False propaganda detection model training method, false propaganda detection method and related equipment

Abstract

The invention relates to a training method of a false propaganda detection model, a false propaganda detection method and related equipment. The training method comprises the following steps: acquiring original propaganda data; inputting each piece of original propaganda data into the large language model to obtain structured propaganda data; inputting the propaganda text content into a false propaganda detection model to obtain a plurality of false propaganda prediction results; calculating a value of a loss function based on the plurality of false publicity prediction results and corresponding tags indicating whether false publicity is involved; and adjusting parameters of the false propaganda detection model according to the value of the loss function so as to train the false propaganda detection model. Compared with a mode of obtaining the training data by manually marking the original propaganda data, the large language model provided by the invention can significantly improve the efficiency of obtaining the structured propaganda data. The false propaganda detection model is trained by using the structured propaganda data, that is, more standard and more normative training data, so that the training effect, accuracy and recall ratio of the model can be ensured.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and more specifically, to a false propaganda detection model training method, a false propaganda detection method, and related equipment. Background Art

[0002] False advertising is a common and potentially harmful behavior in commercial activities, which may mislead consumers and affect market order and social trust. For example, some drug dealers may over-exaggerate the efficacy of drugs in order to increase drug sales, which may involve false advertising. It is difficult for ordinary consumers to distinguish the authenticity of the merchants' propaganda and they are likely to be deceived, which will seriously disrupt the market order.

[0003] Among the related technologies, false propaganda detection mainly relies on manual review and rule-based automated methods. However, these detection methods have low detection efficiency and low completeness. Especially with the advent of the big data era, the speed of information dissemination has accelerated, and the forms of false propaganda have become more diverse and covert. Traditional false propaganda detection methods can no longer meet actual needs. Summary of the invention

[0004] The present disclosure provides a training method for a false propaganda detection model, a false propaganda detection method and related equipment to at least solve the problem of low detection efficiency and low recall rate in the above-mentioned related technologies that rely on manual review and rule-based automated methods to detect false propaganda.

[0005] According to a first aspect of an embodiment of the present disclosure, there is provided a method for training a false advertising detection model, comprising: obtaining a plurality of original advertising data; inputting each of the plurality of original advertising data into a large language model to obtain structured advertising data corresponding to each of the original advertising data, wherein the structured advertising data at least comprises advertising text content extracted from the original advertising data and a label indicating whether false advertising is involved; inputting the advertising text content into the false advertising detection model to obtain a plurality of false advertising prediction results; calculating a value of a loss function based on the plurality of false advertising prediction results and the corresponding labels indicating whether false advertising is involved; and training the false advertising detection model by adjusting the parameters of the false advertising detection model according to the value of the loss function.

[0006] Optionally, before inputting each of the multiple original promotional data into the large language model, the training method also includes: for each of the preset number of original promotional data included in the multiple original promotional data, obtaining structured promotional data obtained through manual construction; inputting each of the preset number of original promotional data and each structured promotional data obtained through manual construction into the large language model, and learning structured promotional data construction rules through the large language model; wherein, inputting each of the multiple original promotional data into the large language model to obtain structured promotional data corresponding to each of the original promotional data includes: inputting other original promotional data among the multiple original promotional data except the preset number of original promotional data into the large language model, and constructing structured promotional data corresponding to the other original promotional data through the large language model using the learned structured promotional data construction rules.

[0007] According to a second aspect of an embodiment of the present disclosure, a false advertising detection method is provided, comprising: obtaining target advertising text content; inputting the target advertising text content into a false advertising detection model trained according to the training method of the present disclosure, and obtaining a false advertising detection result corresponding to the target advertising text content.

[0008] Optionally, obtaining the target promotional text content includes: setting a target filtering condition in response to a filtering condition setting instruction input by a user, wherein the target filtering condition includes at least one of the following items: the promotional text content includes specified keywords, the release time of the promotional text content is within a preset time period, and the promotional text content comes from a specified platform; and obtaining the target promotional text content based on the target filtering condition.

[0009] Optionally, the target promotional text content includes multiple target promotional text contents, each of which corresponds to a false promotional detection result; wherein the false promotional detection method also includes: obtaining a query field input by a user; searching for a false promotional detection result containing the query field from multiple false promotional detection results; determining the target promotional text content corresponding to the found false promotional detection result among the multiple target promotional text contents; and displaying the determined target promotional text content corresponding to the found false promotional detection result.

[0010] According to a third aspect of an embodiment of the present disclosure, there is provided a training device for a false advertising detection model, comprising: a publicity data acquisition module, configured to acquire a plurality of original publicity data; a structured publicity data acquisition module, configured to input each of the plurality of original publicity data into a large language model, and obtain structured publicity data corresponding to the each original publicity data, wherein the structured publicity data at least comprises publicity text content extracted from the original publicity data and a label indicating whether false advertising is involved; a prediction result acquisition module, configured to input the publicity text content into the false advertising detection model, and obtain a plurality of false advertising prediction results; a calculation module, configured to calculate the value of a loss function based on the plurality of false advertising prediction results and the corresponding labels indicating whether false advertising is involved; and a training module, configured to train the false advertising detection model by adjusting the parameters of the false advertising detection model according to the value of the loss function.

[0011] Optionally, the training device also includes: a manual construction module, configured to obtain structured promotional data obtained through manual construction for each of the preset number of original promotional data included in the multiple original promotional data; a construction rule learning module, configured to input each of the preset number of original promotional data and each structured promotional data obtained through manual construction into the large language model, and learn structured promotional data construction rules through the large language model; wherein the structured promotional data acquisition module is configured to: input other original promotional data among the multiple original promotional data except the preset number of original promotional data into the large language model, and construct structured promotional data corresponding to the other original promotional data through the large language model using the learned structured promotional data construction rules.

