False news detection method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a false news detection method and device, electronic equipment and a storage medium, and the false news detection method comprises the following steps: extracting positive evidence and negative evidence related to target news from an internal knowledge base and an external data source of a large language model to obtain a positive and negative evidence set; decomposing an authenticity detection task of the target news into a series of executable sub-problems in combination with the semantic features of the target news, and generating a sub-problem chain; based on the positive and negative evidence set and the sub-problem chain, generating a reasoning step by using a large language model and a thinking chain technology, executing a reasoning process, and outputting a final conclusion; wherein the positive evidence is the evidence supporting that the target news is true, and the negative evidence is the evidence supporting that the target news is false. According to the method, false news can be accurately detected, a detailed detection process and evidence can be provided for a user, and the transparency and credibility of a detection result are improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and particularly to a method, device, electronic device and storage medium for detecting false news. Background Art

[0002] With the rapid development of the Internet, the spread speed and scope of false news have been continuously expanding, causing serious negative impacts on society. Automatic false news detection technology has emerged, aiming to quickly distinguish false news from real news through machine learning models. Currently, most of the existing detection methods are detection models based on deep learning methods (such as GCN) or rely on pre-trained small language models, such as (BERT and RoBERTa). Although these models are relatively accurate in detection results to a certain extent, due to limited parameter quantities, deficiencies in knowledge and ability levels still lead to limited detection effects. In addition, these methods often lack interpretability and are difficult to clearly show the detection process and basis to users. Summary of the Invention

[0003] The present application provides a method, device, electronic device and storage medium for detecting false news to solve the problems in the above background art.

[0004] In a first aspect, the present application provides a method for detecting false news, including:

[0005] Extracting positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model to obtain a set of positive and negative evidence;

[0006] Combining the semantic features of the target news, decomposing the authenticity detection task of the target news into a series of executable sub-problems, and generating a chain of sub-problems;

[0007] Based on the set of positive and negative evidence and the chain of sub-problems, using the large language model and the chain of thought technology to generate reasoning steps and execute the reasoning process, and output a final conclusion;

[0008] Wherein, the positive evidence is evidence supporting the target news as true, and the negative evidence is evidence supporting the target news as false.

[0009] Further, the extracting positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model includes:

[0010] Using the large language model to generate a retrieval query for the target news, and extracting positive evidence and negative evidence related to the target news from the internal knowledge base;

[0011] Within a limited time period, retrieving positive evidence and negative evidence related to the target news from external data sources through a search engine.

[0012] Further, after extracting positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model, the following steps are also included:

[0013] Use natural language inference technology to screen and deduplicate the extracted positive evidence and negative evidence, and score each piece of evidence and calculate the attention weight.

[0014] Further, combining the semantic features of the target news, decomposing the authenticity detection task of the target news into a series of executable sub-problems, including:

[0015] Use the large language model to perform semantic analysis on the target news and extract the key content of the target news, where the key content includes the subject, event, time, and location;

[0016] According to the structure of the target news and the key content, identify the type of the target news, where the types include narrative, commentary, and citation;

[0017] Based on the key content and type of the target news, decompose the authenticity detection task of the target news into a series of executable sub-problems.

[0018] Further, after combining the semantic features of the target news and decomposing the authenticity detection task of the target news into a series of executable sub-problems, the following steps are also included:

[0019] Determine the logical relationship between sub-problems based on the semantic features of the target news, and use the large language model to prioritize the sub-problems.

[0020] Further, based on the positive and negative evidence sets and the sub-problem chain, using the large language model and the thought chain technology to generate reasoning steps, including:

[0021] Design two-way thought chain prompting words adapted to the sub-problem chain based on the positive evidence set and the negative evidence set respectively, and divide the reasoning task into forward reasoning and backward reasoning;

[0022] Generate forward reasoning steps using the large language model based on the forward reasoning prompting words, and generate backward reasoning steps using the large language model based on the backward reasoning prompting words.

[0023] Further, performing the reasoning process and outputting the final conclusion, including:

[0024] Use the large language model to execute the reasoning steps according to the positive and negative evidence sets, and calculate the confidence of each reasoning step;

[0025] Use a large language model to weight positive and negative inferences based on the weight of evidence and the confidence of reasoning steps, and generate a final conclusion.

