Artificial intelligence-based misleading information detection method and system

Abstract

This invention belongs to the field of information detection, specifically referring to a method and system for detecting misleading information based on artificial intelligence. The method includes large language model construction, early hallucination detection, multi-model response fusion graph, and hierarchical conflict detection logic. This solution selects multiple models to generate independent responses, constructs an initial graph, filters a subset of core triples based on confidence levels, mines candidate implicit relationships, and filters effective implicit relationships to complete the graph. It updates the confidence level, classifying triples into consensus, disputed, and isolated facts. The hierarchical conflict detection logic reduces the impact of semantic bias on confidence levels, clarifies conflict judgment criteria, and improves the accuracy of hallucination detection while reducing the false positive rate. By extracting four types of representations from the first word in the model's generated factual response and before the complete response is output, it achieves early hallucination detection, predicting whether hallucinations will occur without waiting for the complete output.

Description

Technical Field

[0001] This invention belongs to the field of information detection, specifically referring to a method and system for detecting misleading information based on artificial intelligence. Background Technology

[0002] The illusion problem refers to the issue where the output generated by a Large Language Model (LLM) is factually incorrect, irrelevant, or completely fictitious. Artificial intelligence-based methods and systems for detecting misleading information refer to methods and systems that use artificial intelligence technology to detect illusions in large data models.

[0003] Among existing approximate solutions, such as CN117688164B's hallucination detection method, system, and storage medium based on a large language model, this solution addresses the technical problems of existing hallucination detection methods' difficulty in identifying fine-grained hallucinations such as digital hallucinations and entity hallucinations in retrieval enhancement scenarios, the need for a large amount of manually labeled data due to reliance on classifiers, and the vulnerability of question-answering methods to detection failure due to the large language model's misunderstanding of concepts such as "contradiction" and "conflict." This solution adopts a method that obtains the content to be detected and related document sets, identifies numbers and non-replaceable entity units in the content to be detected, constructs a question bank consisting of entity fill-in-the-blank questions and true / false questions, and uses the related document set to analyze the fill-in-the-blank questions. The system uses a combination of fill-in-the-blank questions and the original content to be detected, along with judgments based on relevant document sets. If a single question is incorrect in all relevant documents, it is determined that a hallucination exists. It also employs a technique of COT-enhancing prompts to improve reasoning ability. This technique enables hallucination detection tasks to be completed without manual annotation, reducing annotation costs, supporting long text hallucination detection, and improving the accuracy of fine-grained hallucination detection and the model's understanding and execution of the task. However, it has technical problems such as the confidence assessment of a single model's answer being easily affected by differences in expression, lack of multi-model consensus verification, and insufficient mining of implicit relationships in the fact knowledge graph, leading to insufficient accuracy in conflict detection.

[0004] Secondly, for example, CN118839134B's self-iterative training method for a large language model illusion detector based on self-consistent voting addresses the technical problems of existing technologies being stuck in a vicious cycle where large-scale datasets are needed to train high-precision detectors, while building large-scale datasets requires high-precision detectors, resulting in extremely high manual annotation costs, difficulty in expanding datasets, and unsatisfactory annotation accuracy. This method employs a three-stage self-iterative training approach: the first stage trains a basic illusion annotator based on the ANAH dataset; the second stage expands the dataset by adding multi-model response data and using self-consistent voting annotations to train a stronger annotator; the third stage expands the dataset by adding multi-topic questions and corresponding model responses, repeating the annotation expansion and training process. The annotation process is divided into three steps: factual existence check, reference information extraction, and illusion type judgment. This achieves simultaneous improvement in dataset size and detector accuracy, high annotation accuracy, strong scalability, and a dataset that closely matches the natural illusion distribution in normal question-and-answer scenarios of large language models. However, it suffers from technical problems such as semantic diversity leading to triplet statistical bias, distorted confidence calculation, and unclear conflict judgment logic that easily misjudges illusions.

