Public opinion early warning method based on artificial intelligence

Abstract

The present application relates to the field of information and communication technology specifically applicable to the commercial field, and in particular to a public opinion early warning method based on artificial intelligence. The public opinion early warning method based on artificial intelligence includes: obtaining real public opinion materials; predicting the tampering bias of each real public opinion material through a preset real public opinion material tampering bias prediction model; generating corresponding false public opinion pre-training materials according to each real public opinion material through a preset large language model and the tampering bias corresponding to each real public opinion material; training a false public opinion recognition model based on the generated false public opinion pre-training materials; and monitoring online public opinion using the false public opinion recognition model to identify false public opinion from the current online public opinion. The public opinion early warning method based on artificial intelligence provided in the present application can provide more effective public opinion early warning.

Description

Technical Field

[0001] The present application relates to the field of information and communication technology specifically applicable to the commercial field, and in particular to a public opinion early warning method based on artificial intelligence. Background Art

[0002] In the current information age, public opinion warning and management have become an important part of maintaining brand image.

[0003] The existing public opinion warning method based on emotion recognition can only identify whether the public opinion caused by some public opinion materials is positive or negative, and directly warns the staff of negative public opinion. Since the staff of the corresponding brand cannot confirm the authenticity of the public opinion materials that caused the negative public opinion in the first time, they will take the approach of direct denial.

[0004] Existing technologies cannot effectively distinguish between real negative public opinion and fabricated false public opinion in the early stages, resulting in real negative public opinion being incorrectly labeled and covered up in some cases, and unable to be corrected in a timely and accurate manner. When these errors are exposed by social supervision forces, they will increase the public's distrust of brand entities.

[0005] Therefore, the public opinion early warning method provided by the prior art still cannot provide a more effective early warning of public opinion. Summary of the invention

[0006] In order to solve the above technical problems or at least partially solve the above technical problems, the present application provides a public opinion early warning method based on artificial intelligence, which can more effectively warn of public opinion.

[0007] In a first aspect, the present application provides a public opinion early warning method based on artificial intelligence, and the public opinion early warning method based on artificial intelligence includes:

[0008] Obtaining real public opinion materials;

[0009] Predict the tampering bias of each real public opinion material through the preset real public opinion material tampering bias prediction model;

[0010] Through the preset large language model and the tampering bias corresponding to each real public opinion material, the corresponding false public opinion pre-training material is generated according to each real public opinion material;

[0011] A false public opinion recognition model is trained according to the generated false public opinion pre-training materials;

[0012] The false public opinion identification model is used to monitor online public opinion to identify false public opinion from the current online public opinion.

[0013] Optionally, the real public opinion material tampering bias prediction model is trained by the following steps:

[0014] Obtain multiple real public opinion training materials;

[0015] Obtain the known false public opinion training materials corresponding to each real public opinion training material, and mark the type of each false public opinion training material;

[0016] The real public opinion training materials are used as inputs to the real public opinion material tampering bias prediction model framework that has been constructed, and the types of false public opinion training materials corresponding to the real public opinion materials are used as labels, so as to train and obtain the real public opinion material tampering bias prediction model;

[0017] The real public opinion material tampering bias prediction model framework includes:

[0018] A first word vector model extraction layer is used to convert the real public opinion training material into a first word vector matrix through the first word vector model, so as to obtain a representation matrix of the real public opinion training material;

[0019] A first Transformer model framework layer, for inputting the representation matrix of the real public opinion training material into the first Transformer model framework layer for data extraction;

[0020] The first softmax output layer inputs the output of the first Transformer model framework layer into the first softmax output layer to obtain the tampering bias probability matrix of the false public opinion material corresponding to the real public opinion training material, and then obtains the tampering bias of the false public opinion material corresponding to each real public opinion material.

