Video authenticity detection method and system based on large model and retrieval enhancement generation technology

Abstract

The invention belongs to the technical field of computer vision and natural language processing, and discloses a video authenticity detection method and system based on a large model and a retrieval enhancement generation technology. Jointly constructing a to-be-detected text by combining the video title, the brief introduction, the text obtained by audio conversion and the video description text; completing the construction of an external knowledge base; performing fact verification based on a retrieval enhancement generation technology (RAG); performing common sense reasoning based on a large model: constructing cue words, performing common sense and logic reasoning on the to-be-detected text by utilizing the large model ChatGLM, detecting whether common sense mistakes and logic conflicts exist or not, and judging whether news is true or false according to the common sense mistakes and the logic conflicts; and comprehensively judging the authenticity of the to-be-detected video according to the fact check and common sense reasoning. Through the combination of the large model and the RAG technology, the accuracy and interpretability of video content authenticity detection are improved; the detection efficiency is improved; false news videos can be found and processed in time, and the health of a network environment and public trust can be maintained.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and natural language processing, and in particular to a video authenticity detection method and system based on large model and retrieval enhancement generation technology. Background Art

[0002] With the rise of social media and short video platforms, fake news and misleading information are spreading faster and faster, especially fake news about public figures and important events, which has a great negative impact on social order and public trust. Existing video detection technologies often rely on manual review or simple text, image, and multimodal detection technologies. The former is too inefficient, and the latter can only judge the authenticity of the video as a whole, but cannot accurately identify the false content in the video and provide explainable explanations. The fact-checking methods based on external knowledge bases that can accurately identify and locate false content currently rely on specific knowledge graphs. However, knowledge is stored in the knowledge graph in the form of "entity-relationship-entity" triples. When applied to video detection, knowledge stored in this form will lack most of the details, mainly reflected in the excessive emphasis on entity information and the neglect of relationship information, and the easy misjudgment of information for polysemous entities. In addition, the existing knowledge bases are often not comprehensive enough, and the timeliness of knowledge is difficult to guarantee.

[0003] In view of this limitation of the knowledge graph, this application creatively introduces large models and retrieval-augmented generation (RAG) technology. The parameterized knowledge of the large model is more complete and comprehensive than the structured knowledge in the knowledge graph. It can sort out comprehensive and diverse analysis from massive data and realize efficient and explainable video detection, providing new ideas for video detection. Considering the inevitable hallucination problem of large models and the lag of model knowledge, RAG enhances the performance of generation tasks by retrieving external knowledge, and its documented and easy-to-update external knowledge base can complement the large model well.

[0004] Therefore, it is necessary to provide a video detection method based on large models and retrieval enhanced generation technology to improve the detection efficiency, accuracy and interpretability of fake news videos. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a video authenticity detection method and system based on large model and retrieval enhancement generation technology.

[0006] The technical solution of the present application is as follows: A video authenticity detection method based on a large model and retrieval enhancement generation technology, comprising the following steps:

[0007] Step 1: Construction of the text to be detected: Based on the multimodal large model LLaVA, construct prompt words to describe the video to be detected and obtain the video description text; extract the subtitles in the video to be detected through OCR text recognition, and combine the video title, introduction, audio conversion text and video description text to construct the text to be detected;

[0008] Step 2: External knowledge base construction: Collect a large amount of historical news from official news sources, pre-process the collected news text, convert it into news vector representation and create an index for it to complete the construction of the external knowledge base;

[0009] Step 3: Fact-checking based on retrieval-augmented generation technology RAG: The text to be checked is converted into a vector representation, and similarity matching is performed with the news vector stored in the external knowledge base to obtain the basis for fact-checking; the basis is combined with the text to be checked to construct prompt words, and the large model ChatGLM is used to perform fact-checking and judge the authenticity of the news;

[0010] Step 4: Common sense reasoning based on the big model: Construct prompt words, use the big model ChatGLM to reason about the text to be tested from common sense and logic, detect whether there are common sense fallacies and logical conflicts, and judge the authenticity of the news based on this;

[0011] Step 5: News authenticity judgment: Based on the above fact verification and common sense reasoning, comprehensively judge the authenticity of the video to be tested.