[0005] Furthermore, for example, CN120181073B, "A Rumor Detection System and Method Based on a Large Language Model," addresses the technical problems of existing rumor detection methods being susceptible to interference from irrelevant and confusing information in the text, difficulty in effectively utilizing external objective facts, lack of interpretability in detection results, and insufficient persuasiveness. This approach employs a core contradiction extraction module that guides the large language model through dual prompts to extract the core contradictions of entity relationships and entity attributes from the text to be detected. A fact-checking retrieval module retrieves knowledge base data using the RAG algorithm, searches the internet, and filters from a source database to obtain credible retrieval results. It then uses prompts to further filter relevant credible retrieval results. Finally, a rumor judgment decision module guides the large language model through prompts to generate a summary of judgment criteria, analyzes its impact on the core contradictions, and outputs the detection results. This approach achieves the technical effects of reducing interference from irrelevant information, enriching the sources of judgment criteria and improving reliability, avoiding the illusion effect of the large language model, and ensuring that the detection results conform to human judgment logic and possess interpretability. However, it suffers from the technical problem that detection is mostly performed after the large language model generates a complete answer, resulting in a lag and an inability to provide early warnings. Summary of the Invention

[0006] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based method and system for detecting misleading information. Addressing the technical problems of single-model answer confidence assessment being susceptible to variations in expression, lack of multi-model consensus verification, and insufficient mining of implicit relationships in fact knowledge graphs leading to inadequate conflict detection accuracy, this solution selects at least two models to generate independent answers, standardizes and extracts triples with source annotation, merges identical entities and relationships to construct an initial graph, filters a subset of core triples based on confidence, generates candidate implicit relationships through a pre-built logical rule base, and combines sentence embedding algorithms to calculate similarity and filter effective implicit relationships to complete the graph. This achieves multi-model consensus quantification and graph structuring enhancement, improving the technical effects of increasing the accuracy of confidence assessment and the basic reliability of conflict detection. Furthermore, it addresses the issues of semantic expression diversity leading to triple statistical bias, distorted confidence calculation, and unclear conflict determination logic that easily results in misjudgments of illusions. To address the technical issues, this solution uses a sentence embedding algorithm to determine the semantic similarity of triples, merges triples with different expressions but consistent semantics to achieve semantic normalization, and then updates the confidence level to classify triples into consensus, disputed, and isolated facts. A hierarchical conflict detection logic is adopted to reduce the impact of semantic bias on confidence, clarify conflict judgment criteria, improve the accuracy of hallucination detection, and reduce the false positive rate. Addressing the technical problem that existing hallucination detection methods often perform the detection after the large language model generates a complete answer, resulting in lag and the inability to provide early warning, this solution calculates the attribution score from the input layer when the model generates the first lexical unit of the factual answer, before the complete answer is output. It also extracts the distribution probability of lexical units from the output layer, and extracts the attention level quantification value and the internal activation value of the fully connected layer from the internal computation layer to capture four types of representations. Through a binary classification model, it achieves early hallucination detection, predicting whether hallucinations will occur without waiting for the complete output.

[0007] The technical solution adopted by this invention is as follows: The misleading information detection method based on artificial intelligence provided by this invention includes the following steps:

[0008] Step S1: Large language model construction, used to construct a large language model based on the Transformer architecture, which is used to generate corresponding factual answers to the questions raised;

[0009] Step S2: Early hallucination detection, used to detect hallucinations before the large language model makes a factual answer;

[0010] Step S3: Multi-model answer fusion graph, quantify the logical relationships between factual answers from large language models, and initialize the factual knowledge graph;

[0011] Step S4: Layered conflict detection logic, which performs conflict detection on the large language model to achieve illusion detection after the large language model makes a factual answer.

[0012] Furthermore, in step S2, the early detection of hallucination specifically includes the following steps:

[0013] Step S21: Representation capture, specifically, when the large language model has just generated the first word of the factual answer and has not yet output the complete factual answer, extract four types of representations from the three dimensions of the input layer, output layer and internal operation of the large language model respectively.

[0014] Step S22: Using the four types of representations as input, train four binary classification models specifically for judging the existence of hallucinations. The binary classification models directly give the result of judging whether a hallucination exists.