[0026] In a second aspect, the present application provides a fake news detection device, including:

[0027] An evidence extraction module, configured to extract positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model, and obtain a set of positive and negative evidence;

[0028] A problem decomposition module, configured to combine the semantic features of the target news, decompose the authenticity detection task of the target news into a series of executable sub-problems, and generate a chain of sub-problems;

[0029] An inference execution module, configured to generate inference steps and execute an inference process based on the set of positive and negative evidence and the chain of sub-problems, using the large language model and the thought chain technology, and output a final conclusion;

[0030] Wherein, the positive evidence is evidence supporting the target news as true, and the negative evidence is evidence supporting the target news as false.

[0031] In a third aspect, the present application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the fake news detection method described above is implemented.

[0032] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the fake news detection method described above is implemented.

[0033] The above technical solutions of the present application have the following advantages:

[0034] The fake news detection method provided in the first aspect of the present application extracts positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model to obtain a set of positive and negative evidence, combines the semantic features of the target news, decomposes the authenticity detection task of the target news into a series of executable sub-problems, generates a chain of sub-problems, and based on the set of positive and negative evidence and the chain of sub-problems, uses the large language model and the thought chain technology to generate inference steps and execute the inference process, and outputs a final conclusion, which can accurately detect fake news, and can also provide users with a detailed detection process and evidence, improving the transparency and credibility of the detection results.

[0035] It can be understood that the beneficial effects of the above second aspect, third aspect and fourth aspect can refer to the relevant descriptions in the first aspect above, and will not be repeated here. Description of the Drawings

[0036] To more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0037] Figure 1 It is a flowchart of the false news detection method provided by the present application;

[0038] Figure 2 It is a schematic diagram of the principle of the false news detection method provided by the present application;

[0039] Figure 3 It is a structural diagram of the false news detection device provided by the present application;

[0040] Figure 4 It is a schematic structural diagram of the electronic device provided by the present application. Specific Embodiments

[0041] In the following description, for the purpose of illustration rather than limitation, specific details such as specific system architectures and technologies are proposed to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0042] It should be understood that when used in the specification and appended claims of the present application, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0043] In addition, in the description of the specification and appended claims of the present application, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0044] Reference to "one embodiment" or "some embodiments" etc. described in the specification of this application means that in one or more embodiments of this application, specific features, structures or characteristics described in connection with that embodiment are included. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments" etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "comprise", "include", "have" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in other ways. "A plurality" means "two or more".

[0045] This application aims to propose a false news detection method, which can efficiently and accurately identify false news content on the network and generate high-quality result explanation statements that are easy for users to understand, so as to help users quickly and accurately identify false news from a vast amount of network information, thereby helping users effectively control its harm.

[0046] The following will further describe in detail the specific implementation manners of this application in conjunction with the drawings and embodiments. The following embodiments are used to illustrate this application, but are not used to limit the scope of this application.

[0047] As Figure 1 shown, an embodiment of this application provides a false news detection method, which specifically includes the following steps: extracting positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model to obtain a set of positive and negative evidence; combining the semantic features of the target news, decomposing the authenticity detection task of the target news into a series of executable sub-questions to generate a chain of sub-questions; based on the set of positive and negative evidence and the chain of sub-questions, using the large language model and the thought chain technology to generate reasoning steps and execute the reasoning process, and output a final conclusion; wherein, the positive evidence is evidence supporting the target news to be true, and the negative evidence is evidence supporting the target news to be false.

[0048] The goal of interpretable fake news detection technology is not only to accurately detect fake news, but also to provide users with detailed detection processes and evidence, improving the transparency and credibility of detection results. Large Language Models (LLMs) generally refer to models with tens of billions or hundreds of billions of parameters. These models are pre-trained on large-scale corpora and possess powerful natural language understanding capabilities. LLMs represented by GPT-3.5 have demonstrated impressive capabilities in various tasks. By pre-training on large-scale corpora, LLMs have acquired powerful natural language understanding capabilities. This enables them to deeply understand the semantics, grammar, and context information of news texts, showing excellent natural language understanding and generation capabilities, thereby more accurately identifying the characteristics of fake news and generating high-quality explanatory statements for detection results to help users better understand the detection results.