[0012] According to a fourth aspect of an embodiment of the present disclosure, a false advertising detection device is provided, comprising: a promotional text content acquisition module, configured to acquire target promotional text content; a detection module, configured to input the target promotional text content into a false advertising detection model trained according to the training method of the present disclosure, and obtain a false advertising detection result corresponding to the target promotional text content.

[0013] Optionally, the promotional text content acquisition module is configured to: set target filtering conditions in response to filtering condition setting instructions input by the user, wherein the target filtering conditions include at least one of the following items: the promotional text content includes specified keywords, the release time of the promotional text content is within a preset time period, and the promotional text content comes from a specified platform; and acquire the target promotional text content based on the target filtering conditions.

[0014] Optionally, the target promotional text content includes multiple target promotional text contents, each of which corresponds to a false promotional detection result; wherein the false promotional detection device also includes: a query field acquisition module, configured to obtain a query field input by a user; a detection result search module, configured to search for a false promotional detection result containing the query field from multiple false promotional detection results; a determination module, configured to determine the target promotional text content corresponding to the found false promotional detection result among the multiple target promotional text contents; and a display module, configured to display the determined target promotional text content corresponding to the found false promotional detection result.

[0015] According to a fifth aspect of an embodiment of the present disclosure, there is provided an electronic device, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement a training method for a false advertising detection model according to the present disclosure, or to implement a false advertising detection method according to the present disclosure.

[0016] According to a sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the training method of the false advertising detection model according to the present disclosure, or to execute the false advertising detection method according to the present disclosure.

[0017] According to a seventh aspect of an embodiment of the present disclosure, there is provided a computer program product, comprising a computer program, which, when executed by a processor, implements the training method of the false advertising detection model according to the present disclosure, or implements the false advertising detection method according to the present disclosure.

[0018] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0019] In the present disclosure, by introducing a large language model with powerful text understanding and generation capabilities, a more in-depth semantic analysis of the original promotional data can be performed, and more standardized and normalized training data can be obtained for subsequent training. Compared with the method of obtaining training data by manually annotating the original promotional data, the large language model provided by the present disclosure can greatly reduce the huge workload brought by manual annotation, and thus can significantly improve the efficiency of obtaining structured promotional data. In addition, using structured promotional data, that is, using more standard and more standardized training data to train the false propaganda detection model, can ensure the training effect, accuracy and recall rate of the model, and avoid the phenomenon of poor training effect caused by the format of the training data is not standardized enough. It can be seen that the false propaganda detection model provided by the present disclosure can deal with various forms of false propaganda, and can provide a strong technical guarantee for maintaining market order and consumer rights.

[0020] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0022] Figure 1 is a flow chart illustrating a method for training a false advertising detection model according to an exemplary embodiment of the present disclosure;

[0023] Figure 2 is a flow chart illustrating a false advertising detection method according to an exemplary embodiment of the present disclosure;

[0024] Figure 3 is a schematic diagram showing obtaining target promotion text content based on target screening conditions according to an exemplary embodiment of the present disclosure;

[0025] Figure 4 is a schematic diagram showing a detection result obtained by performing false propaganda detection on target propaganda text content according to an exemplary embodiment of the present disclosure;

[0026] Figure 5 is a block diagram illustrating a training apparatus for a false advertising detection model according to an exemplary embodiment of the present disclosure;

[0027] Figure 6 is a block diagram illustrating a false advertising detection apparatus according to an exemplary embodiment of the present disclosure;

[0028] Figure 7 is a block diagram illustrating an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0029] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0030] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The implementation methods described in the following examples do not represent all implementation methods consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the attached claims.

[0031] It should be noted that the phrase "at least one of the items" in the present disclosure includes three types of parallel situations: "any one of the items", "a combination of any number of the items", and "all of the items". For example, "including at least one of A and B" includes the following three parallel situations: (1) including A; (2) including B; (3) including A and B. Another example is "executing at least one of step 1 and step 2" which means the following three parallel situations: (1) executing step 1; (2) executing step 2; (3) executing step 1 and step 2.

[0032] Figure 1 is a flow chart illustrating a method for training a false advertising detection model according to an exemplary embodiment of the present disclosure.

[0033] Reference Figure 1 In step 101, a plurality of original publicity data may be obtained. For example, the “original publicity data” may be, but not limited to, news reports. The “news reports” may include, but not limited to, the following contents: the name of the merchant being reported, the publicity copy of the merchant, i.e., the publicity advertisement, whether the publicity copy of the merchant involves false publicity, if the publicity copy of the merchant involves false publicity, which legal provision is violated and the specific content of the violated legal provision, if the publicity copy of the merchant involves false publicity, the administrative penalty result against the merchant, for example, a fine, revocation of business license, etc.

[0034] For example, a "news report" may be: "Merchant A's promotional copy is: The efficacy of this product has reached the pinnacle and can cure all diseases, and has achieved world-renowned achievements. The promotional slogan of Merchant A obviously involves false advertising and violates the YYY clause. Therefore, Merchant A is fined and warned, and this is used to warn other merchants to promote their products truthfully and not exaggerate to deceive consumers."

[0035] It should be noted that text data related to false advertising can be collected from various sources, such as, but not limited to, web crawling, public databases, industry reports, etc. Then, the obtained text data related to false advertising can be preprocessed, such as, but not limited to, removing irrelevant content, tokenizing, removing stop words, unifying the format, etc. Next, the preprocessed data can be preliminarily labeled to form a preliminary data set covering the characteristics of false advertising.

[0036] In step 102, each of the multiple original promotional data can be input into the large language model to obtain structured promotional data corresponding to each original promotional data, wherein the structured promotional data can at least include promotional text content extracted from the original promotional data and a label indicating whether false advertising is involved, and the "promotional text content" is the "promotional copy" mentioned above.