[0012] The step 2 is specifically performed as follows:

[0013] Step 2.1: Data collection: Collect a large amount of historical news from official news sources to ensure the authenticity and reliability of the information;

[0014] Step 2.2: Data preprocessing: Clean the collected news text, remove irrelevant characters and punctuation marks, and then split it into several text blocks to fit the contextual constraints of the large language model;

[0015] Step 2.3: Vector representation: Use the embedding model BERT to convert the preprocessed text block into a vector representation;

[0016] Step 2.4: Index creation: Create an index to store these text chunks and their vector embeddings as key-value pairs.

[0017] The step 3 is specifically performed as follows:

[0018] Step 3.1: Vector representation of the text to be detected: Use the embedding model BERT to vectorize the text to be detected and convert it into a vector representation;

[0019] Step 3.2: Retrieve fact-checking evidence: Use the BM25 algorithm to perform similarity matching between the text vector to be tested and the historical news vector stored in the external knowledge base. By calculating the similarity between the two, select the top K items with the highest similarity as the basis for fact-checking;

[0020] Step 3.3: Construct the big model prompt words: Combine the retrieved evidence with the text to be tested, prompt the big model to consider the causal logical relationship between the two, whether there is a conflict, whether there is direct evidence to judge whether the text to be tested is true or false, and provide an example to the big model;

[0021] Step 3.4: Send the constructed prompt words into the large model for reasoning, judge the authenticity of the news, and obtain the results of fact verification based on RAG technology.

[0022] The step 4 is specifically performed as follows:

[0023] Step 4.1: Assign roles to the big model: Set roles for the big model, limit the context, narrow the scope of the big model's answers to obtain more accurate answers, and guide the big model to interact in a more professional way;

[0024] Step 4.2: Construct common sense reasoning prompt words: Based on the role setting of the big model, prompt the big model to identify the authenticity of the description in the news from the perspective of historical common sense, and at the same time remind the big model to pay attention to logical errors in time, place, and characters;

[0025] Step 4.3: Construct a thinking chain: Divide the prompt words into four parts: question, news, thinking chain, and answer. The question part is a combination of the role assigned to the big model in step 4.1 and the common sense reasoning prompt words constructed in step 4.2. The news part is the text to be tested constructed in step 1. The thinking chain part is the process of common sense reasoning. The answer part is the judgment of whether the news is true or false. According to this structure, provide an example to the big model.

[0026] Step 4.4: Send the constructed prompt words into the big model for reasoning, judge the authenticity of the news, and obtain the results of common sense reasoning based on the big model.

[0027] A video authenticity detection system based on large model and retrieval enhancement generation technology, comprising:

[0028] The module for constructing text to be detected is used to uniformly convert the video, audio, and text information in the video to be detected into textual information for subsequent processing;

[0029] External knowledge base construction module, used to pre-process a large amount of historical news and convert it into an external knowledge base that can be used by RAG;

[0030] The historical news detection module is used to fact-check the text to be detected and determine its authenticity by using RAG technology and external knowledge base;

[0031] The common sense reasoning module is used to perform common sense reasoning on the text to be tested through a large model to determine its authenticity.

[0032] The beneficial effects of the present invention are that, through the combination of large models and RAG technology, the accuracy and explainability of video content authenticity detection are improved; the automated processing flow reduces the workload of manual review and improves detection efficiency; and fake news videos are discovered and processed in a timely manner, which helps maintain the health of the network environment and public trust. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is an overall flow chart of a video detection method based on a large model and RAG technology in one embodiment;

[0034] Figure 2 Constructing a flow chart for an external knowledge base in one embodiment;

[0035] Figure 3 A flowchart of fact checking based on RAG technology in one embodiment;