[0049] The Chain of Thought (CoT) technique of large language models is a method for solving problems through step-by-step reasoning. The core idea of the CoT technique is to decompose complex tasks into multiple sub-problems and gradually solve these sub-problems to finally reach a conclusion. This method can improve the reasoning ability and interpretability of the model. Specifically, the CoT technique includes the following steps: (1) problem decomposition, (2) generation of reasoning steps, (3) execution of reasoning, (4) result verification, and (5) explanation generation. The CoT technique not only improves the reasoning ability of large language models but also enhances the interpretability of the models, enabling them to better adapt to complex tasks such as the fake news detection concerned in this application. Through the CoT technique, large language models can gradually reason and verify news statements, and finally generate accurate reasoning results on the truth or falsehood of news statements and detailed explanatory texts, improving the transparency and credibility of detection.

[0050] In some embodiments, extracting positive and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model includes: using the large language model to generate retrieval queries for the target news and extracting positive and negative evidence related to the target news from the internal knowledge base; within a limited time period, retrieving positive and negative evidence related to the target news from external data sources through a search engine.

[0051] In some embodiments, after extracting positive and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model, it further includes: using natural language reasoning technology to screen and deduplicate the extracted positive and negative evidence, and scoring each piece of evidence and calculating attention weights.

[0052] This application adopts a two-way knowledge extraction method that combines large language models with external search technology, aiming to extract high-quality evidence related to the target news from the internal knowledge base of the large language model and external data sources, providing comprehensive and reliable support for subsequent reasoning. At the same time, during the extraction, the "positive evidence" that supports the news statement as true and the "negative evidence" that supports the news statement as false are extracted separately, so that in the chain of thought reasoning process, independent logical verification and refutation can be carried out for positive and negative evidence respectively, ensuring the integrity of the reasoning chain and the reliability of the conclusion, and avoiding deviation of the reasoning result due to evidence confusion.

[0053] The internal knowledge extraction part relies on the semantic understanding and knowledge storage capabilities of large language models (such as Llama2 or GPT-4) to retrieve positive and negative evidence directly related to the target news content from the input dataset and knowledge base, and ranks and scores them according to their relevance and importance. The external knowledge extraction part uses an external search engine (supporting time period restrictions) to retrieve relevant information on the network in real time to ensure the timeliness and authenticity of the extracted evidence.

[0054] The specific steps are as follows:

[0055] Using the semantic matching ability of the large language model, generate a retrieval query for the target news and extract relevant positive and negative evidence from the model's internal knowledge base. There are:

[0056]

[0057] Among them, c represents the target news, represents the i-th candidate positive / negative evidence, and θ(e i ) represents the retrieval parameter of the large language model. Within the limited time period, retrieve relevant information of the target news through the search engine. Adopt keyword expansion technology to ensure the breadth and depth of the retrieval, and at the same time improve the timeliness of the retrieved content by setting the time range:

[0058]

[0059] Among them, c represents the target news, keywords represents the expanded keywords, and time_range represents the time range.

[0060] Extract the key information that supports or refutes the news from the search results, use natural language inference (Natural Language Inference, NLI) technology for information screening and deduplication, and integrate the evidence into the initial evidence set:

[0061] E + ={e1 + ,e2 + ,…,e n+}, E - = {e1 - , e2 - , …, e n -}

[0062] where E + represents the integrated set of positive evidence, and e i + represents the i-th piece of positive evidence, and E - represents the integrated set of negative evidence, and e i - represents the i-th piece of negative evidence.

[0063] Based on the extracted evidence, generate an explanatory text to support whether the news is true or false. Train a binary classifier to assign true / false scores to each piece of evidence, and the magnitude of the scores reflects the degree of support or refutation of the evidence for the truthfulness of the news:

[0064]

[0065] where v i represents the truthfulness and falsity scores of the i-th piece of evidence, θ (vi) represents the parameters of the scoring model, and MLP(·) represents a multi-layer perceptron. Calculate the attention weights for each piece of evidence to more accurately evaluate its importance in subsequent reasoning:

[0066]

[0067] where α i represents the attention weight of the i-th piece of evidence, represents the parameters of the attention weight calculation model.

[0068] Combine internal knowledge with external search results to generate a complete evidence set containing positive and negative evidence. This evidence set not only covers the semantic information of the news content but also integrates the results of external fact verification, providing multi-dimensional support for subsequent reasoning:

[0069] E = {{E + , E -}, v i , α i}

[0070] Through the above steps, it is possible to comprehensively and accurately obtain the evidence related to the target news and provide basic support for reasoning. By combining the internal knowledge of the large language model with external real-time information, while ensuring the comprehensiveness of the information, the credibility and accuracy of false news detection are significantly improved.

