invention uses AI modeling methods such as deep learning to restore variant information text to normal information text, and then connects it to AI text classification. On this basis, it fuses multiple information to extract multiple features of the variant information text, and then combines multimodal fusion and multi-task learning methods. At the same time, data enhancement is performed on the training data. This can not only reduce the error accumulation and transmission problem caused by simply using AI text restoration and then connecting it to AI text classification; it can also improve the model recognition effect, enhance the overall recognition ability of the model, and improve the generalization ability and robustness of the model.

[0114] In the second aspect, the present invention also proposes a black and gray market SMS identification system that integrates multiple information, such as Figure 7 As shown, the system specifically includes: an acquisition module 71, a restoration module 72, a feature extraction module 73, a normal feature classification task module 74, a multimodal classification task module 75, a multimodal fusion module 76, a multivariate feature classification task module 77 and a weighted processing module 78.

[0115] In this embodiment, the acquisition module 71 is used to obtain the original text; the restoration module 72 is used to perform AI text restoration on the obtained original text to generate normal text; the feature extraction module 73 is used to extract features of the normal text and perform multi-feature extraction on the obtained original text; the normal feature classification task module 74 is used to input the extracted feature vector into a preset AI text classification model to complete the normal feature classification task task1; the multimodal classification task module 75 is used to input the obtained multi-feature vector into a preset AI text classification model to complete the multimodal classification task task2; the multimodal fusion module 76 is used to perform multimodal fusion on the features of the obtained normal text and the multi-information features obtained by multi-feature fusion; the multi-feature classification task module 77 is used to input the obtained multi-modal vector into a preset AI text classification model to complete the multi-feature classification task task3; the weighted processing module 78 is used to perform weighted loss update on the normal feature classification task task1, the multimodal classification task task2 and the multi-feature classification task task3 to obtain the final multi-task learning model.

[0116] Reference below Figure 8 , which shows an electronic device (eg Figure 1 A schematic structural diagram of a computer device 800 (a server or terminal device as shown). Figure 8 The electronic device shown is only an example and should not limit the functions and scope of use of the embodiments of the present invention.

[0117] like Figure 8 As shown, the computer device 800 includes a central processing unit (CPU) 801 and a graphics processing unit (GPU) 802, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 803 or the program loaded from the storage part 809 to the random access memory (RAM) 806. Various programs and data required for the operation of the device 800 are also stored in the RAM 804. The CPU 801, GPU 802, ROM 803 and RAM 804 are connected to each other via a bus 805. An input / output (I / O) interface 806 is also connected to the bus 805.

[0118] The following components are connected to the I / O interface 806: an input section 807 including a keyboard, a mouse, and the like; an output section 808 including a display such as a liquid crystal display (LCD), a speaker, and the like; a storage section 809 including a hard disk and the like; and a communication section 810 including a network interface card such as a LAN card or a modem. The communication section 810 performs communication processing via a network such as the Internet. A drive 811 may also be connected to the I / O interface 806 as needed. A removable medium 812, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like, is installed in the drive 811 as needed, so that a computer program read therefrom can be installed into the storage section 809 as needed.

[0119] In particular, according to an embodiment disclosed in the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment disclosed in the present invention includes a computer program product comprising a computer program carried on a computer-readable medium, the computer program comprising a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication portion 810, and / or installed from a removable medium 812. When the computer program is executed by a central processing unit (CPU) 801 and a graphics processing unit (GPU) 802, the above-mentioned functions defined in the method of the present invention are performed.

[0120] It should be noted that the computer-readable medium described in the present invention may be a computer-readable signal medium or a computer-readable medium, or any combination thereof. Computer-readable media may be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, apparatuses, or components, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to, an electrical connection having one or more conductors, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution device, apparatus, or component. In the present invention, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable medium that can transmit, propagate, or transport a program for use by or in conjunction with an instruction execution apparatus, device, or device. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wireline, optical cable, RF, or any suitable combination thereof.

[0121] Computer program code for performing the operations of the present invention may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on the remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., through the Internet using an Internet service provider).

[0122] The flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the devices, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the module, program segment, or a part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order than that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of the boxes in the block diagram and / or flowchart, can be implemented with a dedicated hardware-based device that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0123] The modules involved in the embodiments of the present invention may be implemented in software or hardware, and may also be set in a processor.

[0124] As another aspect, the present invention further provides a computer-readable medium, which may be included in the electronic device described in the above embodiments, or may exist independently and not incorporated into the electronic device. The computer-readable medium carries one or more programs, and when executed by the electronic device, the electronic device performs the method steps described in the first aspect of the present invention.

[0125] The above description is merely an illustration of the preferred embodiments of the present invention and the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in the present invention.