[0037] It should be noted that with the rapid development of artificial intelligence technology, especially the technological breakthroughs in deep learning and natural language processing (NLP), large language models (LLMs) have demonstrated powerful capabilities in text understanding, generation, and semantic analysis. By training on massive text data, the model can capture the complex structure and semantic information of the language, thereby achieving a deep understanding of the text content. This provides a new technical path for false propaganda detection, that is, it is possible to use pre-trained large language models and combine them with fine-tuning strategies for specific tasks to generate standard and standardized training data for training false propaganda detection models.

[0038] Specifically, you can first train or select a large language model that already has the ability to understand and generate text, such as the GPT series, Qwen series, or a large language model customized for the legal field, etc. These models have mastered rich language expressions and semantic knowledge through pre-training on large-scale corpora. In order to meet the needs of the false propaganda detection task in this disclosure, it is necessary to build a high-quality annotation instruction set. That is, you can first manually obtain the structured propaganda data corresponding to a small amount of original propaganda data, and each original propaganda data in this small amount of original propaganda data and the corresponding structured propaganda data can be input into the large language model, so that the large language model learns the structured propaganda data construction rules.

[0039] For example, you can first conduct a demand analysis, that is, you can first clarify the requirements and standards for labeling, and then you can write detailed labeling instructions, for example, it can include but is not limited to: identifying false advertising features, classifying different types of promotional texts, handling boundary conditions, etc. For example, a specific labeling instruction can be: "Merchant name; Promotional copy; Whether it involves false advertising; If it involves false advertising, the specific name / number / content of the clause violated; If it involves false advertising, the administrative penalty imposed."

[0040] Then, experts can be invited to review the instruction set to ensure its accuracy and practicality, and the instruction set can be revised based on feedback to cover all important features. Next, representative promotional data can be selected from the preliminary data set obtained by preprocessing the original promotional data for precise annotation, that is, structured promotional data can be constructed, and the annotation results can be further reviewed. Then, the selected representative promotional data and the corresponding annotation results, that is, the structured promotional data obtained through manual construction, can be input into the large language model together, so that the large language model can learn the structured promotional data construction rules.

[0041] Specifically, according to an exemplary embodiment of the present disclosure, structured promotional data obtained through manual construction may be obtained for each of a preset number of original promotional data included in a plurality of original promotional data.

[0042] For example, as mentioned above, the "original promotional data" may be a news report, and the "news report" may be: "Merchant A's promotional copy is: The efficacy of this product has reached the pinnacle, it can cure all diseases, and has achieved world-renowned achievements. The promotional slogan of Merchant A obviously involves false propaganda and violates the YYY clause. Therefore, Merchant A is fined and warned, and other merchants are warned to promote their products in a realistic manner and not to exaggerate to deceive consumers." In this case, the structured promotional data obtained through manual construction corresponding to the original promotional data may be:

[0043] “Business Name: A;

[0044] Promotional copy: The efficacy of this product has reached the pinnacle, curing all diseases and has achieved world-renowned success;

[0045] Does the promotional copy involve false advertising: Yes;

[0046] If false advertising is involved, the specific name / number / content of the clause violated is: YYY,…

[0047] If false advertising is involved, the administrative penalties will be: fines and warnings.

[0048] Next, each of the preset number of original promotional data and each structured promotional data obtained through manual construction may be input into the large language model, and the structured promotional data construction rules may be learned through the large language model.

[0049] For example, the above-mentioned original promotional data, that is, the news report: "Merchant A's promotional copy is: The efficacy of this product has reached the pinnacle, it can cure all diseases, and has achieved world-renowned achievements. The promotional slogan of Merchant A obviously involves false propaganda and violates the YYY clause. Therefore, Merchant A is fined and warned, and other merchants are warned to promote their own products in a realistic manner and not to exaggerate to deceive consumers", and the structured promotional data corresponding to the news report obtained through manual construction: "Merchant name: A; Promotional copy :The efficacy of this product has reached the pinnacle and can cure all diseases, and has achieved world-renowned achievements; Does the promotional copy involve false advertising: Yes; If it involves false advertising, the specific name / number / content of the clause violated: YYY,...; If it involves false advertising, the administrative penalty imposed: fine and warning" are input into the big language model together, so that the big language model learns the rules for building structured promotional data, that is, the big language model learns how to extract structured promotional data from the original promotional data that can reflect the core idea of ​​the original promotional data and has a more standard format.

[0050] Then, the other original promotional data except a preset number of original promotional data among the multiple original promotional data can be input into the large language model, and the structured promotional data corresponding to the other original promotional data can be constructed by the large language model using the learned structured promotional data construction rules.

[0051] Exemplarily, assuming that a total of 10,000 pieces of original promotional data are obtained, 100 pieces of original promotional data can be selected from the 10,000 pieces of original promotional data and 100 pieces of corresponding structured promotional data can be manually constructed. Then, the selected 100 pieces of original promotional data and the corresponding 100 pieces of structured promotional data can be input into the large language model for the large language model to learn the structured promotional data construction rules. Next, the 9,900 pieces of original promotional data other than the 100 pieces of original promotional data selected previously from the 10,000 pieces of original promotional data obtained can be respectively input into the large language model, so as to construct the structured promotional data corresponding to each piece of original promotional data in the 9,900 pieces of original promotional data through the large language model using the learned structured promotional data construction rules. Moreover, the 100 pieces of structured promotional data constructed manually and the 9,900 pieces of structured promotional data constructed by the large language model can be used as training data for training the false propaganda detection model.

