Bad short message identification method and device based on large language model error correction

Abstract

The invention discloses a bad short message recognition method and device based on big language model error correction, belongs to the technical field of content security, and particularly relates to bad short message recognition based on big language model error correction. The problems that according to an existing bad short message recognition method, the recognition accuracy of a trained recognition model is reduced due to the fact that a lawbreaker continuously updates a corpus through the sound-form-meaning replacement technology, and new data needs to be continuously collected and manually corrected are solved. The method comprises the following steps of: performing text error correction on a text with a high text error rate based on a large language model; performing vectorization processing on the text with the low text error rate or the text after error correction; and a step of identifying the text vector by using the MacBERT identification model subjected to fine adjustment, and judging whether the short message is a bad short message or not. The bad short message recognition method and device based on large language model error correction are suitable for recognition of bad short messages.

Description

Technical Field

[0001] The present invention relates to the technical field of content security, and in particular to bad short message identification based on large language model error correction. Background Art

[0002] With the rapid development of information technology and the widespread popularity of smart phones, various social platforms have emerged rapidly. Short messages, as an important form of daily communication, have profoundly changed the way information is transmitted. These short messages can exist in the form of text, pictures, and audio. Their convenience and efficiency make the transmission of information between users faster and more efficient. However, it is precisely because of this convenience that various bad short messages have also flooded, which not only interferes with the normal communication experience of users, but also may cause property losses, personal privacy leaks, and even endanger the normal order of society. Therefore, providing a bad short message identification method with high recognition accuracy is of great significance to protecting the public's information security and property security.

[0003] In view of the significant threat posed by bad short messages, those skilled in the art have proposed a series of bad short message identification methods. Currently, the most widely used short message identification method is the bad short message identification method based on keyword matching. This method predefines a certain number of "risk keywords". When it matches one or more of these risk keywords in a short message, it determines that the message is a bad short message. The accuracy of this method depends on the predefined keyword library, and there are problems such as over-reliance on manual keyword extraction and high false negative rate.

[0004] With the continuous development of machine learning technology and deep learning technology, various bad short message recognition models based on machine learning or deep learning have emerged one after another. At present, methods based on deep learning have been widely used in bad short message recognition, and can effectively recognize traditional bad short messages (i.e. bad short messages with correct text or low text error rate). However, in order to evade detection, criminals generally use near-sound, similar, synonymous replacement techniques and frequently update replacement corpora, so that the replaced text contains a large number of words that have not appeared in the training set, which poses a challenge to the effectiveness of the previously trained recognition model. In addition, the replaced text is generally uncommon characters, special characters, etc. If the existing model wants to improve the recognition rate, it can only continuously collect new data and manually correct and retrain the model. However, criminals can use simple scripts to implement large-scale sound, shape and meaning replacement operations to evade the trained bad short message recognition model, and the technical cost is low. In comparison, the technical cost of building the recognition model of the detection party is very high. The collection, cleaning, error correction and labeling of bad short messages for training require a lot of manpower and material resources, and there is a certain time lag between the start of model training and its actual application. During this period, criminals may have updated their corpus based on the sound, shape and meaning replacement technology, making the newly trained model still unable to accurately identify new bad short messages.

[0005] To sum up, there is an inequality in technical implementation costs between criminals and detection parties, which leads to the inability of existing bad short message identification methods to adapt to the impact of criminals using sound, shape and meaning replacement technology to update the corpus during the confrontation process of bad short message identification and anti-identification, resulting in a continuous decrease in its identification accuracy.

[0006] Meanwhile, existing methods for identifying bad short messages mainly focus on identifying text short messages, but pay insufficient attention to identifying bad short messages carried in the form of voice and pictures.

[0007] In summary, the problem of reduced recognition accuracy of trained recognition models due to the continuous updating of corpus by criminals using the sound, shape and meaning replacement technology, and the need to continuously collect new data and perform manual correction, is a key problem that needs to be solved in the current field of bad short message recognition. Based on this, there is an urgent need for a recognition method or model that can accurately identify bad short messages with high text error rates created by criminals using the sound, shape and meaning replacement technology to frequently update the corpus without retraining. Summary of the invention

[0008] The present invention proposes a method and device for identifying bad short messages based on large language model error correction, which solves the problem that the recognition accuracy of the trained recognition model is reduced due to the fact that criminals use sound, shape and meaning replacement technology to continuously update the corpus, and new data needs to be continuously collected and manually corrected in the existing bad short message recognition method.

