Health rumor detection method and device, equipment and storage medium

Abstract

The invention discloses a health rumor detection method, device and equipment and a storage medium, and relates to the technical field of natural language processing, and the method comprises the steps: constructing a text causal graph based on the causal relationship between medical entities in a to-be-detected health medical rumor text; performing intensity estimation on the text causal graph, and generating a causal intensity graph according to an intensity estimation result; determining the causal strength credibility in the causal strength graph through the medical knowledge graph, and generating an enhanced text causal graph based on the causal strength credibility; performing feature coding on the enhanced text causal graph by adopting an edge enhanced graph attention network to obtain graph structure information; and performing health rumor detection on the to-be-detected health medical rumor text based on the graph structure information to obtain a health rumor detection result. By applying the technical scheme, the technical problem that a health rumor detection method in the prior art depends on a specific text mode, and complex medical causal logic is difficult to understand is solved.

Description

Technical Field

[0001] The present application relates to the field of natural language processing technology, and in particular to health rumor detection methods, devices, equipment and storage media. Background Art

[0002] The Internet has become an important source of health information. However, although this convenience has lowered the threshold for medical consultation, defects in the information supervision system have also led to the rapid spread of health rumors. Such false information can easily lead to personal health decisions and hinder epidemic prevention and control. Therefore, it is urgent to effectively detect it.

[0003] Currently, existing health rumor detection methods mainly rely on label-specific features trained on known rumor instances, and classify them by capturing the shallow semantics and statistical patterns of the text. However, when faced with health rumors that deliberately distort professional terminology and pseudo-scientific logic, this detection method has difficulty understanding complex medical causal logic, resulting in low accuracy in health rumor detection. Summary of the Invention

[0004] The main purpose of this application is to provide a health rumor detection method, device, equipment and storage medium, aiming to solve the technical problem that the health rumor detection method in the existing technology relies on specific text patterns and has difficulty in understanding complex medical causal logic.

[0005] To achieve the above objectives, this application proposes a method for detecting health rumors, which includes:

[0006] Constructing a text causal graph based on the medical entities in the health and medical rumor text to be detected and the causal relationship between the medical entities;

[0007] Performing strength estimation on the text causal graph, and generating a causal strength graph according to the strength estimation result;

[0008] Determining the causal strength credibility in the causal strength graph through a medical knowledge graph, and generating an enhanced text causal graph based on the causal strength credibility;

[0009] Using an edge-enhanced graph attention network to perform feature encoding on the enhanced text causal graph to obtain graph structure information;

[0010] Based on the graph structure information, health rumor detection is performed on the health medical rumor text to be detected to obtain a health rumor detection result.

[0011] In one embodiment, the step of constructing a text causal graph based on the medical entities in the health medical rumor text to be detected and the causal relationships between the medical entities includes:

[0012] Using a preset medical language model to extract medical entities from the health and medical rumor text to be detected, and obtaining medical entities in the health and medical rumor text to be detected;

[0013] Determining the causal relationship between the medical entities by using a preset greedy fast causal inference algorithm;

[0014] A text causal graph is constructed based on the medical entities and the causal relationships.

[0015] In one embodiment, the step of performing strength estimation on the text causal graph and generating a causal strength graph according to the strength estimation result includes:

[0016] Performing strength estimation on the text causal graph, and determining the causal association strength corresponding to each edge in the text causal graph according to the strength estimation result;

[0017] The strength of each edge is assigned based on the causal association strength to generate a causal strength graph.

[0018] In one embodiment, the step of determining the causal strength credibility in the causal strength graph using the medical knowledge graph and generating an enhanced textual causal graph based on the causal strength credibility includes:

[0019] Determining semantic similarity between a knowledge graph entity in the medical knowledge graph and a causal graph entity in the causal strength graph;

[0020] Perform entity alignment on the knowledge graph entity and the causal graph entity according to the semantic similarity to obtain an aligned causal graph entity;

[0021] generating an aligned causal graph according to the aligned causal graph entities;

[0022] Determining the causal strength credibility in the aligned causal graph according to the entity association relationship between the entities in the aligned causal graph;

[0023] determining a comprehensive credibility parameter based on the causal strength credibility;

[0024] The causal strength graph is updated based on the comprehensive credibility parameter to generate an enhanced text causal graph.

[0025] In one embodiment, the step of determining the credibility of the causal strength in the aligned causal graph based on the entity association relationship between entities in the aligned causal graph includes:

[0026] If the entity association relationship between the entities in the aligned causal graph is a direct association relationship, determining the causal strength credibility in the aligned causal graph according to a preset directional consistency index;

[0027] Alternatively, if the entity association relationship between the entities in the aligned causal graph is an indirect association relationship, the semantic coherence and directional consistency corresponding to the entity paths between the entities in the aligned causal graph are determined, and the causal strength credibility in the aligned causal graph is determined based on the semantic coherence and the directional consistency.

[0028] In one embodiment, the step of using an edge-enhanced graph attention network to perform feature encoding on the enhanced text causal graph to obtain graph structure information includes:

[0029] An edge-enhanced graph attention network is used to perform joint feature encoding on the causal strength and knowledge verification credibility in the enhanced text causal graph to obtain edge features;

[0030] Encoding the node features and the edge features in the enhanced text causal graph, and determining an attention score corresponding to the enhanced text causal graph;

[0031] Normalizing the attention scores to generate attention distributions corresponding to all nodes in the enhanced text causal graph;

[0032] Based on the attention distribution and the edge features, neighbor nodes of the central node in the enhanced text causal graph are aggregated, and graph structure information is generated according to the aggregated nodes.

[0033] In one embodiment, the step of performing health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result includes:

[0034] Mapping the graph structure information to a multi-layer perceptron;

[0035] When the mapping is completed, the attention weights corresponding to the node embeddings in the enhanced text causal graph are determined through the attention pooling mechanism;

[0036] Performing a weighted summation on the attention weights and the aggregated nodes to generate a graph-level representation corresponding to the node embedding;

[0037] The graph-level representation is input into a classifier, so that the classifier outputs a health rumor detection result based on the graph-level representation.