[0052] In this way, a small amount of structured promotional data can be constructed manually first, and the large language model can learn the structured promotional data construction rules based on the manually constructed structured promotional data and the corresponding original promotional data, and then the large language model that has learned the construction rules can be directly used to perform structured construction on the remaining large amount of original promotional data to obtain training data. Compared with the method of obtaining training data by manually annotating the original promotional data, the large language model provided by the present disclosure can greatly reduce the huge workload caused by manual annotation, and thus can significantly improve the efficiency of obtaining structured promotional data.

[0053] Furthermore, the incorrect annotations can be corrected based on the feedback from the large language model review to ensure the accuracy of the annotation data, that is, further refinement and tuning can be performed based on the feedback from the large language model. Specifically, the following operations can be performed for refinement and tuning:

[0054] (1) Feedback collection and error correction: Feedback information from channels such as internal testing of the large language model, user feedback, and manual review can be collected, and the feedback data can be sorted and classified. Then, the labeled data can be corrected based on the error information in the feedback (for example, label misalignment, text misclassification, etc.). In addition, the error type can be recorded. Next, the parameters of the large language model can be adjusted or the large language model can be retrained based on the feedback from the large language model to reduce the occurrence of similar errors, thereby ensuring the accuracy of the data.

[0055] (2) Multiple rounds of verification and optimization: A multiple round of verification plan can be developed to repeatedly verify the corrected data to ensure that it meets the expected quality standards. By using strict evaluation criteria for inspection and continuously optimizing the annotation data and large language model settings based on feedback, the annotation accuracy and large language model performance can be improved.

[0056] In this way, by constructing a high-quality annotation instruction set, that is, by manually constructing a small amount of structured propaganda data to enable the large language model to learn, and dynamically adjusting and optimizing under the feedback of the large language model, the accuracy of data annotation is ensured, thereby improving the training effect of the large language model.

[0057] It should be noted that, as mentioned above, a multi-round verification plan can be formulated to repeatedly verify the corrected data to ensure that it meets the expected quality standards. That is, the instruction set can be dynamically iterated and updated through multiple rounds of quality inspection and review by experts, and ultimately the structured output of the subject, promotional copy, whether false propaganda is involved, and legal provisions violated in the false propaganda text corpus can be achieved, that is, false propaganda training data can be generated in batches for subsequent training of false propaganda detection models. Specifically, the implementation process of dynamically iterating and updating the instruction set through multiple rounds of quality inspection and review to obtain structured output can be as follows:

[0058] (1) Expert review and instruction set update: Experts can be organized to conduct multiple rounds of review and evaluation of the annotation instruction set, and then the content and format of the instruction set can be dynamically adjusted and optimized based on the feedback from experts to ensure its applicability in various false propaganda scenarios. The specific content of each update can also be recorded, and the instruction set can be gradually improved to cover a wider range of false propaganda characteristics.

[0059] (2) Structured data generation: Use the final updated instruction set to annotate large-scale false propaganda texts to generate structured data. The structured data may include but is not limited to: subject, propaganda copy, false propaganda identification, i.e., whether it involves false propaganda, legal clauses involved, etc. In addition, these data can be organized into training data sets for subsequent training and testing of false propaganda detection models.

[0060] In this way, through multiple rounds of expert quality inspection, model evaluation and optimization, the robustness and stability of the false propaganda detection model in practical applications can be ensured, and the occurrence of overfitting or underfitting can be greatly reduced.

[0061] Furthermore, you can also perform a quality check on the labeled data set to remove low-quality data such as unclear descriptions and wrong labels. Specifically, you can perform the following operations to perform a quality check on the labeled data set:

[0062] (1) Quality inspection and data processing: Based on pre-established quality inspection standards (e.g., clarity of text descriptions, accuracy and consistency of labels, etc.), the labeled data sets can be inspected by automated tools and manually reviewed to identify and correct quality issues, thereby eliminating low-quality data such as unclear descriptions and incorrect labels. In addition, the cleaned data can be organized and updated to ensure the overall quality of the data set.

[0063] (2) Dataset maintenance and update: A regular update and maintenance plan can be developed to maintain the high quality of the dataset. That is, all update and maintenance work can be recorded to ensure the long-term accuracy and consistency of the dataset.

[0064] In step 103, the promotional text content may be input into the false promotion detection model to obtain multiple false promotion prediction results. That is, the promotional text content contained in each structured promotional data in the multiple structured promotional data corresponding to the multiple original promotional data may be input into the false promotion detection model, and then the false promotion prediction result corresponding to each promotional text content may be obtained.

[0065] Exemplarily, the false advertising detection model in the present disclosure may be, but is not limited to, a Bidirectional Encoder Representations from Transformers (BERT) model.

[0066] In step 104, the value of the loss function may be calculated based on a plurality of false advertising prediction results and corresponding labels indicating whether false advertising is involved.

[0067] In step 105, the false advertising detection model may be trained by adjusting the parameters of the false advertising detection model according to the value of the loss function.

[0068] It should be noted that in the present disclosure, after training the false propaganda detection model, it is also necessary to evaluate the false propaganda detection model in order to timely discover and solve the overfitting or underfitting problem. In addition, it is also necessary to test the trained false propaganda detection model to verify its effect in practical applications.

[0069] (1) Model training and performance evaluation: Training data sets can be prepared and processed, including but not limited to: word segmentation, vectorization, data normalization, etc. In addition, the data set can be divided into training set, validation set and test set. The false propaganda detection model, i.e., the training parameters of the BERT model (e.g., learning rate, batch size, number of training rounds, optimizer selection, etc.) can also be configured to perform model training and monitor the loss value and accuracy during the training process in real time. The validation set can also be used to evaluate the performance of the model. For example, the performance of the model can be evaluated by calculating various evaluation indicators (e.g., accuracy, precision, recall, etc.). Next, the model parameters or structure can be adjusted according to the evaluation results to solve the problem of overfitting or underfitting.

