Large model combined knowledge graph reasoning method for false news detection

Abstract

The invention belongs to the field of combination of a large model and a knowledge graph, and particularly relates to a large model and knowledge graph combined reasoning method for false news detection, which breaks through the limitation of a traditional detection method in the aspects of knowledge instantaneity and reasoning controllability through heterogeneous architecture design. Based on multi-dimensional features such as timeliness, propagation mode and information density of news propagation, a double-engine-driven intelligent reasoning system is constructed, on one hand, important entity extraction and semantic understanding are carried out through a large language model, and on the other hand, structured fact verification is provided through a dynamic knowledge graph. And the inherent factual illusion problem of a large language model is effectively solved. Compared with a traditional knowledge iteration mode depending on single model parameter updating, the method supports an incremental updating strategy based on knowledge graph nodes.

Description

Technical Field

[0001] The present invention belongs to the field of combining large models with knowledge graphs, and in particular relates to a large model combined with knowledge graph reasoning method for fake news detection. Background Art

[0002] With the development of the internet, fake news has spread more rapidly and extensively, posing a real-world threat across a wide range of critical sectors. Its covert and strategic nature makes it a persistent challenge in natural language processing and data analysis. In the era of digital communication, fake news detection is crucial for maintaining the online ecosystem, with applications encompassing public opinion monitoring, content security review, and news credibility rating. The explosive spread of information on social media platforms significantly complicates detection, and malicious content creators often employ tactics like mixing text and images to enhance deceptiveness. This reality has driven technological innovation in fake news detection, effectively addressing the emerging threat of fabrication posed by artificial intelligence-generated content (AIGC) technology. This has strategic value in safeguarding public perception and maintaining the foundation of trust in the digital world.

[0003] Fake news detection faces a number of challenges, one of which is that the creators of fake news may manipulate any part of the news, for example, by using different writing strategies and fabricating non-existent entities and relationships. Research on fake news detection has made some progress in recent years. Early studies explored the use of pre-trained small language models, such as SLM, for detection. Subsequent research has introduced alternative social context and knowledge bases for improvement. However, these limited knowledge and capabilities have limited further improvement in the effectiveness of detectors.

[0004] In recent years, large language models (LLMs) have emerged as a novel alternative to SLMs. Their powerful corpora and super-human reasoning capabilities have enabled significant breakthroughs in fake news detection. However, despite this progress, LLMs often suffer from hallucinations (typically manifesting as models fabricating facts, providing false information, or generating illogical content), particularly in scenarios requiring depth and accountability. Therefore, a key challenge in using large models for fake news detection is making LLMs' reasoning more informed and efficient, while mitigating hallucinations. The following sections introduce some related approaches.

[0005] (1) Large Language Model

[0006] Large Language Models (LLMs) are a revolutionary technology in the field of artificial intelligence. By learning from massive amounts of text data, they achieve a deep understanding and generation of natural language. Unlike traditional natural language processing (NLP) methods, LLMs, based on a self-supervised pre-training paradigm, can capture the complex semantics, grammar, and contextual relationships in language, resulting in near-human-level performance in tasks such as text generation, question answering, and translation. Their core architecture, based on the Transformer, dynamically models long-range dependencies through a self-attention mechanism, breaking through the sequence modeling bottlenecks of traditional recurrent neural networks (RNNs) and convolutional neural networks (CNNs).

[0007] Its core technology, the Transformer architecture, features a self-attention mechanism. Transformer uses multi-head self-attention to parallelly calculate the association weights between words, such as the referential relationship between "it" and multiple candidate nouns in a sentence. The formula is:

[0008]

[0009] Where Q, K, and V represent query, key, and value matrices respectively. k is the dimension scaling factor.

[0010] (2) Knowledge Graph Integration

[0011] In the process of addressing the challenge of large model reasoning, a natural and promising solution is to integrate external knowledge, such as knowledge graphs (KGs), to help improve LLM reasoning. KGs provide structured, explicit, and editable knowledge representations, proposing a complementary strategy to alleviate the limitations of LLMs. KGs convert unstructured data (such as text, images) into "entity-relationship-entity" triples (such as (Canberra, capital_of, Australia)). This supports reasoning based on logical rules (such as "If A is the capital of B, then B is a country"). For example, the Freebase knowledge graph contains more than 40 million entities and 3 billion triples, and each entity is assigned a unique MID. Summary of the Invention

[0012] The present invention proposes a collaborative false news detection framework that integrates a large language model and a dynamic knowledge graph, and breaks through the limitations of traditional detection methods in terms of knowledge real-time and reasoning controllability through heterogeneous architecture design. Based on the multi-dimensional characteristics of news dissemination, such as timeliness, dissemination mode and information density, the present invention constructs a dual-engine driven intelligent reasoning system: on the one hand, a large language model (LLM) is used to extract important entities through semantic understanding, and on the other hand, a dynamic knowledge graph (KG) is used to provide structured fact verification. In particular, the collaborative mechanism established by the knowledge graph entity alignment technology and the implicit knowledge of the large model effectively solves the fact illusion problem inherent in the LLM. Compared with the traditional knowledge iteration method that relies on the update of a single model parameter, the present invention supports an incremental update strategy based on knowledge graph nodes.