[0036] Figure 4 A schematic diagram of the architecture of a video detection system based on a large model and RAG technology in an embodiment; DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0038] In one embodiment, Figure 1 As shown, a video authenticity detection method based on a large model and retrieval enhancement generation technology is provided, comprising the following steps:

[0039] Step 1: Construction of the text to be detected: According to CN118485941A, the large model LLaVA is used to construct prompt words to describe the video to be detected, and the video description text is obtained. At the same time, the subtitles in the video are extracted by OCR text recognition, and the text obtained by combining the video title, introduction, audio conversion and video description text is jointly constructed to construct the text to be detected;

[0040] Step 2: External knowledge base construction: Collect a large amount of historical news from official news sources, pre-process the collected news text, convert it into vector representation and create an index for it to complete the construction of the external knowledge base;

[0041] Step 3: Fact-checking based on retrieval-augmented generation technology RAG: The text to be checked is converted into a vector representation, and similarity matching is performed with the news vector stored in the external knowledge base to obtain the basis for fact-checking. The basis is combined with the text to be checked to construct prompt words, and the large model ChatGLM is used to perform fact-checking and judge the authenticity of the news;

[0042] Step 4: Common sense reasoning based on the big model: Construct prompt words and use the big model ChatGLM to reason about the test text from common sense, logic, etc., to detect whether there are common sense fallacies, logical conflicts, etc., and judge the authenticity of the news based on this;

[0043] Step 5: News authenticity judgment: Based on the above fact verification and common sense reasoning, comprehensively judge the authenticity of the video to be tested.

[0044] In step 1, since both the external knowledge base and the large model input are in text form, in order to use them for reasoning, the information of each modality needs to be converted into text form. For the video to be tested, the large model LLaVA is first used to describe the video, and the video description text that can describe the background knowledge such as celebrity information, national flags, party emblems, company organization logos, location buildings, etc. is obtained.

[0045] Next, the Whisper model is used to process the input video, parse the audio in the video, and then transcribe the audio into text. Next, the transcribed text is subjected to preliminary grammatical and semantic analysis to remove obvious irrelevant information, such as music, background sound, etc., to ensure that the extracted text content focuses on the voice of the person speaking. For audio without specific semantics, its text content is set to empty. The video audio conversion text is obtained.

[0046] Finally, OCR text recognition is used to extract subtitles from the video, and combined with the video title and introduction to obtain the video supplementary information text.

[0047] After the above steps, the obtained texts are integrated and the text to be detected is constructed.

[0048] In step 2, the construction of external knowledge base is a crucial initial step in the data preparation process. Figure 2 As shown in Figure 1, the construction of the external knowledge base mainly includes the following steps:

[0049] Step 2.1: Data collection: Collect a large amount of historical news from official news sources to ensure the authenticity and reliability of the information;

[0050] Step 2.2: Data preprocessing: The collected news text is cleaned to remove irrelevant characters, punctuation marks, etc., and then split into smaller and more manageable chunks to fit the contextual constraints of the large model;

[0051] Step 2.3: Vector representation: Use the embedding model BERT to convert the preprocessed text blocks into vector representations, making storage and subsequent similarity calculations easier;

[0052] Step 2.4: Index creation: Create an index to store these text blocks and their vector embeddings as key-value pairs to enable efficient search capabilities during the detection process.

[0053] In step 3, if Figure 3 As shown in Figure 2, historical news detection based on RAG mainly includes the following steps:

[0054] Step 3.1: Vector representation of the text to be detected: Use the embedding model BERT to vectorize the text to be detected and convert it into a vector representation;

[0055] Step 3.2: Retrieve fact-checking evidence: Use the BM25 algorithm to perform similarity matching between the text vector to be tested and the historical news vector stored in the external knowledge base. By calculating the similarity between the two, select the top K items with the highest similarity as the basis for fact-checking;