[0021] Optionally, training a false public opinion recognition model according to the generated false public opinion pre-training material includes the following steps:

[0022] Input the false public opinion pre-training materials and the real public opinion materials into the first false public opinion recognition model framework that has been constructed, and use whether it is a false public opinion material as a label to train the first false public opinion recognition model;

[0023] The first false public opinion identification model framework includes:

[0024] A second word vector model extraction layer is used to convert the real public opinion material or the false public opinion pre-training material into a second word vector matrix through the second word vector model, so as to obtain a representation matrix of the real public opinion material or the false public opinion pre-training material;

[0025] The second CNN layer includes multiple convolutional layers with convolutional kernels of different sizes, and inputs the representation matrix of the real public opinion material or the false public opinion pre-training material into the CNN layer to capture word patterns of different lengths through multiple convolutional layers;

[0026] The second pooling layer obtains the output of the CNN layer to perform a maximum pooling operation on the output of each convolutional layer;

[0027] The second LSTM layer obtains the output of the second pooling layer, performs feature processing and then outputs it externally;

[0028] The second fully connected layer obtains the output of the LSTM layer to integrate the output of the second LSTM layer and form the final feature representation

[0029] The second softmax output layer obtains the output of the second fully connected layer to obtain a probability matrix of whether the input public opinion material belongs to true public opinion material or false public opinion material, and then determines whether the input public opinion material belongs to false public opinion material.

[0030] Optionally, the public opinion early warning method based on artificial intelligence further includes the following steps:

[0031] Obtain the key public opinion materials of the current brand entity, which are the public opinion materials that the brand entity is about to release;

[0032] The tampering preference of key public opinion materials is predicted through the real public opinion material tampering bias prediction model to obtain the tampering preference of key public opinion materials;

[0033] According to the tampering preference of the key public opinion material, generating key public opinion pre-training materials of the key public opinion material through a preset large language model;

[0034] Obtain a pre-trained model for identifying false public opinion, wherein the pre-trained model for identifying false public opinion is obtained by the following steps:

[0035] Delete the second softmax layer of the first false public opinion recognition model;

[0036] Freeze the weights of the remaining model parts of the first false public opinion identification model;

[0037] A new third fully connected layer and a third softmax layer are added to the tail of the second fully connected layer, and the weights of the third fully connected layer and the third softmax are initialized to obtain a pre-trained model for false public opinion recognition;

[0038] The false public opinion recognition pre-training model is trained by using key public opinion pre-training materials to obtain a second false public opinion recognition model, including the following steps:

[0039] Input the key public opinion pre-training materials into the first false public opinion recognition model framework that has been constructed, use the false public opinion materials as labels, and train the second false public opinion recognition model;

[0040] The second false public opinion identification model is used to monitor online public opinion to identify false public opinion from the current online public opinion.

[0041] Optionally, the public opinion early warning method based on artificial intelligence further includes the following steps:

[0042] After identifying false public opinions from public opinions, the public opinion materials marked as false public opinions are verified through the preset real public opinion identification model;

[0043] The real public opinion recognition model is obtained by the following steps:

[0044] All existing real public opinion materials are input into the preset large language model to make it remember the content of the real public opinion information, thereby obtaining a real public opinion recognition model for verifying the public opinion materials marked as false public opinion.

[0045] Optionally, the tampering tendency of the false public opinion includes distortion of facts, false propaganda, outdated information and fictitious content.

[0046] Compared with the prior art, the technical solution provided by this application has the following advantages:

[0047] First, this application uses a large language model to generate tampered and biased false pre-training materials corresponding to real public opinion materials, so as to prepare possible false public opinion materials in advance, and train a false public opinion recognition model to recognize these false public opinion materials in advance, thereby significantly enhancing the sensitivity and response capabilities of the false public opinion recognition model to various false information techniques. This training strategy enables the recognition model to quickly and accurately identify the false characteristics of similar false information when it encounters it in actual applications.