[0070] (2) Practical application testing and optimization: The trained model can be applied to actual scenarios for testing to evaluate its effect and performance in real applications. Next, the model can be further optimized based on the actual test results to ensure its stability and effectiveness in actual applications.

[0071] In the present disclosure, in terms of data processing, the BERT model can be used to preprocess and extract features from false propaganda texts. As a bidirectional encoder representation model, BERT can capture contextual information through a bidirectional Transformer structure and can keenly identify subtle semantic differences in texts. Therefore, BERT can effectively extract key information from false propaganda texts, and can provide a reliable feature basis for subsequent classification and judgment.

[0072] In this disclosure, in order to prevent problems such as overfitting or underfitting during the training process of the false propaganda detection model, a systematic model evaluation and optimization strategy is proposed, that is, the performance of the model in different scenarios can be verified through multiple rounds of training and testing to ensure its stability and robustness in practical applications. In addition, the accuracy and efficiency of false propaganda detection can be further improved by continuously optimizing the model structure and training data.

[0073] In summary, in this disclosure, by combining the technical advantages of LLM and BERT, an innovative false propaganda detection scheme is proposed. Through dynamic iteration and expert review, this scheme improves the detection accuracy of false propaganda texts, ensures the reliability and applicability of the model in complex tasks, and provides strong support for the automated detection of false propaganda. Moreover, this technology can greatly reduce the workload of manual review in practical applications, and thus significantly improve the efficiency of false propaganda detection, which can provide strong technical support for maintaining market order and consumer rights.

[0074] Figure 2is a flow chart illustrating a false advertising detection method according to an exemplary embodiment of the present disclosure.

[0075] Reference Figure 2 In step 201, the target promotional text content may be obtained. Exemplarily, the target promotional text content may be, but is not limited to, promotional copy of a merchant on an e-commerce platform.

[0076] According to an exemplary embodiment of the present disclosure, in response to a filter condition setting instruction input by a user, a target filter condition may be set, wherein the target filter condition may include at least one of the following items: the promotional text content includes a specified keyword, the release time of the promotional text content is within a preset time period, the promotional text content comes from a specified platform, etc. Next, the target promotional text content may be obtained based on the target filter condition. Figure 3 is a schematic diagram showing how to obtain target promotional text content based on target screening conditions according to an exemplary embodiment of the present disclosure. Figure 3 , shows a filter condition setting box, that is, a filter item setting box, in which users can set specific filter conditions based on their own needs to filter out the target propaganda text content for which they want to detect false propaganda.

[0077] For example, suppose the target screening conditions set by the user are: "the promotional text content contains the keyword "health care products", the release time of the promotional text content is between 2022.01.12 and 2022.02.12, and the promotional text content comes from the e-commerce platform ZZ". In this case, false advertising detection can be performed only on the target promotional text content that meets the above conditions set by the user.

[0078] In this way, in the present disclosure, users can arbitrarily set screening conditions based on actual needs, and thus can implement false propaganda detection for the propaganda text content that users really want to detect, thereby obtaining greater autonomy and flexibility in obtaining the propaganda text content.

[0079] In step 202, the target propaganda text content can be input into the false propaganda detection model trained according to the training method of the present disclosure, and then the false propaganda detection result corresponding to the target propaganda text content can be obtained. The false propaganda detection result can include two types, namely "not involving false propaganda" and "involving false propaganda". In addition, when the type of the false propaganda detection result is "involving false propaganda", the false propaganda detection result can also include the specific legal provisions violated and the specific content of the legal provisions.

[0080] For example, Figure 4is a schematic diagram showing the detection result obtained by performing false propaganda detection on target propaganda text content according to an exemplary embodiment of the present disclosure. Figure 4 , the target promotional text under test is: "The food sold in this store has unprecedented therapeutic effects. When used, just drop one drop under the tongue to bring the dead back to life. It is a well-deserved miracle drug." The false propaganda detection model's test result for the target promotional text is: "It involves false propaganda and violates Article 8 of the Anti-Unfair Competition Law: Business operators shall not make false or misleading commercial propaganda about the performance, function, quality, sales status, user evaluation, honors, etc. of their products, and deceive or mislead consumers."

[0081] According to an exemplary embodiment of the present disclosure, the above-mentioned target promotional text content may include multiple target promotional text contents, and each target promotional text content may correspond to a false propaganda detection result. According to an exemplary embodiment of the present disclosure, the false propaganda detection method may also include the following steps: obtaining a query field input by a user. Exemplarily, the query field may be "Article 28 of the Advertising Law".

[0082] Then, the false propaganda detection result containing the query field can be searched from the multiple false propaganda detection results. For example, assuming that a total of 100 target propaganda text contents are detected, a total of 100 detection results can be obtained. At this time, based on the query field "Article 28 of the Advertising Law" input by the user, the false propaganda detection result containing the query field can be searched from the 100 false propaganda detection results.

[0083] Next, target propaganda text contents corresponding to the found false propaganda detection results can be determined from among the multiple target propaganda text contents. For example, assuming that a total of 3 false propaganda detection results are found, 3 target propaganda text contents corresponding to the 3 found false propaganda detection results can be determined from among the 100 target propaganda text contents.

[0084] Then, the target promotional text content corresponding to the false promotion detection results found can be displayed, that is, the three target promotional text contents corresponding to the three false promotion detection results found can be displayed. At this point, the user can know which target promotional text contents are suspected of violating Article 28 of the Advertising Law.

[0085] In this way, in the present disclosure, users can input any query field according to their own needs, and then directly use the input query field to find the promotional text content that meets the user's actual expectations, which can achieve fast search for specific promotional text content, and the search has good autonomy and flexibility.

[0086] It should be noted that the false propaganda detection system based on LLM and BERT provided by the present disclosure can be mainly composed of a "false propaganda detection management module", a "false propaganda detection module", a "detection result export center module", etc.