Claims

1. A method for identifying black and gray market text messages by integrating multiple information, characterized in that: The method comprises the following steps: S1. Perform AI text restoration on the original text to generate normal text. After feature extraction of the normal text, the extracted feature vector is input into a pre-set AI text classification model to complete the normal feature classification task task 1. S2. Extract multi-features from the original text, perform multi-feature fusion, and then input the obtained multi-feature vector into the pre-set AI text classification model to complete the multimodal classification task task2. S3, performing multimodal fusion on the features of the normal text obtained in S1 and the multimodal information features obtained by multimodal feature fusion in S2, and then inputting the obtained multimodal vector into a pre-set AI text classification model to complete the multimodal feature classification task task3; S4, performing weighted loss update on the normal feature classification task task1, the multimodal classification task task2, and the multivariate feature classification task task3 to obtain a final multi-task learning model; The normal feature classification task task1, the multimodal classification task task2, and the multivariate feature classification task task3 are performed simultaneously; the AI ​​text restoration uses an AI text restoration model based on the FairSeq tool library, and the AI ​​text classification model uses ensemble learning CastBoost; The multivariate feature extraction includes: segmenting the original text according to the Char Tokenizer Rule, extracting the encoding features of each character in the sentence after segmentation, and obtaining the multivariate information features of each character, wherein the multivariate information features include Bert features, pinyin features, stroke features, font structure features, four-corner encoding features, Wubi features, Zhengma features, radical features, Unicode encoding features, GBK encoding features, and TFIDF features; using linear mapping Linear to expand the feature dimensions of the obtained multivariate information features and ensure that the mapping is to the same dimension; performing feature concat on the features after linear mapping Linear to complete the multivariate feature extraction.

2. The method for identifying black and gray market text messages by integrating multiple information according to claim 1 is characterized in that: The multi-feature fusion includes: inputting the multi-features after feature concatenation Feature Concat into the network structure of the multi-feature fusion Multi Feature Fusion to obtain the final multi-feature vector output that fuses the multi-feature information. The network structure of the multi-feature fusion Multi Feature Fusion includes a network structure consisting of a fully connected neural network, a random dropout, and an activation function tanh three-layer network.

3. The method for identifying black and gray market text messages by integrating multiple information according to claim 2 is characterized in that: The multimodal fusion includes: Input the multivariate information features and the normal text features into the feedforward neural network respectively, and set random inactivation Dropout at the same time; Then they are input together into the layer normalization layer normalize for layerNorm normalization, and then go through a layer of linear mapping (Linear); Finally, the activation function tanh is used for nonlinear mapping to obtain the final multimodal vector output.

4. The method for identifying black and gray market text messages by integrating multiple information according to claim 3 is characterized in that: The multi-task learning includes: The loss is weighted and minimized, and the losses of the three tasks are also minimized at the same time; the weighted loss formula is as follows: ， Formula (1) is the weighted loss at the tth moment of gradient update, where represents the weighted loss of the three tasks at time t, represents the loss of the normal feature classification task task1 at time t-1, represents the loss of the multimodal classification task task2 at time t-1, represents the loss of multi-feature classification task task3 at time t-1, are the corresponding weight parameters respectively; formula (2) and formula (3) are weight parameters The constraints of the loss function The binary cross entropy loss BCE Loss is used.

5. The method for identifying black and gray market text messages by integrating multiple information according to claim 1 is characterized in that: Also includes data augmentation steps: Select a normal training sample without variants, randomly select 20% of the data from this sample, and make random variant errors for each sentence in the data, including controlling the number of errors in each sentence based on the sentence length, and selecting keywords in the text based on the number of errors counted; The number of errors count is selected as follows: when the sentence length is less than 4, the number count is 1; when the sentence length is less than 6, the number count is 2; when the sentence length is less than 35, the number count is a random integer value in the range of 2 to 1 / 3 of the sentence length + 1; when the sentence length is greater than or equal to 35, the number count is a random integer value in the range of 2 to 12; Variant forms include: replacing keywords with pinyin, keywords mixed with symbols, replacing keywords with homophones, changing keyword shapes, splitting keywords, combining keyword sound changes with pinyin, replacing keywords with synonyms, replacing keywords with antonyms, and replacing keywords with references.

6. A black and gray market SMS identification system integrating multiple information, characterized by: The method according to any one of claims 1 to 5, wherein the system specifically comprises: Acquisition module: used to obtain original text; Restoration module: used to restore the original text obtained by AI to generate normal text; Feature extraction module: used to extract features from normal text and extract multi-features from the acquired original text; Normal feature classification task module: used to input the extracted feature vector into the pre-set AI text classification model to complete the normal feature classification task task1; Multimodal classification task module: used to input the obtained multivariate feature vector into the pre-set AI text classification model to complete the multimodal classification task task2; Multimodal fusion module: used to perform multimodal fusion of features under these two modalities: the features of the acquired normal text and the multivariate information features obtained by fusion of multivariate features; Multi-feature classification task module: used to input the obtained multimodal vector into the pre-set AI text classification model to complete the multi-feature classification task task3; Weighted processing module: used to perform weighted loss update on the normal feature classification task task1, the multimodal classification task task2 and the multivariate feature classification task task3 to obtain the final multi-task learning model.

7. An electronic device comprising: one or more processors; a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images