[0009] The bad short message identification method based on large language model error correction described in the present invention has the following technical solutions:

[0010] The method comprises the following steps:

[0011] A step for monitoring in real time the short messages carried in the form of text, sound or pictures received in the terminal;

[0012] A step for obtaining text content in the short message;

[0013] A step for evaluating the text error rate of the acquired text content to determine whether the text content is a high text error rate text or a low text error rate text;

[0014] A step for performing text error correction on a text with a high text error rate based on a large language model to obtain a corrected text;

[0015] A step for performing vectorization processing on a text with a low text error rate or a text after error correction to obtain a text vector;

[0016] The step is used to use the fine-tuned MacBERT recognition model to recognize the text vector and determine whether the short message corresponding to the text vector is a bad short message.

[0017] Further, a preferred implementation is provided, wherein the step of obtaining the text content in the short message comprises:

[0018] For short messages carried in the form of sound or picture, the text content is obtained after the short messages are converted into text.

[0019] Furthermore, a preferred embodiment is provided, wherein the step of evaluating the text error rate of the acquired text content and determining whether the text content is a high text error rate text or a low text error rate text comprises:

[0020] Compare the text in the obtained text content with the text in the given text collection:

[0021] If the proportion of words not in the given word set in the acquired text content is greater than a given threshold, the acquired text content is a high text error rate text;

[0022] Otherwise, it is low text error rate text.

[0023] Furthermore, a preferred embodiment is provided, wherein the step of performing text error correction on a text with a high text error rate based on a large language model to obtain the corrected text comprises:

[0024] Get the engineering prompt word template for text error correction;

[0025] Embed high text error rate text into the engineering prompt word template for text error correction to form a complete input instruction input language model;

[0026] The large language model performs text correction on the text with a high text error rate according to the complete input instructions to obtain the corrected text.

[0027] Furthermore, a preferred embodiment is provided, wherein the step of performing text correction on a text with a high text error rate based on a large language model to obtain the corrected text further comprises a filtering method based on text length comparison:

[0028] Compare the error-corrected text output by the large language model with the corresponding high text error rate text. If the difference between the two text lengths is greater than 5 characters, the error-corrected text is an unexpected output.

[0029] When the corrected text is unexpected output:

[0030] The high text error rate text is again embedded into the engineering prompt word template for text error correction to form a complete input instruction input large language model, and emphasizes that only the text content after error correction is output.

[0031] The present invention also proposes a bad short message identification device based on large language model error correction, and its technical solution is as follows:

[0032] The device comprises the following modules:

[0033] Multimodal short message real-time monitoring module: used to monitor the short messages received by the terminal in the form of text, sound or pictures in real time;

[0034] Multimodal short message texting module: used to obtain the text content in the short message;

[0035] A text error rate evaluation module: used to evaluate the text error rate of the acquired text content, and determine whether the text content is a high text error rate text or a low text error rate text;

[0036] Text error correction module based on large language model: used to perform text error correction on text with high text error rate based on large language model to obtain the corrected text;

[0037] Text vectorization module: used to vectorize text with low text error rate or text after error correction to obtain text vectors;

[0038] Bad short message identification module: used to identify text vectors using the fine-tuned MacBERT recognition model, and determine whether the short message corresponding to the text vector is a bad short message.

[0039] Further, a preferred embodiment is provided, wherein the device further comprises a risk warning module;

[0040] The risk warning module is used to send a warning message according to the identified bad short messages.

[0041] In this implementation, a prompt message is sent out by way of a system pop-up window warning prompt to remind the user that a bad short message has been found, thereby enhancing the user's alertness to potential threats and ensuring the security of personal information.

[0042] In this implementation manner, when the identification result is a normal short message, the risk warning module does not take any measures and does not affect the user's normal communication experience.

[0043] The present invention also proposes a computer device, and its technical solution is as follows:

[0044] A computer device comprises: a processor and a memory, wherein the memory is used to store executable instructions of the processor, and the processor is configured to execute the above-mentioned bad short message identification method based on large language model error correction by executing the executable instructions.

[0045] The present invention also proposes a computer storage medium, and its technical solution is as follows:

[0046] A computer storage medium stores a computer program, and when the computer program is run, the bad short message identification method based on large language model error correction is executed.

[0047] The present invention also proposes a computer program product, and its technical solution is as follows:

[0048] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the steps of the above-mentioned bad short message identification method based on large language model error correction.