[0087] In the "False Advertising Detection Management Module", you can enter information such as product attributes and search terms. Then, you can create a new detection task to collect data from a specified website platform. You can also view all detection tasks and schedule detection resources according to actual needs to ensure the effective implementation of detection tasks.

[0088] In the "False Propaganda Detection Module", a large language model can be tuned through a manually constructed instruction set to generate text label data for subsequent training. And parameters can be configured to train a false propaganda detection model, namely the BERT model.

[0089] In the "Test Result Export Center Module", you can query the test results by specifying fields, perform visual data analysis on the test results, and export the test results in batches.

[0090] Figure 5 is a block diagram illustrating a training device 500 for a false advertising detection model according to an exemplary embodiment of the present disclosure.

[0091] Reference Figure 5 The training device 500 of the false advertising detection model may include a publicity data acquisition module 501, a structured publicity data acquisition module 502, a prediction result acquisition module 503, a calculation module 504 and a training module 505.

[0092] The publicity data acquisition module 501 can acquire a plurality of original publicity data. For example, the “original publicity data” can be, but is not limited to, news reports.

[0093] It should be noted that text data related to false advertising can be collected from various sources, such as, but not limited to, web crawling, public databases, industry reports, etc. Then, the obtained text data related to false advertising can be preprocessed, such as, but not limited to, removing irrelevant content, tokenizing, removing stop words, unifying the format, etc. Next, the preprocessed data can be preliminarily labeled to form a preliminary data set covering the characteristics of false advertising.

[0094] The structured promotional data acquisition module 502 can input each of the multiple original promotional data into the large language model to obtain structured promotional data corresponding to each original promotional data, wherein the structured promotional data can at least include the promotional text content extracted from the original promotional data and a label indicating whether false advertising is involved, and the "promotional text content" is the "promotional copy" mentioned above.

[0095] According to an exemplary embodiment of the present disclosure, the training device 500 of the false advertising detection model may further include an artificial construction module and a construction rule learning module.

[0096] The manual construction module may obtain structured promotional data obtained through manual construction for each of the preset number of original promotional data included in the plurality of original promotional data.

[0097] Next, the rule building learning module can input each of the preset number of original promotional data and each structured promotional data obtained through manual construction into the large language model, and learn the structured promotional data building rules through the large language model.

[0098] Then, the structured promotional data acquisition module 502 can input other original promotional data except a preset number of original promotional data into the large language model, and can use the learned structured promotional data construction rules through the large language model to construct structured promotional data corresponding to other original promotional data.

[0099] In this way, a small amount of structured promotional data can be constructed manually first, and the large language model can learn the structured promotional data construction rules based on the manually constructed structured promotional data and the corresponding original promotional data, and then the large language model that has learned the construction rules can be directly used to perform structured construction on the remaining large amount of original promotional data to obtain training data. Compared with the method of obtaining training data by manually annotating the original promotional data, the large language model provided by the present disclosure can greatly reduce the huge workload caused by manual annotation, and thus can significantly improve the efficiency of obtaining structured promotional data.

[0100] The prediction result acquisition module 503 can input the promotional text content into the false promotion detection model to obtain multiple false promotion prediction results. That is, the promotional text content contained in each structured promotional data in the multiple structured promotional data corresponding to the multiple original promotional data can be input into the false promotion detection model, and then the false promotion prediction result corresponding to each promotional text content can be obtained.

[0101] Exemplarily, the false advertising detection model in the present disclosure may be, but is not limited to: a Bidirectional Encoder Representation Transformer (BERT) model.

[0102] The calculation module 504 may calculate the value of the loss function based on a plurality of false advertising prediction results and corresponding labels indicating whether false advertising is involved.

[0103] The training module 505 may train the false advertising detection model by adjusting the parameters of the false advertising detection model according to the value of the loss function.

[0104] Figure 6 is a block diagram illustrating a false advertising detection apparatus 600 according to an exemplary embodiment of the present disclosure.

[0105] Reference Figure 6 The false advertising detection device 600 may include a promotional text content acquisition module 601 and a detection module 602 .

[0106] The promotional text content acquisition module 601 can acquire target promotional text content. For example, the target promotional text content can be, but is not limited to: promotional copy of a merchant on an e-commerce platform.

[0107] According to an exemplary embodiment of the present disclosure, the promotional text content acquisition module 601 can set a target filtering condition in response to a filtering condition setting instruction input by a user, wherein the target filtering condition can include at least one of the following items: the promotional text content includes a specified keyword, the release time of the promotional text content is within a preset time period, the promotional text content comes from a specified platform, etc. Next, the promotional text content acquisition module 601 can acquire the target promotional text content based on the target filtering condition.

[0108] In this way, in the present disclosure, users can arbitrarily set screening conditions based on actual needs, and thus can implement false propaganda detection for the propaganda text content that users really want to detect, thereby obtaining greater autonomy and flexibility in obtaining the propaganda text content.

[0109] The detection module 602 can input the target propaganda text content into the false propaganda detection model trained according to the training method of the present disclosure, and then obtain the false propaganda detection result corresponding to the target propaganda text content. The false propaganda detection result can include two types, namely "not involving false propaganda" and "involving false propaganda". In addition, when the type of the false propaganda detection result is "involving false propaganda", the false propaganda detection result can also include the specific legal provisions violated and the specific content of the legal provisions.

[0110] According to an exemplary embodiment of the present disclosure, the false advertising detection device 600 may further include a query field acquisition module, a detection result search module, a determination module, and a display module.

[0111] The above-mentioned target promotional text content may include multiple target promotional text contents, and each target promotional text content may correspond to a false advertising detection result. The query field acquisition module may also acquire the query field input by the user. Exemplarily, the query field may be "Article 28 of the Advertising Law".