[0038] In addition, to achieve the above objectives, the present application also proposes a health rumor detection device, which includes:

[0039] A causal graph construction module is used to construct a text causal graph based on the medical entities in the health and medical rumor text to be detected and the causal relationship between the medical entities;

[0040] A causal strength estimation module, configured to perform strength estimation on the text causal graph and generate a causal strength graph based on the strength estimation result;

[0041] a credibility determination module, configured to determine the credibility of the causal strength in the causal strength graph through a medical knowledge graph, and generate an enhanced text causal graph based on the causal strength credibility;

[0042] A feature encoding module is used to perform feature encoding on the enhanced text causal graph using an edge-enhanced graph attention network to obtain graph structure information;

[0043] The rumor detection module is used to perform health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result.

[0044] In addition, to achieve the above-mentioned purpose, the present application also proposes a health rumor detection device, which includes: a memory, a processor, and a computer program stored on the memory and runnable on the processor, and the computer program is configured to implement the steps of the health rumor detection method described above.

[0045] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium. A computer program is stored on the storage medium, and when the computer program is executed by the processor, the steps of the health rumor detection method described above are implemented.

[0046] The present application provides a health rumor detection method, which discloses constructing a text causal graph based on medical entities and causal relationships between medical entities in a health medical rumor text to be detected; performing strength estimation on the text causal graph and generating a causal strength graph based on the strength estimation result; determining the causal strength credibility in the causal strength graph through a medical knowledge graph, and generating an enhanced text causal graph based on the causal strength credibility; using an edge-enhanced graph attention network to feature encode the enhanced text causal graph to obtain graph structure information; performing health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result; compared with the health rumor detection method in the prior art that mainly relies on label-specific features trained on known rumor instances, it is difficult to understand the complex medical causal logic when faced with health rumors that deliberately distort professional terms and pseudo-scientific logic. Since the present invention can construct a text causal graph based on the medical entities and their causal relationships in the health medical rumor text to be detected, and perform health rumor detection on the health medical rumor text to be detected through the graph structure information in the enhanced text causal graph corresponding to the text causal graph, it solves the technical problem that the health rumor detection method in the prior art relies on specific text patterns and has difficulty understanding complex medical causal logic, thereby improving the accuracy of health rumor detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0049] Figure 1 A flowchart of the first embodiment of the health rumor detection method of this application is provided;

[0050] Figure 2 This is an entity causal relationship display diagram output by the preset greedy fast causal inference algorithm in the health rumor detection method of this application;

[0051] Figure 3 This is a schematic diagram of the causal strength estimation process in the health rumor detection method of this application;

[0052] Figure 4 This paper proposes a dual-graph fusion framework for the causal graph and medical knowledge graph in the health rumor detection method of this application;

[0053] Figure 5A flow chart illustrating the second embodiment of the health rumor detection method of this application;

[0054] Figure 6 This is a diagram showing the entity alignment strategy and verification status in the health rumor detection method of this application;

[0055] Figure 7 Schematic diagram of the minimum model for credibility estimation in the health rumor detection method of this application;

[0056] Figure 8 This is a flowchart of Example 3 of the health rumor detection method of this application;

[0057] Figure 9 This paper proposes an edge-enhanced dual-graph fusion classifier framework for the health rumor detection method in this application.

[0058] Figure 10 This is a schematic diagram of the module structure of the health rumor detection device according to an embodiment of the present application;

[0059] Figure 11 This is a schematic diagram of the device structure of the hardware operating environment involved in the health rumor detection method in the embodiment of this application.

[0060] The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0061] It should be understood that the specific embodiments described herein are merely used to explain the technical solutions of the present application and are not intended to limit the present application.

[0062] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0063] Since the health rumor detection methods in the existing technology mainly rely on label-specific features trained on known rumor instances, it is difficult to understand complex medical causal logic when faced with health rumors that deliberately distort professional terms and pseudo-scientific logic.

[0064] This application provides a solution that can construct a text causal graph based on the medical entities and their causal relationships in the health and medical rumor text to be detected, and perform health rumor detection on the health and medical rumor text to be detected through the graph structure information in the enhanced text causal graph corresponding to the text causal graph, thereby solving the technical problem that the health rumor detection method in the existing technology relies on specific text patterns and has difficulty in understanding complex medical causal logic, thereby improving the accuracy of health rumor detection.

[0065] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, mobile phone, etc., or an electronic device capable of implementing the above functions, a health rumor detection device, etc. The following uses the health rumor detection device as an example (hereinafter referred to as the device) to illustrate this embodiment and the following embodiments.

[0066] Based on this, the embodiment of the present application provides a method for detecting health rumors. Figure 1 , Figure 1 This is a flow chart of the first embodiment of the health rumor detection method of this application.

[0067] In this embodiment, the health rumor detection method includes steps S10 to S50:

[0068] Step S10: Constructing a text causal graph based on the medical entities in the health medical rumor text to be detected and the causal relationships between the medical entities.

[0069] It is understandable that the above-mentioned health and medical rumor text to be detected may be text content that may contain false, misleading or unverified medical and health information (in line with the laws of nature), such as antibiotics for treating viral infections, bone soup for calcium supplementation, etc. This embodiment does not limit this.

[0070] It should be understood that the above-mentioned medical entities can be specific nouns or terms related to medical health in the health medical rumor text to be detected, which can represent specific concepts, objects or things in the medical field, such as disease names, treatment methods, biomarkers, body parts or organs, nutrients or food ingredients, etc., and this embodiment does not limit this. Correspondingly, the causal relationship between medical entities can be the causal relationship between two or more medical entities in the health medical rumor text to be detected, for example, disease and cause: such as "smoking causes lung cancer"; drugs and efficacy: such as "penicillin treats bacterial infection"; treatment methods and disease improvement: such as "chemotherapy shrinks tumors"; biomarkers and disease risks: such as "high blood sugar increases the risk of diabetes", etc., and this embodiment does not limit this.

[0071] It should be noted that the above-mentioned text causal graph can be a graph used to represent the logical connections and potential causal relationships between medical entities. It can display the medical entities in the text and the causal relationships between them through nodes and edges, where nodes represent medical entities and edges represent the causal relationships between medical entities.

[0072] Specifically, step S10 includes: using a preset medical language model to extract medical entities from the health medical rumor text to be detected, and obtaining the medical entities in the health medical rumor text to be detected; determining the causal relationship between the medical entities through a preset greedy fast causal inference algorithm; and constructing a text causal graph based on the medical entities and the causal relationship.