[0048] Secondly, the public opinion warning method based on artificial intelligence provided by this application realizes dynamic adjustment of the recognition sensitivity of different types of false public opinion materials by combining transfer learning, and realizes a dynamic adaptive way to identify and process false public opinion. The method first obtains the public opinion materials that the brand entity is about to release, and uses the pre-trained model to predict the possible tampering preferences of these materials, and strengthens the recognition sensitivity of the first false public opinion recognition model to the types of false public opinions corresponding to these tampering preferences through transfer learning, thereby obtaining a second false public opinion recognition model that is more suitable for capturing false public opinions in the current time period.

[0049] Finally, since the false public opinion recognition model is trained based on the generated false public opinion pre-training materials. When the new public opinion event is similar in content to the false pre-training materials previously generated through real public opinion materials, for example, a product has significantly improved its performance after multiple iterations, the false public opinion recognition model will confuse the real positive public opinion with the false public opinion of previously simulated false propaganda. In this case, the public opinion early warning method provided by the present application uses a large language model to only store the key information of the real public opinion materials and evaluates the public opinion materials that are preliminarily identified as false public opinion materials based on the information, which can effectively distinguish such coincidental similarities and reduce the possibility of final misjudgment.

[0050] Therefore, the public opinion early warning method based on artificial intelligence provided in this application can provide early warning of public opinion more effectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A flowchart of the public opinion warning method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The technical solution in this application will be described below in conjunction with the accompanying drawings.

[0053] In the following description, many specific details are set forth to facilitate a full understanding of the present application, but the present application can also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only part of the embodiments of the present application, not all of the embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0054] In a first aspect, the present application provides a public opinion early warning method based on artificial intelligence, and the public opinion early warning method based on artificial intelligence includes:

[0055] S101: Obtain real public opinion materials.

[0056] In the embodiment of the present application, the real public opinion material is text material, including the brand entity's past news, financial reports, and product release articles about the brand entity.

[0057] S102: Predicting the tampering bias of each real public opinion material through a preset real public opinion material tampering bias prediction model.

[0058] Specifically, the real public opinion material tampering bias prediction model is trained by the following steps:

[0059] Obtain multiple real public opinion training materials;

[0060] Obtain the known false public opinion training materials corresponding to each real public opinion training material, and mark the type of each false public opinion training material;

[0061] The real public opinion training materials are used as inputs to the real public opinion material tampering bias prediction model framework that has been constructed, and the types of false public opinion training materials corresponding to the real public opinion materials are used as labels, so as to train and obtain the real public opinion material tampering bias prediction model;

[0062] The real public opinion material tampering bias prediction model framework includes:

[0063] A first word vector model extraction layer is used to convert the real public opinion training material into a first word vector matrix through the first word vector model, so as to obtain a representation matrix of the real public opinion training material;

[0064] A first Transformer model framework layer, for inputting the representation matrix of the real public opinion training material into the first Transformer model framework layer for data extraction;

[0065] The first softmax output layer inputs the output of the first Transformer model framework layer into the first softmax output layer to obtain the tampering bias probability matrix of the false public opinion material corresponding to the real public opinion training material, and then obtains the tampering bias of the false public opinion material corresponding to each real public opinion material.

[0066] Specifically, the step of obtaining the known false public opinion training materials corresponding to each real public opinion training material needs to be manually collected on the Internet.

[0067] Specifically, the types of false public opinion training materials and the tampering tendencies of false public opinion include distortion of facts, false propaganda, outdated information and fictitious content.

[0068] S103: Generate corresponding false public opinion pre-training materials according to each real public opinion material through a preset large language model and the tampering bias corresponding to each real public opinion material.

[0069] Specifically, generating corresponding false public opinion pre-training materials according to each real public opinion material through a preset large language model and the tampering bias corresponding to each real public opinion material includes the following steps:

[0070] Load a pre-trained large language model, which is ChatGPT in the embodiment of the present application, and may also be a large language model system such as Wenxinyiyan in other embodiments.

[0071] A plurality of representative real public opinion materials, false public opinion materials and the types of the false public opinion materials are manually selected as prompts and input into the large language model.

[0072] The types of multiple false public opinion materials here need to cover the types of tampering tendencies of false public opinion.