[0071] In some embodiments, by combining the semantic features of the target news, the task of detecting the authenticity of the target news is decomposed into a series of executable sub-problems, including: using a large language model to perform semantic analysis on the target news to extract the key content of the target news, where the key content includes the subject, event, time, and location; identifying the type of the target news according to the structure of the target news and the key content, where the type includes narrative type, commentary type, and citation type; and decomposing the task of detecting the authenticity of the target news into a series of executable sub-problems based on the key content and the type of the target news.

[0072] In some embodiments, after decomposing the task of detecting the authenticity of the target news into a series of executable sub-problems by combining the semantic features of the target news, it further includes: determining the logical relationship between sub-problems based on the semantic features of the target news, and using a large language model to perform priority ranking on the sub-problems.

[0073] This application can decompose the complex task of detecting false news into a series of executable sub-problems, enabling the large language model to gradually complete complex tasks in a chain reasoning manner, improving the accuracy and interpretability of reasoning. By combining the semantic features of news statements with task requirements, decomposition steps are dynamically generated, and prompt words adapted to the large language model are designed to ensure the rationality of sub-problem generation and the coherence of the reasoning chain.

[0074] The specific steps are as follows:

[0075] Use a large language model to perform semantic analysis on the target news statement to extract the key content of the news statement, including the subject, event, time, and location. Subject: Identify the main event participants in the news, such as people, institutions, organizations, etc.; Event: Determine the main event or action described in the news statement; Time: Extract the time information when the event occurred, especially pay attention to the time range or key time nodes involved in the news statement; Location: Identify the geographical location information in the news to ensure that the spatio-temporal background information of the news is fully considered. The extraction of these features lays the foundation for subsequent decomposition steps, enabling each sub-problem to be reasoned based on different dimensions.

[0076] Identify the type of news according to the structure and semantic features of the statement content, including narrative type, commentary type, and citation type. Narrative type: This type of news mainly states facts, is relatively objective, and has no obvious subjective opinions or speculations; Commentary type: This type of news contains the personal views, analyses, or comments of the author or publisher; Citation type: Contains citations from other sources or people, and may involve indirect statements or paraphrases of facts. By identifying the type of news, it is possible to dynamically select a suitable sub-problem processing method to ensure that the direction of subsequent reasoning conforms to the characteristics of the news statement.

[0077] Decompose the goal of news authenticity detection into multiple sub-goals, which can be divided into fact verification, semantic analysis, and background retrieval. Fact verification (Verify): Check whether the facts stated in the news are consistent with the known evidence information. This includes verifying the authenticity, time, location, and other characteristics of the event. The answer format for this type of question is specified as a binary yes or no. Semantic analysis (Semanteme): Analyze whether there are potential ambiguities, contradictions, or ambiguities in the semantics of the news statement. For example, whether the statement is too subjective, whether there is language misleading or unclear expression. The answer format for this type of question is specified as the specific semantic problem that exists. Background retrieval (Background): Verify whether the news statement is consistent with historical data, background knowledge, or external facts. Especially for statements in some specific fields, external knowledge is needed for verification. The answer format for this type of question is specified as relevant background information evidence.

[0078] Based on the statement content, clarify the logical relationship between sub-questions to ensure the internal consistency of the decomposition result. According to the characteristics of the news statement, some questions may need to be solved first (such as verifying key facts), while other questions can be carried out in subsequent steps. Therefore, it is necessary for the large model to sort the priorities of the questions. After meeting the above step requirements, the large model can ensure the internal consistency and logic of the subsequent question chain, ensuring that each sub-question is clear and definite from the initial task goal to the final reasoning result, and can naturally lead to the next question.

[0079] In some embodiments, generating reasoning steps using the large language model and the chain-of-thought technology based on the positive and negative evidence sets and the sub-question chain includes: respectively designing bidirectional chain-of-thought prompt words adapted to the sub-question chain based on the positive evidence set and the negative evidence set, and dividing the reasoning task into forward reasoning and reverse reasoning; generating forward reasoning steps using the large language model based on the forward reasoning prompt words, and generating reverse reasoning steps using the large language model based on the reverse reasoning prompt words.