[0073] It is understandable that the above-mentioned preset medical language model can be a model for identifying and extracting medical entities in the health and medical rumor text to be detected, such as a pre-trained BioBERT model. Among them, the pre-trained BioBERT model is a pre-trained language model based on the BERT (Bidirectional Encoder Representation Transformer) architecture, which can be optimized for the biomedical field and pre-trained on large-scale biomedical literature data to improve performance in biomedical text mining tasks. In practical applications, the pre-trained BioBERT model can accurately identify key medical entities such as disease names, genes, and drugs in the health and medical rumor text to be detected, thereby obtaining medical entities in the health and medical rumor text to be detected.

[0074] It should be noted that the aforementioned preset greedy fast causal inference algorithm can be an algorithm for identifying causal relationships between medical entities by analyzing medical data. In practical applications, the device can learn causal structures from medical data using the preset greedy fast causal inference algorithm, and finally construct a text causal graph using medical entities as nodes and causal relationships as edges. This embodiment uses the preset greedy fast causal inference algorithm to analyze the causal graph, allowing for intuitive visualization of the causal relationships between variables. This helps detectors identify potential logical loopholes or irrationalities in rumors, providing a more in-depth basis for rumor detection.

[0075] In actual applications, the automatically extracted causal graph may contain noisy edges that are inconsistent with medical logic. In order to improve the credibility of the causal graph, this embodiment can further design dual constraint rules to filter out noisy edges, including: a) Medical logic constraints: Some entities and relationships in medicine have clear logic. At this time, symptom entities are not allowed to point to disease, drug or treatment entities as causes. Such relationships violate the inherent logic of cause-symptom in the medical knowledge base. b) Temporal constraints: Considering that causal relationships usually precede results in time, and the description of symptoms is usually written in the chronological order of events, the chronological order of the description can be regarded as a constraint. If factor A appears after B in most cases, the edge from A to B will not be allowed. It should be noted that the temporal constraint does not mean that there is an edge from B to A, because the temporal order is not a sufficient condition for causality.

[0076] In this embodiment, referring to Figure 2 , Figure 2This is the entity causal relationship display diagram output by the preset greedy fast causal inference algorithm in the health rumor detection method of this application. Figure 2 As shown, Figure 2 The four uncertain relationships contained in the partial ancestral graph output by the preset greedy fast causal inference algorithm are described in

[15] , including: ->, <->, Among them, A—>B means that A causes B, A<—>B means that there is an unobserved confounding factor between A and B, Indicates that either A causes B, or there are unobserved confounders. Indicates that either A causes B, or B causes A, or there are unobserved confounding factors. In this embodiment, the clear edges can be retained or removed, and the two types of uncertain edges can be removed. The probability of retaining or removing is 1 / 2. The probability of keeping, removing or reversing is 1 / 3.

[0077] Step S20: performing strength estimation on the text causal graph, and generating a causal strength graph according to the strength estimation result.

[0078] It should be noted that the above-mentioned causal strength graph can be a visual graph used to quantify the strength of causal relationships between medical entities. It can display the causal relationship strength between medical entities (such as diseases, drugs, symptoms, etc.) in the health medical rumor text to be detected in a numerical and visual way, thereby helping the detection personnel to intuitively evaluate the credibility of the causal relationship.

[0079] Specifically, the step S20 includes: performing strength estimation on the text causal graph, and determining the causal association strength corresponding to each edge in the text causal graph according to the strength estimation result; and performing strength allocation on each edge based on the causal association strength to generate a causal strength graph.

[0080] It should be understood that the above-mentioned causal association strength may be the strength of the causal effect between the medical entities corresponding to each edge in the text causal graph.

[0081] In practical applications, since the resulting graph is noisy, the sampled causal graph can be optimized by estimating the learned causal association strength. This embodiment can assign higher strength to edges with strong causal effects, and assign close to zero strength to edges that do not represent causal relationships or have weak effects. This embodiment can regard the strength of T->Y in the text causal graph G as the average processing effect. And use propensity score matching to measure:

[0082]

[0083] in, represents the medical instance most similar to i in the opposite group, t i ,y i , z i are the treatment value, outcome, and confounding factor of instance i, respectively.

[0084] Reference Figure 3 , Figure 3 This is a flow chart of the causal strength estimation process in the health rumor detection method of this application. Figure 3 As shown, first, entities can be extracted from language text (such as the health and medical rumor text to be detected) and the causal relationship between these entities can be learned. Then, causal graph sampling is performed to generate several candidate causal graphs, and the strength of these causal graphs is estimated. Based on the causal association strength corresponding to each edge in the causal graph, the strength is assigned to generate a causal strength graph.

[0085] Step S30: Determine the causal strength credibility in the causal strength graph through the medical knowledge graph, and generate an enhanced text causal graph based on the causal strength credibility.

[0086] It should be noted that the above-mentioned medical knowledge graph can be a knowledge graph in the biomedical field, which can structure entities in the medical field (such as diseases, symptoms, drugs, treatment methods, genes, proteins, etc.) and the relationships between these entities.

[0087] It should be noted that the causal strength credibility may be a parameter used to characterize the authenticity of the causal strength in the causal strength graph. Accordingly, the enhanced text causal graph may be a causal graph obtained by adding the causal strength credibility to the causal strength graph.

[0088] Step S40: Using an edge-enhanced graph attention network to perform feature encoding on the enhanced text causal graph to obtain graph structure information.

[0089] In this embodiment, the Edge-enhanced Graph Attention Network (EGAT) can be an improved model of the Graph Attention Network (GAT), which optimizes information transfer and feature aggregation by enhancing the weight or importance of the edges in the graph, thereby improving the accuracy and robustness of health rumor detection. In practical applications, EGAT can dynamically adjust the weights of the edges so that the model pays more attention to important edges and nodes, thereby more accurately capturing key information in the process of rumor propagation. This dynamic adjustment mechanism helps to reduce noise and interference and improve the robustness of rumor detection. In addition, through the edge enhancement mechanism, EGAT can better aggregate the feature information of neighboring nodes and generate more discriminative node representations, which helps the model to more accurately identify abnormal patterns or inconsistencies in rumor content.