[0049] The present invention has the following beneficial effects:

[0050] 1. The bad short message identification method based on large language model error correction described in the present invention innovatively introduces a text error correction step based on a large language model before the recognition model characterizes the short message to be identified, so as to restore the short message with a high text error rate carefully constructed by the criminals back to the correct text it expresses, and then hand it over to the recognition module based on MacBERT to determine whether it is a bad short message. This can deal with the problem that the recognition accuracy of the trained recognition model is reduced due to the criminals using the sound, shape and meaning replacement technology to continuously update the corpus, and the need to continuously collect new data and perform manual correction.

[0051] 2. The bad short message identification method based on large language model error correction described in the present invention has strong robustness and can accurately identify bad short messages carried in the form of text, sound, and pictures that are processed by criminals using sound, shape, and meaning replacement technology.

[0052] 3. The bad short message identification method based on large language model error correction described in the present invention can identify bad short messages with low text error rate without using sound, shape and meaning replacement with a very high accuracy.

[0053] 4. The bad short message identification method based on large language model error correction described in the present invention can achieve an identification accuracy of 95.20% for bad short messages with high text error rate processed by sound, shape and meaning replacement technology, which is better than the bad short message identification method based on keyword matching and the bad short message identification method based on support vector machine.

[0054] The bad short message identification method and device based on large language model error correction described in the present invention are applied to the field of content security.

[0055] The bad short message identification method and device based on large language model error correction of the present invention are suitable for identifying bad short messages. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 A flowchart of a bad short message identification method based on large language model error correction in one embodiment of the present invention;

[0058] Figure 2 The present invention is a flowchart of performing text error correction on text with high text error rate based on a large language model in one embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the technical solutions and advantages of the present invention more clearly described, the specific implementation methods of the present invention will be further described in detail and completely in conjunction with the accompanying drawings. The various implementation methods described below are only part of the preferred solutions of the present invention, rather than all implementation methods; the various implementation methods described below are intended to explain the present invention and cannot be understood as limitations on the present invention; the reasonable combination of technical features defined in the various implementation methods of the present invention, and all other implementation methods obtained by ordinary technicians in the field without creative work based on the implementation methods of the present invention, are within the scope of protection of the present invention.

[0060] In one embodiment, a bad short message identification method based on large language model error correction is provided:

[0061] The method comprises the following steps:

[0062] A step for monitoring in real time the short messages carried in the form of text, sound or pictures received in the terminal;

[0063] A step for obtaining text content in the short message;

[0064] A step for evaluating the text error rate of the acquired text content to determine whether the text content is a high text error rate text or a low text error rate text;

[0065] A step for performing text error correction on a text with a high text error rate based on a large language model to obtain a corrected text;

[0066] A step for performing vectorization processing on a text with a low text error rate or a text after error correction to obtain a text vector;

[0067] The step is used to use the fine-tuned MacBERT recognition model to recognize the text vector and determine whether the short message corresponding to the text vector is a bad short message.

[0068] In this implementation, the method for identifying inappropriate short messages is mainly used for identifying inappropriate short messages in mobile phones, and can also be used for identifying inappropriate short messages in other terminals.

[0069] In this implementation manner, the fine-tuned MacBERT recognition model can determine (recognize) whether the short message corresponding to the text vector is a bad short message or a normal short message.

[0070] In addition, in one embodiment, the step of obtaining the text content in the short message includes:

[0071] For short messages carried in the form of sound or picture, the text content is obtained after the short messages are converted into text.

[0072] In addition, in one embodiment, for a short message carried in the form of sound, the text content thereof is obtained by converting it into text using a voice recognition technology.

[0073] In addition, in one embodiment, for a short message carried in the form of a picture, the text content thereof is obtained after the short message is converted into text using a picture text recognition technology.

[0074] In addition, in one embodiment, the step of evaluating the text error rate of the acquired text content and determining whether the text content is a high text error rate text or a low text error rate text includes:

[0075] Compare the text in the obtained text content with the text in the given text collection:

[0076] If the proportion of words not in the given word set in the acquired text content is greater than a given threshold, the acquired text content is a high text error rate text;

[0077] Otherwise, it is low text error rate text.

[0078] In this embodiment, the given character set is the commonly used 3500 Chinese characters. The characters in the non-given character set are characters other than the commonly used 3500 characters.

[0079] In this implementation manner, the given threshold is 10%.