[0112] Then, the detection result search module can search for false propaganda detection results containing the query field from multiple false propaganda detection results. For example, assuming that a total of 100 target propaganda text contents are detected, a total of 100 detection results can be obtained. At this time, based on the query field "Article 28 of the Advertising Law" input by the user, the false propaganda detection result containing the query field can be searched from the 100 false propaganda detection results.

[0113] Next, the determination module may determine the target propaganda text content corresponding to the found false propaganda detection results among the multiple target propaganda text contents. For example, assuming that a total of 3 false propaganda detection results are found, 3 target propaganda text contents corresponding to the 3 found false propaganda detection results among the above 100 target propaganda text contents may be determined.

[0114] Then, the display module can display the target promotional text content determined to correspond to the found false promotional detection results, that is, it can display the three target promotional text contents determined to correspond to the three found false promotional detection results. At this point, the user can know which target promotional text contents are suspected of violating Article 28 of the Advertising Law.

[0115] In this way, in the present disclosure, users can input any query field according to their own needs, and then directly use the input query field to find the promotional text content that meets the user's actual expectations, which can achieve fast search for specific promotional text content, and the search has good autonomy and flexibility.

[0116] Figure 7 is a block diagram illustrating an electronic device 700 according to an exemplary embodiment of the present disclosure.

[0117] Reference Figure 7 The electronic device 700 includes at least one memory 701 and at least one processor 702, wherein the at least one memory 701 stores instructions, and when the instructions are executed by the at least one processor 702, a false advertising detection model training method or a false advertising detection method according to an exemplary embodiment of the present disclosure is executed.

[0118] As an example, the electronic device 700 may be a PC, a tablet device, a personal digital assistant, a smart phone, or other device capable of executing the above instructions. Here, the electronic device 700 is not necessarily a single electronic device, but may also be any device or circuit capable of executing the above instructions (or instruction sets) individually or in combination. The electronic device 700 may also be part of an integrated control system or a system manager, or may be configured as a portable electronic device interconnected with a local or remote (e.g., via wireless transmission) interface.

[0119] In the electronic device 700, the processor 702 may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller or a microprocessor. As an example and not limitation, the processor may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0120] The processor 702 may execute instructions or codes stored in the memory 701, which may also store data. Instructions and data may also be sent and received over a network via a network interface device, which may employ any known transmission protocol.

[0121] The memory 701 may be integrated with the processor 702, for example, by placing RAM or flash memory within an integrated circuit microprocessor or the like. In addition, the memory 701 may include a separate device, such as an external disk drive, a storage array, or any other storage device that can be used by a database system. The memory 701 and the processor 702 may be operatively coupled, or may communicate with each other, such as through an I / O port, a network connection, etc., so that the processor 702 can read files stored in the memory.

[0122] In addition, the electronic device 700 may further include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, a mouse, a touch input device, etc.) All components of the electronic device 700 may be connected to each other via a bus and / or a network.

[0123] According to an exemplary embodiment of the present disclosure, a computer-readable storage medium may also be provided, and when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the above-mentioned false propaganda detection model training method or false propaganda detection method. Examples of computer-readable storage media here include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disk storage, hard disk drive (HDD), solid state drive (SSD), card storage (such as, multimedia card, secure digital (SD) card or extreme digital (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid state disk and any other device, any other device is configured to store computer programs and any associated data, data files and data structures in a non-transitory manner and provide the computer programs and any associated data, data files and data structures to a processor or computer so that the processor or computer can execute the computer program. The computer program in the above-mentioned computer-readable storage medium can be run in an environment deployed in a computer device such as a client, a host, an agent device, a server, etc. In addition, in one example, the computer program and any associated data, data files and data structures are distributed on a networked computer system, so that the computer program and any associated data, data files and data structures are stored, accessed and executed in a distributed manner by one or more processors or computers.

[0124] According to an exemplary embodiment of the present disclosure, a computer program product may also be provided, including a computer program, which, when executed by a processor, implements the false propaganda detection model training method or false propaganda detection method according to the present disclosure.

[0125] According to the training method, false propaganda detection method and related equipment of the false propaganda detection model disclosed in the present invention, by introducing a large language model with powerful text understanding and generation capabilities, it is possible to perform a more in-depth semantic analysis of the original propaganda data, and thus obtain more standardized and more normalized training data for subsequent training. Compared with the method of obtaining training data by manually annotating the original propaganda data, the large language model provided by the present invention can greatly reduce the huge workload brought by manual annotation, and thus can significantly improve the efficiency of obtaining structured propaganda data. In addition, using structured propaganda data, that is, using more standard and more standardized training data to train the false propaganda detection model, can ensure the training effect, accuracy and recall rate of the model, and avoid the phenomenon of poor training effect caused by the format of the training data is not standardized enough. It can be seen that the false propaganda detection model provided by the present invention can deal with various forms of false propaganda and can provide a strong technical guarantee for maintaining market order and consumer rights.

[0126] According to the exemplary embodiments of the present disclosure, a small amount of structured promotional data can be constructed manually first, and the large language model can learn the structured promotional data construction rules based on the manually constructed structured promotional data and the corresponding original promotional data, and then the large language model that has learned the construction rules can be directly used to perform structured construction on the remaining large amount of original promotional data to obtain training data. Compared with the method of obtaining training data by manually annotating the original promotional data, the large language model provided by the present disclosure can greatly reduce the huge workload caused by manual annotation, and thus can significantly improve the efficiency of obtaining structured promotional data.

[0127] According to the exemplary embodiments of the present disclosure, users can arbitrarily set screening conditions based on actual needs, thereby achieving false propaganda detection for the propaganda text content that users really want to detect, and having better autonomy and flexibility in obtaining the propaganda text content.