[0090] It should be noted that the above-mentioned graph structure information can be an overall expression of information such as node relationships and key features in the enhanced text causal graph, such as node feature representation, edge weight information, subgraph structure features, and global graph features, etc. This embodiment does not impose any restrictions on this.

[0091] Step S50: performing health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result.

[0092] In this embodiment, after obtaining the graph structure information corresponding to the enhanced text causal graph, the graph structure information can be mapped to the classification space, and the text classification can be performed through the classifier to output the classification results of health rumors, including rumors or non-rumors.

[0093] In the specific implementation, refer to Figure 4 , Figure 4 This is a dual-graph fusion framework of the causal graph and the medical knowledge graph in the health rumor detection method of this application. Figure 4 As shown, Figure 4 This paper presents the overall architecture of a dual-graph fusion framework (CKDG) based on causal logic and knowledge graphs. It aims to detect health rumors through collaborative reasoning between causal graphs and knowledge graphs. First, the framework receives rumor text as input and performs medical entity recognition on it using the pre-trained BioBERT model to extract medical entities and key information from the rumor text. Then, the greedy fast causal inference algorithm (GFCI) is used to determine the causal relationships between these medical entities. A causal graph is constructed using medical entities as nodes and causal relationships as edges. Furthermore, a medical knowledge graph (such as CMeKG) is introduced to align entities in the causal graph with those in the knowledge graph. Prior knowledge in the knowledge graph is used to verify the credibility of causal relationships in the causal graph, including direct relationship verification, multi-hop path verification, and no-association testing. Finally, an edge-enhanced graph attention network (EGRET) and a hierarchical pooling strategy are used to dynamically fuse causal strength and knowledge verification signals, enabling multi-granularity reasoning from local paths to global semantics, ultimately achieving accurate detection of health rumors.

[0094] This embodiment provides a health rumor detection method, which discloses constructing a text causal graph based on medical entities and causal relationships between medical entities in a health medical rumor text to be detected; performing strength estimation on the text causal graph and generating a causal strength graph based on the strength estimation result; determining the causal strength credibility in the causal strength graph through a medical knowledge graph, and generating an enhanced text causal graph based on the causal strength credibility; using an edge-enhanced graph attention network to feature encode the enhanced text causal graph to obtain graph structure information; performing health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result; compared with health rumor detection methods in the prior art that mainly rely on label-specific features trained on known rumor instances and have difficulty understanding complex medical causal logic when faced with health rumors that deliberately distort professional terminology and pseudo-scientific logic, the present invention can construct a text causal graph based on medical entities and their causal relationships in the health medical rumor text to be detected, and perform health rumor detection on the health medical rumor text to be detected based on the graph structure information in the enhanced text causal graph corresponding to the text causal graph, thereby solving the technical problem that health rumor detection methods in the prior art rely on specific text patterns and have difficulty understanding complex medical causal logic, thereby improving the accuracy of health rumor detection.

[0095] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the above embodiment 1 can be referred to the above introduction and will not be described in detail later. Figure 5 , Figure 5 This is a flow chart of Example 2 of the health rumor detection method of this application.

[0096] In this embodiment, step S30 includes steps S301 to S306:

[0097] Step S301: Determine the semantic similarity between the knowledge graph entity in the medical knowledge graph and the causal graph entity in the causal strength graph.

[0098] It can be understood that the above-mentioned knowledge graph entities can be the basic units that constitute the medical knowledge graph, such as disease entities, symptom entities, drug entities, treatment method entities, etc., and this embodiment does not limit this.

[0099] It should be understood that the above-mentioned semantic similarity can be the degree of proximity between the medical concepts, meanings or significances expressed by the knowledge graph entity and the causal graph entity at the semantic level.

[0100] Step S302: aligning the knowledge graph entity and the causal graph entity according to the semantic similarity to obtain an aligned causal graph entity.

[0101] It should be noted that the aligned causal graph entities can be entities aligned to the knowledge graph. In practical applications, the goal of entity alignment is to align the entities e in the causal graph. i and the entity e′ in the knowledge graph i Matching is performed. In this embodiment, the semantic similarity matrix between entities can be defined as in Represents the causal graph entity e i and knowledge graph entity e′ i The semantic similarity between .

[0102] In practical applications, refer to Figure 6 , Figure 6 This is a diagram showing the entity alignment strategy and verification status in the health rumor detection method of this application. Figure 6 As shown, the matching process in this embodiment can be divided into three steps, including: exact matching, fuzzy matching and misaligned marking. In the entity alignment process, first, the exact matching method can be used to directly match entities with exactly the same name through the hash table. Then, fuzzy matching is used to calculate the similarity θ between entities that do not pass the exact match. e,e′ , retain the entity pairs whose matching degree is higher than the set threshold θ, and let is θ e,e′ ,Finally, the unmatched entities are marked as “Missing”, thus completing the entity alignment.

[0103] Step S303: generating an aligned causal graph according to the aligned causal graph entities.

[0104] It can be understood that the above-mentioned aligned causal graph can be a causal graph constructed by aligned causal graph entities.

[0105] Step S304: determining the credibility of the causal strength in the aligned causal graph according to the entity association relationship between the entities in the aligned causal graph.

[0106] It should be noted that after completing the entity alignment, each edge e in the causal graph can be verified through the knowledge graph. i →e j The credibility R of the causal strength can be designed to have a function of the causal strength credibility R. In this embodiment, we can first focus on a minimum model graph, refer to Figure 7 , Figure 7 This is a schematic diagram of the minimum model for estimating credibility in the health rumor detection method of this application. Figure 7 As shown, there are four entities in the figure: e i ,e j is the entity of the causal graph, e′ i ,e′j To align to entities in the knowledge graph, we define φ(e i →e j ) is the external domain knowledge constraint, ψ(e i →e j ) is the original causal strength estimate θ e,e ′ represents the semantic similarity of entity alignment.

[0107] In practical applications, based on the concept of a gating mechanism, we construct a credibility estimation function R(·), whose design adheres to knowledge dominance and directionality. Under the knowledge dominance principle, the external domain knowledge constraint φ, as prior information, should dominate the φ regulation process. When φ = 0, the function should completely suppress the credibility of the causal strength. Furthermore, the function needs to capture the positive and negative correlation between the direction of the knowledge path and the causal direction.