[0080] This application can dynamically generate complete reasoning steps based on the generated sub-question chain and in combination with the positive and negative evidence sets extracted from knowledge, construct a bidirectional reasoning chain through the separation and integration of positive and negative evidence, and achieve the logical balance of forward support and reverse verification. By designing precise prompt words, utilizing the semantic generation ability of the large language model, and combining logical verification and optimization mechanisms, ensure that each sub-question is answered coherently and systematically, and finally form a complete and reliable reasoning chain, providing a clear operation path for the subsequent reasoning program.

[0081] The specific steps are as follows:

[0082] Receive the generated sub-question chain Q = {q1, q2, …, q n} and the positive and negative evidence sets E + and E - . Determine the solution goals and context semantic scopes of each sub-question, providing an input framework for subsequent inference steps.

[0083] A complete Prompt containing unidirectional CoT usually consists of the following three parts. Instruction: Used to describe the problem and inform the output format of the large model. Rationale: Refers to the intermediate reasoning process of CoT, which can include the solution to the problem, intermediate reasoning steps, and any external knowledge related to the problem. Exemplars: Provide the basic format of input-output pairs for the large model in a few-shot manner, and each exemplar includes the problem, reasoning process, and answer.

[0084] According to the above principles, design bidirectional CoT prompts adapted to the sub-question chain based on the positive evidence set and negative evidence set respectively, and clarify the reasoning goals and step logics in both directions. System Prompt: Set the task role and goal of the large language model, for example: "You are a logical reasoning assistant, and you need to prove that the news statement is (true / false), and you are responsible for generating complete reasoning steps based on the decomposed sub-question chain." This corresponds to meeting the instruction part of the design principle. Positive and Negative Evidence Prompt: Combine the sub-question chain to provide background information of the previous task, the corresponding positive and negative evidence sets, and their corresponding scores and weights. This corresponds to meeting the rationale part of the design principle. Dynamically Generated Prompt: According to the characteristics of each sub-question, combine the artificially designed example reasoning paths in the data to supplement and generate personalized reasoning guidance. This corresponds to meeting the exemplar part of the design principle.

[0085] This application divides the reasoning task into two major directions: forward reasoning and backward reasoning. The forward reasoning chain is constructed based on the positive evidence set, and the backward reasoning chain is constructed based on the negative evidence set. For each sub-question q i in both directions, extract relevant evidence from the positive evidence set E + and the negative evidence set E - respectively, and generate corresponding reasoning steps T + and T - :

[0086]

[0087]

[0088] In forward reasoning, the constructed forward reasoning prompt words are input into the large model to generate forward reasoning steps. During this process, the construction of the reasoning chain follows the principle of the chain of thought, and the reasoning path is dynamically adjusted according to the relevance and importance of the evidence during the reasoning process to ensure logical coherence. Similarly, in backward reasoning, by inputting backward prompt words into the large model, backward reasoning steps are generated, and the model can raise doubts and counter-evidence through refuting logic to help reveal potential fallacies or inconsistencies in the news statement.

[0089] In some embodiments, performing the reasoning process and outputting the final conclusion includes: using the large language model to perform reasoning steps based on the positive and negative evidence sets, and calculating the confidence of each reasoning step; using the large language model to weight the forward and backward reasoning based on the weight of the evidence and the confidence of the reasoning steps to generate the final conclusion.

[0090] This application actually performs the reasoning process according to the obtained reasoning chain, combines the positive and negative evidence sets, synthesizes the two-way reasoning chain and outputs the final reasoning result.

[0091] The specific steps are as follows:

[0092] Based on the extracted evidence set, the large model executes the complete reasoning chain generated by the two-way reasoning steps. During the reasoning process, the model calculates the confidence of each reasoning step according to the relevance and importance of the supporting evidence used in each reasoning link. During the synthesis of the reasoning chain, the model weights the forward and backward reasoning according to the weight of the evidence and the confidence of the reasoning steps:

[0093] T = Combine(T + , T - ).

[0094] High-weight evidence in the forward reasoning chain will strengthen the supporting direction in the reasoning process, while high-weight evidence in the backward reasoning chain will strengthen the refuting force in the reasoning process. The model will also check the reliability of the evidence through cross-validation. If the positive evidence supports a certain conclusion while the negative evidence refutes this conclusion, the model will further examine the contradiction between the evidences and analyze which evidence is more authoritative or credible. Finally, the model will generate a comprehensive conclusion based on the weighted reasoning chain.