[0056] BM25 is an algorithm for information retrieval that can be used to measure the relevance of a document. The core of the BM25 algorithm is a scoring function that calculates a score of the degree of match between a document and a query condition. The scoring function calculates the score based on three factors: the weight of the query term frequency, the document length, and the weight of the document frequency. The weight of the query term frequency is calculated based on the frequency of the query term in the document. The weight represents the contribution of the query term to the relevance of the document and decreases as the number of times the query term appears in the document. The document length refers to the number of times the query term appears in the document. The document length affects the weight of the query term. The larger the document length, the smaller the weight of the query term; the smaller the document length, the larger the weight of the query term. The weight of the document frequency represents the prevalence of the query term in the document collection. If the query term appears in many documents, its weight will be reduced. The BM25 algorithm formula is shown below:

[0057]

[0058] Among them, Q represents the query statement, D represents the document, and q i represents the i-th word in the query, n is the number of words in the query, and IDF(q i ) is the inverse document frequency, f(q i ,D) is the number of occurrences of the i-th word in the document, k1 and b are adjustment parameters, |D| is the length of the document, and avgdl is the average length of the documents in the document collection.

[0059] Step 3.3: Construct the big model prompt words: Combine the retrieved evidence with the text to be tested, prompt the big model to consider the causal logical relationship between the two, such as whether there is a conflict, whether there is direct evidence to determine whether the text to be tested is true or false, etc., and provide an example to the big model;

[0060] Here is an example of a prompt word:

[0061] Information: [Full list of 3.15 party exposure] 1. The "mainboard" that creates the water army 2. Fireproof glass that is not fireproof 3. The fire extinguisher that cuts corners 4. The "bad meat" in the pork with preserved vegetables 5. The magical Tinghua wine 6. The marriage and dating platform has deep routines 7. The disturbing BMW transmission shaft 8. The gift card tricks in online lending

[0062] Question: You are a fact-checking expert. Please help me answer a question based on the given information to determine whether the following news contains false information. Please note that you need to pay special attention to information about time, place, people, events, etc., analyze the conflicts or logical errors in time, place, people, events, and point out the specific errors and reasons.

[0063] News: 315 Gala named starch sausage, which contained chicken bone paste. The merchant said: It is not suitable for human consumption

[0064] Thinking chain: Because the complete list of 3.15 Gala exposure given in the data does not include starch sausage, the description of "315 Gala named starch sausage" in the news is wrong. Therefore, the news contains false information.

[0065] Answer: The news contains false information.

[0066] Data: {obtained from similarity matching in step 3.2}

[0067] Question: You are a fact-checking expert. Please help me answer a question based on the given information to determine whether the following news contains false information. Please note that you need to pay special attention to information about time, place, people, events, etc., analyze the conflicts or logical errors in time, place, people, events, and point out the specific errors and reasons.

[0068] News: {constructed from the text to be detected in step 1}

[0069] Thinking chain:

[0070] Answer:

[0071] Here is an example of news detection results:

[0072] Here is an example of news detection results:

[0073] Video introduction: Xi'an has a long history, a long-standing culture, and a wide variety of food. Its mutton soup with steamed buns, herbal tea, roujiamo, cold noodles, steamed beef with rice flour, and sour soup dumplings are well-known throughout the country.

[0074] Video description: In the video, various kinds of food appear, such as mutton soup with steamed buns, roujiamo, cold noodles, steamed beef with rice flour, etc.

[0075] Video subtitles: Mutton soup steamed buns, steamed meat buns, cold flour steamed beef dumplings

[0076] Audio content:

[0077] Historical news shows that "due to history, Xi'an's food culture is also profound and extensive. As a representative of Northwest China's food, Xi'an snacks are the most famous. Because the Hui people live in a relatively concentrated area, halal food accounts for a certain proportion in Xi'an, such as mutton soup steamed buns, cured beef and mutton, and the famous Jia San soup dumplings. Local Han snacks include roujiamo and cold skin. And Biángbiáng noodles are particularly eye-catching with their unique name. Famous foods: beef and mutton soup steamed buns, roujiamo, dumpling banquet, gourd head soup steamed buns, sour soup dumplings, cold skin, meatballs and spicy soup, oil tea twist, zongzi, oil-splashed noodles, mutton skewers, etc. Famous drinks: thick wine Xifeng wine, pomegranate juice, etc." This video does not have obvious human voice audio, so the audio content is empty, and the subtitles are just a simple list of food names. From historical news, we know that Xi'an has rich and diverse cuisine. Mutton soup steamed bun, roujiamo, steamed beef with rice flour, and sour soup dumplings are indeed part of its food culture, but herbal tea and cold noodles are not Xi'an delicacies. Herbal tea is a specialty drink in Guangdong, which has the effect of clearing away heat and removing fire, while cold noodles are mainly a specialty food in the Northeast region. Therefore, the video contains false information. The big model uses its rich parametric knowledge and combines historical materials to successfully determine that the video is a fake video.

[0078] Such a conflict ultimately led to the news being judged as fake news. It also shows that even if it is impossible to find historical news that occurred at the same time, the big model can retrieve relevant historical news from aspects such as time, place, and related people, and make multi-angle logical reasoning to determine the authenticity of the video.

[0079] Step 3.4: Send the constructed prompt words into the large model for reasoning, judge the authenticity of the news, and obtain the results of fact verification based on RAG technology.

[0080] In step 4, in addition to the RAG process of fact-checking historical news, this application also performs common sense detection on the text to be detected, so as to judge the authenticity of the news from common sense, logic and other aspects.

[0081] Step 4.1: Assign roles to the big model: Set roles for the big model, limit the context, narrow the scope of the big model's answers to obtain more accurate answers, and guide the big model to interact in a more professional way.

[0082] Step 4.2: Construct common sense reasoning prompt words: According to the role setting of the big model, prompt the big model to identify the authenticity of the description in the news from the perspective of historical common sense, and at the same time remind the big model to pay attention to logical errors in time, place, and characters.

[0083] Step 4.3: Construct a thinking chain: Divide the prompt words into four parts: question, news, thinking chain, and answer. The question part is the combination of the role assigned to the big model in step 4.1 and the common sense reasoning prompt words constructed in step 4.2. The news part is the text to be tested constructed in step 1. The thinking chain part is the common sense reasoning process. The answer part is the judgment of whether the news is true or false. And according to this structure, provide an example for the big model.

[0084] Here is an example of a prompt word based on common sense reasoning of a large model:

[0085] Question: You are a common sense encyclopedia. Please help me answer a question based on historical and cultural common sense to determine whether the following news contains false information in terms of common sense. Please note that you need to pay special attention to information about time, place, and people, analyze the common sense or logical errors, and point out the specific errors and reasons. Please think step by step.

[0086] News: TV stations reported that African swine fever broke out in many places in China, but the news was concealed. People claimed that "the African swine fever virus is too strong" and "the people are infected and cannot be cured, and they are suffering greatly from it."

[0087] Thinking chain: African swine fever is a virus that is only transmitted in pigs and does not infect humans. Therefore, the news report that "the people are infected and cannot be cured and are suffering greatly" is false information, so the news is fake news.

[0088] Answer: This news is fake news.

[0089] Question: You are a common sense encyclopedia. Please help me answer a question based on historical and cultural common sense to determine whether the following news contains false information in terms of common sense. Please note that you need to pay special attention to information about time, place, and people, analyze the common sense or logical errors, and point out the specific errors and reasons. Please think step by step.

[0090] News: {constructed from the text to be detected in step 1}

[0091] Thinking chain:

[0092] Answer:

[0093] This common sense reasoning process mainly targets historical and cultural common sense reasoning and logical reasoning related to time, place and people. It complements the news fact verification in step 3. News fact verification can only verify similar news that has happened before based on similarity, and is mainly used to identify: ① maliciously tampering with news that has happened before into recent news; ② forged news that replaces important time, place and people; ③ forged news with many related reports and a clear logical chain. Common sense reasoning is not limited to similarity and historical news evidence, and also has a good ability to identify fabricated news.

