Large model forged information detection method based on news emotion and writing style

Abstract

The invention discloses a large model forged information detection method based on news emotion and writing styles, which comprises the following steps of: collecting original news texts, and generating generated news texts with emotion colors and different writing styles by utilizing a generator LLM, reasoning and judging by the news discriminator LLM according to the generated news text to obtain an initial prediction result of news authenticity, and generating adversarial learning between the generator LLM and the news discriminator LLM; judging an expert field to which the collected original news text belongs by utilizing an expert discriminator LLM, and generating feedback for an initial prediction result by utilizing an expert feedback LLM according to the expert field judged by the expert discriminator LLM; and performing final prediction based on expert feedback, comprehensively considering each expert feedback generated by the expert feedback LLM, and finally predicting a final prediction result generated by the LLM. According to the method, the detection requirements of news in various different fields in real life are met, and the experimental performance of LLM false information detection is improved.

Description

Technical Field

[0001] The present invention relates to a large-model forged information detection method based on news sentiment and writing style. Background Art

[0002] With the development of the Internet, people can get information on social media and express their opinions. However, due to the lack of supervision of Internet platforms and citizens' abuse of Internet resources, this can also lead to the spread of false information. The spread of false information will harm social order, endanger the national economy, and bring unnecessary panic and anxiety to people. Therefore, fake information detection methods are needed to identify the authenticity of news.

[0003] Since fake news is more likely to be emotional or inflammatory than real news, and the writing styles of real news and fake news are very different, fake news usually uses a specific writing style to attract and persuade consumers. Therefore, news sentiment and writing style are important factors affecting the truth of news. How to use these factors to judge the truth of news is a big challenge. However, in the existing fake news detection methods using large language models (LLM), LLM has limited ability to identify news sentiment and writing style, and LLM has limited ability to understand news in different fields. Most existing methods only focus on one field, but the news sentiment and writing style in different fields are different. For example, fake news publishers in the political field show more disgust, while fake news publishers in the health field show more anger. Ignoring field differences will reduce experimental performance.

[0004] Therefore, the present invention is just produced based on the above shortcomings. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a large-model forged information detection method based on news sentiment and writing style.

[0006] The present invention is achieved through the following technical solutions:

[0007] A method for detecting fake information based on a large model of news sentiment and writing style, characterized by comprising the following steps:

[0008] S1. Collect original news texts, use the generator LLM to generate news texts with emotional colors and different writing styles, and the news discriminator LLM infers and judges the generated news texts to obtain the initial prediction results of the truth or falsehood of the news, and generate adversarial learning between the generator LLM and the news discriminator LLM;

[0009] S2, using the expert discriminator LLM to determine the expert field to which the collected original news text belongs, and the expert feedback LLM generates feedback on the initial prediction result according to the expert field determined by the expert discriminator LLM;

[0010] S3. Make final prediction based on expert feedback, comprehensively consider each expert feedback generated by the expert feedback LLM, and finally predict the final prediction result generated by the LLM.

[0011] The large model forged information detection method based on news sentiment and writing style is characterized in that: step S1 includes:

[0012] S11. Generator LLM generates news sentiment and writing style. Set generator LLM as , based on the original news text , according to the generation strategy Generate news text with emotional colors and different writing styles and the related generative explanation , the resulting formula is: ;

[0013] S12, discriminator LLM reasoning judgment, set the news discriminator LLM to be , using the generated news text with emotional colors and different writing styles and identification strategies for Design prompts to get the initial prediction results of whether the news is true or false and the corresponding generated explanation , the formula for reasoning is: ;

[0014] S13, generation and recognition strategy update, a strategic confrontation balance is established between the generator LLM and the news discriminator LLM, so that the generation strategy and recognition strategy can be automatically improved and upgraded; when the news discriminator LLM can correctly predict, the generator LLM needs to update the generation strategy to generate more complex emotions and writing style news, the calculation formula is: = ( ); When the news discriminator LLM makes an incorrect recognition, the news discriminator LLM needs to update the recognition strategy, and the update formula is: = ( ).