[0128] According to an exemplary embodiment of the present disclosure, a user can input any query field according to his or her own needs, and then directly use the input query field to find promotional text content that meets the user's actual expectations, thereby achieving fast search for specific promotional text content, and the search has good autonomy and flexibility.

[0129] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present disclosure is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0130] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A training method for a false advertising detection model, characterized in that: include: Obtain multiple original promotional data; Inputting each of the plurality of original promotional data into a large language model to obtain structured promotional data corresponding to each of the original promotional data, wherein the structured promotional data at least includes promotional text content extracted from the original promotional data and a label indicating whether false promotion is involved; Inputting the content of the promotional text into the false promotion detection model to obtain multiple false promotion prediction results; Calculating a value of a loss function based on the plurality of false advertising prediction results and corresponding labels indicating whether false advertising is involved; The false advertising detection model is trained by adjusting the parameters of the false advertising detection model according to the value of the loss function.

2. The training method according to claim 1, characterized in that: Before inputting each of the plurality of original promotional data into the large language model, the training method further includes: For each of the preset number of original promotional data included in the plurality of original promotional data, obtaining structured promotional data obtained through manual construction; Input each of the preset number of original promotional data and each structured promotional data obtained through manual construction into the large language model, and learn the structured promotional data construction rules through the large language model; The step of inputting each of the plurality of original promotional data into the large language model to obtain structured promotional data corresponding to each of the original promotional data includes: The other original promotional data among the multiple original promotional data except the preset number of original promotional data are input into the large language model, and the structured promotional data corresponding to the other original promotional data are constructed by the large language model using the learned structured promotional data construction rules.

3. A false advertising detection method, characterized in that: include: Obtain the target promotional text content; The target promotional text content is input into a false promotion detection model trained by the training method according to claim 1 or 2 to obtain a false promotion detection result corresponding to the target promotional text content.

4. The false advertising detection method according to claim 3, characterized in that: The obtaining of target promotional text content includes: In response to the filter condition setting instruction input by the user, a target filter condition is set, wherein the target filter condition includes at least one of the following items: the promotional text content includes a specified keyword, the release time of the promotional text content is within a preset time period, and the promotional text content comes from a specified platform; The target promotional text content is obtained based on the target screening condition.

5. The false advertising detection method according to claim 3, characterized in that: The target propaganda text content includes a plurality of target propaganda text contents, each of which corresponds to a false propaganda detection result; The false advertising detection method further includes: Get the query field entered by the user; Searching for a false advertising detection result containing the query field from multiple false advertising detection results; Determining target promotional text content corresponding to the found false promotion detection result among the multiple target promotional text contents; The determined target propaganda text content corresponding to the found false propaganda detection result is displayed.

6. A training device for a false advertising detection model, characterized in that: include: A promotional data acquisition module is configured to acquire a plurality of original promotional data; a structured publicity data acquisition module, configured to input each of the plurality of original publicity data into a large language model, and obtain structured publicity data corresponding to each of the original publicity data, wherein the structured publicity data at least includes publicity text content extracted from the original publicity data and a label indicating whether false publicity is involved; A prediction result acquisition module is configured to input the promotional text content into the false propaganda detection model to obtain multiple false propaganda prediction results; A calculation module configured to calculate a value of a loss function based on the plurality of false advertising prediction results and corresponding labels indicating whether false advertising is involved; The training module is configured to train the false advertising detection model by adjusting the parameters of the false advertising detection model according to the value of the loss function.

7. The training device according to claim 6, characterized in that The training device also includes: A manual construction module is configured to obtain structured promotional data obtained through manual construction for each of a preset number of original promotional data included in the plurality of original promotional data; A construction rule learning module is configured to input each of the preset number of original promotional data and each structured promotional data obtained through manual construction into the large language model, and learn the structured promotional data construction rule through the large language model; Wherein, the structured publicity data acquisition module is configured as follows: The other original promotional data among the multiple original promotional data except the preset number of original promotional data are input into the large language model, and the structured promotional data corresponding to the other original promotional data are constructed by the large language model using the learned structured promotional data construction rules.

8. A false advertising detection device, characterized in that: include: A promotional text content acquisition module is configured to acquire target promotional text content; The detection module is configured to input the target propaganda text content into a false propaganda detection model trained according to the training method according to claim 1 or 2, and obtain a false propaganda detection result corresponding to the target propaganda text content.

9. The false advertising detection device according to claim 8, characterized in that: The promotional text content acquisition module is configured as follows: In response to the filter condition setting instruction input by the user, a target filter condition is set, wherein the target filter condition includes at least one of the following items: the promotional text content includes a specified keyword, the release time of the promotional text content is within a preset time period, and the promotional text content comes from a specified platform; The target promotional text content is obtained based on the target screening condition.

10. The false advertising detection device according to claim 8, characterized in that: The target propaganda text content includes a plurality of target propaganda text contents, each of which corresponds to a false propaganda detection result; Wherein, the false advertising detection device also includes: A query field acquisition module is configured to acquire a query field input by a user; A detection result search module is configured to search for a false propaganda detection result containing the query field from a plurality of false propaganda detection results; A determination module is configured to determine the target propaganda text content corresponding to the found false propaganda detection result among the multiple target propaganda text contents; The display module is configured to display the target propaganda text content determined to correspond to the found false propaganda detection result.

11. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the training method of the false advertising detection model as described in claim 1 or 2, or to implement the false advertising detection method as described in any one of claims 3 to 5.

12. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the training method of the false advertising detection model as described in claim 1 or 2, or to execute the false advertising detection method as described in any one of claims 3 to 5.

13. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the training method of the false advertising detection model as described in claim 1 or 2, or implements the false advertising detection method as described in any one of claims 3 to 5.

Images