Instant detection method for error information

Abstract

The invention provides an error information instant detection method which can provide more accurate, timely and more credible detection results in a dynamic and complex error information environment. The method comprises the following steps: (1) a retrieval enhancement generation step: (1) retrieving context information related to declarations needing to be classified on a network; (2) the context information is further extracted, and several contents with the highest ranking are extracted; (3) inputting several extracted contents with the highest ranking into a large language model, and summarizing context information by the large language model; (4) enabling the large language model to output a prediction result of statement classification and related evidence, naming the content output by the large language model as priori knowledge, and then combining the statement with the priori knowledge to complete the expansion of the statement; (2) an information classification step; and inputting the expanded declaration into a large language model for classification to obtain a final detection result.

Description

Technical Field

[0001] The invention relates to a method for instant error information detection. Background Art

[0002] Subtle but persistent misinformation has long disrupted the post-truth society. Studies have shown that among all categories of information, false information spreads farther, faster, deeper, and wider than true information. Therefore, real-time misinformation detection is an urgent task to reduce its negative impact on society.

[0003] There are many methods for misinformation detection, which can be divided into content-based, social context-based, and external knowledge-based detection methods. Content-based methods include recurrent neural networks, convolutional neural networks, and transformer models, which can learn features from the content of information and classify it. User interaction on social media gives information additional features, such as using user information for misinformation detection, and the propagation pattern of information can also be used to detect misinformation. In addition, external knowledge can also effectively help detect misinformation. However, these methods may not be fully effective when faced with new data whose features are different from the learned data. Even if external knowledge is used, most current studies are based on static knowledge bases. Once there is no corresponding corpus in the knowledge base, the detection ability will decrease.

[0004] Large language models (LLMs) have gradually attracted widespread attention due to their powerful performance. The extensive knowledge and powerful reasoning ability of large language models have led to corresponding research in many fields, and have achieved good results in solving various complex problems. The emergence of large language models has also brought new opportunities for error message detection. Retrieval-augmented generation (RAG) technology allows large language models to use external knowledge to improve the accuracy and reliability of their output. Recently, more and more studies have used large language models (LLMs) to solve the problem of error message detection. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned deficiencies in the prior art and to provide a reasonably designed method for real-time error information detection, which can provide more accurate, timely and more reliable detection results in the face of dynamic and complex error information environments.

[0006] The technical solution adopted by the present invention to solve the above problem is: a method for real-time detection of error information, characterized in that it includes the following steps: (I) The search enhancement generation step includes the following steps: (1) For the statement that needs to be classified, retrieve the relevant context information on the Internet; (2) further refining the context information to extract the top-ranked pieces of content; (3) The extracted top-ranked pieces of content are fed into the large language model to allow it to summarize the contextual information; (4) Let the large language model output the prediction results of the statement classification and related evidence. This part of the large language model output is called prior knowledge. Then, the statement is combined with the prior knowledge to complete the expansion of the statement. (ii) Information classification steps; The expanded statement in step (a) (4) is input into the large language model for classification to obtain the final detection result.

[0007] In step (1) of the present invention, a search is performed on the Internet through the Brave Search API.

[0008] In step (1) of the present invention, the context information is continuously iterated.

[0009] In step (1) (2) of the present invention, the top five contents are extracted.

[0010] In step (i) of the present invention, the large language model used is the Llama-3-70B-Instruct model.

[0011] The present invention evaluates the retrieval enhancement generation step: the t-SNE method is used to reduce the dimension of the statement and the retrieved context information and visualize it, and the correlation between the context information and the statement is examined; the RAGAs method is also used to evaluate the generation performance of the retrieval enhancement generation step.

[0012] In step (ii) of the present invention, the large language model used is the Llama-3-8B-Instruct model.

[0013] The present invention fine-tunes the large language model in step (ii).

[0014] The present invention uses the PEFT method to fine-tune the large language model in step (ii).

[0015] The present invention uses the DoRA method in the PEFT method to fine-tune the large language model in step (ii).

[0016] Compared with the prior art, the present invention has the following advantages and effects: it can provide more accurate, timely and reliable detection results in the face of dynamic and complex error information environments. The present invention also gets rid of the reliance on local knowledge bases and third-party fact-checking information, and to a certain extent solves the cold start problem of traditional error information detection methods. The present invention can not only make the detection results clearer and more intuitive, but also reduce people's confusion about it, making error information detection more meaningful. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The following is a flowchart of an embodiment of the present invention.

[0018] Figure 2 This is a comparison result diagram of the credibility of the three LLM models in the experiment of the present invention.