[0015] The large model forged information detection method based on news sentiment and writing style is characterized in that: step S2 includes:

[0016] S21, expert domain discrimination, designing a prompt learning template for the expert discriminator LLM, and the output of the expert discriminator LLM is all the experts involved in the news;

[0017] S22, expert feedback generation, giving experts feedback on LLM design tips, allowing them to generate feedback on the news discriminator LLM prediction results. The feedback calculation formula generated by each expert is:

[0018]

[0019] in, represents the i-th expert;

[0020] When there are k experts, the feedback generated by all experts is expressed as .

[0021] Compared with the prior art, the present invention has the following advantages:

[0022] 1. This large model forged information detection method has the ability to identify news emotions and writing styles. Based on the idea of ​​generative adversarial learning, a generative adversarial large model (LLM-GAN) is proposed, which allows the LLM as a generator to generate news emotions and writing styles, and makes the LLM as a news discriminator as much as possible unable to recognize them, so that the news discriminator LLM tries hard to recognize them. LLM-GAN allows the generator LLM and the news discriminator LLM to continuously update strategies and learn how to judge the authenticity of news based on the user's emotions and positions; then let experts judge LLM and determine the expert field to which the news belongs, and then let news experts generate feedback on the prediction results of the news discriminator LLM. Finally, the feedback of all experts is comprehensively considered to give the final prediction, which meets the detection needs of news in various fields in real life and improves the experimental performance of LLM in detecting false information. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the present invention. DETAILED DESCRIPTION

[0024] The present invention will be further described below in conjunction with the accompanying drawings:

[0025] like Figure 1 As shown, a large model forged information detection method based on news sentiment and writing style includes the following steps:

[0026] S1. Collect original news texts, use the generator LLM to generate news texts with emotional colors and different writing styles, and the news discriminator LLM infers and judges the generated news texts to obtain the initial prediction results of the truth or falsehood of the news, and generate adversarial learning between the generator LLM and the news discriminator LLM;

[0027] S2, using the expert discriminator LLM to determine the expert field to which the collected original news text belongs, and the expert feedback LLM generates feedback on the initial prediction result according to the expert field determined by the expert discriminator LLM;

[0028] S3. Make final prediction based on expert feedback, comprehensively consider each expert feedback generated by the expert feedback LLM, and finally predict the final prediction result generated by the LLM.

[0029] In this method, step S1 includes three parts: the generator LLM generates news sentiment and writing style, and the news discriminator LLM infers, judges, generates and updates the recognition strategy.

[0030] S11. Generator LLM generates news sentiment and writing style. Set generator LLM as , based on the original news text , according to the generation strategy Generate news text with emotional colors and different writing styles and the related generative explanation , the resulting formula is: ;

[0031] S12, discriminator LLM reasoning judgment, set the news discriminator LLM to be , using the generated news text with emotional colors and different writing styles and identification strategies for Design prompts to get the initial prediction results of whether the news is true or false and the corresponding generated explanation , the formula for reasoning is: ;

[0032] S13, generate and identify strategy updates. Better promote By learning the impact of news sentiment and writing style on the truth of news, a strategic confrontation balance is established between the generator LLM and the news discriminator LLM, so that they can automatically improve and upgrade the generation strategy and recognition strategy respectively; when the news discriminator LLM can correctly predict, the generator LLM needs to update the generation strategy to generate more complex sentiment and writing style news. The calculation formula is: = ( );

[0033] On the contrary, when the news discriminator LLM makes an incorrect recognition, the news discriminator LLM needs to update the recognition strategy, and the update formula is:

[0034] = ( );

[0035] The preliminary prediction results are:

[0036] ;

[0037] in, They represent the discriminator’s preliminary prediction results and the explanation of the preliminary predictions respectively.