[0013] The technical solutions of the present invention are as follows:

[0014] A large model combined with knowledge graph reasoning method for fake news detection, the steps are as follows:

[0015] Step 1: News Text Entity Extraction

[0016] During the news text preprocessing phase, a multi-level entity recognition architecture was designed. This architecture leverages the context-awareness of a large language model (LLM) to perform named entity recognition (NER). A hierarchical annotation strategy based on an attention mechanism was employed to simultaneously capture important entities (such as people and organizations) in the text. The details are as follows:

[0017] First, the news text is sequence labeled, assuming that the news text is a word sequence T = {w1, w2, ..., w n}, the entity tag set is ε={e1,e2,...,e k The LLM generates a context-aware hidden state H through a multi-layer Transformer encoder:

[0018]

[0019] Where d is the hidden state dimension and n is the length of the word sequence. The probability distribution of the entity label is obtained by linear projection of the hidden state and softmax normalization:

[0020] P(e i |w i )=Softmax(W e h i +b e )

[0021] in, is a learnable parameter, and k is the number of all possible entity types. iis the vector representation of the i-th word.

[0022] After labeling, important entities are selected and attention weighting is performed through the self-attention mechanism during the selection process, that is, the self-attention weight is used Identify key entities Importance(w i ):

[0023]

[0024] Among them, A i,j is the i-th row and j-th column element of the self-attention weight matrix, representing the word w i to w j The intensity of attention. j is the vector representation of the j-th word.

[0025] Filter words whose importance is higher than the threshold τ as candidate entities

[0026]

[0027] By setting it up, the large model can extract the top K important entities among the candidate entities.

[0028] Step 2: Building a multi-level prompting system for face-to-face fact-checking

[0029] By designing domain-adaptive prompt templates and dynamic reasoning control mechanisms, we optimize the knowledge activation path and output of the Large Language Model (LLM) in fact-checking tasks. The multi-level prompting project for face-based fact-checking includes the following parts:

[0030] (1) Relationship mapping layer:

[0031] Objective: Identify the core relationships between entities (such as event causality, institutional affiliation, etc.).

[0032] The input text is the news text to be verified, and a predefined entity recognition template (see the example below) is used to extract a set of key entities from the text.

[0033] Entity recognition template: "Analyze the core relations in the text, select the three most relevant items from the candidate list [relation 1; relation 2; ...], and output them as {'relation':'score'} (total score = 1.0):"

[0034] Technical point: Dynamically assign relationship scores based on attention weights to filter the evidence chain most relevant to the statement.

[0035] (2) Verification decision layer

[0036] Goal: Output a final truth judgment based on the evidence triples (entity, relationship, entity). This is divided into steps such as determining whether the current knowledge is sufficient to support the judgment and directly making a truth judgment based on the current triples. The input is the triple path generated after the search, and the truth judgment is made using a predefined multi-path reasoning template.

[0037] Predefined multi-path reasoning template: "Verify the truth of a statement based on the following chain of evidence, and only output {'results': True / False}:

[0038] Chain of evidence (sorted by relevance):

[0039] 1. {path_1_text}(confidence:{score_1}) ...

[0041] n.{path_n_text}(confidence:{score_n})”

[0042] Technical point: Limit the model generation space through output constraints (only True / False is allowed) to avoid invalid interpretations.

[0043] (3) Structured output control

[0044] Standardized output is achieved through syntax tree constraints (in order to obtain standardized output during decoding,

[0045] A syntax tree template must be used to normalize the output of LLM):

[0046] {

[0047] "result":"false or true",

[0048] "Confidence": ["High", "Medium", "Low"],

[0049] "Error Evidence":"…",

[0050] "Supporting evidence":"……"

[0051] }

[0052] Verify the decoding process and check the field type in real time when the model is generated to normalize its output.

[0053] In the design of the prompting project, the LLM dynamically adjusts its prompt template construction strategy based on the language characteristics of the input text (Chinese / English). This structured prompting strategy effectively enhances the large language model's understanding of news texts, improving its reasoning capabilities in fact-checking, source verification, and contradiction detection, thereby establishing a more comprehensive mechanism for distinguishing authenticity from falsity.