[0095] Finally, the final result of the inference is output, which includes the following. 1. Detection result of the question: For each sub-question, the model will output the inference conclusion, including the inference process, evidence support, and the final judgment result. 2. Explanation of the interpretability of the inference chain: The model will generate a detailed explanation of the inference chain based on the executed inference steps, clearly elaborating how each inference step depends on the evidence for derivation. 3. Evidence source and weight: For each conclusion, the model will list the relevant evidence used in the inference steps and explain the role of this evidence in the inference process, especially which positive and negative evidences have influenced the final judgment of the model.

[0096] The following is illustrated by specific examples.

[0097] Example

[0098] In this example, taking the detection of the news statement "James Cameron, the director of 'Avatar 3' expected to be released by the end of this year, is from the same country as the director of 'Interstellar'" as an example, the principle of the detection method is as Figure 2 shown and includes the following steps:

[0099] Step 1: Knowledge extraction. In this stage, first, evidence related to the target news statement is obtained to ensure that the inference process has sufficient support. For example:

[0100] 1. Internal dataset extraction: Use a large language model to analyze the news statement and extract evidence from its internal knowledge base. According to the content of the news statement, extract relevant information such as the background information of James Cameron and the director of 'Interstellar', and their nationalities.

[0101] Positive evidence 1: James Cameron is from Canada;

[0102] Positive evidence 2: The director of 'Avatar 3' is James Cameron;

[0103] Positive evidence 3: 'Avatar 3' is expected to be released by the end of this year;

[0104] Negative evidence 1: There is no direct indication that the director of 'Interstellar' is from Canada;

[0105] 2. Introduction of an external search engine

[0106] Further verify the accuracy of the news content through a search engine, especially verify the nationality information of Christopher Nolan, the director of 'Interstellar'. Confirm through querying external data sources that Christopher Nolan is indeed British.

[0107] Negative evidence 2: The director of 'Interstellar' is Christopher Nolan;

[0108] Negative Evidence 3: Christopher Nolan is British.

[0109] 3. Integration and Duplicate Removal

[0110] When extracting external data, NLI (Natural Language Inference) technology is used to screen and remove duplicates from information to ensure that the collected evidence is of high quality and non-duplicate.

[0111] 4. Evidence Evaluation

[0112] For Positive Evidence 1: James Cameron is indeed Canadian, according to multiple reliable resources (such as movie databases, director profiles, etc.). Calculate its authenticity score as 0.95 and attention weight as 0.73, indicating strong support for the claim and high relevance to the claim.

[0113] Evaluate Positive Evidence 2 and 3 and Negative Evidence 1, 2, and 3 respectively according to the same steps.

[0114] 5. Evidence Set Generation

[0115] Finally, integrate all positive and negative evidence into evidence set E, where:

[0116] E + : Includes Positive Evidence 1, 2, 3 and related scores.

[0117] E - : Includes Negative Evidence 1, 2, 3 and related scores.

[0118] Step 2: Problem Decomposition

[0119] Next, decompose the news statement authenticity detection task into a series of executable sub-problems and perform reasoning in combination with the previously extracted evidence.

[0120] 1. News Statement Feature Analysis

[0121] Subject: Avatar, Interstellar, James Cameron.

[0122] Event: Analyze the events involved in the news: the director of Avatar, the director of Interstellar, the nationalities of James Cameron and the director of Interstellar, the release time of Avatar 3.

[0123] Time: The time when the events mentioned in the news statement occurred "expected by the end of this year".

[0124] Location: The news mentions the nationalities of the two directors, mainly focusing on the feature of "Canadian".

[0125] 2. Content Type Identification

[0126] This statement is factual news, aiming to clarify the release time of "Avatar 3" and the nationalities of the two directors. Therefore, the task mainly focuses on verifying the accuracy of the two factual time information.

[0127] 3. Sub-question generation

[0128] Fact verification (Verify): Verify whether James Cameron is Canadian, whether Christopher Nolan is Canadian, and the expected release time of "Avatar 3";

[0129] Semantic analysis (Semanteme): Check whether the statement has subjective or inaccurate language, especially whether the expression "both being Canadians" is misleading or ambiguous.

[0130] Background retrieval (Background): Verify the background information of the two directors and the identity of the director of "Interstellar".