[0015] Further, in step S21, the characterization capture specifically includes the following steps:

[0016] Step S211: Input angle dimension extraction, specifically, the proposed question is divided into the smallest semantic units, the gradient integral method is used to calculate the attribution score of each of the smallest semantic units for the factual answer, and the attribution score is compressed into a single scalar using the L2 norm;

[0017] Step S212: Output angle dimension extraction, specifically, extract the distribution probability of the first word element;

[0018] Step S213: Extraction of internal operation angle dimension. Specifically, firstly, based on the self-attention principle, when the large language model generates the first word, the attention degree quantization value of the last smallest semantic unit in the question is extracted. Secondly, the output of the fully connected layer in the large language model is extracted and recorded as the internal activation value. When generating non-illusion, the attention degree quantization value will be concentrated on the key information words of the question, and the internal activation value will be focused on the dimension related to the facts. However, when generating illusion, the attention degree quantization value will be lower than the mean, and the distribution of the internal activation value will deviate from the fact dimension.

[0019] Step S214: The attribution score, the distribution probability, the attention level quantification value, and the internal activation value are abbreviated as four types of representations.

[0020] Furthermore, in step S3, the multi-model response fusion graph specifically includes the following steps:

[0021] Step S31: Information standardization collection. For the same question, select at least two large language models to generate independent factual answers. Use keyword filtering to filter speculative words and subjective evaluation words. Use named entity recognition algorithm to extract entities in each factual answer. Use relation extraction technology to extract triples of entities and relations. Label each triple with the source identifier of the large language model. Finally, obtain a standardized multi-source fact set.

[0022] Step S32: Merge triples, which are used to treat triples as graph data. Specifically, identical entities in the standardized multi-source fact set are merged into unified nodes, and identical relations are merged into unified edges. A fact knowledge graph is built based on the Neo4j graph database.

[0023] Step S33: Calculate the first confidence level and build consensus. Specifically, calculate the confidence level of triples in the fact knowledge graph, preset a confidence threshold, and take all triples greater than or equal to the confidence threshold as basic accepted facts among the large language models. All basic accepted facts form a subset of core triples. The confidence level calculation formula is as follows: ;

[0024] In the formula, This indicates the confidence level of the triplet. This indicates the number of large language models that support the triple, i.e., the number of large language models that can be directly extracted from the factual responses. This represents the total number of large language models involved in generating factual answers;

[0025] Step S34: Implicit relation mining, used to complete the implicit relations of the fact knowledge graph and form structured associations. Specifically, it identifies implicit relations between entities that are not directly expressed but are logically necessary, and adds the implicit relations as new edges to the fact knowledge graph.

[0026] Step S35: Calculate the second confidence level and initialize the graph. Specifically, repeat the confidence level calculation method in step S33, update the confidence level, and divide all triples in the fact knowledge graph into consensus facts, disputed facts, and isolated facts based on the confidence level.

[0027] Furthermore, in step S34, the implicit relationship mining specifically includes the following steps:

[0028] Step S341: Construct a logical rule base. The logical rule base consists of deducible rules, including transitive relationship rules, attribute inheritance rules, and mutually exclusive relationship exclusion rules. Based on the core triplet subset, the logical rule base is traversed through the rule matching engine to automatically generate new triplet combinations that conform to the deducible rules. The new triplet combinations are recorded as candidate relation triplets.

[0029] Step S342: Using a sentence embedding algorithm, perform relation embedding calculations on the candidate relation triples and the subset of core triples, convert the triples into embedding vectors, set a similarity threshold, calculate the cosine similarity between the embedding vectors corresponding to the candidate relation triples and the embedding vectors corresponding to the subset of core triples, and filter out candidate relation triples with a cosine similarity greater than or equal to the similarity threshold, which are the final implicit relation triples. Add the implicit relation triples as new edges to the fact knowledge graph.

[0030] Further, in step S4, the hierarchical conflict detection logic specifically includes the following steps:

[0031] Step S41: Calculate the third confidence and semantic normalization. Specifically, combine cosine similarity to determine the semantic similarity between triples in the fact knowledge graph. Merge triples with different expressions but consistent semantics into a unified triple. Repeat the confidence calculation method in step S33 to update the confidence of the triple again.