[0080] In addition, in one embodiment, the step of evaluating the text error rate of the acquired text content and determining whether the text content is a high text error rate text or a low text error rate text includes:

[0081] Compare the text in the obtained text content with the 3500 most commonly used Chinese characters:

[0082] If the proportion of the commonly used 3,500 non-Chinese characters in the acquired text content is greater than 10%, the acquired text content is a high text error rate text;

[0083] Otherwise, it is low text error rate text.

[0084] In addition, in one embodiment, the step of performing text error correction on a text with a high text error rate based on a large language model to obtain a corrected text includes:

[0085] Get the engineering prompt word template for text error correction;

[0086] The text with high text error rate (i.e., the text to be recognized) is embedded into the engineering prompt word template used for text error correction to form a complete input instruction input language model;

[0087] The large language model performs text correction on the text with a high text error rate according to the complete input instructions to obtain the corrected text.

[0088] In this implementation, the domestic large language model "Wenxinyiyan" can be used as the large language model for error correction, and text error correction can be achieved by calling the large language model API method.

[0089] In this implementation, the background of the officially provided editable prompt engineering prompt word template can be used to customize the corresponding engineering prompt word template for text error correction according to the requirements of the text error correction task.

[0090] The engineering prompt word template for text error correction is shown as follows, where sentence is a variable representing the text to be corrected:

[0091] Roles and responsibilities: Text correction expert.

[0092] Background overview: A scholar discovered many typos in the text of messages while conducting research on identifying bad short messages. As a text correction expert, your task is to assist the scholar in correcting the typos in these texts.

[0093] Operation instructions: Please carefully review the text provided below and correct any typos; after correction, directly output the corrected text.

[0094] Text to be corrected: {sentence}.

[0095] Output Specification:

[0096] 1Content: Output only the corrected text without any additional explanation, instructions or prompts.

[0097] 2. Format: Make sure the corrected text is on the first line of the output and convert Traditional Chinese to Simplified Chinese.

[0098] 3. Integrity: Keep all links in the original text and do not delete any.

[0099] 4. Security disclaimer: No security or legality checks are performed on the text content, and only typos are corrected.

[0100] 5. No obvious errors: If there are no obvious typos in the text, the original text will be returned directly.

[0101] In addition, in one embodiment, the step of performing text correction on a text with a high text error rate based on a large language model to obtain the corrected text further includes the following filtering method based on text length comparison:

[0102] Compare the corrected text with the corresponding high text error rate text. If the difference between the two text lengths is greater than a given number of characters (e.g., 5 characters), the corrected text is an unexpected output.

[0103] When the corrected text is unexpected output:

[0104] The high text error rate text is again embedded into the engineering prompt word template for text error correction, and it is emphasized that only the text content after error correction is output to form a complete input instruction input language model.

[0105] In this implementation, unexpected outputs of a large language model are controlled by a filtering method based on text length comparison.

[0106] In addition, in one embodiment, the fine-tuned MacBERT recognition model is obtained by constructing and training the MacBERT recognition model using Python language in combination with keras, tensorflow, and transformers.

[0107] In addition, in one embodiment, one training set and two test sets are used when constructing and training the MacBERT recognition model.

[0108] The one training set and two test sets are constructed based on short message sets from at least two different sources, wherein:

[0109] A collection of short messages from one source is referred to as source 1, which includes several bad short messages that do not adopt the sound, shape and meaning replacement and several normal short messages;

[0110] Another short message collection includes several bad short messages from source 2 that use sound, shape and meaning substitution and several normal short messages collected in daily life;

[0111] The training set and a test set are obtained by dividing according to source No. 1;

[0112] Another test set was obtained based on source 2 and self-collection;

[0113] The training set is used to train the MacBERT recognition model to obtain a fine-tuned MacBERT recognition model;

[0114] Using the test set obtained from source 1, the recognition capability of the method for bad short messages with low text error rate without using phonetic, morphological and semantic substitution is evaluated;

[0115] The test set obtained from source 2 and self-collection is used to evaluate the recognition ability of the method for bad short messages with high text error rate using sound, form and meaning substitution.