[0131] 4. Logical relationship analysis

[0132] The logical relationships between sub-questions are as follows:

[0133] Fact verification: First, verify the director and release time of "Avatar 3", then verify whether James Cameron is Canadian (positive evidence), and then verify whether Christopher Nolan is Canadian (negative evidence).

[0134] Semantic analysis: The statement of this news is misleading because it does not accurately mention the nationality of Christopher Nolan.

[0135] Background retrieval: Background retrieval confirms that the nationality information of the two directors is inconsistent, providing evidence to support the negation of the news authenticity.

[0136] Generate the sub-question chain as Q = {q1, q2, q3, q4}, where q1 is fact verification: Whether the release time of "Avatar 3" is expected to be the end of this year and whether the director is James Cameron; q2 is fact verification: Whether James Cameron is Canadian; q3 is background retrieval: Who is the director of "Interstellar"; q4 is fact verification: Whether the director of "Interstellar" is Canadian.

[0137] Step 3: Generate two-way reasoning steps. After inputting the sub-question chain and positive and negative data sets, according to the prompts in Table 1 and Table 2, make the large model generate reasoning steps.

[0138]

[0139]

[0140]

[0141] Step 4: Inference detection execution. According to the inference steps in Step 3, make the large model execute the generated inference steps, and the output result is shown in Table 3; comprehensively make a final judgment to obtain the final conclusion: There are some inaccuracies in this news statement, especially the description about the nationality of the director of "Interstellar" is misleading. Negative evidence 3 indicates that his nationality should be the UK. Therefore, this news statement is a false news as a whole.

[0142]

[0143]

[0144] The false news detection method provided by the embodiment of the present application extracts positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model to obtain a set of positive and negative evidence, combines the semantic features of the target news, decomposes the authenticity detection task of the target news into a series of executable sub-problems, generates a sub-problem chain, and based on the set of positive and negative evidence and the sub-problem chain, uses the large language model and the chain of thought technology to generate inference steps and execute the inference process, and outputs the final conclusion, which can accurately detect false news, and can also provide users with a detailed detection process and evidence, improving the transparency and credibility of the detection results.

[0145] Corresponding to the false news detection method described in the above embodiment, as Figure 3 shown, the embodiment of the present application also provides a false news detection device, and this false news detection device includes:

[0146] An evidence extraction module, configured to extract positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model to obtain a set of positive and negative evidence;

[0147] A problem decomposition module, configured to combine the semantic features of the target news, decompose the authenticity detection task of the target news into a series of executable sub-problems, and generate a sub-problem chain;

[0148] An inference execution module, configured to generate inference steps and execute the inference process based on the set of positive and negative evidence and the sub-problem chain, using the large language model and the chain of thought technology, and output the final conclusion;

[0149] Wherein, the positive evidence is the evidence supporting the target news as true, and the negative evidence is the evidence supporting the target news as false.

[0150] It should be noted that the information interaction, execution process, etc. between the above modules / units, due to being based on the same concept as the method embodiment of the present application, for their specific functions and the technical effects brought, please refer to the method embodiment part specifically, and will not be elaborated here.

[0151] Those skilled in the art can clearly understand that, for the convenience and conciseness of description, only the above-mentioned division of each functional unit and module is used as an example. In practical applications, the above functions can be allocated to different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction and do not limit the protection scope of this application. The specific working processes of the units and modules in the above system can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated here.

[0152] The embodiment of this application also provides an electronic device, such as Figure 4 shown, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the false news detection method provided in the first aspect.

[0153] In applications, the electronic device may include, but is not limited to, a processor and a memory. Figure 4 This is only an example of an electronic device and does not limit the electronic device. It may include more or fewer components than shown in the figure, or combine some components, or different components. For example, input / output devices, network access devices, etc. The input / output devices may include a camera, an audio acquisition / playback device, a display screen, etc. The network access device may include a network module for performing wireless network with external devices.

[0154] In applications, the processor may be a central processing unit (CPU), and this processor may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or this processor may also be any conventional processor, etc.

[0155] In an application, in some embodiments, the memory may be an internal storage unit of an electronic device, such as a hard disk or memory of the electronic device. In other embodiments, the memory may also be an external storage device of the electronic device. For example, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc., equipped on the electronic device. The memory may also include both an internal storage unit and an external storage device of the electronic device. The memory is used to store an operating system, application programs, a BootLoader, data, and other programs, such as program codes of computer programs. The memory may also be used to temporarily store data that has been output or will be output.