[0032] Step S42: Conflict determination branch, specifically, the triples of the large language model to be detected in the fact knowledge graph are denoted as target triples, and the conflict between the target triples and the consensus facts in the fact knowledge graph is detected;

[0033] If the target triple conflicts with the consensus fact, and the target triple is not supported by any large oracle model in the fact knowledge graph after semantic normalization in step S41, i.e., the confidence level is 0, then it is determined that the fact answer of the large language model to be detected is an illusion.

[0034] If the target triple does not conflict with the consensus fact but only with the disputed fact, then the target triple is marked as pending verification. Additional true and correct consensus facts need to be added to the fact knowledge graph before a final judgment is made.

[0035] The present invention provides an artificial intelligence-based method for detecting misleading information, which is applied to an artificial intelligence-based system for detecting misleading information. The artificial intelligence-based system for detecting misleading information includes a large language model construction module, an early hallucination detection module, a multi-model response fusion graph module, and a hierarchical conflict detection logic module.

[0036] The large language model construction module constructs a large language model based on the Transformer architecture, which is used to generate corresponding factual answers to the questions raised.

[0037] The hallucination early detection module is used to detect hallucinations before the large language model makes a factual answer.

[0038] The multi-model answer fusion graph module is used to quantify the logical relationships between factual answers from large language models and initialize the factual knowledge graph.

[0039] The hierarchical conflict detection logic module detects conflicts in the large language model, thereby enabling illusion detection after the large language model has provided a factual answer.

[0040] The beneficial effects achieved by adopting the above solution are as follows:

[0041] (1) To address the technical issues that the confidence assessment of a single model's answer is easily affected by differences in expression, lacks multi-model consensus verification, and the implicit relationships in the fact knowledge graph are not fully explored, resulting in insufficient accuracy in conflict detection, this solution selects at least two models to generate independent answers, standardizes and extracts triples and labels their sources, merges the same entities and relationships to build an initial graph, filters the core triple subset based on confidence, generates candidate implicit relationships through a pre-built logical rule base, and combines sentence embedding algorithm to calculate similarity and filter effective implicit relationships to complete the graph. This achieves multi-model consensus quantification and graph structuring enhancement, improving the technical effect of increasing the accuracy of confidence assessment and the basic reliability of conflict detection.

[0042] (2) In response to the technical problems of statistical bias and confidence calculation distortion caused by the diversity of semantic expressions, and the unclear conflict judgment logic which easily leads to misjudgment of hallucinations, this solution uses sentence embedding algorithm to determine the semantic similarity of triples, merges triples with different expressions but consistent semantics to complete semantic normalization, and then updates the confidence to classify triples into consensus, dispute, and isolated facts. By adopting a hierarchical conflict detection logic, the solution achieves the technical effects of reducing the impact of semantic bias on confidence, clarifying the conflict judgment criteria, improving the accuracy of hallucination detection, and reducing the misjudgment rate.

[0043] (3) In view of the technical problem that existing hallucination detection is mostly performed after the large language model generates a complete answer, which has a lag and cannot provide early warning, this solution captures four types of representations by calculating the attribution score from the input layer when the model generates the first word of the factual answer and before the complete answer is output, extracting the distribution probability of the word in the output layer, and extracting the attention level quantification value and the internal activation value of the fully connected layer from the internal operation layer. Through the binary classification model, early hallucination detection is achieved, and the technical effect of predicting whether hallucination will occur later without waiting for the complete output is realized. Attached Figure Description

[0044] Figure 1 A flowchart illustrating the artificial intelligence-based misleading information detection method provided by the present invention;

[0045] Figure 2 The module connection diagram of the AI-based misleading information detection system provided by the present invention is shown.