[0108] In this embodiment, starting from the product gate structure, R = ψ·g(φ, θ e,e′ ), where the gating function g(·) must satisfy the control dominance of φ. Considering the semantic drift problem generated during entity alignment, the two endpoint alignment quality indicators can be integrated here. and The arithmetic mean of is used as the overall alignment confidence, mitigating the impact of single-endpoint errors on global confidence. To further address the directional issue, this embodiment allows φ to take a negative value when the knowledge path direction conflicts with the causal direction, and then normalizes it after the calculation is completed. The above derivation ultimately yields the analytical expression for the gating function: g(φ, θ) = avg(θ)·φ.

[0109] This derives the complete credibility estimation function:

[0110]

[0111] After completing the design of the causal strength credibility estimation function R, it is necessary to further explore the φ(e′ i →e′ j ) situation.

[0112] Specifically, step S304 includes: if the entity association relationship between the entities in the aligned causal graph is a direct association relationship, then determining the causal strength credibility in the aligned causal graph according to a preset directional consistency index; or, if the entity association relationship between the entities in the aligned causal graph is an indirect association relationship, then determining the semantic coherence and directional consistency corresponding to the entity path between the entities in the aligned causal graph, and determining the causal strength credibility in the aligned causal graph based on the semantic coherence and the directional consistency.

[0113] It should be noted that the above-mentioned direct association relationship may be a relationship in which the aligned causal graph entities are directly associated; the above-mentioned indirect association relationship may be a relationship in which the aligned causal graph entities do not have a direct association but have a multi-hop path association.

[0114] In practical applications, for the causal edge e i →e j , if there is a knowledge graph triple (e′ i ,r,e′ j )∈R k And r is a direct causal chain, that is, entity e i and e j If there is a direct correlation, a directional consistency index can be defined, and the causal strength credibility in the aligned causal graph can be determined based on the directional consistency index. In this embodiment, the directional consistency index is as follows:

[0115]

[0116] Among them, λ pos and λ neg is the penalty coefficient, which is used to reflect the credibility difference of forward or reverse verification.

[0117] When the aligned causal graph entities do not have direct connections in the knowledge graph but have multi-hop paths (such as ), that is, when the entity association relationship between the entities in the aligned causal graph is an indirect association relationship, the semantic coherence and directional consistency of the path can be evaluated, and the causal strength credibility in the aligned causal graph can be determined based on the semantic coherence and directional consistency. In this embodiment, a bidirectional breadth-first search algorithm can be used to efficiently search for multi-hop paths. The steps of this algorithm are as follows:

[0118] a) Path search: From e′ i and e′ j Start at the same time and expand along the knowledge graph relationship until the two search domains intersect or the maximum number of hops L is reached max .

[0119] b) Path filtering: only paths with the same direction are retained (i.e., e′ i →…→e′ j or e′ j ←…←e′ i ). Path φ(e′) i →e′ j ) is modeled by a semantic decay factor:

[0120]

[0121] Among them, σ(r k )∈[0,1] is the relationship type weight, rk is the relationship type, γ is the path length attenuation coefficient, |P| is the number of path hops, L max is the maximum allowed path length, 1-|P| / L max represents the dilution effect of penalizing long paths on causal strength. At the same time, φ path It takes a positive value when the path direction is consistent with the causal direction, otherwise it takes a negative value.

[0122] In addition, if there is no explicit association between the knowledge graph entities corresponding to the causal edges (i.e., the aligned causal graph entities), the causal strength credibility can be dynamically adjusted by the global knowledge graph coverage:

[0123]

[0124] Among them, Deg(·) represents the entity degree, N k is the total number of knowledge graph entities, and α is the baseline penalty term. This formula can reflect the fact that no relationship is observed in a highly connected subgraph (such as no association between two common disease entities), which is more falsifiable than a sparse subgraph (such as rare drugs and symptoms). Specifically, if φ null If the value is low, it means that the corresponding entity e′ i and e′ j In the knowledge graph, it is dense data, and in this case, there is no multi-hop path between entities, which may indicate that the association between entities is very low. On the contrary, if φ null The value is high, which means the corresponding entity e′ i and e′ j It is sparse data in the knowledge graph, and there is a certain probability that this is because the data in the knowledge graph is incomplete, not because the correlation between entities is low.

[0125] In this example, the labeling of missing entities has been completed by the upstream task. Here, the penalty function for specific missing entities is implemented:

[0126] φ null =exp(-β·ψ)·I missing ;

[0127] Among them, β is the missing entity penalty coefficient, I missing It is an indicator function when there are misaligned entities at the endpoints of the causal edge.

[0128] Step S305: Determine a comprehensive credibility parameter based on the causal strength credibility.

[0129] It should be noted that, considering that there may be a direct relationship and multiple multi-hop paths between two entities, we will i →e′ j) Integrate the following situations and adjust through contradictory evidence:

[0130]

[0131] Where K = {φ dir ,φ path} is the credibility set of each verification scenario, the variance term Var(·) is used to penalize contradictory evidence, and the logarithmic enhancement term log(1+ψ) is used to amplify the influence of high causal strength edges.

[0132] Step S306: updating the causal strength graph based on the comprehensive credibility parameter to generate an enhanced text causal graph.

[0133] In this embodiment, after calculating the comprehensive credibility parameter, φ can be substituted into R(·) and normalized to obtain the credibility of the corresponding edge. On this basis, the credibility parameter (i.e. the above-mentioned comprehensive credibility parameter) is added to the original causal strength estimation graph to form an enhanced graph structure expression G c (i.e. the enhanced text causal graph mentioned above), in

[0134] In this embodiment, the method of determining the semantic similarity between knowledge graph entities in a medical knowledge graph and causal graph entities in a causal strength graph is disclosed; entity alignment of knowledge graph entities and causal graph entities is performed based on the semantic similarity to obtain aligned causal graph entities; an aligned causal graph is generated based on the aligned causal graph entities; the causal strength credibility in the aligned causal graph is determined based on the entity association relationship between the aligned causal graph entities; a comprehensive credibility parameter is determined based on the causal strength credibility; the causal strength graph is updated based on the comprehensive credibility parameter to generate an enhanced text causal graph; since this embodiment can deeply integrate the causal graph with the medical knowledge graph through a knowledge-guided credibility estimation method, and verify the credibility of the relationship in the causal graph through the prior verification knowledge in the medical knowledge graph, the accuracy of health rumor detection can be further improved.