[0019] Figure 3 This is a comparison result diagram of the correctness of the answers of the three LLM models in the experiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and by way of examples. The following examples are intended to explain the present invention but the present invention is not limited to the following examples.

[0021] 1. The instant error information detection method of the embodiment of the present invention comprises the following steps: (I) Retrieval enhancement generation steps.

[0022] The retrieval enhancement generation step aims to enhance the capabilities of the large language model (LLM) by integrating external knowledge in the generation process, with the main purpose of improving the accuracy, credibility, and relevance of the generated content. It includes the following steps: (1) For claims that need to be classified, we first retrieve relevant contextual information on the Internet through the Brave Search API; Brave Search API is a privacy-based search engine.

[0023] In this step, the context information is continuously iterated in a loop to enhance the detection capability and adaptability to new data of the present invention, thereby overcoming the limitations of human cognition.

[0024] (2) Since the searched context information may contain noise that is not helpful for the final classification, it is necessary to further refine the context information and extract the top five highest-ranked contents.

[0025] (3) The five highest-ranked extracted contents are input into the large language model (LLM) to summarize the contextual information.

[0026] (4) Let the large language model LLM output the prediction results and related evidence for the classification of the claim. This is mainly to provide a preliminary judgment and reasoning process for the subsequent classification task. This part of the content output by the large language model LLM is called prior knowledge. Subsequently, the claim is combined with the prior knowledge to complete the expansion of the claim.

[0027] (5) In order to ensure the effectiveness of the retrieval enhancement generation step, a comprehensive evaluation is performed on it. The t-SNE method is used to reduce the dimension of the claim and the retrieved context information and visualize it to examine the correlation between the context information and the claim. RAGAs is also used to evaluate the generation performance of the retrieval enhancement generation step, and one of the multiple models in the large language model (LLM) that performs best on this task is selected as the basis for subsequent classification experiments.

[0028] In this step, the large language model LLM used is the Llama-3-70B-Instruct model.

[0029] (B) Information classification steps.

[0030] The expanded claim in step (a) (4) is input into the large language model LLM for classification to obtain the final detection result.

[0031] Prior knowledge can expand the claim, but there may be some erroneous content that affects the large language model LLM from making the correct judgment. Therefore, the large language model LLM can be fine-tuned, and the fine-tuning can also make up for the impact of these errors on the final classification results. In order to standardize the final output of the large language model LLM, that is, to output the classification result TRUE or FALSE, the PEFT (Parameter-Efficient Fine-Tuning) method is used to fine-tune the large language model LLM. The PEFT methods include LoRA, LoRA+, VeRA, rsLoRA, DoRA and other methods. In this embodiment, the DoRA method is used to fine-tune the large language model of step (ii).

[0032] In this step, the large language model LLM used is the Llama-3-8B-Instruct model.

[0033] 2. Experiment.

[0034] (I) In order to ensure the reliability of the retrieval augmentation generation step, a series of experiments were conducted on the retrieval augmentation generation step using RAGAs (Retrieval Augmented Generation Assessment), which verified the effectiveness of the retrieval augmentation generation step and enhanced the credibility of the model's decision.

[0035] For the generation performance of the retrieval enhancement generation step, experiments were conducted on the SOLAR-10.7B-Instruct, Mixtral-8x7B-Instruct, and Llama-3-70B-Instruct models. Figure 2 and Figure 3The comparison results of the generation performance of the three LLM models can be obtained in . The horizontal axis represents the interval of a certain indicator, and the vertical axis represents the number of samples falling into the interval. The experimental results show that in the retrieval enhancement generation step, the three LLM models have different degrees of differences in the scores of credibility and answer correctness. In terms of credibility, it can be clearly seen that the scores of Llama-3-70B-Instruct are mostly distributed in a higher interval ((0.8, 1.0]), while the scores of the other two LLM models are less distributed in this interval. From the perspective of average credibility score, Llama-3-70B-Instruct has the highest average score of 0.923, Mixtral-8x7B-Instruct has an average of 0.900, and SOLAR-10.7B-Instruct has the lowest average of 0.851. It can be seen that in credibility evaluation, Llama-3-70B-Instruct performs the best among these LLM models. In terms of answer correctness, in the higher interval ((0.8, 1.0]), the number of Llama-3-70B-Instruct scores is also higher than that of the other two LLMs. From the perspective of average answer correctness score, Llama-3-70B-Instruct has the highest average score of 0. 767, Mixtral-8x7B-Instruct has an average score of 0.650, and SOLAR-10.7B-Instruct has the lowest average score of 0.626. Therefore, in the evaluation of answer correctness, Llama-3-70B-Instruct is still the best performing LLM model. Considering these two indicators comprehensively, Llama-3-70B-Instruct performs well in both credibility and answer correctness, ranking first; Mixtral-8x7B-Instruct ranks second; SOLAR-10.7B-Instruct- has the worst performance among these LLMs. The overall generation performance ranking is: Llama-3-70B-Instruct>Mixtral-8x7B-Instruct>SOLAR-10.7B-Instruct. Finally, Llama-3-70B-Instruct was selected as the LLM model for generating prior knowledge in subsequent experiments.