[0054] Step 3: Large model combined with knowledge graph for exploration and reasoning

[0055] Based on the entities extracted in step one and the prompt engineering constructed in step two, through the collaborative reasoning of the large model (LLM) and the knowledge graph (KG), the logical path is dynamically constructed and the completeness of the answer is verified, and finally a credible conclusion is output.

[0056] Relationship Exploration: First, based on the entity nodes extracted in step 1, a breadth-first relationship search is performed in the knowledge graph. Relationship discovery begins: Starting from the current entity, all associated relationship edges (such as kinship and affiliation) are retrieved to form an initial candidate set. Relationship screening then utilizes the semantic understanding capabilities of the large model to rank candidate relationships by relevance, selecting the top N relationships with the highest confidence (N is user-defined). Finally, path construction is performed: candidate relationships are integrated into reasoning branches based on logical relevance to form a preliminary hypothesis path.

[0057] Entity exploration: Then, a deep traversal of the knowledge graph is performed based on the selected relationships. First, adjacent node retrieval is performed to locate related entities along the selected relationship edges to obtain a set of candidate entities. Next, entity verification is performed: the contextual reasoning capabilities of the large model are used to evaluate the relevance of the candidate entities to the reasoning target. Because there are multiple candidate entities, path optimization is required: all preliminary hypothesis paths are dynamically pruned using the LLM to retain the optimal reasoning path, ensuring the coherence of the logical chain and the factual consistency of the knowledge graph. This design leverages the complementary advantages of the LLM and the KG to achieve an effective combination of dynamic focus on reasoning and knowledge constraints while maintaining the simplicity of the core process. Finally, through selection, the top-N most relevant relationships can be retained.

[0058] In the reasoning stage, it is necessary to determine whether the currently accumulated path information is sufficient to answer the question. Let the reasoning path set be The subscripts a and c are entity indexes, used to distinguish different entities; b is the relationship index, which identifies the relationship type. Each triple (e a ,r b ,e c ) represents the independent paths retrieved from KG. For each path LLM calculates the confidence score (p i ):

[0059] Score(p i )=LLM reason ([Claim]+[Path:p i ])

[0060] Among them, LLM reason Represents a large model used for reasoning; [Claim] represents a text claim to be verified, [Path:pi ] represents the textual representation of the reasoning path extracted from the knowledge graph.

[0061] By customizing the confidence level, if it is sufficient, the path information is used to generate an answer. If it is insufficient, the system returns to "relationship exploration" and "entity exploration" to continue exploring new paths until the termination conditions are met: a) obtaining a positive or negative news evaluation result; b) reaching the set maximum exploration depth D_max. If D_max is reached and no valid path is generated, the system relies solely on the internal knowledge of the LLM to generate an answer.

[0062] Beneficial effects of the present invention:

[0063] The present invention constructs a framework for false news detection by combining a large model with a knowledge graph, based on the collaborative working mechanism of a large language model (LLM) and a dynamic knowledge graph (KG). First, the named entity recognition (NER) technology of the LLM is used to accurately extract key entities in the news (such as people, institutions, and places). Then, combined with the relationship and entity reasoning of the knowledge graph, these entities are analyzed for relationship associations - that is, various relationships such as direct associations (such as "employed in") and indirect associations (such as "belonging to the same industry") between entities are retrieved through the knowledge graph. And through the semantic understanding ability of the LLM (processing the implicit logic of the text) and the structured fact verification provided by the KG, a deep integration is carried out.

[0064] The present invention enhances the deep reasoning ability of LLM for knowledge-intensive tasks by not only relying on its external knowledge during the reasoning process of the large model, but also enabling different reasoning paths to be extracted from KG as the basis of LLM reasoning during the reasoning process.

[0065] For the task of false news detection, this approach leverages LLM reasoning capabilities while incorporating external knowledge from KGs to improve detection performance. This also enhances the interpretability of the LLM reasoning process, addressing issues such as the presence of nonexistent content in false news. Furthermore, the judgment and reasoning process takes into account the textual patterns and sentiment of the news itself. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is the overall framework diagram of the present invention.

[0067] Figure 2 This is a graph of reasoning instances that combines a large model with a knowledge graph. DETAILED DESCRIPTION

[0068] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0069] The overall framework of the large model for fake news detection combined with the knowledge graph reasoning method of the present invention is as follows Figure 1 shown.