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0048] Example 1, see Figure 1 and Figure 2 The present invention provides an artificial intelligence-based method for detecting misleading information, which includes the following steps:

[0049] Step S1: Large language model construction, used to construct a large language model based on the Transformer architecture, which is used to generate corresponding factual answers to the questions raised;

[0050] Step S2: Early hallucination detection, used to detect hallucinations before the large language model makes a factual answer;

[0051] Step S3: Multi-model answer fusion graph, quantify the logical relationships between factual answers from large language models, and initialize the factual knowledge graph;

[0052] Step S4: Layered conflict detection logic, which performs conflict detection on the large language model to achieve illusion detection after the large language model makes a factual answer.

[0053] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S2, the early detection of hallucinations specifically includes the following steps:

[0054] Step S21: Representation capture, specifically, when the large language model has just generated the first word of the factual answer and has not yet output the complete factual answer, extract four types of representations from the three dimensions of the input layer, output layer and internal operation of the large language model respectively.

[0055] Step S22: Using the four types of representations as input, train four binary classification models specifically for judging the existence of hallucinations. The binary classification models directly give the result of judging whether a hallucination exists.

[0056] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S21, the characterization capture specifically includes the following steps:

[0057] Step S211: Input angle dimension extraction, specifically, the proposed question is divided into the smallest semantic units, the gradient integral method is used to calculate the attribution score of each of the smallest semantic units for the factual answer, and the attribution score is compressed into a single scalar using the L2 norm;

[0058] Step S212: Output angle dimension extraction, specifically, extract the distribution probability of the first word element;

[0059] Step S213: Extraction of internal operation angle dimension. Specifically, firstly, based on the self-attention principle, when the large language model generates the first word, the attention degree quantization value of the last smallest semantic unit in the question is extracted. Secondly, the output of the fully connected layer in the large language model is extracted and recorded as the internal activation value. When generating non-illusion, the attention degree quantization value will be concentrated on the key information words of the question, and the internal activation value will be focused on the dimension related to the facts. However, when generating illusion, the attention degree quantization value will be lower than the mean, and the distribution of the internal activation value will deviate from the fact dimension.

[0060] Step S214: The attribution score, the distribution probability, the attention level quantification value, and the internal activation value are abbreviated as four types of representations.

[0061] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the multi-model response fusion graph specifically includes the following steps:

[0062] Step S31: Information standardization collection. For the same question, select at least two large language models to generate independent factual answers. Use keyword filtering to filter speculative words and subjective evaluation words. Use named entity recognition algorithm to extract entities in each factual answer. Use relation extraction technology to extract triples of entities and relations. Label each triple with the source identifier of the large language model. Finally, obtain a standardized multi-source fact set.

[0063] Step S32: Merge triples, which are used to treat triples as graph data. Specifically, identical entities in the standardized multi-source fact set are merged into unified nodes, and identical relations are merged into unified edges. A fact knowledge graph is built based on the Neo4j graph database.

[0064] Step S33: Calculate the first confidence level and build consensus. Specifically, calculate the confidence level of triples in the fact knowledge graph, preset a confidence threshold, and take all triples greater than or equal to the confidence threshold as basic accepted facts among the large language models. All basic accepted facts form a subset of core triples. The confidence level calculation formula is as follows: ;

[0065] In the formula, This indicates the confidence level of the triplet. This indicates the number of large language models that support the triple, i.e., the number of large language models that can be directly extracted from the factual responses. This represents the total number of large language models involved in generating factual answers;

[0066] Step S34: Implicit relation mining, used to complete the implicit relations of the fact knowledge graph and form structured associations. Specifically, it identifies implicit relations between entities that are not directly expressed but are logically necessary, and adds the implicit relations as new edges to the fact knowledge graph.

[0067] Step S35: Calculate the second confidence level and initialize the graph. Specifically, repeat the confidence level calculation method in step S33, update the confidence level, and divide all triples in the fact knowledge graph into consensus facts, disputed facts, and isolated facts based on the confidence level.

[0068] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S34, the hidden relationship mining specifically includes the following steps:

[0069] Step S341: Construct a logical rule base. The logical rule base consists of deducible rules, including transitive relationship rules, attribute inheritance rules, and mutually exclusive relationship exclusion rules. Based on the core triplet subset, the logical rule base is traversed through the rule matching engine to automatically generate new triplet combinations that conform to the deducible rules. The new triplet combinations are recorded as candidate relation triplets.