[0073] Input new real public opinion materials and request the large language model to generate fake public opinion material text based on the specified tampering bias type;

[0074] Finally, adjust the input prompts based on the results until satisfactory results are obtained, and instruct the large language model to remember this way of generating false public opinion materials.

[0075] In the future, as long as the real public opinion material is input into this adjusted large language model, the corresponding false public opinion pre-training material can be generated through the adjusted large language model.

[0076] S104: A false public opinion recognition model is trained according to the generated false public opinion pre-training material.

[0077] Specifically, training a false public opinion recognition model according to the generated false public opinion pre-training material includes the following steps:

[0078] Input the false public opinion pre-training materials and the real public opinion materials into the first false public opinion recognition model framework that has been constructed, and use whether it is a false public opinion material as a label to train the first false public opinion recognition model;

[0079] The first false public opinion identification model framework includes:

[0080] A second word vector model extraction layer is used to convert the real public opinion material or the false public opinion pre-training material into a second word vector matrix through the second word vector model, so as to obtain a representation matrix of the real public opinion material or the false public opinion pre-training material;

[0081] The second CNN layer includes multiple convolutional layers with convolutional kernels of different sizes, and inputs the representation matrix of the real public opinion material or the false public opinion pre-training material into the CNN layer to capture word patterns of different lengths through multiple convolutional layers;

[0082] The second pooling layer obtains the output of the CNN layer to perform a maximum pooling operation on the output of each convolutional layer;

[0083] The second LSTM layer obtains the output of the second pooling layer, performs feature processing and then outputs it externally;

[0084] The second fully connected layer obtains the output of the LSTM layer to integrate the output of the second LSTM layer and form the final feature representation

[0085] The second softmax output layer obtains the output of the second fully connected layer to obtain a probability matrix of whether the input public opinion material belongs to true public opinion material or false public opinion material, and then determines whether the input public opinion material belongs to false public opinion material.

[0086] S105: Monitor online public opinion using the false public opinion identification model to identify false public opinion from current online public opinion.

[0087] In one embodiment, the first false public opinion identification model is directly used to monitor online public opinion.

[0088] Its specific working principle and beneficial effects are:

[0089] First, this application uses a large language model to generate tampered and biased false pre-training materials corresponding to real public opinion materials, so as to prepare possible false public opinion materials in advance, and train a false public opinion recognition model to recognize these false public opinion materials in advance, thereby significantly enhancing the sensitivity and response capabilities of the false public opinion recognition model to various false information techniques. This training strategy enables the recognition model to quickly and accurately identify the false characteristics of similar false information when it encounters it in actual applications.

[0090] In another embodiment, the public opinion early warning method based on artificial intelligence further includes the following steps:

[0091] Obtain the key public opinion materials of the current brand entity, which are the public opinion materials that the brand entity is about to release;

[0092] The tampering preference of key public opinion materials is predicted through the real public opinion material tampering bias prediction model to obtain the tampering preference of key public opinion materials;

[0093] According to the tampering preference of the key public opinion material, generating key public opinion pre-training materials of the key public opinion material through a preset large language model;

[0094] Obtain a pre-trained model for identifying false public opinion, wherein the pre-trained model for identifying false public opinion is obtained by the following steps:

[0095] Delete the second softmax layer of the first false public opinion recognition model;

[0096] Freeze the weights of the remaining model parts of the first false public opinion identification model;

[0097] A new third fully connected layer and a third softmax layer are added to the tail of the second fully connected layer, and the weights of the third fully connected layer and the third softmax are initialized to obtain a pre-trained model for false public opinion recognition;

[0098] The false public opinion recognition pre-training model is trained by using key public opinion pre-training materials to obtain a second false public opinion recognition model, including the following steps:

[0099] Input the key public opinion pre-training materials into the first false public opinion recognition model framework that has been constructed, use the false public opinion materials as labels, and train the second false public opinion recognition model;

[0100] The second false public opinion identification model is used to monitor online public opinion to identify false public opinion from the current online public opinion.