[0156] An embodiment of the present application also provides a computer-readable storage medium storing a computer program, which when executed by a processor can implement the steps in the above-mentioned method embodiments.

[0157] To implement all or part of the processes in the above-mentioned method embodiments of the present application, it can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program codes, and the computer program codes can be in the form of source code, object code, executable files, or some intermediate forms, etc. The computer-readable medium can at least include: any entity or device capable of carrying the computer program code to an electronic device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk, or an optical disc, etc.

[0158] Those of ordinary skill in the art can realize that the devices and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0159] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. Another point is that the couplings or direct couplings or communication connections shown or discussed among each other can be through some interfaces. The devices are indirectly coupled or communication-connected, and can be in electrical, mechanical or other forms.

[0160] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. And these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A false news detection method, characterized in that, Including: Extracting positive and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model to obtain a set of positive and negative evidence; Combining the semantic features of the target news, decomposing the authenticity detection task of the target news into a series of executable sub-questions, and generating a chain of sub-questions; Based on the set of positive and negative evidence and the chain of sub-questions, using the large language model and the chain of thought technology to generate reasoning steps and execute the reasoning process, and output a final conclusion; Among them, the positive evidence is evidence that supports the target news as true, and the negative evidence is evidence that supports the target news as false.

2. The false news detection method according to claim 1, characterized in that, The extracting positive and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model includes: Using the large language model to generate a retrieval query for the target news, and extracting positive and negative evidence related to the target news from the internal knowledge base; Within a limited time period, retrieving positive and negative evidence related to the target news from external data sources through a search engine.

3. The false news detection method according to claim 1, wherein After extracting positive and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model, it further includes: Using natural language inference technology to screen and deduplicate the extracted positive and negative evidence, and scoring each piece of evidence and calculating the attention weight.

4. The false news detection method according to claim 1, wherein The combining the semantic features of the target news and decomposing the authenticity detection task of the target news into a series of executable sub-questions includes: Using the large language model to perform semantic analysis on the target news, and extracting the key content of the target news, where the key content includes the subject, event, time, and location; According to the structure of the target news and the key content, identifying the type of the target news, where the type includes narrative, commentary, and citation; Based on the key content and type of the target news, decomposing the authenticity detection task of the target news into a series of executable sub-questions.

5. The fake news detection method according to claim 1, characterized in that, After combining the semantic features of the target news and decomposing the authenticity detection task of the target news into a series of executable sub-questions, it further includes: Determining the logical relationship between sub-questions based on the semantic features of the target news, and using the large language model to prioritize the sub-questions.

6. The fake news detection method according to claim 1, characterized in that, The generating reasoning steps based on the set of positive and negative evidence and the chain of sub-questions, using the large language model and the chain of thought technology includes: Respectively based on the set of positive evidence and the set of negative evidence, designing bidirectional chain of thought prompt words adapted to the chain of sub-questions, and dividing the reasoning task into forward reasoning and backward reasoning; Generating forward reasoning steps using the large language model based on the forward reasoning prompt words, and generating backward reasoning steps using the large language model based on the backward reasoning prompt words.

7. The fake news detection method according to claim 6, wherein The executing the reasoning process and outputting the final conclusion includes: Using the large language model to execute the reasoning steps according to the set of positive and negative evidence, and calculating the confidence of each reasoning step; Using the large language model to weight the positive and negative reasoning according to the weight of the evidence and the confidence of the reasoning steps, and generating a final conclusion.

8. A false news detection device, characterized in that, Including: An evidence extraction module, configured to extract positive evidence and negative evidence related to the target news from the internal knowledge base and external data sources of the large language model, so as to obtain a set of positive and negative evidence; A problem decomposition module, configured to combine the semantic features of the target news, decompose the authenticity detection task of the target news into a series of executable sub-problems, and generate a chain of sub-problems; An inference execution module, configured to generate inference steps and execute the inference process based on the set of positive and negative evidence and the chain of sub-problems, using the large language model and the chain of thought technique, and output a final conclusion; Wherein, the positive evidence is evidence supporting that the target news is true, and the negative evidence is evidence supporting that the target news is false.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the fake news detection method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the fake news detection method according to any one of claims 1 to 7.

Images