[0070] Step S342: Using a sentence embedding algorithm, perform relation embedding calculations on the candidate relation triples and the subset of core triples, convert the triples into embedding vectors, set a similarity threshold, calculate the cosine similarity between the embedding vectors corresponding to the candidate relation triples and the embedding vectors corresponding to the subset of core triples, and filter out candidate relation triples with a cosine similarity greater than or equal to the similarity threshold, which are the final implicit relation triples. Add the implicit relation triples as new edges to the fact knowledge graph.

[0071] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S4, the hierarchical conflict detection logic specifically includes the following steps:

[0072] Step S41: Calculate the third confidence and semantic normalization. Specifically, combine cosine similarity to determine the semantic similarity between triples in the fact knowledge graph. Merge triples with different expressions but consistent semantics into a unified triple. Repeat the confidence calculation method in step S33 to update the confidence of the triple again.

[0073] Step S42: Conflict determination branch, specifically, the triples of the large language model to be detected in the fact knowledge graph are denoted as target triples, and the conflict between the target triples and the consensus facts in the fact knowledge graph is detected;

[0074] If the target triple conflicts with the consensus fact, and the target triple is not supported by any large oracle model in the fact knowledge graph after semantic normalization in step S41, i.e., the confidence level is 0, then it is determined that the fact answer of the large language model to be detected is an illusion.

[0075] If the target triple does not conflict with the consensus fact but only with the disputed fact, then the target triple is marked as pending verification. Additional true and correct consensus facts need to be added to the fact knowledge graph before a final judgment is made.

[0076] Example 7, see Figure 1 and Figure 2 Based on the above embodiments, the artificial intelligence-based misleading information detection system provided by the present invention includes a large language model construction module, an early hallucination detection module, a multi-model response fusion graph module, and a hierarchical conflict detection logic module.

[0077] The large language model construction module constructs a large language model based on the Transformer architecture, which is used to generate corresponding factual answers to the questions raised.

[0078] The hallucination early detection module is used to detect hallucinations before the large language model makes a factual answer.

[0079] The multi-model answer fusion graph module is used to quantify the logical relationships between factual answers from large language models and initialize the factual knowledge graph.

[0080] The hierarchical conflict detection logic module detects conflicts in the large language model, thereby enabling illusion detection after the large language model has provided a factual answer.

[0081] Example 8, see Figure 1 This embodiment is based on the above embodiment. In step S31, for the same question, this embodiment selects three large language models to generate independent factual answers. The selected large language models are GPT, deepseek and Zhipu Qingyan.

[0082] Example 9, see Figure 1 This embodiment is based on the above embodiment. In step S22, the four binary classification models specifically used to determine the existence of hallucinations include the following steps:

[0083] Step S221: For the attribution scores, a binary classification model is constructed using a 4-layer recurrent neural network, with each layer undergoing 25% pruning.

[0084] Step S222: For the distribution probability, the attention level quantization value, and the internal activation value, use a single-layer neural network as a binary classification model.

[0085] Example 10, see Figure 1 This embodiment is based on the above embodiment. In step S22, the Adam optimizer is used to adjust the binary classification model during training, and the learning speed is set to... The weight constraint is Each training iteration uses 128 samples, for a total of 1000 training rounds. The dataset is divided into training and test sets in an 8:2 ratio.

[0086] Example 11, see Figure 1 This embodiment is based on the above embodiment. In step S22, the probability that the first word is Beijing is 60% and the probability that it is Shanghai is 40%.

[0087] Example 12, see Figure 1 This embodiment is based on the above embodiment. In step S213, extracting the attention level quantization value of the last smallest semantic unit in the problem specifically includes the following steps:

[0088] Step S2131: Vector transformation, convert the word embedding of the last word of the question into a key vector through the weight matrix, and convert the hidden state before generating the first answer word into a query vector;

[0089] Step S2132: Calculate the original similarity by taking the dot product of the query vector and the key vector to obtain the original score;

[0090] Step S2133: Standardization. The original score is divided by the square root of the key vector dimension, and then normalized by the Softmax function to obtain a value between 0 and 1, which is the attention level quantification value.