[0101] Specifically, the public opinion materials that the brand entity is about to release in the embodiment of the present application are the financial report and new product articles that the brand entity is about to release.

[0102] Its specific working principle and beneficial effects lie in that the artificial intelligence-based public opinion warning method provided in this application realizes the adjustment of the dynamic recognition sensitivity of different types of false public opinion materials by combining transfer learning, thereby realizing a dynamic and adaptive way to identify and process false public opinions.

[0103] This method first obtains the public opinion materials that the brand entity is about to release, and uses a pre-trained model to predict the possible tampering preferences of these materials, and then uses transfer learning to enhance the recognition sensitivity of the first false public opinion recognition model to the false public opinions of the types corresponding to these tampering preferences, thereby obtaining a second false public opinion recognition model that is more suitable for capturing false public opinions in the current time period.

[0104] Because the types of false public opinions will change dynamically with the different public opinion information to be released, such as product releases or financial reports, this model can dynamically adjust the sensitivity of different types of false public opinions through the public opinion materials to be released. This is to cope with the increase in data traffic of a certain type of false public opinion on the network due to the specificity of the event, thereby ensuring that high performance and accuracy can be maintained during high traffic periods.

[0105] Therefore, the present application can achieve more effective management of computing resources by dynamically adjusting the sensitivity, especially when the data traffic of a certain type of false public opinion may increase significantly in the future. The present application can concentrate resources to process the corresponding high-risk public opinion, thereby optimizing the overall resource allocation and utilization.

[0106] In one embodiment, the public opinion early warning method based on artificial intelligence further includes the following steps:

[0107] After identifying false public opinions from public opinions, the public opinion materials marked as false public opinions are verified through the preset real public opinion identification model;

[0108] The real public opinion recognition model is obtained by the following steps:

[0109] All existing real public opinion materials are input into the preset large language model to make it remember the content of the real public opinion information, thereby obtaining a real public opinion recognition model for verifying the public opinion materials marked as false public opinion.

[0110] Specifically, the preset large language model is the ChatGPT large language model. In other embodiments, other large language models such as Wenxinyiyan may also be used.

[0111] Its specific working principle and beneficial effect are that the false public opinion recognition model is trained based on the generated false public opinion pre-training materials. When the new public opinion event is highly similar in content to the false pre-training materials previously generated through real public opinion materials, for example, the performance of a product is significantly improved after multiple iterations, the false public opinion recognition model will confuse the real positive public opinion with the false public opinion of previously simulated false propaganda. In this case, the public opinion early warning method provided by the present application uses a large language model to store the key information of the real public opinion material and can be evaluated based on the implied information, which can effectively distinguish such coincidental similarities and reduce the possibility of final misjudgment.

[0112] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. In addition, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. Moreover, in the description of the embodiments of the present application, unless otherwise specified, " / " means or, for example, A / B can represent A or B; "and / or" in this article is only a kind of association relationship describing the associated objects, indicating that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. Furthermore, in the description of the embodiments of the present application, “plurality” refers to two or more than two.

[0113] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments described herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. The public opinion early warning method based on artificial intelligence is characterized by: The public opinion early warning method based on artificial intelligence includes: Obtaining real public opinion materials; Predict the tampering bias of each real public opinion material through the preset real public opinion material tampering bias prediction model; Through the preset large language model and the tampering bias corresponding to each real public opinion material, the corresponding false public opinion pre-training material is generated according to each real public opinion material; A false public opinion recognition model is obtained by training the generated false public opinion pre-training materials; Using the false public opinion identification model to monitor online public opinion, so as to identify false public opinion from the current online public opinion; The real public opinion material tampering bias prediction model is trained by the following steps: Obtain multiple real public opinion training materials; Obtain the known false public opinion training materials corresponding to each real public opinion training material, and mark the type of each false public opinion training material; The real public opinion training materials are used as inputs to the real public opinion material tampering bias prediction model framework that has been constructed, and the types of false public opinion training materials corresponding to the real public opinion materials are used as labels, so as to train and obtain the real public opinion material tampering bias prediction model; The real public opinion material tampering bias prediction model framework includes: A first word vector model extraction layer is used to convert the real public opinion training material into a first word vector matrix through the first word vector model, so as to obtain a representation matrix of the real public opinion training material; A first Transformer model framework layer, for inputting the representation matrix of the real public opinion training material into the first Transformer model framework layer for data extraction; The first softmax output layer inputs the output of the first Transformer model framework layer into the first softmax output layer to obtain the tampering bias probability matrix of the false public opinion material corresponding to the real public opinion training material, and then obtains the tampering bias of the false public opinion material corresponding to each real public opinion material; The tampering tendencies of the false public opinion materials include distortion of facts, false propaganda, outdated information and fictitious content.