[0094] In terms of historical and cultural common sense reasoning, for example, the example news given in the above prompt words: TV stations reported that African swine fever broke out in many places in China, and the news was concealed. People claimed that "the African swine fever virus is too strong" and "the people are infected and cannot be eliminated, and they are deeply affected." If only news fact checking is used for detection, it is easy to retrieve news related to various infectious diseases. Since there is no prompt for the big model to consider the truth or falsity of common sense, the big model will easily judge that the news is true news because of the similarity of sentences such as virus outbreak and people are deeply affected. However, since African swine fever only infects pigs and no human infection cases have been found so far, "the people are infected and cannot be eliminated, and they are deeply affected" is obviously forged. This news should be false news.

[0095] In terms of logical reasoning related to time, place and people, the sample news is as follows: On the streets of Japan, snowflakes are falling and dancing with cherry blossom petals, decorating the entire street. People are bustling, wearing heavy coats, and their faces are filled with admiration and joy for the beautiful scenery, as if the whole world is intoxicated by it. On the surface, it is a normal description of the environment. However, through common sense detection, the big model replied: Cherry blossoms usually bloom in spring, and the flower season usually lasts from late March to mid-April, and some may last until May. The news describes "snowflakes are falling and dancing with cherry blossom petals" and "people are wearing heavy coats", which obviously means that the time period is in winter. The timelines described in these two scenes are inconsistent, which may be due to the fusion of elements of different seasons through editing or creative techniques. News fact checking does not perform well for this type of false news, but the advantages of common sense reasoning in this field can be better utilized, which shows that the two are highly complementary.

[0096] At the same time, no matter whether it is finally judged to be true news or false news, the big model will give explainable reasons, which also provides decision-making materials for our comprehensive judgment of the authenticity of the news in step 5.

[0097] Step 4.4: Send the constructed prompt words into the big model for reasoning, judge the authenticity of the news, and obtain the results of common sense reasoning based on the big model.

[0098] In step 5, the application combines the results of fact-checking and common sense reasoning to make the following comprehensive judgment on the authenticity of the news:

[0099] If the results of fact-checking and common-sense reasoning are both true, the news is judged to be true news, and explainable reasons are given.

[0100] If the results of fact-checking and common sense reasoning are both false, the news is judged to be false news, and explainable reasons are given.

[0101] If the results of fact checking and common sense reasoning are one true and one false, the news is also judged as false news. This is because fact checking and common sense reasoning can only judge the truth of news in their respective fields. If there is no false common sense in a piece of news, but fact checking detects it as false news (or fact checking detects it as true news, but common sense reasoning detects it as false news), it is still false news. To avoid disputes, explainable reasons for fact checking and common sense reasoning to judge the truth of news will be given at the end.

[0102] Based on the above embodiments, Figure 4 As shown, the present application also proposes a video detection system based on a large model and RAG technology, including:

[0103] The module for constructing text to be detected is used to uniformly convert the video, audio, and text information in the video to be detected into textual information for subsequent processing;

[0104] External knowledge base construction module, used to pre-process a large amount of historical news and convert it into an external knowledge base that can be used by RAG;

[0105] The historical news detection module is used to fact-check the text to be detected and determine its authenticity by using RAG technology and external knowledge base;

[0106] The common sense reasoning module is used to perform common sense reasoning on the text to be tested through a large model to determine its authenticity.