[0070] This example uses the cleaned Weibo21 dataset, which consists of both Chinese and English sections. Both datasets contain both real and fake news. The goal is to determine whether a piece of news is fabricated or contains false information. Table 1 shows dataset statistics, including sample size and the number of real and fake news stories.

[0071] Table 1 Statistics of fake news dataset

[0072]

[0073] In this method's evaluation system, a multi-dimensional performance evaluation framework was constructed, encompassing four core metrics: accuracy (Accuracy), macro-average F1 score (macF1), true positive F1 score (F1real), and false positive F1 score (F1fake). This metric system was designed according to the following principles: Accuracy (Accuracy = (TP+TN) / (TP+TN+FP+FN)) measures overall classification accuracy, providing a global performance benchmark for the model; macro-average F1 is calculated by independently calculating the F1 score for each class and taking the arithmetic average (macF1 = (F1real+F1fake) / 2), effectively mitigating evaluation bias caused by class imbalance; and dual-class F1 is measured by harmonic mean (F1 = 2*(P*R) / (P+R)), quantifying the model's sensitivity to true and false positives based on two key metrics: precision (TP / (TP+FP)) and recall (Recall = TP / (TP+FN)). Experiments show that the evaluation system can comprehensively reflect the comprehensive performance of the model in scenarios with uneven data distribution, especially in fake news detection, it can effectively distinguish the model's fact capture ability and misjudgment control level.

[0074] This implementation uses a dual-language model architecture, using Llama3-6B for English data and DeepSeek-R1-6B for Chinese data to perform inference tasks. The knowledge graph system integrates Freebase with an entity-aligned subset of Wikidata (approximately 1.9 billion triples). The maximum generation length of the language model is set to 1024 tokens to avoid redundant content generation.

[0075] The knowledge graph exploration process implements a dual constraint mechanism: both the path width and depth thresholds are set at 3, limiting the number of candidate entities retrieved per iteration to a maximum of five. During the relationship selection phase, the top five highest-confidence relationships are retained, and the search process is dynamically controlled by real-time evaluation of the sufficiency of acquired evidence. When the criteria are met, the search is immediately terminated and a conclusion is output; otherwise, a combined search strategy combining breadth-first and depth-first search continues. A specially designed bidirectional relationship retrieval mode simultaneously explores relationships initiated by an entity and relationships pointing to that entity, ensuring the integrity of the entity's contextual information.

[0076] The device system for running the method of the present invention is Ubuntu 16.04, the CPU model is Intel Xeon CPU E5-2650v4@2.20GHz, the GPU used is NVIDIA GeForce 4090, and the video memory is 24G.

[0077] by Figure 2 For further explanation, let's take an example. The text is in English and refers to: "To celebrate France's National Day, international superstar Taylor Swift announced today that a special performance will be added to the Stade de France on July 14..." First, LLM is used to extract important entities (only two are used as examples here) "France's National Day" and "Stade de France" from the text and locate them in the knowledge graph. Secondly, starting with the above important entities, through relationship exploration in the knowledge graph and relationship pruning through LLM, the relationship of "France's National Day" that is most relevant to this question is "forbidden_events". Since this relationship is connected to multiple entities, the prompt template enables LLM to select the entity most relevant to this question, namely "commercial_performances". This completes the first depth of exploration. From this, the second depth of exploration continues, and during the exploration process, it is determined whether the information is sufficient to generate an answer. In this embodiment, since there is no relationship connection to the "commercial_performances" entity in the knowledge graph, LLM will take this triple as a consideration and finally make a false news judgment. The triple content is "Commercial performances are prohibited on France's National Day" and LLM will judge this news as false news after integrating other paths and its internal content.