[0135] Based on the first embodiment and / or the second embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those in the above embodiments can be referred to the above introduction and will not be described in detail later. Figure 8 , Figure 8 This is a flow chart of Example 3 of the health rumor detection method provided in this application.

[0136] In this embodiment, step S40 includes steps S401 to S404:

[0137] Step S401: Using an edge-enhanced graph attention network to perform joint feature encoding on the causal strength and knowledge verification credibility in the enhanced text causal graph to obtain edge features.

[0138] It should be noted that the enhanced text causal graph Both node features and edge features exist in the graph. However, the standard graph attention network (GAT) framework does not explicitly incorporate edge features into its calculation process. To overcome this limitation and fully utilize the structural information encoded by edge features, this embodiment can adopt a variant of GAT, namely the edge-enhanced graph attention network (EGRET). Unlike the traditional GAT, the EGRET network can effectively utilize the structural information encoded by edge features in the aggregation stage and the attention score calculation stage, thereby systematically improving the model's ability to utilize edge feature information.

[0139] In this embodiment, in order to reduce the dimension of subsequent calculations of the solution, the causal strength ψ in the causal strength graph of the credibility estimate can be ij and knowledge verification credibility R ij Joint encoding as edge feature ξ ij , whose dimensions correspond to ψ ij With R ij . Then the two types of features are combined and normalized as follows:

[0140] ξ ij =σ(α·ψ ij +(1-α)·R ij );

[0141] Among them, α is a learnable parameter and σ is a Sigmoid function, which is used to dynamically balance the two types of signals.

[0142] Step S402: Encode the node features and the edge features in the enhanced text causal graph to determine the attention score corresponding to the enhanced text causal graph.

[0143] It should be noted that this embodiment can use the last hidden state of BioBERT as the causal graph node v i ∈V c The initial characteristics This method inherits the semantic understanding ability of BioBERT in medical text. At the same time, in order to further enhance the semantic representation of nodes, this embodiment can be used to align the knowledge graph entity e′ i ∈G k Use TransE for embedding and fuse it with BioBERT’s representation to obtain the above node features:

[0144]

[0145] in, is the entity e′ i TransE embedding vector, W i It is a learnable parameter matrix used to align knowledge graph embeddings into the semantic space of BioBERT.

[0146] In this embodiment, for each edge e ij When calculating the attention score, in addition to the node feature h i , h j In addition, the edge feature ξ of the directed edge from node i to node j is also included ij The node-dependent and edge-dependent attention scores are calculated as follows:

[0147]

[0148] Among them, d ij is the non-normalized representation of the attention score, || represents the concatenation operation, and W a , W v , W e Both are learnable parameters, which are used to linearly change the features of nodes and edges respectively.

[0149] Step S403: normalize the attention scores to generate attention distributions corresponding to all nodes in the enhanced text causal graph.

[0150] It should be noted that when calculating the attention score d for each edge ij After that, you can continue to ij |j∈N i} Perform softmax normalization to generate all nodes j∈N i (where i is the central node), that is, the probability distribution of attention distribution {α ij |j∈N i}:

[0151]

[0152] Step S404: performing neighbor node aggregation on the central node in the enhanced text causal graph based on the attention distribution and the edge features, and generating graph structure information based on the aggregated nodes.

[0153] It should be understood that the neighbor node aggregation of the central node is the operation of aggregating the features of the neighbor nodes corresponding to the central node to the central node. ij The potential of aggregating neighbor node features {h j |j∈N i} while incorporating edge features, where node i is the central node:

[0154]

[0155] Edge eigenvector ξ ij Before polymerization ij Apply linear transformation. Where σ(·) is the activation function, the output of σ(·) is related to h i (feature representations before node i) are connected together to generate a new representation That is, edge features enhance the graph structure information output by the graph attention layer.

[0156] Furthermore, step S50 includes: mapping the graph structure information to a multi-layer perceptron; when the mapping is completed, determining the attention weight corresponding to the node embedding in the enhanced text causal graph through the attention pooling mechanism; performing weighted summation on the attention weight and the aggregated nodes to generate a graph-level representation corresponding to the node embedding; inputting the graph-level representation into a classifier so that the classifier outputs a health rumor detection result based on the graph-level representation.

[0157] It should be noted that this embodiment can use a multi-layer perceptron (MLP) as a classifier. In practical applications, considering the hierarchical characteristics of graph-level semantics in the health rumor detection task, it is necessary to first aggregate the node-level embedding into a graph-level representation. In addition, this embodiment can introduce an attention pooling mechanism to dynamically capture the contribution differences of key nodes. Specifically, for the causal graph G c All nodes in Through learnable query vectors Calculate node-level attention weights:

[0158]

[0159] in, and is the projection parameter, the tanh function is used to enhance the nonlinear expression ability, and the graph level representation h G Generated by weighted summation:

[0160]

[0161] This machine gives higher weight to nodes with higher causal strength, which is consistent with the task characteristics of rumor detection that needs to focus on the core causal chain. GThe input to the MLP is transformed nonlinearly through two fully connected layers. The first layer uses a dimensionality reduction strategy to compress the d-dimensional features to d / 4 dimensions to eliminate redundant information. The second layer maps them to a binary probability space. At the same time, the GeLU activation function and batch normalization are introduced between layers. The mathematical expression is:

[0162] y pred =Softmax(W2·GeLU(BN(W1h G +b1)+b2));

[0163] Among them, y ored represents the predicted probability of rumor detection, and It is a trainable parameter and BN represents batch normalization operation.

[0164] This design enhances the model's ability to discriminate complex causal logic through hierarchical feature abstraction while avoiding the risk of overfitting.