[0116] More specifically:

[0117] The at least two short message sets from different sources, short message set 1 is from the public Chinese short message dataset on Git-hub (as source 1), the bad short messages in short message set 2 are from the 360 ​​malicious short message variant word restoration competition dataset (as source 2), and the normal short messages in short message set 2 are collected from daily life;

[0118] The public Chinese SMS dataset on Git-hub, after removing 2 invalid data, has a total of 799,998 messages, including 719,998 normal messages and 80,000 bad messages that do not use phonetic, morphological and semantic substitution;

[0119] Short message set 2 includes 1,000 bad short messages with high text error rates after using phonetic, morphological and semantic substitution from the 360 ​​malicious SMS variant word restoration competition dataset, and 1,000 normal short messages collected in daily life;

[0120] Suppose the two test sets are test set T1 and test set T2;

[0121] The training set and test set T1 are obtained from source 1 in a ratio of 7:3;

[0122] The test set T2 comes from source 2 and is collected from daily life.

[0123] The test set T1 is used to evaluate the effectiveness of the method on bad short messages with low text error rate without using sound, form and meaning substitution;

[0124] The test set T2 is used to evaluate the recognition ability of the method for bad short messages with high text error rate using sound, shape and meaning replacement.

[0125] The specific distribution of the data sets of the training set, test set T1, and test set T2 is shown in Table 1.

[0126] Table 1 Detailed distribution table of data sets

[0127] Serial number Training set Test set T1 Test set T2 sum Normal short message 503998 216000 1000 720998 Bad SMS 56000 24000 (no sound, shape and meaning substitution) 1000 (replace the sound, shape and meaning) 81000 sum 559998 240000 2000 801998

[0128] In addition, in one embodiment, evaluation indicators such as accuracy, precision, recall, and F1-Score are used to evaluate and test the evaluation ability of the method.

[0129] Table 2 shows the evaluation and test results of the method on the test set T1. As shown in Table 2, the method can obtain an accuracy rate of 99.66% on the test set T1, indicating that it can identify low text error rate bad short messages without using sound, form and meaning replacement with a high accuracy rate.

[0130] Table 2 Test results of this method on test set T1

[0131] Accuracy Precision Recall F1-Score 99.66％ 99.21％ 97.38％ 98.28％

[0132] Table 3 shows the experimental results of the comparison between the recognition method based on keyword matching and the method based on support vector machine and this method on the test set T2. As can be seen from Table 3, this method is significantly superior to the recognition method based on keyword matching and the method based on support vector machine in terms of accuracy, precision, recall and F1 value. The reason is not only that the deep learning model used in this method can learn deeper information of the text, but also that the large language model used in this method for text error correction can effectively correct the high text error rate text processed by the criminals using the sound, shape and meaning replacement technology. The recognition method based on keyword matching is difficult to identify the risky keywords in the bad short messages after the corpus is replaced, resulting in missed recognition. The method based on support vector machine is due to the fact that the short messages after the sound, shape and meaning replacement contain a large number of rare characters and special characters that have not appeared in the training set, which affects the recognition accuracy of this method.

[0133] Table 3 Comparative experimental results of different methods and this method on test set T2

[0134] method Accuracy Precision Recall F1Score Keyword matching based method 69.59％ 80.50％ 54.05％ 64.68％ Support vector machine based method 68.65％ 96.74％ 38.60％ 55.18％ This method 95.20％ 99.24％ 91.10％ 94.99％

[0135] The above results show that this method, by adding a text correction step based on a large language model to the bad short message identification process, can restore the short messages with high text error rate carefully constructed by criminals back to the correct text of their meaning, and then hand them over to the recognition module based on MacBERT to determine whether they are bad short messages. This method can deal with the problem that the recognition accuracy of the trained recognition model is reduced due to the criminals using the sound, shape and meaning replacement technology to continuously update the corpus, and the need to continuously collect new data and perform manual correction.

[0136] In one embodiment, a bad short message identification device based on large language model error correction is provided:

[0137] The device comprises the following modules:

[0138] Multimodal short message real-time monitoring module: used to monitor the short messages received by the terminal in the form of text, sound or pictures in real time;

[0139] Multimodal short message texting module: used to obtain the text content in the short message;

[0140] A text error rate evaluation module: used to evaluate the text error rate of the acquired text content, and determine whether the text content is a high text error rate text or a low text error rate text;

[0141] Text error correction module based on large language model: used to perform text error correction on text with high text error rate based on large language model to obtain the corrected text;

[0142] Text vectorization module: used to vectorize text with low text error rate or text after error correction to obtain text vectors;

[0143] Bad short message identification module: used to identify text vectors using the fine-tuned MacBERT recognition model, and determine whether the short message corresponding to the text vector is a bad short message.