[0091] Example 13, see Figure 1 This embodiment is based on the above embodiment. In step S33, the confidence threshold is 0.5.

[0092] Example 14, see Figure 1 This embodiment is based on the above embodiment. In step S35, all triples in the fact knowledge graph are divided into consensus facts, disputed facts and isolated facts according to the confidence level. Confidence level ≥ 0.6 is marked as consensus facts, 0.2 ≤ confidence level < 0.6 is marked as disputed facts, and confidence level < 0.2 is marked as isolated facts.

[0093] Example 15, see Figure 1 This embodiment is based on the above embodiment. In step S341, the triple "China-Capital-Beijing" from GPT and the triple "Beijing-Continent-Asia" from deepseek are matched and deduced according to the transitive relation rule: the two triples satisfy the transitive logic, and the candidate implicit relation "China-Continent-Asia" is automatically generated.

[0094] Example 16, see Figure 1 This embodiment is based on the above embodiment. In step S342, the similarity threshold is 0.85.

[0095] Example 17, see Figure 1 This embodiment is based on the above embodiment. In step S42, the conflict between the target triple and the consensus facts in the fact knowledge graph, and the conflict between the target triple and the disputed facts in the fact knowledge graph, specifically refers to traversing and detecting the attribute conflicts and relational logic conflicts between entities in the target triple in the fact knowledge graph. In this embodiment, in the triple "Zhang San - born in -1984", "1984" belongs to the time attribute, which conflicts with the location attribute of the entity.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0098] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for detecting misleading information based on artificial intelligence, characterized in that: The method includes the following steps: Step S1: Large language model construction, used to construct a large language model based on the Transformer architecture, which is used to generate corresponding factual answers to the questions raised; Step S2: Early hallucination detection, used to detect hallucinations before the large language model makes a factual answer; Step S3: Multi-model answer fusion graph, quantify the logical relationships between factual answers from large language models, and initialize the factual knowledge graph; Step S4: Layered conflict detection logic, which performs conflict detection on the large language model to achieve illusion detection after the large language model makes a factual answer.

2. The method for detecting misleading information based on artificial intelligence according to claim 1, characterized in that: In step S2, the early detection of hallucinations specifically includes the following steps: Step S21: Representation capture, specifically, when the large language model has just generated the first word of the factual answer and has not yet output the complete factual answer, extract four types of representations from the three dimensions of the input layer, output layer and internal operation of the large language model respectively. Step S22: Using the four types of representations as input, train four binary classification models specifically for judging the existence of hallucinations. The binary classification models directly give the result of judging whether a hallucination exists.

3. The method for detecting misleading information based on artificial intelligence according to claim 2, characterized in that: In step S21, the characterization capture specifically includes the following steps: Step S211: Input angle dimension extraction, specifically, the proposed question is divided into the smallest semantic units, the gradient integral method is used to calculate the attribution score of each of the smallest semantic units for the factual answer, and the attribution score is compressed into a single scalar using the L2 norm; Step S212: Output angle dimension extraction, specifically, extract the distribution probability of the first word element; Step S213: Extraction of internal operation angle dimension. Specifically, firstly, based on the self-attention principle, when the large language model generates the first word, extract the attention degree quantization value of the last smallest semantic unit in the problem. Secondly, extract the output of the fully connected layer in the large language model and record it as the internal activation value. Step S214: The attribution score, the distribution probability, the attention level quantification value, and the internal activation value are abbreviated as four types of representations.