[0165] In this embodiment, the training objective of the model is jointly optimized by multi-task losses, including the main classification loss, knowledge consistency constraint, and sparse regularization term. The main loss adopts Label Smoothing Cross Entropy Loss (LS-CE), and the overconfidence problem caused by category imbalance is alleviated by introducing a smoothing factor ∈:

[0166]

[0167] Among them, y c is the true label, p c is the predicted probability. To further enhance the semantic alignment of the causal path and the knowledge graph, the knowledge consistency loss term L is designed kg , used to constrain the attention weight α ij and causal credibility The distribution consistency of :

[0168]

[0169] Among them, ∈ = le-8 is a numerical stability term. This loss term forces the model to rely more on high-confidence edges of knowledge verification in the decision-making process. In addition, L2 regularization is applied to the MLP parameters To control the complexity of the model. The final total loss function is the weighted sum of the three:

[0170] L total =λ1L cls +λ2L kg +λ3L reg ;

[0171] Among them, the weight coefficients λ1, λ2, and λ3 will be determined through Pareto frontier analysis to balance the classification performance and the robustness of knowledge logic.

[0172] In practical applications, refer to Figure 9 , Figure 9 This is the edge-enhanced dual-graph fusion classifier framework used in the health rumor detection method of this application. Figure 9 As shown, for edge-enhanced graph attention calculation, we can first input the causal graph G c and knowledge graph G k , the input causal graph Among them, V c Represents the medical entities in the health and medical rumor text to be detected, and the edge set E c Represents the causal relationship between entities, represents the causal strength, Represents the credibility of knowledge verification; the input knowledge graph G k =(V k ,P k ), where V k Represents knowledge entity, P k Represents the causal relationship between knowledge entities. Then, this embodiment can perform node feature encoding and edge feature encoding on the causal graph. Specifically, the BioBERT model can be used to extract the initial features of the causal graph nodes. Combined with the knowledge graph entity features embedded by TransE, the learnable parameter matrix V align Align it to the semantic space of BioBERT to generate enhanced node feature representation; at the same time, the causal strength and knowledge verification credibility Joint encoding as edge feature ξ ij , as shown in the figure ξ1, ξ2, ξ3. When calculating the attention score, for each edge e ij , can be combined with node feature h i and h j and edge features ξ ij Calculate the attention score d ij , use the LeakyReLU activation function and get the attention weight α through softmax normalization ij However, the device can incorporate edge features while aggregating neighbor node features. ij , generate new node representations through linear transformation and activation function Finally, the node-level embedding can be aggregated into a graph-level representation through a multi-layer perceptron, and the graph-level representation is input into a two-layer fully connected network. The first layer adopts a dimensionality reduction strategy, and the second layer maps it to a binary probability space. The GeLU activation function and batch normalization are used to enhance the nonlinear expression ability of the model. Finally, the output layer outputs the predicted probability y of rumor detection pred , and use the Softmax function to map the output to the [0, 1] interval, indicating the probability of rumor and non-rumor.

[0173] In this embodiment, it is disclosed that an edge-enhanced graph attention network is used to jointly encode the causal strength and knowledge verification credibility in the enhanced text causal graph to obtain edge features; the node features and edge features in the enhanced text causal graph are encoded to determine the attention score corresponding to the enhanced text causal graph; the attention score is normalized to generate the attention distribution corresponding to all nodes in the enhanced text causal graph; the neighbor nodes of the central node in the enhanced text causal graph are aggregated based on the attention distribution and edge features, and graph structure information is generated based on the aggregated nodes; because this embodiment can introduce the edge-enhanced attention mechanism and hierarchical pooling strategy, dynamically fuse the causal strength and knowledge verification signals, and combine local path feature aggregation with global semantic distillation, it can improve the model's representation ability and detection robustness for complex causal topologies.

[0174] It should be noted that the above examples are only used to understand this application and do not constitute a limitation on the health rumor detection method of this application. More simple transformations based on this technical concept are all within the scope of protection of this application.

[0175] This application also provides a health rumor detection device, please refer to Figure 10 , the health rumor detection device includes:

[0176] A causal graph construction module 10 is used to construct a text causal graph based on the medical entities in the health and medical rumor text to be detected and the causal relationship between the medical entities;

[0177] A causal strength estimation module 20 is used to perform strength estimation on the text causal graph and generate a causal strength graph based on the strength estimation result;

[0178] a credibility determination module 30 for determining the credibility of the causal strength in the causal strength graph through the medical knowledge graph, and generating an enhanced text causal graph based on the causal strength credibility;

[0179] A feature encoding module 40 is configured to perform feature encoding on the enhanced text causal graph using an edge-enhanced graph attention network to obtain graph structure information;

[0180] The rumor detection module 50 is used to perform health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result.

[0181] The health rumor detection device provided in this application, which utilizes the health rumor detection method described in the aforementioned embodiments, can address the technical issue of existing health rumor detection methods, which rely on specific text patterns and have difficulty understanding complex medical causal logic. Compared to the existing technology, the beneficial effects of the health rumor detection device provided in this application are the same as those of the health rumor detection method described in the aforementioned embodiments. Other technical features of the health rumor detection device are the same as those disclosed in the aforementioned embodiments and are not further elaborated here.

[0182] The present application provides a health rumor detection device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the health rumor detection method in the above-mentioned embodiment one.

[0183] Reference below Figure 11 , which shows a schematic diagram of the structure of a health rumor detection device suitable for implementing the embodiments of the present application. The health rumor detection device in the embodiments of the present application may include, but is not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 11 The health rumor detection device shown is merely an example and should not limit the functions and scope of use of the embodiments of the present application.

[0184] like Figure 11As shown, the health rumor detection device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes based on programs stored in a read-only memory 1002 or programs loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the health rumor detection device. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An input / output interface 1006 is also connected to the bus. Typically, the following systems can be connected to the input / output interface 1006: an input device 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; an output device 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the health rumor detection device to communicate with other devices wirelessly or wired to exchange data. Although the figure shows a health rumor detection device with various systems, it should be understood that it is not required to implement or have all of the systems shown. More or fewer systems can be implemented or provided instead.

[0185] The health rumor detection device provided in this application utilizes the health rumor detection method described in the aforementioned embodiments to address the technical challenges of health rumor detection. Compared to the prior art, the beneficial effects of the health rumor detection device provided in this application are the same as those of the health rumor detection method described in the aforementioned embodiments. Other technical features of the health rumor detection device are the same as those disclosed in the aforementioned embodiments and are not further detailed here.