[0038] In this method, step S2 includes two parts: expert domain identification and expert feedback generation.

[0039] S21, expert domain identification, design a prompt learning template for the expert discriminator LLM, let LLM determine the expert domain to which the news belongs, and the prompt learning template is: "You are given a piece of news, please identify the expert area of ​​the news". The output result of the expert discriminator LLM is all the experts involved in the news;

[0040] S22, expert feedback generation, giving experts feedback on LLM design tips, allowing them to generate feedback on the news discriminator LLM prediction results. The feedback calculation formula generated by each expert is:

[0041]

[0042] in, represents the i-th expert;

[0043] When there are k experts, the feedback generated by all experts is expressed as .

[0044] In general, this large model forged information detection method has the ability to identify news emotions and writing styles. Based on the idea of ​​generative adversarial learning, a generative adversarial large model is proposed, denoted as LLM-GAN, which allows LLM as a generator to generate news emotions and writing styles, and makes LLM as a news discriminator as much as possible unable to recognize them, so that the news discriminator LLM tries hard to recognize them. LLM-GAN allows the generator LLM and the news discriminator LLM to continuously update strategies and learn how to judge the authenticity of news based on the user's emotions and positions; then let experts judge LLM and judge the expert field to which the news belongs, and then let news experts generate feedback on the prediction results of the news discriminator LLM. Finally, the feedback of all experts is comprehensively considered to give the final prediction, which meets the detection needs of news in various fields in real life and improves the experimental performance of LLM in detecting false information.

Claims

1. A large model forged information detection method based on news sentiment and writing style, characterized in that: The following steps are involved: S1. Collect original news texts, use the generator LLM to generate news texts with emotional colors and different writing styles, and the news discriminator LLM infers and judges the generated news texts to obtain the initial prediction results of the truth or falsehood of the news, and generate adversarial learning between the generator LLM and the news discriminator LLM; S2, using the expert discriminator LLM to determine the expert field to which the collected original news text belongs, and the expert feedback LLM generates feedback on the initial prediction result according to the expert field determined by the expert discriminator LLM; S3. Make final prediction based on expert feedback, comprehensively consider each expert feedback generated by the expert feedback LLM, and finally predict the final prediction result generated by the LLM.

2. The method for detecting fake information based on a large model of news sentiment and writing style according to claim 1 is characterized by: The step S1 includes: S11. Generator LLM generates news sentiment and writing style. Set generator LLM as , based on the original news text , according to the generation strategy Generate news text with emotional colors and different writing styles and the related generative explanation , the resulting formula is: ; S12, discriminator LLM reasoning judgment, set the news discriminator LLM to be , using the generated news text with emotional colors and different writing styles and identification strategies for Design prompts to get the initial prediction results of whether the news is true or false and the corresponding generated explanation , the formula for reasoning is: ; S13, generation and recognition strategy update, establishes a strategic confrontation balance between the generator LLM and the news discriminator LLM, so that the generation strategy and recognition strategy can be automatically improved and upgraded; When the news discriminator LLM can correctly predict, the generator LLM needs to update the generation strategy to generate more complex emotional and writing style news. The calculation formula is: = ( ); When the news discriminator LLM makes an incorrect recognition, the news discriminator LLM needs to update the recognition strategy, and the update formula is: = ( ).

3. The method for detecting fake information based on a large model of news sentiment and writing style according to claim 1 is characterized by: The step S2 includes: S21, expert domain discrimination, designing a prompt learning template for the expert discriminator LLM, and the output of the expert discriminator LLM is all the experts involved in the news; S22, expert feedback generation, giving experts feedback on LLM design tips, allowing them to generate feedback on the news discriminator LLM prediction results. The feedback calculation formula generated by each expert is: ； in, represents the i-th expert; When there are k experts, the feedback generated by all experts is expressed as .

Images