2. The public opinion early warning method based on artificial intelligence according to claim 1 is characterized in that: The false public opinion recognition model is trained according to the generated false public opinion pre-training material, including the following steps: Input the false public opinion pre-training materials and the real public opinion materials into the first false public opinion recognition model framework that has been constructed, and use whether it is a false public opinion material as a label to train the first false public opinion recognition model; The first false public opinion identification model framework includes: A second word vector model extraction layer is used to convert the real public opinion material or the false public opinion pre-training material into a second word vector matrix through the second word vector model, so as to obtain a representation matrix of the real public opinion material or the false public opinion pre-training material; The second CNN layer includes multiple convolutional layers with convolutional kernels of different sizes, and inputs the representation matrix of the real public opinion material or the false public opinion pre-training material into the CNN layer to capture word patterns of different lengths through multiple convolutional layers; The second pooling layer obtains the output of the CNN layer to perform a maximum pooling operation on the output of each convolutional layer; The second LSTM layer obtains the output of the second pooling layer, performs feature processing and then outputs it externally; The second fully connected layer obtains the output of the LSTM layer to integrate the output of the second LSTM layer and form the final feature representation The second softmax output layer obtains the output of the second fully connected layer to obtain a probability matrix of whether the input public opinion material belongs to true public opinion material or false public opinion material, and then determines whether the input public opinion material belongs to false public opinion material.

3. The public opinion early warning method based on artificial intelligence according to claim 2 is characterized in that: The public opinion early warning method based on artificial intelligence also includes the following steps: Obtain the key public opinion materials of the current brand entity, which are the public opinion materials that the brand entity is about to release; The tampering preference of key public opinion materials is predicted through the real public opinion material tampering bias prediction model to obtain the tampering preference of key public opinion materials; According to the tampering preference of the key public opinion material, generating key public opinion pre-training materials of the key public opinion material through a preset large language model; Obtain a pre-trained model for identifying false public opinion, wherein the pre-trained model for identifying false public opinion is obtained by the following steps: Delete the second softmax layer of the first false public opinion recognition model; Freeze the weights of the remaining model parts of the first false public opinion identification model; A new third fully connected layer and a third softmax layer are added to the tail of the second fully connected layer, and the weights of the third fully connected layer and the third softmax are initialized to obtain a pre-trained model for false public opinion recognition; The false public opinion recognition pre-training model is trained by using key public opinion pre-training materials to obtain a second false public opinion recognition model, including the following steps: Input the key public opinion pre-training materials into the first false public opinion recognition model framework that has been constructed, use the false public opinion materials as labels, and train the second false public opinion recognition model; The second false public opinion identification model is used to monitor online public opinion to identify false public opinion from the current online public opinion.

4. The public opinion early warning method based on artificial intelligence according to claim 3 is characterized in that: The public opinion early warning method based on artificial intelligence also includes the following steps: After identifying false public opinions from public opinions, the public opinion materials marked as false public opinions are verified through the preset real public opinion identification model; The real public opinion recognition model is obtained by the following steps: All existing real public opinion materials are input into the preset large language model to make it remember the content of the real public opinion information, thereby obtaining a real public opinion recognition model for verifying the public opinion materials marked as false public opinion.

Images