Claims

1. A large model combined with knowledge graph reasoning method for fake news detection, characterized by: Here are the steps: Step 1: News Text Entity Extraction During the news text preprocessing phase, a multi-level entity recognition architecture was designed. This architecture leverages the context-awareness of the Large Language Model (LLM) to perform named entity recognition, while simultaneously capturing important entities in the text through a hierarchical annotation strategy based on an attention mechanism. The details are as follows: First, the news text is sequence labeled, assuming that the news text is a word sequence T = {w1, w2, ..., w n }, the entity tag set is LLM generates context-aware hidden states H through a multi-layer Transformer encoder: Where d is the hidden state dimension and n is the length of the word sequence. The probability distribution of the entity label is obtained by linear projection of the hidden state and Softmax normalization: P(e i |w i )=Softmax(W e h i +b e ) in, is a learnable parameter, k is the number of all possible entity types; h i is the vector representation of the i-th word; After labeling, important entities are selected and attention weighting is performed through the self-attention mechanism during the selection process, that is, the self-attention weight is used Identify key entities Importance(w i ): Among them, A i,j is the i-th row and j-th column element of the self-attention weight matrix, representing the word w i to w j Attention intensity; h j is the vector representation of the j-th word; Filter words whose importance is higher than the threshold τ as candidate entities By setting it up, the large model can extract the top K important entities among the candidate entities; Step 2: Building a multi-level prompting system for face-to-face fact-checking By designing domain-adaptive prompt templates and dynamic reasoning control mechanisms, we optimize the knowledge activation path and output of the Large Language Model (LLM) in fact-checking tasks. The multi-level prompting project for face-to-face fact-checking includes the following parts: (1) Relationship mapping layer: Identify the core relationships between entities, dynamically assign relationship scores based on attention weights, and filter the evidence chain most relevant to the claim. The input text is the news text to be verified, and a predefined entity recognition template is used to extract a set of key entities from the text. (2) Verification decision layer Output the final authenticity judgment based on the evidence triple (entity, relationship, entity); input is the triple path generated after the search, and the authenticity judgment is performed using the predefined multi-path reasoning template; (3) Structured output control Standardized output is achieved through syntax tree constraints. In order to obtain standardized output during the decoding process, a predefined syntax tree template is used to standardize the output of the LLM so that evidence can be obtained later to improve the interpretability of model detection. Step 3: Large model combined with knowledge graph for exploration and reasoning Based on the entities extracted in step 1 and the hint project constructed in step 2, through the collaborative reasoning of the large model LLM and the knowledge graph KG, the logical path is dynamically constructed and the completeness of the answer is verified, and finally a credible conclusion is output; the details are as follows: Relationship exploration: First, based on the entity nodes extracted in step 1, perform breadth-first relationship search in the knowledge graph. First, perform relationship discovery: starting from the current entity, retrieve all its associated relationship edges to form an initial candidate set. Next, perform relationship screening: use the semantic understanding ability of the large model to sort the candidate relationships by relevance, and select the top N relationships with the highest confidence. Finally, perform path construction: integrate the candidate relationships into reasoning branches according to logical relevance to form a preliminary hypothesis path. Entity exploration: Deeply traverse the knowledge graph based on the selected relationships. First, perform adjacent node retrieval, locating related entities along the selected relationship edges to obtain a set of candidate entities. Next, perform entity verification: Use the contextual reasoning capabilities of the large model to evaluate the relevance of candidate entities to the reasoning target. Because there are multiple candidate entities, path optimization is required: Use LLM to dynamically prune all preliminary hypothesis paths to retain the optimal reasoning path, ensuring the coherence of the logical chain and the factual consistency of the knowledge graph. Finally, select and retain the top-N most relevant relationships. In the reasoning stage, it is necessary to determine whether the currently accumulated path information is sufficient to answer the question; let the reasoning path set be The subscripts a and c are entity indexes, used to distinguish different entities; b is the relationship index, identifying the relationship type; each triple (e a ,r b ,e c ) represents an independent path retrieved from KG; for each path LLM calculates the confidence score (p i ): Score(p i )=LLM reason ([Claim]+[Path:p i ]) Among them, LLM reason Represents a large model used for reasoning; [Claim] represents a text claim to be verified, [Path:p i ] represents the text representation of the reasoning path extracted from the knowledge graph; By customizing the confidence level, if it is sufficient, the path information is used to generate the answer; if it is insufficient, it returns to "relationship exploration" and "entity exploration" and continues to explore new paths until the termination conditions are met: a) obtaining a positive or negative news evaluation result; b) reaching the set maximum exploration depth D_max; when D_max is reached and still no valid path can be generated, the answer is generated only by relying on the internal knowledge of the LLM.

2. The large model combined with knowledge graph reasoning method for fake news detection according to claim 1 is characterized in that: In the relationship mapping layer, the entity recognition template is: Analyze the core relations in the text, select the three most relevant ones from the candidate list [relation 1; relation 2; ...], and output them as {'relation':'score'} (total score = 1.0).

3. The large model combined with knowledge graph reasoning method for fake news detection according to claim 1 is characterized in that: In the verification decision layer, the multi-path reasoning template is: Verify the authenticity of the claim based on the following chain of evidence, and only output {"result":"false or true"}: Chain of evidence:

1. {path_1_text}(confidence:{score_1}) ... n.{path_n_text}(confidence:{score_n}).

4. The method for detecting fake news using a large model combined with a knowledge graph according to claim 1, characterized in that: In structured output control, the syntax tree template is: {"Result":"False or True";"Credibility":["High","Medium","Low"];"Evidence for Error":"…";"Supporting Evidence":"…"}.

Images