[0144] The device is used to implement the bad short message identification method based on large language model error correction described in the above implementation manner.

[0145] In addition, in one embodiment, the device further includes a risk warning module;

[0146] The risk warning module is used to send a warning message according to the identified bad short messages.

[0147] In this implementation, a prompt message is sent out by way of a system pop-up window warning prompt to remind the user that a bad short message has been found, thereby enhancing the user's alertness to potential threats and ensuring the security of personal information.

[0148] In this implementation manner, when the identification result is a normal short message, the risk warning module does not take any measures and does not affect the user's normal communication experience.

[0149] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable changes and improvements to the present invention, reasonable combinations of implementation modes and equivalent substitutions based on the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A bad short message identification method based on large language model error correction, characterized in that: The method comprises the following steps: A step for monitoring in real time the short messages carried in the form of text, sound or pictures received in the terminal; A step for obtaining text content in the short message; A step for evaluating the text error rate of the acquired text content to determine whether the text content is a high text error rate text or a low text error rate text; A step for performing text error correction on a text with a high text error rate based on a large language model to obtain a corrected text; A step for performing vectorization processing on a text with a low text error rate or a text after error correction to obtain a text vector; The step is used to use the fine-tuned MacBERT recognition model to recognize the text vector and determine whether the short message corresponding to the text vector is a bad short message.

2. The bad short message identification method based on large language model error correction according to claim 1 is characterized in that: The step of obtaining the text content in the short message comprises: For short messages carried in the form of sound or picture, the text content is obtained after the short messages are converted into text.

3. The bad short message identification method based on large language model error correction according to claim 1 is characterized in that: The step of evaluating the text error rate of the acquired text content and determining whether the text content is a high text error rate text or a low text error rate text comprises: Compare the text in the obtained text content with the text in the given text collection: If the proportion of words not in the given word set in the acquired text content is greater than a given threshold, the acquired text content is a high text error rate text; Otherwise, it is low text error rate text.

4. The bad short message identification method based on large language model error correction according to claim 1 is characterized in that: The step of performing text error correction on a text with a high text error rate based on a large language model to obtain the corrected text comprises: Get the engineering prompt word template for text error correction; Embed high text error rate text into the engineering prompt word template for text error correction to form a complete input instruction input language model; The large language model performs text correction on the text with a high text error rate according to the complete input instructions to obtain the corrected text.

5. The bad short message identification method based on large language model error correction according to claim 4 is characterized in that: The step of performing text error correction on a text with a high text error rate based on a large language model to obtain the corrected text also includes a filtering method based on text length comparison: Compare the error-corrected text output by the large language model with the corresponding high text error rate text. If the difference between the two text lengths is greater than 5 characters, the error-corrected text is an unexpected output. When the corrected text is unexpected output: The high text error rate text is again embedded into the engineering prompt word template for text error correction to form a complete input instruction input large language model, and emphasizes that only the text content after error correction is output.

6. A bad short message identification device based on large language model error correction, characterized in that: The device comprises the following modules: Multimodal short message real-time monitoring module: used to monitor the short messages received by the terminal in the form of text, sound or pictures in real time; Multimodal short message texting module: used to obtain the text content in the short message; A text error rate evaluation module: used to evaluate the text error rate of the acquired text content, and determine whether the text content is a high text error rate text or a low text error rate text; Text error correction module based on large language model: used to perform text error correction on text with high text error rate based on large language model to obtain the corrected text; Text vectorization module: used to vectorize text with low text error rate or text after error correction to obtain text vectors; Bad short message identification module: used to identify text vectors using the fine-tuned MacBERT recognition model, and determine whether the short message corresponding to the text vector is a bad short message.

7. The bad short message identification device based on large language model error correction according to claim 6 is characterized in that: The device also includes a risk warning module; The risk warning module is used to send a warning message according to the identified bad short messages.

8. A computer device, comprising: A processor and a memory, characterized in that the memory is used to store executable instructions of the processor, and the processor is configured to execute the method for identifying bad short messages based on large language model error correction as described in any one of claims 1 to 5 by executing the executable instructions.

9. A computer storage medium, characterized in that The storage medium stores a computer program, and when the computer program is run, the bad short message identification method based on large language model error correction described in any one of claims 1 to 5 is executed.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the bad short message identification method based on large language model error correction described in any one of claims 1 to 6 are implemented.

Images