[0036] (ii) Table 1 shows the classification results of different models and fine-tuning strategies. In the retrieval enhancement generation step, Llama-3-70B-Instruct is used to generate prior knowledge for text amplification.

[0037] Table 1.

[0038] .

[0039] Using data labeled true and false in COVMIS-2024, we conducted a binary classification combination experiment of multiple LLM models and fine-tuning strategies, in which Llama-3-70B-Instruct was used to generate prior knowledge. From Table 1, it can be seen that for different LLM models, the performance differences of each fine-tuning strategy are not exactly the same. For example, when Llama-3-8B-Instruct is fine-tuned for rsLoRA, its effect is better than LoRA, but the effect of rsLoRA fine-tuning for the other two models is not as good as LoRA. From the comprehensive performance point of view, Llama-3-8B-Instruct performs more outstandingly. From the perspective of fine-tuning strategy, DoRA performs more outstandingly. If the LLM model and fine-tuning strategy are considered comprehensively, the combination of Llama-3-8B-Instruct and DoRA strategy achieves the best results. So far, the current optimal settings have been obtained: the large language model LLM of the retrieval enhancement generation step is Llama-3-70B-Instruct, the large language model LLM of the information classification step is Llama-3-8B-Instruct, and the fine-tuning strategy for classification is the DoRA method, called CEMDo.

[0040] Table 2.

[0041] .

[0042] Table 2 shows the results of using CEMDo for binary classification tasks on the LIAR2 dataset and the COVMIS-2024 dataset. From the experimental results in Table 2, it can be seen that CEMDo is better than FDHN on both the COVMIS-2024 dataset and the LIAR2 dataset.

[0043] In addition, it should be noted that the shapes and names of the parts and components of the specific embodiments described in this specification may be different, and the above content described in this specification is only an example of the structure of the present invention. All equivalent changes or simple changes made based on the structure, features and principles described in the patent concept of the present invention are included in the protection scope of the patent of the present invention. Technicians in the technical field of the present invention can make various modifications or supplements to the specific embodiments described or replace them in a similar manner, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A method for real-time detection of error information, characterized in that: The steps include: (I) The search enhancement generation step includes the following steps: (1) For the statement that needs to be classified, retrieve the relevant context information on the Internet; (2) further refining the context information to extract the top-ranked pieces of content; (3) The extracted top-ranked pieces of content are fed into the large language model to allow it to summarize the contextual information; (4) Let the large language model output the prediction results of the statement classification and related evidence. This part of the large language model output is called prior knowledge. Then, the statement is combined with the prior knowledge to complete the expansion of the statement. (ii) Information classification steps; The expanded statement in step (a) (4) is input into the large language model for classification to obtain the final detection result.

2. The method for real-time error information detection according to claim 1, characterized in that: In step (1) (1), search on the Internet through the Brave Search API.

3. The method for real-time error information detection according to claim 1, characterized in that: In step (i) (1), the context information is continuously iterated.

4. The method for real-time error information detection according to claim 1, characterized in that: In step (1) (2), extract the top five pieces of content.

5. The method for real-time error information detection according to claim 1, characterized in that: In step (i), the large language model used is the Llama-3-70B-Instruct model.

6. The method for real-time error information detection according to claim 1, characterized in that: Evaluation of the retrieval enhancement generation step: The t-SNE method is used to reduce the dimension of the statement and the retrieved context information and visualize it to examine the relevance of the context information to the statement; the RAGAs method is also used to evaluate the generation performance of the retrieval enhancement generation step.

7. The method for real-time error information detection according to claim 1, characterized in that: In step (2), the large language model used is the Llama-3-8B-Instruct model.

8. The method for real-time error information detection according to claim 1, characterized in that: Fine-tune the large language model from step (2).

9. The method for real-time error information detection according to claim 8, characterized in that: Use the PEFT method to fine-tune the large language model in step (2).

10. The method for real-time error information detection according to claim 9, characterized in that: Use the DoRA method in the PEFT method to fine-tune the large language model in step (2).

Images