[0186] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0187] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0188] The present application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, and the computer-readable program instructions are used to execute the health rumor detection method in the above-mentioned embodiment.

[0189] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0190] The above-mentioned computer-readable storage medium may be included in the health rumor detection device; or it may exist independently without being assembled into the health rumor detection device.

[0191] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the health rumor detection device, the health rumor detection device enables the health rumor detection device to: construct a text causal graph based on the medical entities in the health medical rumor text to be detected and the causal relationship between the medical entities; perform strength estimation on the text causal graph, and generate a causal strength graph based on the strength estimation result; determine the causal strength credibility in the causal strength graph through the medical knowledge graph, and generate an enhanced text causal graph based on the causal strength credibility; use the edge-enhanced graph attention network to feature encode the enhanced text causal graph to obtain graph structure information; perform health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain health rumor detection results.

[0192] Computer program code for performing the operations of the present application may be written in one or more programming languages, or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a stand-alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0193] The modules described in the embodiments of the present application may be implemented in software or hardware, wherein the name of a module does not necessarily limit the unit itself.

[0194] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-mentioned health rumor detection method. This can solve the technical problem that the health rumor detection method in the prior art relies on specific text patterns and has difficulty understanding complex medical causal logic. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the health rumor detection method provided in the above-mentioned embodiment, and will not be repeated here.

[0195] The above description is only part of the embodiments of the present application and does not limit the patent scope of the present application. All equivalent structural transformations made by using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A health rumor detection method, characterized in that: The method includes: Constructing a text causal graph based on the medical entities in the health and medical rumor text to be detected and the causal relationship between the medical entities; Performing strength estimation on the text causal graph, and generating a causal strength graph according to the strength estimation result; Determining the causal strength credibility in the causal strength graph through a medical knowledge graph, and generating an enhanced text causal graph based on the causal strength credibility; Using an edge-enhanced graph attention network to perform feature encoding on the enhanced text causal graph to obtain graph structure information; Based on the graph structure information, health rumor detection is performed on the health medical rumor text to be detected to obtain a health rumor detection result.

2. The method according to claim 1, wherein The step of constructing a text causal graph based on the medical entities in the health medical rumor text to be detected and the causal relationship between the medical entities includes: Using a preset medical language model to extract medical entities from the health and medical rumor text to be detected, and obtaining medical entities in the health and medical rumor text to be detected; Determining the causal relationship between the medical entities by using a preset greedy fast causal inference algorithm; A text causal graph is constructed based on the medical entities and the causal relationships.

3. The method according to claim 1, wherein The step of performing strength estimation on the text causal graph and generating a causal strength graph according to the strength estimation result includes: Performing strength estimation on the text causal graph, and determining the causal association strength corresponding to each edge in the text causal graph according to the strength estimation result; The strength of each edge is assigned based on the causal association strength to generate a causal strength graph.

4. The method according to claim 1, wherein The step of determining the causal strength credibility in the causal strength graph through the medical knowledge graph and generating an enhanced text causal graph based on the causal strength credibility includes: Determining semantic similarity between a knowledge graph entity in the medical knowledge graph and a causal graph entity in the causal strength graph; Perform entity alignment on the knowledge graph entity and the causal graph entity according to the semantic similarity to obtain an aligned causal graph entity; generating an aligned causal graph according to the aligned causal graph entities; Determining the causal strength credibility in the aligned causal graph according to the entity association relationship between the entities in the aligned causal graph; determining a comprehensive credibility parameter based on the causal strength credibility; The causal strength graph is updated based on the comprehensive credibility parameter to generate an enhanced text causal graph.

5. The method according to claim 4, wherein The step of determining the credibility of the causal strength in the aligned causal graph according to the entity association relationship between the entities in the aligned causal graph comprises: If the entity association relationship between the entities in the aligned causal graph is a direct association relationship, determining the causal strength credibility in the aligned causal graph according to a preset directional consistency index; Alternatively, if the entity association relationship between the entities in the aligned causal graph is an indirect association relationship, the semantic coherence and directional consistency corresponding to the entity paths between the entities in the aligned causal graph are determined, and the causal strength credibility in the aligned causal graph is determined based on the semantic coherence and the directional consistency.

6. The method according to any one of claims 1 to 5, characterized in that The step of using an edge-enhanced graph attention network to perform feature encoding on the enhanced text causal graph to obtain graph structure information includes: An edge-enhanced graph attention network is used to perform joint feature encoding on the causal strength and knowledge verification credibility in the enhanced text causal graph to obtain edge features; Encoding the node features and the edge features in the enhanced text causal graph, and determining an attention score corresponding to the enhanced text causal graph; Normalizing the attention scores to generate attention distributions corresponding to all nodes in the enhanced text causal graph; Based on the attention distribution and the edge features, neighbor nodes of the central node in the enhanced text causal graph are aggregated, and graph structure information is generated according to the aggregated nodes.

7. The method according to claim 6, wherein The step of performing health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result includes: Mapping the graph structure information to a multi-layer perceptron; When the mapping is completed, the attention weights corresponding to the node embeddings in the enhanced text causal graph are determined through the attention pooling mechanism; Performing a weighted summation on the attention weights and the aggregated nodes to generate a graph-level representation corresponding to the node embedding; The graph-level representation is input into a classifier, so that the classifier outputs a health rumor detection result based on the graph-level representation.

8. A health rumor detection device, characterized in that: The device comprises: A causal graph construction module is used to construct a text causal graph based on the medical entities in the health and medical rumor text to be detected and the causal relationship between the medical entities; A causal strength estimation module, configured to perform strength estimation on the text causal graph and generate a causal strength graph based on the strength estimation result; a credibility determination module, configured to determine the credibility of the causal strength in the causal strength graph through a medical knowledge graph, and generate an enhanced text causal graph based on the causal strength credibility; A feature encoding module is used to perform feature encoding on the enhanced text causal graph using an edge-enhanced graph attention network to obtain graph structure information; The rumor detection module is used to perform health rumor detection on the health medical rumor text to be detected based on the graph structure information to obtain a health rumor detection result.

9. A health rumor detection device, characterized in that: The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the health rumor detection method according to any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the health rumor detection method according to any one of claims 1 to 7 are implemented.

Images