4. The method for detecting misleading information based on artificial intelligence according to claim 1, characterized in that: In step S3, the multi-model response fusion graph specifically includes the following steps: Step S31: Information standardization collection. For the same question, select at least two large language models to generate independent factual answers. Use keyword filtering to filter speculative words and subjective evaluation words. Use named entity recognition algorithm to extract entities in each factual answer. Use relation extraction technology to extract triples of entities and relations. Label each triple with the source identifier of the large language model. Finally, obtain a standardized multi-source fact set. Step S32: Merge triples, which are used to treat triples as graph data. Specifically, identical entities in the standardized multi-source fact set are merged into unified nodes, and identical relations are merged into unified edges. A fact knowledge graph is built based on the Neo4j graph database. Step S33: Calculate the first confidence level and build consensus. Specifically, calculate the confidence level of triples in the fact knowledge graph, preset a confidence threshold, and take all triples greater than or equal to the confidence threshold as basic accepted facts among the large language models. All basic accepted facts form a subset of core triples. The confidence level calculation formula is as follows: ； In the formula, This indicates the confidence level of the triplet. This indicates the number of large language models that support the triple, i.e., the number of large language models that can be directly extracted from the factual responses. This represents the total number of large language models involved in generating factual answers; Step S34: Implicit relation mining, used to complete the implicit relations of the fact knowledge graph and form structured associations. Specifically, it identifies implicit relations between entities that are not directly expressed but are logically necessary, and adds the implicit relations as new edges to the fact knowledge graph. Step S35: Calculate the second confidence level and initialize the graph. Specifically, repeat the confidence level calculation method in step S33, update the confidence level, and divide all triples in the fact knowledge graph into consensus facts, disputed facts, and isolated facts based on the confidence level.

5. The method for detecting misleading information based on artificial intelligence according to claim 4, characterized in that: In step S34, the implicit relationship mining specifically includes the following steps: Step S341: Construct a logical rule base. The logical rule base consists of deducible rules, including transitive relationship rules, attribute inheritance rules, and mutually exclusive relationship exclusion rules. Based on the core triplet subset, the logical rule base is traversed through the rule matching engine to automatically generate new triplet combinations that conform to the deducible rules. The new triplet combinations are recorded as candidate relation triplets. Step S342: Using a sentence embedding algorithm, perform relation embedding calculations on the candidate relation triples and the subset of core triples, convert the triples into embedding vectors, set a similarity threshold, calculate the cosine similarity between the embedding vectors corresponding to the candidate relation triples and the embedding vectors corresponding to the subset of core triples, and filter out candidate relation triples with a cosine similarity greater than or equal to the similarity threshold, which are the final implicit relation triples. Add the implicit relation triples as new edges to the fact knowledge graph.

6. The method for detecting misleading information based on artificial intelligence according to claim 5, characterized in that: In step S4, the hierarchical conflict detection logic specifically includes the following steps: Step S41: Calculate the third confidence and semantic normalization. Specifically, combine cosine similarity to determine the semantic similarity between triples in the fact knowledge graph. Merge triples with different expressions but consistent semantics into a unified triple. Repeat the confidence calculation method in step S33 to update the confidence of the triple again. Step S42: Conflict determination branch, specifically, the triples of the large language model to be detected in the fact knowledge graph are denoted as target triples, and the conflict between the target triples and the consensus facts in the fact knowledge graph is detected; If the target triple conflicts with the consensus fact, and the target triple is not supported by any large oracle model in the fact knowledge graph after semantic normalization in step S41, i.e., the confidence level is 0, then it is determined that the fact answer of the large language model to be detected is an illusion. If the target triple does not conflict with the consensus fact but only with the disputed fact, then the target triple is marked as pending verification. Additional true and correct consensus facts need to be added to the fact knowledge graph before a final judgment is made.

7. An artificial intelligence-based misleading information detection system, used to implement the artificial intelligence-based misleading information detection method as described in any one of claims 1-6, characterized in that: It includes a large language model construction module, an early hallucination detection module, a multi-model response fusion graph module, and a hierarchical conflict detection logic module; The large language model construction module constructs a large language model based on the Transformer architecture, which is used to generate corresponding factual answers to the questions raised. The hallucination early detection module is used to detect hallucinations before the large language model makes a factual answer. The multi-model answer fusion graph module is used to quantify the logical relationships between factual answers from large language models and initialize the factual knowledge graph. The hierarchical conflict detection logic module performs conflict detection on the large language model to achieve illusion detection after the large language model has made a factual answer.

Images