[0107] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0108] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A video authenticity detection method based on large model and retrieval enhancement generation technology, characterized in that: The following steps are involved: Step 1: Construction of the text to be detected: Based on the multimodal large model LLaVA, construct prompt words to describe the video to be detected and obtain the video description text; extract the subtitles in the video to be detected through OCR text recognition, and combine the video title, introduction, audio conversion text and video description text to construct the text to be detected; Step 2: External knowledge base construction: Collect a large amount of historical news from official news sources, pre-process the collected news text, convert it into news vector representation and create an index for it to complete the construction of the external knowledge base; Step 3: Fact-checking based on retrieval-augmented generation technology RAG: The text to be checked is converted into a vector representation, and similarity matching is performed with the news vector stored in the external knowledge base to obtain the basis for fact-checking; the basis is combined with the text to be checked to construct prompt words, and the large model ChatGLM is used to perform fact-checking and judge the authenticity of the news; Step 4: Common sense reasoning based on the big model: Construct prompt words, use the big model ChatGLM to reason about the text to be tested from common sense and logic, detect whether there are common sense fallacies and logical conflicts, and judge the authenticity of the news based on this; Step 5: News authenticity judgment: Based on the above fact verification and common sense reasoning, comprehensively judge the authenticity of the video to be tested.

2. The video authenticity detection method based on large model and retrieval enhancement generation technology according to claim 1 is characterized in that: The step 2 is specifically performed as follows: Step 2.1: Data collection: Collect a large amount of historical news from official news sources to ensure the authenticity and reliability of the information; Step 2.2: Data preprocessing: Clean the collected news text, remove irrelevant characters and punctuation marks, and then split it into several text blocks to fit the contextual constraints of the large language model; Step 2.3: Vector representation: Use the embedding model BERT to convert the preprocessed text block into a vector representation; Step 2.4: Index creation: Create an index to store these text chunks and their vector embeddings as key-value pairs.

3. The video authenticity detection method based on large model and retrieval enhancement generation technology according to claim 1 is characterized in that: The step 3 is specifically performed as follows: Step 3.1: Vector representation of the text to be detected: Use the embedding model BERT to vectorize the text to be detected and convert it into a vector representation; Step 3.2: Retrieve fact-checking evidence: Use the BM25 algorithm to perform similarity matching between the text vector to be tested and the historical news vector stored in the external knowledge base. By calculating the similarity between the two, select the top K articles with the highest similarity as the basis for fact-checking; Step 3.3: Construct the big model prompt words: Combine the retrieved evidence with the text to be tested, prompt the big model to consider the causal logical relationship between the two, whether there is a conflict, whether there is direct evidence to judge whether the text to be tested is true or false, and provide an example to the big model; Step 3.4: Send the constructed prompt words into the large model for reasoning, judge the authenticity of the news, and obtain the results of fact verification based on RAG technology.

4. The video authenticity detection method based on large model and retrieval enhancement generation technology according to claim 1 is characterized in that: The step 4 is specifically performed as follows: Step 4.1: Assign roles to the big model: Set roles for the big model, limit the context, narrow the scope of the big model's answers to obtain more accurate answers, and guide the big model to interact in a more professional way; Step 4.2: Construct common sense reasoning prompt words: Based on the role setting of the big model, prompt the big model to identify the authenticity of the description in the news from the perspective of historical common sense, and at the same time remind the big model to pay attention to logical errors in time, place, and characters; Step 4.3: Construct a thinking chain: Divide the prompt words into four parts: question, news, thinking chain, and answer. The question part is the combination of the role assigned to the big model in step 4.1 and the common sense reasoning prompt words constructed in step 4.

2. The news part is the text to be tested constructed in step 1. The thinking chain part is the common sense reasoning process. The answer part is the judgment of whether the news is true or false. According to this structure, provide an example to the big model. Step 4.4: Send the constructed prompt words into the big model for reasoning, judge the authenticity of the news, and obtain the results of common sense reasoning based on the big model.

5. A video authenticity detection system based on large model and retrieval enhancement generation technology, characterized in that: include: The module for constructing text to be detected is used to uniformly convert the video, audio, and text information in the video to be detected into textual information for subsequent processing; External knowledge base construction module, used to pre-process a large amount of historical news and convert it into an external knowledge base that can be used by RAG; The historical news detection module is used to fact-check the text to be detected and determine its authenticity by using RAG technology and external knowledge base; The common sense reasoning module is used to perform common sense reasoning on the text to be tested through a large model to determine its authenticity.

Images